技术领域technical field
本发明涉及互联网技术领域，尤其涉及一种多媒体数据处理方法以及装置。The invention relates to the technical field of the Internet, in particular to a multimedia data processing method and device.
背景技术Background technique
随着互联网技术的发展，各式各样的应用层出不穷，例如，即时通信应用、游戏应用、多媒体数据应用等等。以多媒体数据应用为例，用户可以通过多媒体数据应用收听各式各样的歌曲，而且也可以通过推测用户所喜欢的歌曲，以将相应歌曲推荐给用户。目前，推测用户所喜欢的歌曲的方式可以包括：将用户所收藏(或下载)的歌曲认定为用户所喜欢的歌曲，因此，可以推测用户所喜欢的歌曲包括与所收藏(或下载)的歌曲相类似的歌曲，进而向用户推荐这些相类似的歌曲。当用户没有收藏(或下载)歌曲时，将完整播放完的歌曲认定为用户所喜欢的歌曲，进而进行相似歌曲的推荐。但是完整播放完的歌曲并不代表是用户在听的歌曲(如用户临时离开电脑，而电脑中的音乐播放器继续播放)，进而也无法代表是用户所喜欢的歌曲，所以若直接将完整播放完的歌曲认定为用户所喜欢的歌曲，则无法保证所推荐的歌曲是用户所喜欢的歌曲，导致推荐效果不佳。With the development of Internet technology, various applications emerge in an endless stream, for example, instant messaging applications, game applications, multimedia data applications, and so on. Taking the multimedia data application as an example, the user can listen to various songs through the multimedia data application, and can also recommend corresponding songs to the user by inferring the user's favorite songs. At present, the way of guessing the songs liked by the user may include: identifying the songs favorited by the user (or downloaded) as the songs liked by the user, therefore, it can be speculated that the songs liked by the user include Similar songs, and then recommend these similar songs to users. When the user does not collect (or download) the song, the song that has been completely played is identified as the song that the user likes, and then similar songs are recommended. However, the completely played song does not represent the song that the user is listening to (for example, the user temporarily leaves the computer, while the music player in the computer continues to play), and it cannot represent the song that the user likes, so if the complete playback is performed directly If the completed song is identified as the user's favorite song, there is no guarantee that the recommended song is the user's favorite song, resulting in poor recommendation effect.
发明内容Contents of the invention
本发明实施例提供一种用于多媒体数据处理方法以及装置，可保证所推荐的歌曲是用户所喜欢的歌曲，以提高推荐效果。Embodiments of the present invention provide a method and device for processing multimedia data, which can ensure that the recommended songs are favorites of users, so as to improve the recommendation effect.
本发明实施例提供了一种多媒体数据处理方法，包括：The embodiment of the present invention provides a kind of multimedia data processing method, comprising:
根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵；Generate a multimedia data operation behavior matrix according to the operation behaviors of the historical user groups on the multiple multimedia data in the preset multimedia database;
基于稀疏自编码神经网络，并根据所述多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量；一个隐含特征向量表征所述历史用户群对一个多媒体数据的喜好程度信息；一个用户特征向量表征一个历史用户对所述多个多媒体数据的喜好程度信息；Based on the sparse self-encoding neural network, and according to the multimedia data operation behavior matrix, the hidden feature vectors corresponding to each multimedia data and the user feature vectors corresponding to each historical user are calculated respectively; a hidden feature vector represents the pair of historical user groups A preference degree information of multimedia data; a user feature vector characterizes preference degree information of a historical user for the plurality of multimedia data;
当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，获取所述目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理。When a recommendation request corresponding to the target user is received and the historical user group includes the target user, multiple multimedia data in the personal operation behavior information of the target user are obtained, and according to the user corresponding to the target user The feature vectors and the implicit feature vectors respectively corresponding to the multimedia data in the personal operation behavior information perform recommendation processing on the multiple multimedia data in the personal operation behavior information.
相应地，本发明实施例还提供了一种多媒体数据处理装置，包括：Correspondingly, an embodiment of the present invention also provides a multimedia data processing device, including:
矩阵生成模块，用于根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵；A matrix generation module is used to generate a multimedia data operation behavior matrix according to the operation behavior of a plurality of multimedia data in the preset multimedia database according to the historical user group;
特征计算模块，用于基于稀疏自编码神经网络，并根据所述多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量；一个隐含特征向量表征所述历史用户群对一个多媒体数据的喜好程度信息；一个用户特征向量表征一个历史用户对所述多个多媒体数据的喜好程度信息；The feature calculation module is used to calculate the hidden feature vectors corresponding to each multimedia data and the user feature vectors corresponding to each historical user respectively according to the multimedia data operation behavior matrix based on the sparse self-encoding neural network; a hidden feature vector representation The preference information of the historical user group to a piece of multimedia data; a user feature vector characterizes the preference information of a historical user to the plurality of multimedia data;
推荐模块，用于当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，获取所述目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理。A recommendation module, configured to obtain a plurality of multimedia data in the personal operation behavior information of the target user when a recommendation request corresponding to the target user is received and the historical user group includes the target user, and according to the The user feature vector corresponding to the target user and the implicit feature vectors respectively corresponding to each multimedia data in the personal operation behavior information perform recommendation processing on a plurality of multimedia data in the personal operation behavior information.
本发明实施例通过根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵，并基于稀疏自编码神经网络，并根据多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量，由此可见，隐含特征向量可以准确表征历史用户群对一个多媒体数据的喜好程度信息，且用户特征向量可以准确表征一个历史用户对多个多媒体数据的喜好程度信息，所以通过隐含特征向量和用户特征向量可以对目标用户实现准确的个性化推荐，即可保证所推荐的歌曲是目标用户所喜欢的歌曲，以提高推荐效果。The embodiment of the present invention generates a multimedia data operation behavior matrix based on the operation behavior of historical user groups on multiple multimedia data in the preset multimedia database, and calculates each multimedia data operation behavior matrix based on the sparse self-encoding neural network. The hidden eigenvectors corresponding to the data and the user eigenvectors corresponding to each historical user respectively. It can be seen that the hidden eigenvectors can accurately represent the preference information of a historical user group for a piece of multimedia data, and the user eigenvectors can accurately represent a Historical users’ preferences for multiple multimedia data, so accurate personalized recommendations can be made to target users through hidden feature vectors and user feature vectors, which can ensure that the recommended songs are songs that target users like, so as to improve Recommended effect.
附图说明Description of drawings
为了更清楚地说明本发明实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings that need to be used in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only These are some embodiments of the present invention. Those skilled in the art can also obtain other drawings based on these drawings without creative work.
图1是本发明实施例提供的一种网络架构的示意图；FIG. 1 is a schematic diagram of a network architecture provided by an embodiment of the present invention;
图2是本发明实施例提供的一种多媒体数据处理方法的流程示意图；FIG. 2 is a schematic flowchart of a multimedia data processing method provided by an embodiment of the present invention;
图3是本发明实施例提供的另一种多媒体数据处理方法的流程示意图；FIG. 3 is a schematic flowchart of another multimedia data processing method provided by an embodiment of the present invention;
图4是本发明实施例提供的一种多媒体数据处理装置的结构示意图；4 is a schematic structural diagram of a multimedia data processing device provided by an embodiment of the present invention;
图5是本发明实施例提供的一种特征计算模块的结构示意图；Fig. 5 is a schematic structural diagram of a feature calculation module provided by an embodiment of the present invention;
图6是本发明实施例提供的一种推荐模块的结构示意图；FIG. 6 is a schematic structural diagram of a recommendation module provided by an embodiment of the present invention;
图7是本发明实施例提供的一种隐含特征生成单元的结构示意图；FIG. 7 is a schematic structural diagram of a hidden feature generation unit provided by an embodiment of the present invention;
图8是本发明实施例提供的一种用户特征生成单元的结构示意图；Fig. 8 is a schematic structural diagram of a user feature generating unit provided by an embodiment of the present invention;
图9是本发明实施例提供的另一种多媒体数据处理装置的结构示意图。FIG. 9 is a schematic structural diagram of another multimedia data processing device provided by an embodiment of the present invention.
具体实施方式detailed description
下面将结合本发明实施例中的附图，对本发明实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例仅仅是本发明一部分实施例，而不是全部的实施例。基于本发明中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本发明保护的范围。The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some, not all, embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without creative efforts fall within the protection scope of the present invention.
请参见图1，是本发明实施例提供的一种网络架构的示意图。所述网络架构可以包括后台服务器和多个客户端，各个客户端均可以通过网络与所述后台服务器连接，所述后台服务器可以将各个客户端分别对应的用户确定为历史用户群，所述后台服务器还可以收集各个客户端分别对应的操作行为(如某客户端的听歌行为，具体包括该客户端对应的用户所收听过的歌曲)。因此，所述后台服务器可以根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵，并基于稀疏自编码神经网络，并根据所述多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量；一个隐含特征向量表征所述历史用户群对一个多媒体数据的喜好程度信息；一个用户特征向量表征一个历史用户对所述多个多媒体数据的喜好程度信息。当某个客户端向所述后台服务器发送推荐请求时，所述后台服务器可以获取多个与该客户端对应的目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理。Please refer to FIG. 1 , which is a schematic diagram of a network architecture provided by an embodiment of the present invention. The network architecture may include a background server and a plurality of clients, and each client may be connected to the background server through a network, and the background server may determine the users corresponding to each client as a historical user group, and the background The server can also collect the operation behaviors corresponding to each client (such as the song listening behavior of a certain client, specifically including the songs that the user corresponding to the client has listened to). Therefore, the background server can generate a multimedia data operation behavior matrix based on the operation behavior of the historical user group on the multiple multimedia data in the preset multimedia database, and based on the sparse self-encoding neural network, and according to the multimedia data operation behavior The matrix calculates the hidden eigenvectors corresponding to each multimedia data and the user eigenvectors corresponding to each historical user respectively; one hidden eigenvector represents the preference information of the historical user group for a piece of multimedia data; one user eigenvector represents a history The user's preferences for the plurality of multimedia data. When a certain client sends a recommendation request to the backend server, the backend server can acquire a plurality of multimedia data in the personal operation behavior information of a plurality of target users corresponding to the client, and The user feature vectors and the implicit feature vectors respectively corresponding to the multimedia data in the personal operation behavior information perform recommendation processing on a plurality of multimedia data in the personal operation behavior information.
请参见图2，是本发明实施例提供的一种多媒体数据处理方法的流程示意图，所述方法可以包括：Please refer to FIG. 2, which is a schematic flowchart of a multimedia data processing method provided by an embodiment of the present invention, the method may include:
S101，根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵；S101. Generate a multimedia data operation behavior matrix according to the operation behaviors of the historical user groups on the multiple multimedia data in the preset multimedia database;
具体的，基于多媒体数据应用的后台服务器可以获取历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为。所述多媒体数据库可以为音乐库，每一个多媒体数据均可以为音乐库中的一首歌，因此，所述操作行为可以包括所述历史用户群中的各历史用户对各个多媒体数据的收听行为，所述收听行为包括已收听行为和未收听行为，并且所述后台服务器可以为不同的收听行为设置不同的特征值，以生成与所述操作行为相应的多媒体数据操作行为矩阵。如下表1所示的历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为的特征表：Specifically, the background server based on the multimedia data application can acquire the operation behaviors of the historical user groups on multiple pieces of multimedia data in the preset multimedia database. The multimedia database can be a music library, and each piece of multimedia data can be a song in the music library. Therefore, the operation behavior can include the listening behavior of each historical user in the historical user group to each multimedia data, The listening behavior includes listening behavior and non-listening behavior, and the background server can set different characteristic values for different listening behaviors, so as to generate a multimedia data operation behavior matrix corresponding to the operation behavior. The characteristic table of the operation behavior of the historical user group to the multiple multimedia data in the preset multimedia database as shown in Table 1 below:
表1Table 1
其中，表1中的特征值“1”表示该用户收听过这首歌曲，特征值“0”表示该用户未收听过这首歌曲，如用户1与歌曲1之间的特征值为“1”，则说明用户1收听过歌曲1；又如用户3与歌曲2之间的特征值为“0”，则说明用户3未收听过歌曲2。因此，根据表1中的所有特征值即可生成与所述操作行为相应的多媒体数据操作行为矩阵，即所述多媒体数据操作行为矩阵中的元素Pui可以表示用户u对歌曲i的收听行为对应的特征值(Pui＝1，说明用户u收听过歌曲i；Pui＝0，说明用户u未收听歌曲i)。Among them, the feature value "1" in Table 1 indicates that the user has listened to this song, and the feature value "0" indicates that the user has not listened to this song. For example, the feature value between user 1 and song 1 is "1". , it means that user 1 has listened to song 1; and if the feature value between user 3 and song 2 is "0", it means that user 3 has not listened to song 2. Therefore, according to all the eigenvalues in Table 1, the multimedia data operation behavior matrix corresponding to the operation behavior can be generated, that is, the element Pui in the multimedia data operation behavior matrix can represent that user u corresponds to the listening behavior of song i (Pui =1, indicating that user u has listened to song i; Pui =0, indicating that user u has not listened to song i).
S102，基于稀疏自编码神经网络，并根据所述多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量；一个隐含特征向量表征所述历史用户群对一个多媒体数据的喜好程度信息；一个用户特征向量表征一个历史用户对所述多个多媒体数据的喜好程度信息；S102, based on the sparse self-encoding neural network, and according to the multimedia data operation behavior matrix, calculate the hidden feature vectors corresponding to each multimedia data and the user feature vectors corresponding to each historical user; one hidden feature vector represents the historical user Information about the preference degree of a group to a piece of multimedia data; a user feature vector characterizes information about the preference degree of a historical user to the plurality of multimedia data;
具体的，所述后台服务器可以将所述多媒体数据操作行为矩阵输入到所述稀疏自编码神经网络对应的稀疏自编码器的输入层，即将所述多媒体数据操作行为矩阵中的各个收听行为对应的特征值输入到所述输入层中；所述稀疏自编码器包括所述输入层、隐藏层、输出层以及所述隐藏层与所述输出层之间的目标参数；所述隐藏层包括预设数量的隐藏节点；所述隐藏节点的数量可以通过平衡所述后台服务器的计算效率和用户/歌曲的特征描述准确性，以推测出来(可以根据经验值推测)。其中，所述输入层的维度与所述输出层的维度相同。Specifically, the background server may input the multimedia data operation behavior matrix to the input layer of the sparse autoencoder corresponding to the sparse autoencoder neural network, that is, the input layer corresponding to each listening behavior in the multimedia data operation behavior matrix. The feature value is input into the input layer; the sparse self-encoder includes the input layer, the hidden layer, the output layer, and the target parameters between the hidden layer and the output layer; the hidden layer includes a preset The number of hidden nodes; the number of hidden nodes can be estimated by balancing the computing efficiency of the background server and the accuracy of user/song feature descriptions (can be estimated based on empirical values). Wherein, the dimension of the input layer is the same as the dimension of the output layer.
所述稀疏自编码器可以根据所述输入层中的参数(即所述多媒体数据操作行为矩阵)以及预设的用于训练所述目标参数和所述隐藏节点的隐藏参数的目标函数，对所述目标参数和所述隐藏节点的隐藏参数进行偏导数训练；其中，所述目标函数可以为：The sparse self-encoder can be used to train the target parameters and the hidden parameters of the hidden nodes according to the parameters in the input layer (ie, the multimedia data operation behavior matrix) and the preset objective function, for all The target parameters and the hidden parameters of the hidden nodes are used for partial derivative training; wherein, the target function can be:
其中，x是所述多媒体数据操作行为矩阵；A是由W1(1)、W2(1)、W3(1)、……、Wk+1(1)组成的矩阵，W1(1)、W2(1)、W3(1)、……、Wk+1(1)分别为K+1个隐藏节点中的隐藏参数(如W1(1)表示第一个隐藏节点的隐藏参数)，且K+1个隐藏节点中的最后一个隐藏节点为截距项，该截距项的隐藏参数为1，通过保留该截距项可以重构出所述输出层。s是由bN1、bN2、bN3、……、bNK组成的矩阵，bN1、bN2、bN3、……、bNK是所述隐藏层到所述输出层之间的目标参数，N为输入层的维度，本发明实施例中，已指定了隐藏节点数K+1，可对隐藏层稀疏要求放宽，因此将s的1范数近似为2范数。其中，由于有些用户虽然没有对某些歌曲进行收听行为，但是并不代表这些用户不喜欢这些歌曲，同样的，即使有些用户听过某些歌曲，也不能直接表示这些用户喜欢这些歌曲，所以本发明实施例通过在所述目标函数中的增加了一个用户兴趣因子项C，以提高所训练出来的所述隐藏层中的s矩阵与用户对歌曲的喜好程度之间的关联关系；所述用户兴趣因子项C包括所述各历史用户分别对所述多媒体数据库中各多媒体数据的兴趣值cui(cui表示用户u对歌曲i的兴趣值)；一个兴趣值是基于一个历史用户对一个多媒体数据的操作行为类型、操作次数以及完整操作率计算得到的。其中，对所述用户兴趣因子项C中的各个兴趣值cui的计算公式可以为：cui＝1+αlog(1+εrui)；其中，当用户u直接收藏/下载歌曲i时，rui＝1；当用户u仅仅是收听过歌曲i时(即没有收藏/下载操作)，rui＝[min(nui，5)/5]*fui，其中，nui是指用户u收听歌曲i的次数，fui是完整收听率(即所述完整操作率)，fui＝完整收听歌曲i的人数/所有听歌人数(即所述历史用户群)；当用户u没有收听过歌曲i时，cui＝0。Wherein, x is the multimedia data operation behavior matrix; A is a matrix composed of W1(1) , W2(1) , W3(1) , ..., Wk+1(1) , W1( 1) , W2(1) , W3(1) , ..., Wk+1(1) are hidden parameters in K+1 hidden nodes respectively (for example, W1(1) represents the first hidden node hidden parameter), and the last hidden node in the K+1 hidden nodes is an intercept item, and the hidden parameter of the intercept item is 1, and the output layer can be reconstructed by retaining the intercept item. s is a matrix composed of bN1 , bN2 , bN3 , ..., bNK , bN1 , bN2 , bN3 , ..., bNK are the target parameters between the hidden layer and the output layer , N is the dimension of the input layer. In the embodiment of the present invention, the number of hidden nodes K+1 has been specified, and the requirement for sparseness of the hidden layer can be relaxed, so the 1-norm of s is approximated as a 2-norm. Among them, although some users do not listen to certain songs, it does not mean that these users do not like these songs. Similarly, even if some users have listened to certain songs, it cannot directly indicate that these users like these songs, so this In the embodiment of the invention, a user interest factor item C is added in the objective function to improve the correlation between the trained s matrix in the hidden layer and the user's preference for songs; the user Interest factor item C comprises described each history user respectively to the interest value cui of each multimedia data in the described multimedia database (cu ui represents user u to the interest value of song i); The operation behavior type, number of operations, and complete operation rate of the data are calculated. Wherein, the calculation formula for each interest value cui in the user interest factor item C may be: cui =1+αlog(1+εrui ); wherein, when user u directly collects/downloads song i, rui = 1; when user u has only listened to song i (that is, there is no collection/download operation), rui = [min(nui , 5)/5]*fui , where nui means that user u has listened to song i The number of times of song i, fui is the complete listening rate (i.e. the complete operation rate), fui =the number of people who have listened completely to song i/all the number of people listening to songs (i.e. the historical user group); when user u has not listened to the song When i, cui =0.
所述稀疏自编码器在进行偏导数训练的过程中可以重复交替执行步骤S1和步骤S2；步骤S1为：固定A，对s求导，用最小二乘法得到最优解；步骤S2为：固定s，对A求导，用最小二乘法得到最优解。The sparse autoencoder can repeat and alternately execute step S1 and step S2 in the process of partial derivative training; step S1 is: fixing A, deriving s, and obtaining the optimal solution with the method of least squares; step S2 is: fixing s, take the derivative of A, and use the least square method to get the optimal solution.
进一步的，当所述稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近(即输出层的参数与所述输入层的参数之间的匹配度达到预设匹配度阈值)时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，此时所述稀疏自编码器停止执行步骤S1和步骤S2，并将满足所述收敛条件的各隐藏节点的隐藏参数组合成隐含特征矩阵，即所述隐含特征矩阵为训练后的矩阵A。为了保证后台服务器的工作效率，因此，所述稀疏自编码器中的隐藏层的维度是要小于所述输入层的维度，所以所述隐藏层中的隐藏参数也可以视为所述输入层中的参数的压缩形式，即所述隐含特征矩阵为所述多媒体数据操作行为矩阵对应的压缩矩阵；其中，所述隐含特征矩阵包括各多媒体数据分别对应的隐含特征向量，例如，可以将所述隐含特征矩阵的第一列特征确定为歌曲1的隐含特征向量，将所述隐含特征矩阵的第二列特征确定为歌曲2的隐含特征向量等等，其中，歌曲1的隐含特征向量可以表征为所述历史用户群对歌曲1的喜好程度信息。Further, when the parameters in the output layer of the sparse autoencoder are similar to the parameters in the input layer (that is, the matching degree between the parameters of the output layer and the parameters of the input layer reaches a preset matching degree Threshold), it is determined that the target parameters and the hidden parameters of the hidden nodes meet the convergence conditions, at this time, the sparse autoencoder stops executing steps S1 and S2, and hides the hidden nodes of the hidden nodes that meet the convergence conditions The parameters are combined into a hidden feature matrix, that is, the hidden feature matrix is the matrix A after training. In order to ensure the working efficiency of the background server, therefore, the dimension of the hidden layer in the sparse autoencoder is smaller than the dimension of the input layer, so the hidden parameters in the hidden layer can also be regarded as The compressed form of the parameters, that is, the hidden feature matrix is a compressed matrix corresponding to the multimedia data operation behavior matrix; wherein, the hidden feature matrix includes hidden feature vectors corresponding to each multimedia data, for example, can be The first column feature of the hidden feature matrix is determined as the hidden feature vector of song 1, and the second column feature of the hidden feature matrix is determined as the hidden feature vector of song 2, etc., wherein the song 1 The hidden feature vector can be characterized as the preference degree information of the historical user group for the song 1 .
进一步的，由于所述稀疏自编码器的数量为一个，所以可以从所述稀疏自编码器中训练后的目标参数对应的参数矩阵s中提取出所述历史用户群中各历史用户分别对应的用户特征向量。所述后台服务器计算出所述历史用户群中各历史用户分别对应的用户特征向量以及所述多媒体数据库中各多媒体数据分别对应的隐含特征向量后，可以存储各个用户特征向量和各个隐含特征向量，以便于后续根据用户特征向量和隐含特征向量进行用户个性化的歌曲推荐。Further, since the number of the sparse autoencoder is one, the parameters corresponding to each historical user in the historical user group can be extracted from the parameter matrix s corresponding to the target parameter after training in the sparse autoencoder. User feature vector. After the background server calculates the user feature vectors corresponding to the historical users in the historical user group and the hidden feature vectors corresponding to the multimedia data in the multimedia database, each user feature vector and each hidden feature vector can be stored. vector, so as to facilitate subsequent user-personalized song recommendation based on user feature vectors and hidden feature vectors.
S103，当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，获取所述目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理；S103. When a recommendation request corresponding to the target user is received, and the historical user group includes the target user, obtain a plurality of multimedia data in the personal operation behavior information of the target user, and The user feature vector and the implicit feature vector corresponding to each multimedia data in the personal operation behavior information respectively perform recommendation processing on a plurality of multimedia data in the personal operation behavior information;
具体的，当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，可以检测目标用户对应的个人操作行为信息中是否包含已收藏的多媒体数据；若检测为包含已收藏的多媒体数据，则获取所述已收藏的多媒体数据(即推荐源)对应的第一相似多媒体数据，并将所述第一相似多媒体数据作为所述目标用户的推荐数据；若检测为未包含已收藏的多媒体数据，则进一步判断所述个人操作行为信息中是否包含已完整操作的多媒体数据(即完整收听的歌曲)。当所述目标用户对应的个人操作行为信息中包含已完整操作的多媒体数据时，将所述目标用户对应的用户特征向量与所述已完整操作的多媒体数据对应的隐含特征向量进行点乘运算，得到个性化特征值，并当所述个性化特征值大于预设特征值阈值时，获取所述已完整操作的多媒体数据(即推荐源)对应的第二相似多媒体数据，并将所述第二相似多媒体数据作为所述目标用户的推荐数据。Specifically, when a recommendation request corresponding to a target user is received, and the historical user group includes the target user, it may be detected whether the personal operation behavior information corresponding to the target user contains the collected multimedia data; Bookmarked multimedia data, then obtain the first similar multimedia data corresponding to the bookmarked multimedia data (i.e. recommendation source), and use the first similar multimedia data as the recommended data for the target user; if detected as not If the collected multimedia data is included, it is further judged whether the personal operation behavior information includes the fully operated multimedia data (that is, the completely listened-to song). When the personal operation behavior information corresponding to the target user includes fully-operated multimedia data, perform a dot product operation on the user feature vector corresponding to the target user and the hidden feature vector corresponding to the fully-operated multimedia data , obtain the personalized feature value, and when the personalized feature value is greater than the preset feature value threshold, obtain the second similar multimedia data corresponding to the fully operated multimedia data (ie, the recommendation source), and convert the second similar multimedia data to Two similar multimedia data are used as recommendation data for the target user.
可选的，当所述目标用户对应的个人操作行为信息不包含已完整操作的多媒体数据时，获取多个候选多媒体数据，并将所述标用户对应的用户特征向量分别与各候选多媒体数据对应的隐含特征向量进行点乘运算，得到所述各候选多媒体数据分别对应的个性化特征值；按照各个性化特征值从大到小的顺序对所述多个候选多媒体数据进行排序，并根据排序结果将预设推荐数量的候选多媒体数据作为所述目标用户的推荐数据。例如，若用户A的个人操作行为信息不包含已完整收听的歌曲，且候选歌曲包括歌曲A、歌曲B以及歌曲C，则可以将用户A的用户特征向量与歌曲A的隐含特征向量进行点乘，得到个性化特征值a；将用户A的用户特征向量与歌曲B的隐含特征向量进行点乘，得到个性化特征值b；将用户A的用户特征向量与歌曲C的隐含特征向量进行点乘，得到个性化特征值c；对个性化特征值进行排序后，得到b>c>a，若预设推荐数量为2，则可以将个性化特征值b对应的歌曲B和个性化特征值c对应的歌曲C推荐给用户A。Optionally, when the personal operation behavior information corresponding to the target user does not contain the multimedia data that has been fully operated, a plurality of candidate multimedia data is obtained, and the user feature vectors corresponding to the target user are respectively corresponding to each candidate multimedia data Dot multiplication of the hidden feature vectors of the hidden feature vectors to obtain the individualized feature values corresponding to the respective candidate multimedia data; sort the plurality of candidate multimedia data according to the order of each personalized feature value from large to small, and according to The sorting result takes a preset recommended number of candidate multimedia data as the recommended data for the target user. For example, if the personal operation behavior information of user A does not include songs that have been completely listened to, and the candidate songs include song A, song B, and song C, then the user feature vector of user A can be compared with the hidden feature vector of song A. Multiply to get the personalized feature value a; dot product the user feature vector of user A with the hidden feature vector of song B to get the personalized feature value b; combine the user feature vector of user A with the hidden feature vector of song C Perform dot multiplication to obtain the personalized feature value c; after sorting the personalized feature values, b>c>a is obtained. If the preset number of recommendations is 2, the song B corresponding to the personalized feature value b and the personalized Song C corresponding to feature value c is recommended to user A.
可选的，在计算所述个性化特征值时，还可以参考显性特征值，即用户u对歌曲i的个性化特征值＝用户u对歌曲i的显性特征值+用户u的用户特征向量与歌曲i的隐含特征向量的点乘结果。所述显示特征值可以根据歌曲的标签和用户的标签之间的匹配度计算得到；歌曲的标签可以包括该歌曲对应的曲风类型、语种类型、节奏类型等等；用户的标签可以包括该用户所喜欢的曲风类型、语种类型、节奏类型等等。通过增加所述显性特征值，可以使所述个性化特征值能够更准确的描述用户u对歌曲i的喜好程度。Optionally, when calculating the personalized feature value, you can also refer to the dominant feature value, that is, the personalized feature value of user u for song i=the dominant feature value of user u for song i+user feature of user u The result of the dot product of the vector and the hidden feature vector of song i. The display feature value can be calculated according to the matching degree between the label of the song and the label of the user; the label of the song can include the genre type, language type, rhythm type, etc. corresponding to the song; the label of the user can include the user's label Favorite music style, language type, rhythm type, etc. By increasing the dominant feature value, the personalized feature value can more accurately describe user u's preference for song i.
可选的，当接收到与目标用户对应的推荐请求，且所述历史用户群不包含所述目标用户时，说明所述目标用户还未在后台服务器所提供的多媒体数据平台上收听过歌曲，则可以按照所述各候选多媒体数据对应的隐含特征向量的向量值从大到小的顺序对所述多个候选多媒体数据进行排序，并根据排序结果将预设推荐数量的候选多媒体数据作为所述目标用户的推荐数据。Optionally, when a recommendation request corresponding to the target user is received, and the historical user group does not include the target user, it means that the target user has not listened to songs on the multimedia data platform provided by the background server, Then, the plurality of candidate multimedia data can be sorted according to the order of the vector values of the hidden feature vectors corresponding to the candidate multimedia data from large to small, and according to the sorting result, the preset recommended number of candidate multimedia data can be used as the selected Describe the recommended data of target users.
可选的，所述后台服务器可以周期性的根据新的操作行为计算和更新各历史用户分别对应的用户特征向量，以及各多媒体数据分别对应的隐含特征向量，使得在对目标用户进行推荐时，可以保证所获取到的与所述目标用户相关的隐含特征向量和用户特征向量是始终贴合用户的兴趣变化，即可以保证所计算出的个性化特征值的准确性。其中，所述目标用户每听一首歌曲，均可更新一次所述目标用户的推荐源，以便于后续在向所述目标用户推荐歌曲时，可以直接获取与最近更新的推荐源相似的多媒体数据，以将相似的多媒体数据推荐给用户，以提高推荐效率；而在客户端的用户界面上可以显示“歌曲B是根据你试听的歌曲A推荐”，其中，歌曲A即为推荐源，歌曲B即为与推荐源相似的多媒体数据。Optionally, the background server can periodically calculate and update the user feature vectors corresponding to each historical user and the implicit feature vectors corresponding to each multimedia data according to the new operation behavior, so that when recommending target users , it can be ensured that the acquired implicit feature vectors and user feature vectors related to the target user are always in line with the user's interest changes, that is, the accuracy of the calculated personalized feature values can be guaranteed. Wherein, each time the target user listens to a song, the recommendation source of the target user can be updated once, so that when recommending songs to the target user, multimedia data similar to the latest updated recommendation source can be directly obtained , to recommend similar multimedia data to users to improve the recommendation efficiency; and on the user interface of the client, it can display "Song B is recommended based on the song A you listened to", wherein, song A is the recommendation source, and song B is the It is multimedia data similar to the recommendation source.
本发明实施例通过根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵，并基于稀疏自编码神经网络，并根据多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量，由此可见，隐含特征向量可以准确表征历史用户群对一个多媒体数据的喜好程度信息，且用户特征向量可以准确表征一个历史用户对多个多媒体数据的喜好程度信息，所以通过隐含特征向量和用户特征向量可以对目标用户实现准确的个性化推荐，即可保证所推荐的歌曲是目标用户所喜欢的歌曲，以提高推荐效果。The embodiment of the present invention generates a multimedia data operation behavior matrix based on the operation behavior of historical user groups on multiple multimedia data in the preset multimedia database, and calculates each multimedia data operation behavior matrix based on the sparse self-encoding neural network. The hidden eigenvectors corresponding to the data and the user eigenvectors corresponding to each historical user respectively. It can be seen that the hidden eigenvectors can accurately represent the preference information of a historical user group for a piece of multimedia data, and the user eigenvectors can accurately represent a Historical users’ preferences for multiple multimedia data, so accurate personalized recommendations can be made to target users through hidden feature vectors and user feature vectors, which can ensure that the recommended songs are songs that target users like, so as to improve Recommended effect.
再请参见图3，是本发明实施例提供的另一种多媒体数据处理方法的流程示意图，所述方法可以包括：Referring to FIG. 3 again, it is a schematic flowchart of another multimedia data processing method provided by an embodiment of the present invention, and the method may include:
S201，根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵；S201. Generate a multimedia data operation behavior matrix according to the operation behaviors of the historical user groups on the multiple multimedia data in the preset multimedia database;
具体的，S201步骤的具体实现方式可以参见上述图2对应实施例中的S101，这里不再进行赘述。Specifically, for the specific implementation manner of step S201, reference may be made to S101 in the above-mentioned embodiment corresponding to FIG. 2 , and details are not repeated here.
S202，将所述多媒体数据操作行为矩阵输入到所述稀疏自编码神经网络对应的稀疏自编码器的输入层；所述稀疏自编码器包括所述输入层、隐藏层、输出层以及所述隐藏层与所述输出层之间的目标参数；所述隐藏层包括预设数量的隐藏节点；S202, input the multimedia data operation behavior matrix to the input layer of the sparse autoencoder corresponding to the sparse autoencoder neural network; the sparse autoencoder includes the input layer, hidden layer, output layer and the hidden A target parameter between the layer and the output layer; the hidden layer includes a preset number of hidden nodes;
具体的，所述后台服务器可以将所述多媒体数据操作行为矩阵输入到所述稀疏自编码神经网络对应的稀疏自编码器的输入层，即将所述多媒体数据操作行为矩阵中的各个收听行为对应的特征值输入到所述输入层中；所述稀疏自编码器包括所述输入层、隐藏层、输出层以及所述隐藏层与所述输出层之间的目标参数；所述隐藏层包括预设数量的隐藏节点；所述隐藏节点的数量可以通过平衡所述后台服务器的计算效率和用户/歌曲的特征描述准确性，以推测出来(可以根据经验值推测)。其中，所述输入层的维度与所述输出层的维度相同。Specifically, the background server may input the multimedia data operation behavior matrix to the input layer of the sparse autoencoder corresponding to the sparse autoencoder neural network, that is, the input layer corresponding to each listening behavior in the multimedia data operation behavior matrix. The feature value is input into the input layer; the sparse self-encoder includes the input layer, the hidden layer, the output layer, and the target parameters between the hidden layer and the output layer; the hidden layer includes a preset The number of hidden nodes; the number of hidden nodes can be estimated by balancing the computing efficiency of the background server and the accuracy of user/song feature descriptions (can be estimated based on empirical values). Wherein, the dimension of the input layer is the same as the dimension of the output layer.
S203，所述稀疏自编码器根据所述输入层中的参数以及预设的用于训练所述目标参数和所述隐藏节点的隐藏参数的目标函数，对所述目标参数和所述隐藏节点的隐藏参数进行偏导数训练；S203, the sparse autoencoder performs the target parameter and the hidden node's Hidden parameters for partial derivative training;
所述稀疏自编码器可以根据所述输入层中的参数(即所述多媒体数据操作行为矩阵)以及预设的用于训练所述目标参数和所述隐藏节点的隐藏参数的目标函数，对所述目标参数和所述隐藏节点的隐藏参数进行偏导数训练；其中，所述目标函数可以为：The sparse self-encoder can be used to train the target parameters and the hidden parameters of the hidden nodes according to the parameters in the input layer (ie, the multimedia data operation behavior matrix) and the preset objective function, for all The target parameters and the hidden parameters of the hidden nodes are used for partial derivative training; wherein, the target function can be:
其中，x是所述多媒体数据操作行为矩阵；A是由W1(1)、W2(1)、W3(1)、……、Wk+1(1)组成的矩阵，W1(1)、W2(1)、W3(1)、……、Wk+1(1)分别为K+1个隐藏节点中的隐藏参数(如W1(1)表示第一个隐藏节点的隐藏参数)，且K+1个隐藏节点中的最后一个隐藏节点为截距项，该截距项的隐藏参数为1，通过保留该截距项可以重构出所述输出层。s是由bN1、bN2、bN3、……、bNK组成的矩阵，bN1、bN2、bN3、……、bNK是所述隐藏层到所述输出层之间的目标参数，N为输入层的维度，本发明实施例中，已指定了隐藏节点数K+1，可对隐藏层稀疏要求放宽，因此将s的1范数近似为2范数。其中，由于有些用户虽然没有对某些歌曲进行收听行为，但是并不代表这些用户不喜欢这些歌曲，同样的，即使有些用户听过某些歌曲，也不能直接表示这些用户喜欢这些歌曲，所以本发明实施例通过在所述目标函数中的增加了一个用户兴趣因子项C，以提高所训练出来的所述隐藏层中的s矩阵与用户对歌曲的喜好程度之间的关联关系；所述用户兴趣因子项C包括所述各历史用户分别对所述多媒体数据库中各多媒体数据的兴趣值cui(cui表示用户u对歌曲i的兴趣值)；一个兴趣值是基于一个历史用户对一个多媒体数据的操作行为类型、操作次数以及完整操作率计算得到的。其中，对所述用户兴趣因子项C中的各个兴趣值cui的计算公式可以为：cui＝1+αlog(1+εrui)；其中，当用户u直接收藏/下载歌曲i时，rui＝1；当用户u仅仅是收听过歌曲i时(即没有收藏/下载操作)，rui＝[min(nui，5)/5]*fui，其中，nui是指用户u收听歌曲i的次数，fui是完整收听率(即所述完整操作率)，fui＝完整收听歌曲i的人数/所有听歌人数(即所述历史用户群)；当用户u没有收听过歌曲i时，cui＝0。Wherein, x is the multimedia data operation behavior matrix; A is a matrix composed of W1(1) , W2(1) , W3(1) , ..., Wk+1(1) , W1( 1) , W2(1) , W3(1) , ..., Wk+1(1) are hidden parameters in K+1 hidden nodes respectively (for example, W1(1) represents the first hidden node hidden parameter), and the last hidden node in the K+1 hidden nodes is an intercept item, and the hidden parameter of the intercept item is 1, and the output layer can be reconstructed by retaining the intercept item. s is a matrix composed of bN1 , bN2 , bN3 , ..., bNK , bN1 , bN2 , bN3 , ..., bNK are the target parameters between the hidden layer and the output layer , N is the dimension of the input layer. In the embodiment of the present invention, the number of hidden nodes K+1 has been specified, and the requirement for sparseness of the hidden layer can be relaxed, so the 1-norm of s is approximated as a 2-norm. Among them, although some users do not listen to certain songs, it does not mean that these users do not like these songs. Similarly, even if some users have listened to certain songs, it cannot directly indicate that these users like these songs, so this In the embodiment of the invention, a user interest factor item C is added in the objective function to improve the correlation between the trained s matrix in the hidden layer and the user's preference for songs; the user Interest factor item C comprises described each history user respectively to the interest value cui of each multimedia data in the described multimedia database (cu ui represents user u to the interest value of song i); The operation behavior type, number of operations, and complete operation rate of the data are calculated. Wherein, the calculation formula for each interest value cui in the user interest factor item C may be: cui =1+αlog(1+εrui ); wherein, when user u directly collects/downloads song i, rui = 1; when user u has only listened to song i (that is, there is no collection/download operation), rui = [min(nui , 5)/5]*fui , where nui means that user u has listened to song i The number of times of song i, fui is the complete listening rate (i.e. the complete operation rate), fui =the number of people who have listened completely to song i/all the number of people listening to songs (i.e. the historical user group); when user u has not listened to the song When i, cui =0.
所述稀疏自编码器在进行偏导数训练的过程中可以重复交替执行步骤S1和步骤S2；步骤S1为：固定A，对s求导，用最小二乘法得到最优解；步骤S2为：固定s，对A求导，用最小二乘法得到最优解。The sparse autoencoder can repeat and alternately execute step S1 and step S2 in the process of partial derivative training; step S1 is: fixing A, deriving s, and obtaining the optimal solution with the method of least squares; step S2 is: fixing s, take the derivative of A, and use the least square method to get the optimal solution.
S204，当所述稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，并将满足所述收敛条件的各隐藏节点的隐藏参数确定为目标输入源；S204. When the parameter in the output layer of the sparse autoencoder is similar to the parameter in the input layer, determine that the target parameter and the hidden parameter of the hidden node meet the convergence condition, and will satisfy the The hidden parameters of each hidden node of the convergence condition are determined as the target input source;
具体的，所述后台服务器也可以预设至少两个稀疏自编码器，当第一个稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，即第一个稀疏自编码器完成训练，此时，可以将第一个稀疏自编码器中满足所述收敛条件的各隐藏节点的隐藏参数确定为目标输入源。Specifically, the background server can also preset at least two sparse autoencoders, and when the parameters in the output layer of the first sparse autoencoder are similar to the parameters in the input layer, determine the target parameters and the hidden parameters of the hidden nodes meet the convergence condition, that is, the first sparse autoencoder completes the training, at this time, the hidden parameters of each hidden node in the first sparse autoencoder that meets the convergence condition can be determined Enter the source for the target.
S205，根据预设数量的稀疏自编码器，将所述目标输入源输入到下一个稀疏自编码器的输入层，所述下一个稀疏自编码器根据所述目标函数训练所述目标输入源对应的隐藏参数，并将所述下一个稀疏自编码器中训练后的隐藏参数作为目标输入源，重复执行本步骤，直至最后一个稀疏自编码器训练出隐藏参数；S205. According to a preset number of sparse autoencoders, input the target input source to the input layer of the next sparse autoencoder, and the next sparse autoencoder trains the corresponding target input source according to the objective function and using the hidden parameters trained in the next sparse autoencoder as the target input source, repeating this step until the last sparse autoencoder trains the hidden parameters;
具体的，所述后台服务器可以进一步根据预设数量的稀疏自编码器，将所述目标输入源输入到下一个稀疏自编码器的输入层，所述下一个稀疏自编码器根据所述目标函数训练所述目标输入源对应的隐藏参数，并将所述下一个稀疏自编码器中训练后的隐藏参数作为目标输入源，重复执行本步骤，直至最后一个稀疏自编码器训练出隐藏参数。其中，第一个稀疏自编码器所训练出的隐藏参数可以称为一阶特征参数，第二个稀疏自编码器所训练出的隐藏参数可以称为二阶特征参数，以此类推，第n个稀疏自编码器所训练出的隐藏参数可以称为n阶特征参数，n值越大，则n阶特征参数可以越精确的描述用户对多媒体数据的喜好程度。其中，所述第一个稀疏自编码器中的目标函数可以包含所述用户兴趣因子项C，而除了所述第一个稀疏自编码器以外的其他稀疏自编码器中的目标函数可以不包含所述用户兴趣因子项C。其中，每个稀疏自编码器的收敛条件都是相同的，即收敛条件都是为所述输出层中的参数与所述输入层中的参数相近。Specifically, the background server may further input the target input source to the input layer of the next sparse autoencoder according to the preset number of sparse autoencoders, and the next sparse autoencoder is based on the objective function Train the hidden parameters corresponding to the target input source, and use the trained hidden parameters in the next sparse autoencoder as the target input source, and repeat this step until the hidden parameters are trained by the last sparse autoencoder. Among them, the hidden parameters trained by the first sparse autoencoder can be called first-order feature parameters, the hidden parameters trained by the second sparse autoencoder can be called second-order feature parameters, and so on, the nth The hidden parameters trained by the sparse autoencoders can be called n-order feature parameters. The larger the value of n, the more accurately the n-order feature parameters can describe the user's preference for multimedia data. Wherein, the objective function in the first sparse autoencoder may include the user interest factor item C, while the objective functions in other sparse autoencoders except the first sparse autoencoder may not include The user interest factor item C. Wherein, the convergence condition of each sparse autoencoder is the same, that is, the convergence condition is that the parameters in the output layer are similar to the parameters in the input layer.
S206，将所述最后一个稀疏自编码器训练出的隐藏参数组合为隐含特征矩阵；S206. Combine the hidden parameters trained by the last sparse autoencoder into a hidden feature matrix;
S207，当所述稀疏自编码器的数量为至少两个时，获取各历史用户分别对应的个人操作行为信息，分别将各个人操作行为信息中已操作的多媒体数据对应的隐含特征向量进行向量平均计算，并将计算出的各平均向量分别作为各历史用户分别对应的用户特征向量；S207. When the number of the sparse autoencoders is at least two, obtain the individual operation behavior information corresponding to each historical user, and perform a vectorization of the hidden feature vectors corresponding to the operated multimedia data in each individual operation behavior information. Calculate the average, and use the calculated average vectors as the user feature vectors corresponding to the historical users respectively;
具体的，当所述稀疏自编码器的数量为至少两个时，可以获取各历史用户分别对应的个人操作行为信息，分别将各个人操作行为信息中已操作的多媒体数据对应的隐含特征向量进行向量平均计算，并将计算出的各平均向量分别作为各历史用户分别对应的用户特征向量。例如，历史用户A对应的个人操作行为信息可以包括用户A所收听过的歌曲，被用户A收听过的歌曲可以包括歌曲1、歌曲2、歌曲3，则历史用户A对应的用户特征向量可以为歌曲1的隐含特征向量、歌曲2的隐含特征向量以及歌曲3的隐含特征向量的平均向量。Specifically, when the number of sparse autoencoders is at least two, the individual operation behavior information corresponding to each historical user can be obtained, and the hidden feature vector corresponding to the operated multimedia data in each individual operation behavior information can be respectively Carry out vector average calculation, and use the calculated average vectors as user feature vectors corresponding to each historical user respectively. For example, the personal operation behavior information corresponding to historical user A may include songs that user A has listened to, and the songs that user A has listened to may include song 1, song 2, and song 3, then the user feature vector corresponding to historical user A may be The average vector of the hidden feature vector of song 1, the hidden feature vector of song 2, and the hidden feature vector of song 3.
S208，当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，获取所述目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理；S208. When a recommendation request corresponding to the target user is received and the historical user group includes the target user, obtain a plurality of multimedia data in the personal operation behavior information of the target user, and The user feature vector and the implicit feature vector corresponding to each multimedia data in the personal operation behavior information respectively perform recommendation processing on a plurality of multimedia data in the personal operation behavior information;
其中，S208步骤的具体实现方式可以参见上述图2对应实施例中的S101，这里不再进行赘述。Wherein, for the specific implementation manner of step S208, reference may be made to S101 in the above-mentioned embodiment corresponding to FIG. 2 , which will not be repeated here.
本发明实施例通过根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵，并基于稀疏自编码神经网络，并根据多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量，由此可见，隐含特征向量可以准确表征历史用户群对一个多媒体数据的喜好程度信息，且用户特征向量可以准确表征一个历史用户对多个多媒体数据的喜好程度信息，所以通过隐含特征向量和用户特征向量可以对目标用户实现准确的个性化推荐，即可保证所推荐的歌曲是目标用户所喜欢的歌曲，以提高推荐效果。The embodiment of the present invention generates a multimedia data operation behavior matrix based on the operation behavior of historical user groups on multiple multimedia data in the preset multimedia database, and calculates each multimedia data operation behavior matrix based on the sparse self-encoding neural network. The hidden eigenvectors corresponding to the data and the user eigenvectors corresponding to each historical user respectively. It can be seen that the hidden eigenvectors can accurately represent the preference information of a historical user group for a piece of multimedia data, and the user eigenvectors can accurately represent a Historical users’ preferences for multiple multimedia data, so accurate personalized recommendations can be made to target users through hidden feature vectors and user feature vectors, which can ensure that the recommended songs are songs that target users like, so as to improve Recommended effect.
请参见图4，是本发明实施例提供的一种多媒体数据处理装置的结构示意图。所述多媒体数据处理装置1可以应用于后台服务器中，所述多媒体数据处理装置1可以包括：矩阵生成模块10、特征计算模块20、推荐模块30；Please refer to FIG. 4 , which is a schematic structural diagram of a multimedia data processing device provided by an embodiment of the present invention. The multimedia data processing device 1 can be applied in a background server, and the multimedia data processing device 1 can include: a matrix generation module 10, a feature calculation module 20, and a recommendation module 30;
所述矩阵生成模块10，用于根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵；The matrix generation module 10 is used to generate a multimedia data operation behavior matrix according to the operation behavior of a plurality of multimedia data in the preset multimedia database according to historical user groups;
具体的，所述矩阵生成模块10可以获取历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为。所述多媒体数据库可以为音乐库，每一个多媒体数据均可以为音乐库中的一首歌，因此，所述操作行为可以包括所述历史用户群中的各历史用户对各个多媒体数据的收听行为，所述收听行为包括已收听行为和未收听行为，所述矩阵生成模块10可以为不同的收听行为设置不同的特征值，以生成与所述操作行为相应的多媒体数据操作行为矩阵。再以上述图2对应实施例中的表1为例，表1中的特征值“1”表示该用户收听过这首歌曲，特征值“0”表示该用户未收听过这首歌曲，如用户1与歌曲1之间的特征值为“1”，则说明用户1收听过歌曲1；又如用户3与歌曲2之间的特征值为“0”，则说明用户3未收听过歌曲2。因此，所述矩阵生成模块10根据所述表1中的所有特征值即可生成与所述操作行为相应的多媒体数据操作行为矩阵，即所述多媒体数据操作行为矩阵中的元素Pui可以表示用户u对歌曲i的收听行为对应的特征值(Pui＝1，说明用户u收听过歌曲i；Pui＝0，说明用户u未收听歌曲i)。Specifically, the matrix generation module 10 may acquire the operation behaviors of historical user groups on multiple pieces of multimedia data in the preset multimedia database. The multimedia database can be a music library, and each piece of multimedia data can be a song in the music library. Therefore, the operation behavior can include the listening behavior of each historical user in the historical user group to each multimedia data, The listening behavior includes listening behavior and non-listening behavior, and the matrix generation module 10 can set different feature values for different listening behaviors to generate a multimedia data operation behavior matrix corresponding to the operation behavior. Taking Table 1 in the above-mentioned embodiment corresponding to Figure 2 as an example, the feature value "1" in Table 1 indicates that the user has listened to this song, and the feature value "0" indicates that the user has not listened to this song, such as the user If the feature value between 1 and song 1 is "1", it means that user 1 has listened to song 1; if the feature value between user 3 and song 2 is "0", it means that user 3 has not listened to song 2. Therefore, the matrix generation module 10 can generate the multimedia data operation behavior matrix corresponding to the operation behavior according to all the eigenvalues in the table 1, that is, the element Pui in the multimedia data operation behavior matrix can represent the user Feature values corresponding to u's listening behavior to song i (Pui =1, indicating that user u has listened to song i; Pui =0, indicating that user u has not listened to song i).
所述特征计算模块20，用于基于稀疏自编码神经网络，并根据所述多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量；一个隐含特征向量表征所述历史用户群对一个多媒体数据的喜好程度信息；一个用户特征向量表征一个历史用户对所述多个多媒体数据的喜好程度信息；The feature calculation module 20 is used to calculate the hidden feature vectors corresponding to each multimedia data and the user feature vectors corresponding to each historical user respectively according to the multimedia data operation behavior matrix based on the sparse self-encoding neural network; A characteristic vector characterizes the preference degree information of the historical user group to a piece of multimedia data; a user characteristic vector characterizes the preference degree information of a historical user group to the plurality of multimedia data;
具体的，请一并参见图5，是所述特征计算模块20的结构示意图，所述特征计算模块20可以包括：输入单元201、训练单元202、隐含特征生成单元203、用户特征生成单元204；Specifically, please refer to FIG. 5 , which is a schematic structural diagram of the feature calculation module 20. The feature calculation module 20 may include: an input unit 201, a training unit 202, a hidden feature generation unit 203, and a user feature generation unit 204. ;
所述输入单元201，用于将所述多媒体数据操作行为矩阵输入到所述稀疏自编码神经网络对应的稀疏自编码器的输入层；所述稀疏自编码器包括所述输入层、隐藏层、输出层以及所述隐藏层与所述输出层之间的目标参数；所述隐藏层包括预设数量的隐藏节点；The input unit 201 is configured to input the multimedia data operation behavior matrix to the input layer of the sparse autoencoder corresponding to the sparse autoencoder neural network; the sparse autoencoder includes the input layer, hidden layer, an output layer and target parameters between the hidden layer and the output layer; the hidden layer includes a preset number of hidden nodes;
所述训练单元202，用于控制所述稀疏自编码器根据所述输入层中的参数以及预设的用于训练所述目标参数和所述隐藏节点的隐藏参数的目标函数，对所述目标参数和所述隐藏节点的隐藏参数进行偏导数训练；The training unit 202 is configured to control the sparse autoencoder to perform the training on the target according to the parameters in the input layer and the preset objective function for training the target parameters and the hidden parameters of the hidden nodes. Parameter and the hidden parameter of described hidden node carry out partial derivative training;
所述隐含特征生成单元203，用于当所述稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，并将满足所述收敛条件的各隐藏节点的隐藏参数组合成隐含特征矩阵；所述隐含特征矩阵为所述多媒体数据操作行为矩阵对应的压缩矩阵，所述隐含特征矩阵包括各多媒体数据分别对应的隐含特征向量；The hidden feature generation unit 203 is configured to determine the target parameter and the hidden parameter of the hidden node when the parameter in the output layer of the sparse autoencoder is similar to the parameter in the input layer Satisfy the convergence condition, and combine the hidden parameters of each hidden node satisfying the convergence condition into a hidden feature matrix; the hidden feature matrix is a compression matrix corresponding to the multimedia data operation behavior matrix, and the hidden feature matrix including hidden feature vectors corresponding to each multimedia data;
所述用户特征生成单元204，用于根据训练后的目标参数或所述隐含特征矩阵计算所述历史用户群中各历史用户分别对应的用户特征向量；The user feature generating unit 204 is configured to calculate user feature vectors corresponding to each historical user in the historical user group according to the trained target parameters or the hidden feature matrix;
其中，当所述稀疏自编码器的数量为一个时，所计算出来的隐藏参数为一阶特征参数，即所述隐含特征矩阵包括所述一阶特征参数。所述输入单元201、所述训练单元202、所述隐含特征生成单元203、所述用户特征生成单元204的具体实现方式可以参见上述图2对应实施例中的S102，这里不再进行赘述。Wherein, when the number of the sparse autoencoder is one, the calculated hidden parameters are first-order feature parameters, that is, the hidden feature matrix includes the first-order feature parameters. For specific implementations of the input unit 201 , the training unit 202 , the hidden feature generation unit 203 , and the user feature generation unit 204 , refer to S102 in the above-mentioned embodiment corresponding to FIG. 2 , which will not be repeated here.
所述推荐模块30，用于当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，获取所述目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理；The recommendation module 30 is configured to acquire a plurality of multimedia data in the personal operation behavior information of the target user when a recommendation request corresponding to the target user is received and the historical user group includes the target user, and performing recommendation processing on a plurality of multimedia data in the personal operation behavior information according to the user feature vector corresponding to the target user and the implicit feature vectors respectively corresponding to each multimedia data in the personal operation behavior information;
具体的，请一并参见图6，是所述推荐模块30的结构示意图，所述推荐模块30可以包括：检测单元301、相似推荐单元302、获取判断单元303、点乘运算单元304、排序推荐单元305；Specifically, please refer to FIG. 6, which is a schematic structural diagram of the recommendation module 30. The recommendation module 30 may include: a detection unit 301, a similar recommendation unit 302, an acquisition judgment unit 303, a dot product operation unit 304, and a sorting recommendation Unit 305;
所述检测单元301，用于当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，检测目标用户对应的个人操作行为信息中是否包含已收藏的多媒体数据；The detection unit 301 is configured to detect whether the personal operation behavior information corresponding to the target user contains the favorite multimedia data when receiving a recommendation request corresponding to the target user and the historical user group includes the target user;
所述相似推荐单元302，用于若所述检测单元301检测为包含已收藏的多媒体数据，则获取所述已收藏的多媒体数据对应的第一相似多媒体数据，并将所述第一相似多媒体数据作为所述目标用户的推荐数据；The similar recommending unit 302 is configured to obtain first similar multimedia data corresponding to the collected multimedia data if the detection unit 301 detects that it contains the collected multimedia data, and store the first similar multimedia data As the recommendation data of the target user;
所述获取判断单元303，用于若所述检测单元301检测为未包含已收藏的多媒体数据，则进一步判断所述个人操作行为信息中是否包含已完整操作的多媒体数据；The acquisition judging unit 303 is configured to further judge whether the personal operation behavior information includes multimedia data that has been completely operated if the detection unit 301 detects that the collected multimedia data is not included;
所述点乘运算单元304，用于当所述个人操作行为信息中包含已完整操作的多媒体数据时，将所述目标用户对应的用户特征向量与所述已完整操作的多媒体数据对应的隐含特征向量进行点乘运算，得到个性化特征值；The dot product calculation unit 304 is configured to, when the personal operation behavior information includes multimedia data that has been fully operated, implicitly represent the user feature vector corresponding to the target user with the multimedia data that has been fully operated. Dot multiplication of eigenvectors to obtain personalized eigenvalues;
所述相似推荐单元302，还用于当所述个性化特征值大于预设特征值阈值时，获取所述已完整操作的多媒体数据对应的第二相似多媒体数据，并将所述第二相似多媒体数据作为所述目标用户的推荐数据；The similar recommendation unit 302 is further configured to obtain second similar multimedia data corresponding to the fully operated multimedia data when the personalized feature value is greater than a preset feature value threshold, and store the second similar multimedia data The data is used as the recommendation data of the target user;
所述点乘运算单元304，还用于当所述个人操作行为信息不包含已完整操作的多媒体数据时，获取多个候选多媒体数据，将所述标用户对应的用户特征向量分别与各候选多媒体数据对应的隐含特征向量进行点乘运算，得到所述各候选多媒体数据分别对应的个性化特征值；The dot product operation unit 304 is also used to obtain a plurality of candidate multimedia data when the personal operation behavior information does not contain fully operated multimedia data, and compare the user feature vector corresponding to the marked user with each candidate multimedia data respectively. performing dot multiplication on the implicit feature vectors corresponding to the data, to obtain individualized feature values corresponding to the respective candidate multimedia data;
所述排序推荐单元305，用于按照各个性化特征值对所述多个候选多媒体数据进行排序，并根据排序结果将预设推荐数量的候选多媒体数据作为所述目标用户的推荐数据；The sorting and recommending unit 305 is configured to sort the multiple candidate multimedia data according to each personalized feature value, and use a preset recommended number of candidate multimedia data as the recommended data for the target user according to the sorting result;
其中，所述检测单元301、所述相似推荐单元302、所述获取判断单元303、所述点乘运算单元304、所述排序推荐单元305的具体实现方式可以参见上述图2对应实施例中的S103，这里不再进行赘述。Wherein, the specific implementation manners of the detection unit 301, the similarity recommendation unit 302, the acquisition judgment unit 303, the dot product operation unit 304, and the ranking recommendation unit 305 can be referred to in the above-mentioned embodiment corresponding to FIG. 2 S103, no more details here.
可选的，所述推荐模块30在计算所述个性化特征值时，还可以参考显性特征值，即用户u对歌曲i的个性化特征值＝用户u对歌曲i的显性特征值+用户u的用户特征向量与歌曲i的隐含特征向量的点乘结果。所述显示特征值可以根据歌曲的标签和用户的标签之间的匹配度计算得到；歌曲的标签可以包括该歌曲对应的曲风类型、语种类型、节奏类型等等；用户的标签可以包括该用户所喜欢的曲风类型、语种类型、节奏类型等等。通过增加所述显性特征值，可以使所述个性化特征值能够更准确的描述用户u对歌曲i的喜好程度。Optionally, when calculating the personalized feature value, the recommendation module 30 may also refer to the dominant feature value, that is, the personalized feature value of user u for song i=the dominant feature value of user u for song i+ The dot product result of the user feature vector of user u and the hidden feature vector of song i. The display feature value can be calculated according to the matching degree between the label of the song and the label of the user; the label of the song can include the genre type, language type, rhythm type, etc. corresponding to the song; the label of the user can include the user's label Favorite music style, language type, rhythm type, etc. By increasing the dominant feature value, the personalized feature value can more accurately describe user u's preference for song i.
可选的，当接收到与目标用户对应的推荐请求，且所述历史用户群不包含所述目标用户时，说明所述目标用户还未在后台服务器所提供的多媒体数据平台上收听过歌曲，则所述推荐模块30可以按照所述各候选多媒体数据对应的隐含特征向量的向量值从大到小的顺序对所述多个候选多媒体数据进行排序，并根据排序结果将预设推荐数量的候选多媒体数据作为所述目标用户的推荐数据。Optionally, when a recommendation request corresponding to the target user is received, and the historical user group does not include the target user, it means that the target user has not listened to songs on the multimedia data platform provided by the background server, Then the recommendation module 30 can sort the plurality of candidate multimedia data according to the order of the vector values of the implicit feature vectors corresponding to the candidate multimedia data from large to small, and according to the sorting results, the preset recommended number Candidate multimedia data is used as recommendation data for the target user.
可选的，所述特征计算模块20可以周期性的根据新的操作行为计算和更新各历史用户分别对应的用户特征向量，以及各多媒体数据分别对应的隐含特征向量，使得在对目标用户进行推荐时，可以保证所获取到的与所述目标用户相关的隐含特征向量和用户特征向量是始终贴合用户的兴趣变化，即可以保证所述推荐模块30所计算出的个性化特征值的准确性。其中，所述目标用户每听一首歌曲，所述推荐模块30均可更新一次所述目标用户的推荐源，以便于后续在向所述目标用户推荐歌曲时，所述推荐模块30可以直接获取与最近更新的推荐源相似的多媒体数据，以将相似的多媒体数据推荐给用户，以提高推荐效率；而在客户端的用户界面上可以显示“歌曲B是根据你试听的歌曲A推荐”，其中，歌曲A即为推荐源，歌曲B即为与推荐源相似的多媒体数据。Optionally, the feature calculation module 20 can periodically calculate and update the user feature vectors corresponding to each historical user and the hidden feature vectors corresponding to each multimedia data according to the new operation behavior, so that when the target user is When recommending, it can be ensured that the acquired implicit feature vectors and user feature vectors related to the target user are always in line with the user's interest changes, that is, it can be guaranteed that the personalized feature values calculated by the recommendation module 30 are consistent with each other. accuracy. Wherein, each time the target user listens to a song, the recommendation module 30 can update the recommendation source of the target user once, so that when recommending songs to the target user, the recommendation module 30 can directly obtain Similar multimedia data to the recently updated recommendation source to recommend similar multimedia data to users to improve recommendation efficiency; and on the user interface of the client, it can be displayed that "song B is recommended according to the song A you tried to listen to", wherein, Song A is the recommendation source, and song B is multimedia data similar to the recommendation source.
进一步的，请一并参见图7，是本发明实施例提供的一种隐含特征生成单元203的结构示意图，所述隐含特征生成单元203包括：确定子单元2031、深度学习子单元2032、组合子单元2033；Further, please refer to FIG. 7, which is a schematic structural diagram of a hidden feature generation unit 203 provided by an embodiment of the present invention. The hidden feature generation unit 203 includes: a determination subunit 2031, a deep learning subunit 2032, Combination subunit 2033;
所述确定子单元2031，用于当所述稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，并将满足所述收敛条件的各隐藏节点的隐藏参数确定为目标输入源；The determining subunit 2031 is configured to determine that the target parameter and the hidden parameter of the hidden node satisfy convergence when the parameter in the output layer of the sparse autoencoder is similar to the parameter in the input layer condition, and determine the hidden parameters of each hidden node satisfying the convergence condition as the target input source;
所述深度学习子单元2032，用于根据预设数量的稀疏自编码器，将所述目标输入源输入到下一个稀疏自编码器的输入层，所述下一个稀疏自编码器根据所述目标函数训练所述目标输入源对应的隐藏参数，并将所述下一个稀疏自编码器中训练后的隐藏参数作为目标输入源，重复执行本步骤，直至最后一个稀疏自编码器训练出隐藏参数；The deep learning subunit 2032 is configured to input the target input source to the input layer of the next sparse autoencoder according to the preset number of sparse autoencoders, and the next sparse autoencoder is based on the target The function trains the hidden parameters corresponding to the target input source, and uses the trained hidden parameters in the next sparse autoencoder as the target input source, and repeats this step until the last sparse autoencoder trains the hidden parameters;
所述组合子单元2033，用于将所述最后一个稀疏自编码器训练出的隐藏参数组合为隐含特征矩阵；The combining subunit 2033 is configured to combine the hidden parameters trained by the last sparse autoencoder into a hidden feature matrix;
其中，第一个稀疏自编码器所训练出的隐藏参数可以称为一阶特征参数，第二个稀疏自编码器所训练出的隐藏参数可以称为二阶特征参数，以此类推，第n个稀疏自编码器所训练出的隐藏参数可以称为n阶特征参数，n值越大，则n阶特征参数可以越精确的描述用户对多媒体数据的喜好程度。所述确定子单元2031、所述深度学习子单元2032、所述组合子单元2033的具体实现方式可以参见上述图3对应实施例中的S204-S206，这里不再进行赘述。Among them, the hidden parameters trained by the first sparse autoencoder can be called first-order feature parameters, the hidden parameters trained by the second sparse autoencoder can be called second-order feature parameters, and so on, the nth The hidden parameters trained by the sparse autoencoders can be called n-order feature parameters. The larger the value of n, the more accurately the n-order feature parameters can describe the user's preference for multimedia data. For specific implementation manners of the determination subunit 2031 , the deep learning subunit 2032 , and the combination subunit 2033 , refer to S204-S206 in the above-mentioned embodiment corresponding to FIG. 3 , which will not be repeated here.
进一步的，再请一并参见图8，是本发明实施例提供的一种用户特征生成单元204的结构示意图，所述用户特征生成单元204包括：提取子单元2041、平均计算子单元2042；Further, please refer to FIG. 8 , which is a schematic structural diagram of a user feature generation unit 204 provided by an embodiment of the present invention. The user feature generation unit 204 includes: an extraction subunit 2041 and an average calculation subunit 2042;
所述提取子单元2041，用于当所述稀疏自编码器的数量为一个时，从所述稀疏自编码器中训练后的目标参数对应的参数矩阵中提取出所述历史用户群中各历史用户分别对应的用户特征向量；The extracting subunit 2041 is configured to extract each history in the historical user group from the parameter matrix corresponding to the target parameter after training in the sparse autoencoder when the number of the sparse autoencoder is one. The user feature vectors corresponding to the users respectively;
所述平均计算子单元2042，用于当所述稀疏自编码器的数量为至少两个时，获取各历史用户分别对应的个人操作行为信息，分别将各个人操作行为信息中已操作的多媒体数据对应的隐含特征向量进行向量平均计算，并将计算出的各平均向量分别作为各历史用户分别对应的用户特征向量；The average calculation subunit 2042 is used to obtain the individual operation behavior information corresponding to each historical user when the number of the sparse autoencoder is at least two, and respectively convert the operated multimedia data in the operation behavior information of each individual to Perform vector average calculation on the corresponding hidden feature vectors, and use the calculated average vectors as user feature vectors corresponding to each historical user respectively;
具体的，当所述稀疏自编码器的数量为至少两个时，所述平均计算子单元2042可以获取各历史用户分别对应的个人操作行为信息，分别将各个人操作行为信息中已操作的多媒体数据对应的隐含特征向量进行向量平均计算，并将计算出的各平均向量分别作为各历史用户分别对应的用户特征向量。例如，历史用户A对应的个人操作行为信息可以包括用户A所收听过的歌曲，被用户A收听过的歌曲可以包括歌曲1、歌曲2、歌曲3，则历史用户A对应的用户特征向量可以为歌曲1的隐含特征向量、歌曲2的隐含特征向量以及歌曲3的隐含特征向量的平均向量。Specifically, when the number of the sparse autoencoders is at least two, the average calculation subunit 2042 can obtain the individual operation behavior information corresponding to each historical user, and respectively combine the operated multimedia The hidden feature vectors corresponding to the data are calculated as vector averages, and the calculated average vectors are respectively used as user feature vectors corresponding to each historical user. For example, the personal operation behavior information corresponding to historical user A may include songs that user A has listened to, and the songs that user A has listened to may include song 1, song 2, and song 3, then the user feature vector corresponding to historical user A may be The average vector of the hidden feature vector of song 1, the hidden feature vector of song 2, and the hidden feature vector of song 3.
本发明实施例通过根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵，并基于稀疏自编码神经网络，并根据多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量，由此可见，隐含特征向量可以准确表征历史用户群对一个多媒体数据的喜好程度信息，且用户特征向量可以准确表征一个历史用户对多个多媒体数据的喜好程度信息，所以通过隐含特征向量和用户特征向量可以对目标用户实现准确的个性化推荐，即可保证所推荐的歌曲是目标用户所喜欢的歌曲，以提高推荐效果。The embodiment of the present invention generates a multimedia data operation behavior matrix based on the operation behavior of historical user groups on multiple multimedia data in the preset multimedia database, and calculates each multimedia data operation behavior matrix based on the sparse self-encoding neural network. The hidden eigenvectors corresponding to the data and the user eigenvectors corresponding to each historical user respectively. It can be seen that the hidden eigenvectors can accurately represent the preference information of a historical user group for a piece of multimedia data, and the user eigenvectors can accurately represent a Historical users’ preferences for multiple multimedia data, so accurate personalized recommendations can be made to target users through hidden feature vectors and user feature vectors, which can ensure that the recommended songs are songs that target users like, so as to improve Recommended effect.
请参见图9，是本发明实施例提供的另一种多媒体数据处理装置的结构示意图。如图9所示，所述多媒体数据处理装置1000可以应用于后台服务器中，所述多媒体数据处理装置1000可以包括：至少一个处理器1001，例如CPU，至少一个网络接口1004，用户接口1003，存储器1005，至少一个通信总线1002。其中，通信总线1002用于实现这些组件之间的连接通信。其中，用户接口1003可以包括显示屏(Display)、键盘(Keyboard)，可选用户接口1003还可以包括标准的有线接口、无线接口。网络接口1004可选的可以包括标准的有线接口、无线接口(如WI-FI接口)。存储器1005可以是高速RAM存储器，也可以是非不稳定的存储器(non-volatile memory)，例如至少一个磁盘存储器。存储器1005可选的还可以是至少一个位于远离前述处理器1001的存储装置。如图9所示，作为一种计算机存储介质的存储器1005中可以包括操作系统、网络通信模块、用户接口模块以及设备控制应用程序。Please refer to FIG. 9 , which is a schematic structural diagram of another multimedia data processing device provided by an embodiment of the present invention. As shown in Figure 9, the multimedia data processing device 1000 can be applied to a background server, and the multimedia data processing device 1000 can include: at least one processor 1001, such as a CPU, at least one network interface 1004, a user interface 1003, and a memory 1005. At least one communication bus 1002. Wherein, the communication bus 1002 is used to realize connection and communication between these components. Wherein, the user interface 1003 may include a display screen (Display) and a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. Optionally, the network interface 1004 may include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a non-volatile memory, such as at least one disk memory. Optionally, the memory 1005 may also be at least one storage device located away from the aforementioned processor 1001 . As shown in FIG. 9 , the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a device control application program.
在图9所示的多媒体数据处理装置1000中，网络接口1004主要用于连接客户端，以向客户端推荐多媒体数据；而用户接口1003主要用于为用户提供输入的接口，获取用户输出的数据；而处理器1001可以用于调用存储器1005中存储的设备控制应用程序，以实现In the multimedia data processing device 1000 shown in FIG. 9 , the network interface 1004 is mainly used to connect to the client to recommend multimedia data to the client; and the user interface 1003 is mainly used to provide the user with an input interface and obtain the data output by the user. ; and the processor 1001 can be used to call the device control application program stored in the memory 1005 to realize
根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵；Generate a multimedia data operation behavior matrix according to the operation behaviors of the historical user groups on the multiple multimedia data in the preset multimedia database;
基于稀疏自编码神经网络，并根据所述多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量；一个隐含特征向量表征所述历史用户群对一个多媒体数据的喜好程度信息；一个用户特征向量表征一个历史用户对所述多个多媒体数据的喜好程度信息；Based on the sparse self-encoding neural network, and according to the multimedia data operation behavior matrix, the hidden feature vectors corresponding to each multimedia data and the user feature vectors corresponding to each historical user are calculated respectively; a hidden feature vector represents the pair of historical user groups A preference degree information of multimedia data; a user feature vector characterizes preference degree information of a historical user for the plurality of multimedia data;
当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，获取所述目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理。When a recommendation request corresponding to the target user is received and the historical user group includes the target user, multiple multimedia data in the personal operation behavior information of the target user are obtained, and according to the user corresponding to the target user The feature vectors and the implicit feature vectors respectively corresponding to the multimedia data in the personal operation behavior information perform recommendation processing on the multiple multimedia data in the personal operation behavior information.
在一个实施例中，所述处理器1001在执行基于稀疏自编码神经网络，并根据所述多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量时，具体执行以下步骤：In one embodiment, the processor 1001 is executing a neural network based on sparse self-encoding, and calculates the hidden feature vectors corresponding to each multimedia data and the user feature vectors corresponding to each historical user according to the multimedia data operation behavior matrix. , specifically perform the following steps:
将所述多媒体数据操作行为矩阵输入到所述稀疏自编码神经网络对应的稀疏自编码器的输入层；所述稀疏自编码器包括所述输入层、隐藏层、输出层以及所述隐藏层与所述输出层之间的目标参数；所述隐藏层包括预设数量的隐藏节点；The multimedia data operation behavior matrix is input to the input layer of the sparse autoencoder corresponding to the sparse autoencoder neural network; the sparse autoencoder includes the input layer, hidden layer, output layer, and the hidden layer and target parameters between the output layers; the hidden layer includes a preset number of hidden nodes;
控制所述稀疏自编码器根据所述输入层中的参数以及预设的用于训练所述目标参数和所述隐藏节点的隐藏参数的目标函数，对所述目标参数和所述隐藏节点的隐藏参数进行偏导数训练；Controlling the sparse autoencoder to hide the target parameters and the hidden nodes according to the parameters in the input layer and the preset objective function for training the hidden parameters of the target parameters and the hidden nodes Parameters for partial derivative training;
当所述稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，并将满足所述收敛条件的各隐藏节点的隐藏参数组合成隐含特征矩阵；所述隐含特征矩阵为所述多媒体数据操作行为矩阵对应的压缩矩阵，所述隐含特征矩阵包括各多媒体数据分别对应的隐含特征向量；When the parameters in the output layer of the sparse autoencoder are similar to the parameters in the input layer, it is determined that the target parameters and the hidden parameters of the hidden nodes satisfy a convergence condition and will satisfy the convergence condition The hidden parameters of each hidden node are combined into a hidden feature matrix; the hidden feature matrix is a compression matrix corresponding to the multimedia data operation behavior matrix, and the hidden feature matrix includes hidden feature vectors corresponding to each multimedia data ;
根据训练后的目标参数或所述隐含特征矩阵计算所述历史用户群中各历史用户分别对应的用户特征向量。Calculate user feature vectors corresponding to each historical user in the historical user group according to the trained target parameters or the hidden feature matrix.
在一个实施例中，所述处理器1001在执行当所述稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，并将满足所述收敛条件的各隐藏节点的隐藏参数组合成隐含特征矩阵时，具体执行以下步骤：In one embodiment, the processor 1001 determines the target parameters and the hidden nodes when the parameters in the output layer of the sparse autoencoder are similar to the parameters in the input layer. When the hidden parameters meet the convergence conditions, and the hidden parameters of each hidden node satisfying the convergence conditions are combined into a hidden feature matrix, the following steps are specifically performed:
当所述稀疏自编码器的所述输出层中的参数与所述输入层中的参数相近时，确定所述目标参数和所述隐藏节点的隐藏参数满足收敛条件，并将满足所述收敛条件的各隐藏节点的隐藏参数确定为目标输入源；When the parameters in the output layer of the sparse autoencoder are similar to the parameters in the input layer, it is determined that the target parameters and the hidden parameters of the hidden nodes satisfy a convergence condition and will satisfy the convergence condition The hidden parameters of each hidden node of are determined as the target input source;
根据预设数量的稀疏自编码器，将所述目标输入源输入到下一个稀疏自编码器的输入层，所述下一个稀疏自编码器根据所述目标函数训练所述目标输入源对应的隐藏参数，并将所述下一个稀疏自编码器中训练后的隐藏参数作为目标输入源，重复执行本步骤，直至最后一个稀疏自编码器训练出隐藏参数；According to the preset number of sparse autoencoders, the target input source is input to the input layer of the next sparse autoencoder, and the next sparse autoencoder trains the hidden layer corresponding to the target input source according to the objective function parameters, and using the hidden parameters trained in the next sparse autoencoder as the target input source, repeating this step until the hidden parameters are trained by the last sparse autoencoder;
将所述最后一个稀疏自编码器训练出的隐藏参数组合为隐含特征矩阵。Combining the hidden parameters trained by the last sparse autoencoder into a hidden feature matrix.
在一个实施例中，所述目标函数包括预设的用户兴趣因子项，所述用户兴趣因子项包括所述各历史用户分别对所述多媒体数据库中各多媒体数据的兴趣值；In one embodiment, the objective function includes a preset user interest factor item, and the user interest factor item includes the interest values of each historical user for each multimedia data in the multimedia database;
一个兴趣值是基于一个历史用户对一个多媒体数据的操作行为类型、操作次数以及完整操作率计算得到的。An interest value is calculated based on a historical user's operation behavior type, operation times and complete operation rate on a piece of multimedia data.
在一个实施例中，所述处理器1001在执行根据训练后的目标参数或所述隐含特征矩阵计算所述历史用户群中各历史用户分别对应的用户特征向量时，具体执行以下步骤：In one embodiment, when the processor 1001 calculates the user feature vectors corresponding to the historical users in the historical user group according to the trained target parameters or the hidden feature matrix, the following steps are specifically performed:
当所述稀疏自编码器的数量为一个时，从所述稀疏自编码器中训练后的目标参数对应的参数矩阵中提取出所述历史用户群中各历史用户分别对应的用户特征向量；When the number of the sparse self-encoder is one, the user feature vector corresponding to each historical user in the historical user group is extracted from the parameter matrix corresponding to the target parameter after training in the sparse self-encoder;
当所述稀疏自编码器的数量为至少两个时，获取各历史用户分别对应的个人操作行为信息，分别将各个人操作行为信息中已操作的多媒体数据对应的隐含特征向量进行向量平均计算，并将计算出的各平均向量分别作为各历史用户分别对应的用户特征向量。When the number of the sparse self-encoder is at least two, the individual operation behavior information corresponding to each historical user is obtained, and the hidden feature vectors corresponding to the operated multimedia data in the operation behavior information of each individual are respectively subjected to vector average calculation. , and the calculated average vectors are respectively used as user feature vectors corresponding to each historical user.
在一个实施例中，所述处理器1001在执行当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，获取所述目标用户的个人操作行为信息中的多个多媒体数据，并根据所述目标用户对应的用户特征向量以及所述个人操作行为信息中各多媒体数据分别对应的隐含特征向量对所述个人操作行为信息中的多个多媒体数据进行推荐处理时，具体执行以下步骤：In one embodiment, when the processor 1001 receives a recommendation request corresponding to the target user and the historical user group includes the target user, acquire more than one part of the target user's personal operation behavior information. multimedia data, and according to the user feature vector corresponding to the target user and the implicit feature vector corresponding to each multimedia data in the personal operation behavior information, when recommending multiple multimedia data in the personal operation behavior information. , perform the following steps:
当接收到与目标用户对应的推荐请求，且所述历史用户群包含所述目标用户时，检测目标用户对应的个人操作行为信息中是否包含已收藏的多媒体数据；When a recommendation request corresponding to a target user is received, and the historical user group includes the target user, it is detected whether the personal operation behavior information corresponding to the target user includes favorite multimedia data;
若检测为包含已收藏的多媒体数据，则获取所述已收藏的多媒体数据对应的第一相似多媒体数据，并将所述第一相似多媒体数据作为所述目标用户的推荐数据；If it is detected as including the collected multimedia data, then obtain the first similar multimedia data corresponding to the collected multimedia data, and use the first similar multimedia data as the recommended data for the target user;
若检测为未包含已收藏的多媒体数据，则进一步判断所述个人操作行为信息中是否包含已完整操作的多媒体数据；If it is detected that the multimedia data that has been collected is not included, it is further judged whether the multimedia data that has been fully operated is included in the personal operation behavior information;
当所述个人操作行为信息中包含已完整操作的多媒体数据时，将所述目标用户对应的用户特征向量与所述已完整操作的多媒体数据对应的隐含特征向量进行点乘运算，得到个性化特征值；When the personal operation behavior information includes multimedia data that has been fully operated, dot product operation is performed on the user feature vector corresponding to the target user and the hidden feature vector corresponding to the multimedia data that has been fully operated to obtain personalized Eigenvalues;
当所述个性化特征值大于预设特征值阈值时，获取所述已完整操作的多媒体数据对应的第二相似多媒体数据，并将所述第二相似多媒体数据作为所述目标用户的推荐数据。When the personalized feature value is greater than a preset feature value threshold, acquire second similar multimedia data corresponding to the fully manipulated multimedia data, and use the second similar multimedia data as recommendation data for the target user.
在一个实施例中，所述处理器1001还执行以下步骤：In one embodiment, the processor 1001 also performs the following steps:
当所述个人操作行为信息不包含已完整操作的多媒体数据时，获取多个候选多媒体数据，将所述标用户对应的用户特征向量分别与各候选多媒体数据对应的隐含特征向量进行点乘运算，得到所述各候选多媒体数据分别对应的个性化特征值；When the personal operation behavior information does not include multimedia data that has been fully operated, multiple candidate multimedia data are obtained, and the user feature vector corresponding to the target user is respectively subjected to a dot product operation with the hidden feature vector corresponding to each candidate multimedia data. , obtaining individualized feature values corresponding to each candidate multimedia data;
按照各个性化特征值对所述多个候选多媒体数据进行排序，并根据排序结果将预设推荐数量的候选多媒体数据作为所述目标用户的推荐数据。Sorting the plurality of candidate multimedia data according to individualized feature values, and using a preset recommended number of candidate multimedia data as recommended data for the target user according to the ranking results.
本发明实施例通过根据历史用户群对预设的多媒体数据库中的多个多媒体数据的操作行为，生成多媒体数据操作行为矩阵，并基于稀疏自编码神经网络，并根据多媒体数据操作行为矩阵计算各多媒体数据分别对应的隐含特征向量和各历史用户分别对应的用户特征向量，由此可见，隐含特征向量可以准确表征历史用户群对一个多媒体数据的喜好程度信息，且用户特征向量可以准确表征一个历史用户对多个多媒体数据的喜好程度信息，所以通过隐含特征向量和用户特征向量可以对目标用户实现准确的个性化推荐，即可保证所推荐的歌曲是目标用户所喜欢的歌曲，以提高推荐效果。The embodiment of the present invention generates a multimedia data operation behavior matrix based on the operation behavior of historical user groups on multiple multimedia data in the preset multimedia database, and calculates each multimedia data operation behavior matrix based on the sparse self-encoding neural network. The hidden eigenvectors corresponding to the data and the user eigenvectors corresponding to each historical user respectively. It can be seen that the hidden eigenvectors can accurately represent the preference information of a historical user group for a piece of multimedia data, and the user eigenvectors can accurately represent a Historical users’ preferences for multiple multimedia data, so accurate personalized recommendations can be made to target users through hidden feature vectors and user feature vectors, which can ensure that the recommended songs are songs that target users like, so as to improve Recommended effect.
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，所述的程序可存储于一计算机可读取存储介质中，该程序在执行时，可包括如上述各方法的实施例的流程。其中，所述的存储介质可为磁碟、光盘、只读存储记忆体(Read-Only Memory，ROM)或随机存储记忆体(Random AccessMemory，RAM)等。Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented through computer programs to instruct related hardware, and the programs can be stored in a computer-readable storage medium. During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a read-only memory (Read-Only Memory, ROM) or a random access memory (Random Access Memory, RAM) and the like.
以上所揭露的仅为本发明较佳实施例而已，当然不能以此来限定本发明之权利范围，因此依本发明权利要求所作的等同变化，仍属本发明所涵盖的范围。The above disclosures are only preferred embodiments of the present invention, and certainly cannot limit the scope of rights of the present invention. Therefore, equivalent changes made according to the claims of the present invention still fall within the scope of the present invention.
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610821282.0ACN106484777B (en) | 2016-09-12 | 2016-09-12 | Multimedia data processing method and device |
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610821282.0ACN106484777B (en) | 2016-09-12 | 2016-09-12 | Multimedia data processing method and device |
| Publication Number | Publication Date |
|---|---|
| CN106484777Atrue CN106484777A (en) | 2017-03-08 |
| CN106484777B CN106484777B (en) | 2020-09-08 |
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610821282.0AActiveCN106484777B (en) | 2016-09-12 | 2016-09-12 | Multimedia data processing method and device |
| Country | Link |
|---|---|
| CN (1) | CN106484777B (en) |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107577736A (en)* | 2017-08-25 | 2018-01-12 | 上海斐讯数据通信技术有限公司 | A kind of file recommendation method and system based on BP neural network |
| CN108345697A (en)* | 2018-03-22 | 2018-07-31 | 山东财经大学 | Wisdom course towards group of college students recommends method, system and storage medium |
| CN108846097A (en)* | 2018-06-15 | 2018-11-20 | 北京搜狐新媒体信息技术有限公司 | The interest tags representation method of user, article recommended method and device, equipment |
| CN108959603A (en)* | 2018-07-13 | 2018-12-07 | 北京印刷学院 | Personalized recommendation system and method based on deep neural network |
| CN109034953A (en)* | 2018-07-02 | 2018-12-18 | 西南交通大学 | A Movie Recommendation Method |
| CN109491704A (en)* | 2018-11-08 | 2019-03-19 | 北京字节跳动网络技术有限公司 | Method and apparatus for handling information |
| CN109522333A (en)* | 2018-11-23 | 2019-03-26 | 北京锐安科技有限公司 | Data analysing method, device, equipment and medium |
| CN110213325A (en)* | 2019-04-02 | 2019-09-06 | 腾讯科技（深圳）有限公司 | Data processing method and data push method |
| CN110520858A (en)* | 2017-03-29 | 2019-11-29 | 阿里巴巴集团控股有限公司 | Method and apparatus for generating push notifications |
| CN111061907A (en)* | 2019-12-10 | 2020-04-24 | 腾讯科技（深圳）有限公司 | Media data processing method, device and storage medium |
| CN112733014A (en)* | 2020-12-30 | 2021-04-30 | 上海众源网络有限公司 | Recommendation method, device, equipment and storage medium |
| CN114358122A (en)* | 2021-12-03 | 2022-04-15 | 南京莱斯信息技术股份有限公司 | Method and system for distributing information on demand |
| CN114637907A (en)* | 2021-07-30 | 2022-06-17 | 广州亚信技术有限公司 | Media file recommendation method and device, electronic equipment and computer storage medium |
| CN115174997A (en)* | 2022-06-29 | 2022-10-11 | Vidaa国际控股(荷兰)公司 | Display device and medium resource recommendation method |
| WO2023115974A1 (en)* | 2021-12-22 | 2023-06-29 | 北京达佳互联信息技术有限公司 | Multimedia resource recommendation method and apparatus and object representation network generation method and apparatus |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102135991A (en)* | 2011-03-11 | 2011-07-27 | 桂林电子科技大学 | Backward learning-based dynamic multi-attribute service selecting method |
| CN102426686A (en)* | 2011-09-29 | 2012-04-25 | 南京大学 | Internet information product recommending method based on matrix decomposition |
| CN102831234A (en)* | 2012-08-31 | 2012-12-19 | 北京邮电大学 | Personalized news recommendation device and method based on news content and theme feature |
| CN103377250A (en)* | 2012-04-27 | 2013-10-30 | 杭州载言网络技术有限公司 | Top-k recommendation method based on neighborhood |
| CN103383702A (en)* | 2013-07-17 | 2013-11-06 | 中国科学院深圳先进技术研究院 | Method and system for recommending personalized news based on ranking of votes of users |
| CN104063481A (en)* | 2014-07-02 | 2014-09-24 | 山东大学 | Film individuation recommendation method based on user real-time interest vectors |
| CN104123315A (en)* | 2013-04-28 | 2014-10-29 | 百度在线网络技术（北京）有限公司 | Multi-media file recommendation method and recommendation server |
| CN104504055A (en)* | 2014-12-19 | 2015-04-08 | 常州飞寻视讯信息科技有限公司 | Commodity similarity calculation method and commodity recommending system based on image similarity |
| CN104731954A (en)* | 2015-04-01 | 2015-06-24 | 天翼爱音乐文化科技有限公司 | Music recommendation method and system based on group perspective |
| CN104765728A (en)* | 2014-01-08 | 2015-07-08 | 富士通株式会社 | Method and device for training neural network and method for determining sparse feature vector |
| CN104834686A (en)* | 2015-04-17 | 2015-08-12 | 中国科学院信息工程研究所 | Video recommendation method based on hybrid semantic matrix |
| WO2016040211A1 (en)* | 2014-09-12 | 2016-03-17 | Microsoft Technology Licensing, Llc | Modified matrix factorization of content-based model for recommendation system |
| US20160099854A1 (en)* | 2014-10-01 | 2016-04-07 | Adobe Systems Incorporated | Digital content consumption analysis |
| CN105512681A (en)* | 2015-12-07 | 2016-04-20 | 北京信息科技大学 | Method and system for acquiring target category picture |
| CN105915949A (en)* | 2015-12-23 | 2016-08-31 | 乐视网信息技术（北京）股份有限公司 | Video content recommending method, device and system |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102135991A (en)* | 2011-03-11 | 2011-07-27 | 桂林电子科技大学 | Backward learning-based dynamic multi-attribute service selecting method |
| CN102426686A (en)* | 2011-09-29 | 2012-04-25 | 南京大学 | Internet information product recommending method based on matrix decomposition |
| CN103377250A (en)* | 2012-04-27 | 2013-10-30 | 杭州载言网络技术有限公司 | Top-k recommendation method based on neighborhood |
| CN102831234A (en)* | 2012-08-31 | 2012-12-19 | 北京邮电大学 | Personalized news recommendation device and method based on news content and theme feature |
| CN104123315A (en)* | 2013-04-28 | 2014-10-29 | 百度在线网络技术（北京）有限公司 | Multi-media file recommendation method and recommendation server |
| CN103383702A (en)* | 2013-07-17 | 2013-11-06 | 中国科学院深圳先进技术研究院 | Method and system for recommending personalized news based on ranking of votes of users |
| CN104765728A (en)* | 2014-01-08 | 2015-07-08 | 富士通株式会社 | Method and device for training neural network and method for determining sparse feature vector |
| CN104063481A (en)* | 2014-07-02 | 2014-09-24 | 山东大学 | Film individuation recommendation method based on user real-time interest vectors |
| WO2016040211A1 (en)* | 2014-09-12 | 2016-03-17 | Microsoft Technology Licensing, Llc | Modified matrix factorization of content-based model for recommendation system |
| US20160099854A1 (en)* | 2014-10-01 | 2016-04-07 | Adobe Systems Incorporated | Digital content consumption analysis |
| CN104504055A (en)* | 2014-12-19 | 2015-04-08 | 常州飞寻视讯信息科技有限公司 | Commodity similarity calculation method and commodity recommending system based on image similarity |
| CN104731954A (en)* | 2015-04-01 | 2015-06-24 | 天翼爱音乐文化科技有限公司 | Music recommendation method and system based on group perspective |
| CN104834686A (en)* | 2015-04-17 | 2015-08-12 | 中国科学院信息工程研究所 | Video recommendation method based on hybrid semantic matrix |
| CN105512681A (en)* | 2015-12-07 | 2016-04-20 | 北京信息科技大学 | Method and system for acquiring target category picture |
| CN105915949A (en)* | 2015-12-23 | 2016-08-31 | 乐视网信息技术（北京）股份有限公司 | Video content recommending method, device and system |
| Title |
|---|
| 孙光福等: "基于时序行为的协同过滤推荐算法", 《软件学报》* |
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110520858A (en)* | 2017-03-29 | 2019-11-29 | 阿里巴巴集团控股有限公司 | Method and apparatus for generating push notifications |
| CN107577736A (en)* | 2017-08-25 | 2018-01-12 | 上海斐讯数据通信技术有限公司 | A kind of file recommendation method and system based on BP neural network |
| CN107577736B (en)* | 2017-08-25 | 2021-12-17 | 武汉数字智能信息科技有限公司 | File recommendation method and system based on BP neural network |
| CN108345697A (en)* | 2018-03-22 | 2018-07-31 | 山东财经大学 | Wisdom course towards group of college students recommends method, system and storage medium |
| CN108846097A (en)* | 2018-06-15 | 2018-11-20 | 北京搜狐新媒体信息技术有限公司 | The interest tags representation method of user, article recommended method and device, equipment |
| CN109034953A (en)* | 2018-07-02 | 2018-12-18 | 西南交通大学 | A Movie Recommendation Method |
| CN108959603A (en)* | 2018-07-13 | 2018-12-07 | 北京印刷学院 | Personalized recommendation system and method based on deep neural network |
| CN108959603B (en)* | 2018-07-13 | 2022-03-29 | 北京印刷学院 | Personalized recommendation system and method based on deep neural network |
| CN109491704A (en)* | 2018-11-08 | 2019-03-19 | 北京字节跳动网络技术有限公司 | Method and apparatus for handling information |
| CN109522333A (en)* | 2018-11-23 | 2019-03-26 | 北京锐安科技有限公司 | Data analysing method, device, equipment and medium |
| CN110213325B (en)* | 2019-04-02 | 2021-09-24 | 腾讯科技（深圳）有限公司 | Data processing method and data pushing method |
| CN110213325A (en)* | 2019-04-02 | 2019-09-06 | 腾讯科技（深圳）有限公司 | Data processing method and data push method |
| CN111061907A (en)* | 2019-12-10 | 2020-04-24 | 腾讯科技（深圳）有限公司 | Media data processing method, device and storage medium |
| CN111061907B (en)* | 2019-12-10 | 2023-06-20 | 腾讯科技（深圳）有限公司 | Media data processing method, device and storage medium |
| CN112733014A (en)* | 2020-12-30 | 2021-04-30 | 上海众源网络有限公司 | Recommendation method, device, equipment and storage medium |
| CN114637907A (en)* | 2021-07-30 | 2022-06-17 | 广州亚信技术有限公司 | Media file recommendation method and device, electronic equipment and computer storage medium |
| CN114358122A (en)* | 2021-12-03 | 2022-04-15 | 南京莱斯信息技术股份有限公司 | Method and system for distributing information on demand |
| WO2023115974A1 (en)* | 2021-12-22 | 2023-06-29 | 北京达佳互联信息技术有限公司 | Multimedia resource recommendation method and apparatus and object representation network generation method and apparatus |
| CN115174997A (en)* | 2022-06-29 | 2022-10-11 | Vidaa国际控股(荷兰)公司 | Display device and medium resource recommendation method |
| CN115174997B (en)* | 2022-06-29 | 2023-11-28 | Vidaa国际控股(荷兰)公司 | Display device and media asset recommendation method |
| Publication number | Publication date |
|---|---|
| CN106484777B (en) | 2020-09-08 |
| Publication | Publication Date | Title |
|---|---|---|
| CN106484777A (en) | A multimedia data processing method and device | |
| CN107451894B (en) | Data processing method, device and computer readable storage medium | |
| CN105701191B (en) | Pushed information click rate estimation method and device | |
| CN106651542B (en) | Article recommendation method and device | |
| WO2019233360A1 (en) | Deep learning-based audio equalization method, device and system | |
| CN109492180A (en) | Resource recommendation method, device, computer equipment and computer readable storage medium | |
| CN104574192B (en) | Method and device for identifying same user in multiple social networks | |
| CN103970802B (en) | A kind of method and device of song recommendations | |
| US11216518B2 (en) | Systems and methods of providing recommendations of content items | |
| CN107423430B (en) | Data processing method, device and computer readable storage medium | |
| CN106599047B (en) | Method and device for pushing information | |
| CN109410001A (en) | A kind of Method of Commodity Recommendation, system, electronic equipment and storage medium | |
| KR20180122829A (en) | Method, apparatus, and computer program for selecting music based on image | |
| CN110472659B (en) | Data processing method, device, computer readable storage medium and computer equipment | |
| Patel et al. | A comparative study of music recommendation systems | |
| CN111046285B (en) | Method, device, server and storage medium for determining recommendation ranking | |
| Gossi et al. | Lyric-based music recommendation | |
| CN110020152A (en) | Using recommended method and device | |
| CN108932262B (en) | Song recommendation method and device | |
| CN106844504A (en) | A kind of method and apparatus for sending the single mark of song | |
| CN116401394B (en) | Object set, image generation method and device, electronic equipment and storage medium | |
| CN112541010A (en) | User gender prediction method based on logistic regression | |
| CN107656927A (en) | A kind of feature selection approach and equipment | |
| CN114637926A (en) | Content recommendation method and device, electronic equipment and storage medium | |
| JP5654977B2 (en) | Product recommendation device, method and program |
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |