Disclosure of Invention
The invention provides a financing lease risk prediction method and system based on a knowledge graph, comprising the following steps: and acquiring various data from the financing lease business system to form an initial knowledge unit set, integrating structured data, unstructured data and real-time transaction stream data, breaking an information barrier and providing a comprehensive data base for risk prediction. The dynamic updated financing lease knowledge graph is constructed, an entity alignment algorithm and a knowledge fusion technology are introduced, a multi-hop association network covering multiple parties is established, the graph is updated dynamically through real-time data, the hidden risk propagation path is identified, the change of the service and new information can be reflected timely, and the timeliness and accuracy of risk assessment are improved. Capturing dynamic association features among entities, extracting risk evolution modes, generating a composite risk feature vector, and realizing feature coupling of static attributes and dynamic behaviors to more comprehensively and deeply describe risk features. The integrated model of the multi-layer perceptron and XGBoost fusion is adopted for prediction, the multi-layer perceptron learns high-dimensional nonlinear characteristics, XGBoost enhances analysis of sparse characteristics and time sequence trends, the advantages of the two models are fully exerted, the generalization capability of the model is improved, and the prediction accuracy and reliability are improved. And introducing an industry risk conduction intensity index, quantifying risk conduction probability based on an industry node topological structure in a knowledge graph, and realizing cross-level prediction from microscopic enterprise default to macroscopic industry risk. And a risk prediction result is generated based on various risk indexes, so that a more comprehensive and accurate reference basis is provided for financing lease decisions. The overall scheme solves the defects of the traditional method in three aspects of data dimension, entity association modeling and risk dynamic evolution through dynamic association mining of knowledge maps and multi-mode data fusion; meanwhile, the integrated model design gives consideration to prediction precision and interpretability, and provides panoramic early warning support from individual default to systematic risk for financing leasing business. The scientificity, the accuracy and the comprehensiveness of financing and leasing risk prediction are improved.
The invention provides a financing lease risk prediction method based on a knowledge graph, which comprises the following steps:
s1, obtaining structured data, unstructured data and real-time transaction stream data from a financing lease business system and forming an initial knowledge unit set;
S2, identifying entity and semantic relation from the initial knowledge unit set, and carrying out entity alignment and knowledge fusion to construct a dynamically updated financing lease knowledge graph;
S3, capturing dynamic association features among entities in the financing lease knowledge graph, extracting a time sequence risk evolution mode of historical transaction data, and generating a composite risk feature vector fusing static properties and dynamic behaviors;
s4, inputting the composite risk feature vector into an integrated model fused by the multi-layer perceptron and XGBoost, and outputting the default probability of the lessee, the equipment asset devaluation rate and the industry risk conduction intensity;
and S5, generating a lessee financing lease risk prediction result based on the lessee default probability, the equipment asset devaluation rate and the industry risk conduction intensity.
Preferably, the method for predicting financing lease risk based on knowledge graph comprises the following steps of S1, obtaining structured data, unstructured data and real-time transaction stream data from a financing lease business system and forming an initial knowledge unit set, wherein the method comprises the following steps:
The method comprises the steps of obtaining structured data, unstructured data and real-time transaction stream data from a financing lease business system, wherein the structured data, the unstructured data and the real-time transaction stream data comprise lessee credit records, equipment asset state data, industry economic indexes, contract clause texts and historical default records;
And cleaning, denoising and standardizing the obtained structured data, unstructured data and real-time transaction stream data, extracting entity, attribute and relation triples, and generating an initial knowledge unit set.
Preferably, the financing lease risk prediction method based on the knowledge graph comprises the following steps of S2, identifying entity and semantic relation from an initial knowledge unit set, and carrying out entity alignment and knowledge fusion to construct a dynamically updated financing lease knowledge graph, wherein the method comprises the following steps:
Defining core entity types and relation types in the financing and leasing field based on an ontology modeling method, wherein the entity types comprise lessees, leasing equipment, guarantee parties, industry classification and contract terms, and the relation types comprise guarantee association, equipment mortgage states and industry risk conduction paths;
Identifying entity and semantic relation from the initial knowledge unit set by adopting a BERT-based combined entity relation extraction model, calculating entity semantic similarity by utilizing a pre-trained domain word vector, and aggregating neighborhood entity information by combining a graph attention network to generate entity pairs Ji Quan;
and voting decision is carried out on attribute conflicts of the same entity in the multiple data sources through a cross-source conflict resolution algorithm, and attribute values with highest confidence are reserved to obtain a dynamically updated financing lease knowledge graph.
Preferably, the financing lease risk prediction method based on the knowledge graph includes the steps of S3, capturing dynamic association features among entities in the financing lease knowledge graph, extracting a time sequence risk evolution mode of historical transaction data, and generating a composite risk feature vector fusing static attributes and dynamic behaviors, wherein the method comprises the following steps:
Analyzing the financing lease knowledge graph by adopting a graph embedding algorithm to generate a low-dimensional vector representation of the entity and the relationship;
Embedding entity node characteristics in the financing leasing knowledge graph and a time stamp into vector splicing, inputting the vector splicing into a gating graph volume lamination, and capturing the time-varying association strength between the entities as dynamic association characteristics between the entities;
Extracting a local time sequence mode of historical transaction data through a sliding time window mechanism, and performing cross-modal fusion with global map features to obtain a time sequence risk evolution mode of the historical transaction data;
based on the low-dimensional vector representation of the entities and the relations, the dynamic association characteristics among the entities and the time sequence risk evolution mode of the historical transaction data, a composite risk characteristic vector fusing the static attribute and the dynamic behavior is generated.
Preferably, the method for predicting the financing lease risk based on the knowledge graph comprises the following steps of S5, generating a lessee financing lease risk prediction result based on the lessee default probability, the equipment asset devaluation rate and the industry risk conduction intensity, wherein the method comprises the following steps:
Obtaining three-party definition weights of the user on the default probability of the lessees, the equipment asset devaluation rate and the industry risk conduction intensity;
Calculating a lessee financing lease risk prediction value based on the user's lessee default probability, equipment asset devaluation rate, three-party definition weight of industry risk conduction intensity, lessee default probability, equipment asset devaluation rate, industry risk conduction intensity;
when the predicted value of the financing and renting risk of the lessee does not exceed the preset prediction threshold, the predicted value of the financing and renting risk of the lessee is taken as a predicted result of the financing and renting risk of the lessee;
when the financing lease risk prediction value of the lessee exceeds a preset prediction threshold, extracting a guarantee chain, a device mortgage state and industry upstream and downstream enterprises associated with the lessee from the financing lease knowledge graph to generate a visualized risk transmission sub-graph;
calculating the risk influence weight of each node in the risk propagation sub-graph, marking key risk nodes and conducting paths based on the risk influence weights of all the nodes in the risk propagation sub-graph to form a key risk path, and generating a risk tracing report as a lessee financing lease risk prediction result by combining lessee financing lease risk prediction values.
Preferably, in the financing lease risk prediction method based on a knowledge graph, a risk influence weight of each node in a risk propagation sub-graph is calculated, and key risk nodes and conduction paths are marked based on the risk influence weights of all the nodes in the risk propagation sub-graph to form a key risk path, including:
determining a risk influence weight initial value of all nodes in the risk propagation subgraph based on the total number of nodes in the risk propagation subgraph;
Acquiring a risk correlation value and a risk propagation strength of adjacent nodes in a risk propagation subgraph;
Performing iterative computation on risk influence weight initial values of all nodes in the risk propagation subgraph based on risk correlation values and risk propagation intensities of adjacent nodes in the risk propagation subgraph and a preset iteration formula until the difference between the risk influence weights obtained after the latest iteration process and the risk influence weights obtained after the last iteration process of all nodes in the risk propagation subgraph is smaller than a preset threshold value, marking all nodes in the risk propagation subgraph, wherein the risk influence weights obtained after the latest iteration process of all nodes are not smaller than the preset risk influence weight threshold value, as all key risk nodes;
Conducting path fitting is conducted based on all key risk nodes in the risk propagation subgraph, and a key risk path is formed.
Preferably, the method for predicting financing lease risk based on knowledge graph obtains risk correlation values and risk propagation intensities of adjacent nodes in a risk propagation subgraph, including:
generating a risk related feature vector of each node based on business transaction data of each node in the risk propagation subgraph;
taking the similarity between the risk related feature vectors of the adjacent nodes in the risk propagation subgraph as a risk correlation value of the adjacent nodes;
and calculating the risk propagation intensity of the adjacent nodes based on the total transaction amount and the transaction frequency between the adjacent nodes in the risk propagation subgraph.
Preferably, the financing lease risk prediction method based on the knowledge graph presets an iteration formula, which comprises the following steps:
In the formula, RIWk+1 (i) is risk influence weight of an ith node in a risk propagation sub-graph after k+1 iteration, alpha is a risk propagation trend coefficient, n is the total number of nodes in the risk propagation sub-graph, relevance (j, i) is a risk correlation value of the jth node and the ith node in the risk propagation sub-graph, M (i) is a set of all nodes pointing to the ith node in the risk propagation sub-graph, RIWk (j) is risk influence weight of the ith node in the risk propagation sub-graph after k iteration, strength (j, i) is risk propagation Strength from the jth node to the ith node in the risk propagation sub-graph, O (j) is a set of all nodes pointing to the jth node in the risk propagation sub-graph, and Strength (j, k) is risk propagation Strength from the jth node to the kth node in the risk propagation sub-graph.
Preferably, the financing lease risk prediction method based on the knowledge graph carries out conducting path fitting based on all key risk nodes in the risk propagation subgraph to form a key risk path, and comprises the following steps:
fitting at least one conductive path based on edges between all critical risk nodes in the risk propagation subgraph;
When there is only one conduction path, then regard only one conduction path as the critical risk path;
When there is more than one conduction path, then the overall conduction probability for each conduction path is calculated based on the risk correlation values and the risk propagation strengths between all adjacent critical risk nodes in each conduction path, and the conduction path of the largest overall conduction path among all conduction paths is taken as the critical risk path.
The invention provides a financing lease risk prediction system based on a knowledge graph, which is used for executing any one of the above financing lease risk prediction methods based on the knowledge graph, and comprises the following steps:
The initial knowledge unit construction module is used for acquiring structured data, unstructured data and real-time transaction stream data from the financing lease business system and forming an initial knowledge unit set;
the knowledge graph construction module is used for identifying the entity and the semantic relation from the initial knowledge unit set, carrying out entity alignment and knowledge fusion, and constructing a dynamically updated financing leasing knowledge graph;
The composite risk feature vector generation module is used for capturing dynamic association features among entities in the financing lease knowledge graph, extracting a time sequence risk evolution mode of historical transaction data and generating a composite risk feature vector integrating static attributes and dynamic behaviors;
the model prediction evaluation module is used for inputting the composite risk feature vector into an integrated model fused by the multi-layer perceptron and XGBoost and outputting the default probability of the lessee, the equipment asset devaluation rate and the industry risk conduction intensity;
And the risk prediction module is used for generating a lessee financing lease risk prediction result based on the lessee default probability, the equipment asset devaluation rate and the industry risk conduction intensity.
Compared with the prior art, the method has the beneficial effects that various data are acquired from the financing leasing business system to form the initial knowledge unit set, the structured data, the unstructured data and the real-time transaction stream data are integrated, the information barrier is broken, and a comprehensive data basis is provided for risk prediction. The dynamic updated financing lease knowledge graph is constructed, an entity alignment algorithm and a knowledge fusion technology are introduced, a multi-hop association network covering multiple parties is established, the graph is updated dynamically through real-time data, the hidden risk propagation path is identified, the change of the service and new information can be reflected timely, and the timeliness and accuracy of risk assessment are improved. Capturing dynamic association features among entities, extracting risk evolution modes, generating a composite risk feature vector, and realizing feature coupling of static attributes and dynamic behaviors to more comprehensively and deeply describe risk features. The integrated model of the multi-layer perceptron and XGBoost fusion is adopted for prediction, the multi-layer perceptron learns high-dimensional nonlinear characteristics, XGBoost enhances analysis of sparse characteristics and time sequence trends, the advantages of the two models are fully exerted, the generalization capability of the model is improved, and the prediction accuracy and reliability are improved. And introducing an industry risk conduction intensity index, quantifying risk conduction probability based on an industry node topological structure in a knowledge graph, and realizing cross-level prediction from microscopic enterprise default to macroscopic industry risk. And a risk prediction result is generated based on various risk indexes, so that a more comprehensive and accurate reference basis is provided for financing lease decisions. The overall scheme solves the defects of the traditional method in three aspects of data dimension, entity association modeling and risk dynamic evolution through dynamic association mining of knowledge maps and multi-mode data fusion; meanwhile, the integrated model design gives consideration to prediction precision and interpretability, and provides panoramic early warning support from individual default to systematic risk for financing leasing business. The scientificity, the accuracy and the comprehensiveness of financing and leasing risk prediction are improved.
Additional features and advantages of the application will be set forth in the description which follows, and in part will be obvious from the description, or may be learned by practice of the application. The objects and other advantages of the application may be realized and obtained by means of the instrumentalities particularly pointed out in the specification.
The technical scheme of the invention is further described in detail through the drawings and the embodiments.
Detailed Description
The preferred embodiments of the present invention will be described below with reference to the accompanying drawings, it being understood that the preferred embodiments described herein are for illustration and explanation of the present invention only, and are not intended to limit the present invention.
Example 1:
The invention provides a financing lease risk prediction method based on a knowledge graph, which comprises the following steps of:
s1, obtaining structured data, unstructured data and real-time transaction stream data from a financing lease business system and forming an initial knowledge unit set;
S2, identifying entity and semantic relation from the initial knowledge unit set, and carrying out entity alignment and knowledge fusion to construct a dynamically updated financing lease knowledge graph;
S3, capturing dynamic association features among entities in the financing lease knowledge graph, extracting a time sequence risk evolution mode of historical transaction data, and generating a composite risk feature vector fusing static properties and dynamic behaviors;
s4, inputting the composite risk feature vector into an integrated model fused by the multi-layer perceptron and XGBoost, and outputting the default probability of the lessee, the equipment asset devaluation rate and the industry risk conduction intensity;
s5, generating a lessee financing lease risk prediction result based on the lessee default probability, the equipment asset devaluation rate and the industry risk conduction intensity;
and the entity relationship and the risk weight in the knowledge graph can be updated by adopting an incremental learning algorithm according to the newly added service data and the early warning feedback result, so that the online self-adaptive optimization of the risk prediction model is realized.
In this embodiment, the financing lease service system is an information system dedicated to processing and managing financing lease related services, and covers a series of processes and data from lease application, contract signing to asset tracking, and the like.
In this embodiment, the structured data is data having a well-defined format and structure, such as rental transaction records, customer information, etc. in tabular form, which can be conveniently processed and analyzed by a computer program.
Unstructured data in this embodiment refers to data that does not have a fixed format or structure, such as a textual description of a rental contract, feedback comments from a customer, etc., and is relatively complex to process and analyze.
In this embodiment, the real-time transaction stream data is continuous transaction related data generated in real time along with the financing lease service, reflecting the instant status and changes of the service.
In this embodiment, the initial knowledge unit set is a set formed by preliminary processing and integration of structured, unstructured, and real-time transaction stream data obtained from the financing rental business system.
In this embodiment, the dynamically updated financing lease knowledge graph is a graph that can be continuously adjusted and perfected along with the acquisition of new data and the change of business, and is used for describing various entities and relationships in the financing lease field.
In this embodiment, the time-series risk evolution pattern of the historical transaction data is the law and characteristics of the time-dependent change and development of the risk summarized from the past rental transaction data.
In this embodiment, the composite risk feature vector that fuses the static attribute and the dynamic behavior is a vector for representing risk that is formed by integrating the fixed attribute describing the financing rental object and its behavior feature over time.
In the embodiment, the integrated model of the multi-layer perceptron and XGBoost is a model for combining the multi-layer perceptron and XGBoost machine learning models for risk prediction, wherein the robustness of the model to noise data is enhanced by adopting countermeasure training (ADVERSAR I A L TRA I N I NG) in the training process, and the feature contribution degree analysis is performed based on Shap l ey values, so that the model interpretability is optimized. Adopting Foca l Loss functions to solve the problem of imbalance of positive and negative samples in financing leasing scenes, introducing meta-path constraint based on knowledge graph to limit the logic rationality of the model on an industry risk conduction path.
In this embodiment, inputting the composite risk feature vector into the integrated model fused by the multi-layer perceptron and XGBoost, outputting the default probability, the equipment asset devaluation rate and the industry risk conduction intensity of the lessee, that is, putting the risk feature vector fused with the static attribute and the dynamic behavior into the model formed by combining the multi-layer perceptron and XGBoost, and then calculating and analyzing the model to give three results, namely, the numerical value of the default probability of the lessee, the numerical value of the proportion of the reduced value of the leasing equipment asset, and the numerical value of the conduction force of the industry risk to the financing leasing service.
In this embodiment, the lessee's probability of default refers to the likelihood that the lessee cannot contract to fulfill obligations in the financing lease service.
In this embodiment, the equipment asset devaluation rate is the rate at which the rental equipment value decreases during use.
In this embodiment, industry risk conduction intensity is the degree and extent of impact that risks present in the industry are transferred into the financing rental business.
In this embodiment, the lessee financing lease risk prediction results are prediction conclusions based on various data and models regarding the risk that the lessee may be exposed to in the financing lease business.
The method has the beneficial effects that various data are acquired from the financing lease business system to form an initial knowledge unit set, structured data, unstructured data and real-time transaction stream data are integrated, an information barrier is broken, and a comprehensive data base is provided for risk prediction. The dynamic updated financing lease knowledge graph is constructed, an entity alignment algorithm and a knowledge fusion technology are introduced, a multi-hop association network covering multiple parties is established, the graph is updated dynamically through real-time data, the hidden risk propagation path is identified, the change of the service and new information can be reflected timely, and the timeliness and accuracy of risk assessment are improved. Capturing dynamic association features among entities, extracting risk evolution modes, generating a composite risk feature vector, and realizing feature coupling of static attributes and dynamic behaviors to more comprehensively and deeply describe risk features. The integrated model of the multi-layer perceptron and XGBoost fusion is adopted for prediction, the multi-layer perceptron learns high-dimensional nonlinear characteristics, XGBoost enhances analysis of sparse characteristics and time sequence trends, the advantages of the two models are fully exerted, the generalization capability of the model is improved, and the prediction accuracy and reliability are improved. And introducing an industry risk conduction intensity index, quantifying risk conduction probability based on an industry node topological structure in a knowledge graph, and realizing cross-level prediction from microscopic enterprise default to macroscopic industry risk. And a risk prediction result is generated based on various risk indexes, so that a more comprehensive and accurate reference basis is provided for financing lease decisions. The overall scheme solves the defects of the traditional method in three aspects of data dimension, entity association modeling and risk dynamic evolution through dynamic association mining of knowledge maps and multi-mode data fusion; meanwhile, the integrated model design gives consideration to prediction precision and interpretability, and provides panoramic early warning support from individual default to systematic risk for financing leasing business. The scientificity, the accuracy and the comprehensiveness of financing and leasing risk prediction are improved.
Example 2:
Based on embodiment 1, the method for predicting financing and renting risk based on a knowledge graph comprises the following steps of S1, obtaining structured data, unstructured data and real-time transaction stream data from a financing and renting business system and forming an initial knowledge unit set, wherein the method comprises the following steps:
The method comprises the steps of obtaining structured data, unstructured data and real-time transaction stream data from a financing lease business system, wherein the structured data, the unstructured data and the real-time transaction stream data comprise lessee credit records, equipment asset state data, industry economic indexes, contract clause texts and historical default records;
And cleaning, denoising and standardizing the obtained structured data, unstructured data and real-time transaction stream data, extracting entity, attribute and relation triples, and generating an initial knowledge unit set.
In this embodiment, the lessee credit record refers to a record related to the credit status of the lessee, including credit score, past loan repayment, and other information.
In this embodiment, the equipment asset status data is data describing various conditions of the rental equipment, such as age, wear level, maintenance condition, etc. of the equipment.
In this embodiment, the industry economic index is various data reflecting the economic operation condition and development trend of the industry, such as industry growth rate, market saturation, etc.
In this embodiment, the contract term text is a textual description of the specific terms contained in the contract entered into by both parties in the financing lease service.
In this embodiment, the history violations record is a record of the past occurrences of the lessees failing to fulfill the financing lease contract agreement.
In this embodiment, the cleaning, denoising and standardization processing of the obtained structured data, unstructured data and real-time transaction stream data refers to removing error, repeated or useless information of all kinds of obtained data, reducing noise interference, and converting the data into a unified format and standard for subsequent analysis and processing.
In this embodiment, the entity, attribute and relationship triples are extracted, that is, specific objects (entities), their characteristics (attributes) and their associations (relationships) with each other are identified from data, and are sorted into a set composed of these three elements.
The technical scheme has the beneficial effects that various comprehensive data including lessee credit records and the like are obtained, and a rich information source is provided for risk prediction. And the acquired data are subjected to cleaning, denoising and standardization treatment, so that the quality and usability of the data are improved. Extracting the entity, attribute and relation triples to generate an initial knowledge unit set, and laying a good foundation for the subsequent construction of a knowledge graph. The accuracy and consistency of the data can be ensured, and the influence of noise and errors on risk prediction is reduced. The method provides a high-quality and standardized data base for financing lease risk prediction, and is beneficial to improving the accuracy and reliability of prediction.
Example 3:
on the basis of embodiment 1, the financing lease risk prediction method based on the knowledge graph includes the steps of S2, identifying entities and semantic relations from an initial knowledge unit set, and carrying out entity alignment and knowledge fusion to construct a dynamically updated financing lease knowledge graph, wherein the method comprises the following steps:
Defining core entity types and relation types in the financing and leasing field based on an ontology modeling method, wherein the entity types comprise lessees, leasing equipment, guarantee parties, industry classification and contract terms, and the relation types comprise guarantee association, equipment mortgage states and industry risk conduction paths;
Identifying entity and semantic relation from the initial knowledge unit set by adopting a BERT-based combined entity relation extraction model, calculating entity semantic similarity by utilizing a pre-trained domain word vector, and aggregating neighborhood entity information by combining a graph attention network to generate entity pairs Ji Quan;
and voting decision is carried out on attribute conflicts of the same entity in the multiple data sources through a cross-source conflict resolution algorithm, and attribute values with highest confidence are reserved to obtain a dynamically updated financing lease knowledge graph.
In this embodiment, defining the core entity type and the relationship type in the financing and renting field based on the ontology modeling method refers to defining key entity types (such as lessees, renting equipment, etc.) in the financing and renting field by using the ontology modeling method, and association types (such as guaranty association, device mortgage status, etc.) between them.
In this embodiment, an industry risk conduction path refers to a way and manner in which risks present in the industry propagate and transfer between different subjects, links. For example, in the automotive industry, because of the significant increase in raw material prices (which is an industry risk), they may first affect component suppliers, which may increase component prices, which is a risk of being conducted from the raw material supply link to the component production link. The cost of the automobile manufacturer is then increased, possibly reducing the production, increasing the automobile sales price, which is a risk to the automobile manufacturing link. Consumers may then reduce purchases due to rising prices of the cars, and the sales of the cars may decrease, affecting the profits of the car dealers, which is a risk to the sales link. The whole process from the rising of raw material price to the influence of automobile sales is the path and mode of risk conduction in the automobile manufacturing industry.
In this embodiment, using a BERT-based federated entity relationship extraction model to identify entities and semantic relationships from an initial knowledge element set is to use BERT-based specific models to identify specific entities (e.g., lessees, vouchers, etc.) and their semantic associations from a preliminarily constructed knowledge element set.
In this embodiment, the semantic similarity of the entities is calculated by using pre-trained domain word vectors, and the neighborhood entity information is aggregated by combining the graph attention network, so that the generation of the entity pair Ji Quan means that the semantic similarity of the entities is measured by using pre-trained domain word vectors, and the information of the surrounding entities is integrated by using the graph attention network, so as to obtain the weight value of entity alignment. It is assumed that in the financial field, there are two entities, "stock investments" and "securities investments". Through the pre-trained financial domain word vectors, the semantic similarity of the word vectors can be calculated. Meanwhile, in the knowledge graph, the neighborhood entities such as stock trading, stock analysis and the like are arranged around the entity of stock investment. With the graph attention network, it is possible to pay more attention to neighborhood entities closely related to "stock investment", such as "stock exchanges", and aggregate more information. And finally obtaining the aligned weight values of the stock investment and the securities investment by combining the semantic similarity and the aggregation result of the neighborhood entity information, wherein the weight values are higher, for example, the weight values indicate that the two entities have stronger association and alignment degree under specific situations.
In this embodiment, voting decision is performed on attribute conflicts of the same entity in multiple data sources through a cross-source conflict resolution algorithm, attribute values with highest confidence are reserved, a dynamically updated financing lease knowledge graph is obtained, namely an algorithm capable of solving conflicts from different data sources is used, attributes of the same entity but conflicts in multiple data sources are determined through voting, and finally, attribute values considered to be most reliable by people are reserved, so that the financing lease knowledge graph capable of being updated continuously is obtained.
The technical scheme has the beneficial effects that the core entity and the relationship type in the financing and renting field are defined, and a clear framework and structure are provided for the construction of the knowledge graph. Advanced combined entity relation extraction models and technologies are adopted, and accuracy and efficiency of entity and semantic relation identification are improved. And more accurate entity alignment is realized by utilizing entity semantic similarity calculation and neighborhood entity information aggregation. And attribute conflicts are solved through a cross-source conflict resolution algorithm, so that the consistency and reliability of data in the knowledge graph are ensured. The dynamically updated knowledge graph is obtained, the service change can be reflected in time, and the latest and most accurate information is provided for risk prediction. The scheme improves the scientificity, accuracy and dynamic adaptability of financing lease knowledge graph construction, and is beneficial to more effectively carrying out risk prediction.
Example 4:
On the basis of embodiment 1, the financing lease risk prediction method based on the knowledge graph includes the steps of S3, capturing dynamic association features among entities in the financing lease knowledge graph, extracting a time sequence risk evolution mode of historical transaction data, and generating a composite risk feature vector fusing static attributes and dynamic behaviors, wherein the method comprises the following steps:
Analyzing the financing lease knowledge graph by adopting a graph embedding algorithm to generate a low-dimensional vector representation of the entity and the relationship;
Embedding entity node characteristics in the financing leasing knowledge graph and a time stamp into vector splicing, inputting the vector splicing into a gating graph volume lamination, and capturing the time-varying association strength between the entities as dynamic association characteristics between the entities;
Extracting a local time sequence mode of historical transaction data through a sliding time window mechanism, and performing cross-modal fusion with global map features to obtain a time sequence risk evolution mode of the historical transaction data;
based on the low-dimensional vector representation of the entities and the relations, the dynamic association characteristics among the entities and the time sequence risk evolution mode of the historical transaction data, a composite risk characteristic vector fusing the static attribute and the dynamic behavior is generated.
In this embodiment, the entity node features refer to various characteristics and attributes representing the entity in the financing lease knowledge base. For example, in a financing rental knowledge graph, a "company" is an entity node that may include company name, time of establishment, registered capital, industry of interest, credit rating, past rental transaction record, and so forth. Also, as for example, a physical node such as "a device" may be characterized by a model number, a value, a service life, a maintenance condition, etc. of the device.
In this embodiment, the timestamp embedding vector is a way to translate the time information into a vector form for model processing and analysis.
In this embodiment, the gating map convolution layer is a neural network layer that performs convolution operations on map structure data, and has a gating mechanism to control the flow and filtering of information.
In this embodiment, the strength of the association between entities over time is a measure that describes how tight or important the relationship between entities is over time. Assume that in the financing rental area, there are two entities, "tenant A" and "rental device B". In the first year, the lessee A pays the rent on time, and the association strength between the lessee A and the leasing equipment B is strong, which indicates that the lessees are closely related and the equipment is normally used. By the next year, tenant a begins to pay a lease overdue, at which point the strength of association with rental equipment B may be weakened, meaning that the relationship between the two parties becomes tense, and there may be a risk. And half a year later, the lessee A breaks down thoroughly, the association strength becomes very weak, and the relationship between the lessees is endangered to be broken. The dynamic change condition of the relationship between the two entities can be clearly understood through the time-varying association strength measurement.
In this embodiment, embedding the entity node features and the time stamps in the financing lease knowledge graph into the vector concatenation, inputting the vector concatenation into the gating graph roll-up layer, capturing the time-varying association strength between the entities as the dynamic association feature between the entities, which means that the vector composed of the features and the time information of the entities in the graph is put into the gating graph roll-up layer, so as to obtain the time-varying tightness of the relationship between the entities, and taking the relationship as the dynamic association feature.
In this embodiment, the sliding time window mechanism is a method of selecting data by moving for a certain time interval.
In this embodiment, the local time sequence pattern of the historical transaction data is extracted by a sliding time window mechanism, and the time sequence rule in a specific time period is obtained from the historical transaction data by using the moving mode of the time window.
In this embodiment, the local timing pattern of the historical transaction data is a regular time series characteristic exhibited by the historical transaction data over a local time horizon.
In this embodiment, the global profile features are the overall features that the entire financing lease knowledge base has. For example, in a financing leasing knowledge graph, the global graph features may include a large number of entities, indicating a large service scale, a relatively tight connection between most of the entities, reflecting a relatively close cooperation relationship in the trip industry, relatively balanced distribution of different types of entities (such as lessees, equipment suppliers, etc.), relatively reasonable display market structure, and a plurality of core entities in the graph, which are connected with a large number of other entities, showing the importance of the core entities in the whole financing leasing service. These are the overall features presented by the overall financing lease knowledge graph.
In this embodiment, a local time sequence mode of historical transaction data is extracted through a sliding time window mechanism, and cross-modal fusion is performed with global map features, and obtaining a time sequence risk evolution mode of the historical transaction data refers to obtaining a local time sequence mode by utilizing a sliding time window, and then fusing the local time sequence mode with features of the whole map in different modes, so that a rule of risk change along with time in the historical transaction data is obtained. Assume that there is a period of financing lease history transaction data for 5 years. A sliding time window was set to 1 year. In the window of the first year, lessees in a certain industry are found to have high willingness to pay when renting equipment in the beginning of the year, but the willingness to pay in the end of the year is obviously reduced. This is the local timing mode. Meanwhile, the global map features show that the whole credit rating of the industry is low, and the industry is highly competitive. The local time sequence mode is fused with the global map feature, for example, the fact that the renter is difficult to operate due to the fact that industry competition is intense is found, so that the payment willingness of renting equipment is influenced, the rule that the risk changes with time in historical transaction data is obtained, and the fact that the industry renter is more prone to occurrence of default risks at the end of year is obtained.
In this embodiment, based on the low-dimensional vector representation of the entities and the relationships, the dynamic association features between the entities, and the time-sequence risk evolution mode of the historical transaction data, the composite risk feature vector integrating the static attribute and the dynamic behavior is generated, that is, the risk feature vector integrating the static attribute (such as the inherent feature of the entities) and the dynamic behavior (such as the change of the relationships and the evolution of the time sequence) is constructed according to the multiple features and modes.
The technical scheme has the beneficial effects that the low-dimensional vector representation is generated by using the graph embedding algorithm, so that the data dimension is reduced, and the calculation efficiency and the data processing capacity are improved. The dynamic association features are captured by using the gating graph volume lamination, so that the relationship strength of the entity changing along with time can be more accurately captured. And extracting a time sequence mode through a sliding time window mechanism and performing cross-mode fusion, so that risk evolution information in historical transaction data is fully mined. And generating a composite risk feature vector fusing the static attribute and the dynamic behavior, so that the description of the risk features is more comprehensive and accurate. The risk in the financing lease business is better understood and analyzed, and the accuracy and reliability of risk prediction are improved. The risk feature extraction and fusion effect is improved, and powerful support is provided for accurate risk prediction.
Example 5:
Based on the embodiment 1, the financing lease risk prediction method based on the knowledge graph, S5, generates a lessee financing lease risk prediction result based on the lessee default probability, the equipment asset devaluation rate and the industry risk conduction intensity, and comprises the following steps:
Obtaining three-party definition weights of the user on the default probability of the lessees, the equipment asset devaluation rate and the industry risk conduction intensity;
Calculating a lessee financing lease risk prediction value based on the user's lessee default probability, equipment asset devaluation rate, three-party definition weight of industry risk conduction intensity, lessee default probability, equipment asset devaluation rate, industry risk conduction intensity;
when the predicted value of the financing and renting risk of the lessee does not exceed the preset prediction threshold, the predicted value of the financing and renting risk of the lessee is taken as a predicted result of the financing and renting risk of the lessee;
when the financing lease risk prediction value of the lessee exceeds a preset prediction threshold, extracting a guarantee chain, a device mortgage state and industry upstream and downstream enterprises associated with the lessee from the financing lease knowledge graph to generate a visualized risk transmission sub-graph;
calculating the risk influence weight of each node in the risk propagation sub-graph, marking key risk nodes and conducting paths based on the risk influence weights of all the nodes in the risk propagation sub-graph to form a key risk path, and generating a risk tracing report as a lessee financing lease risk prediction result by combining lessee financing lease risk prediction values.
In this embodiment, the three-party definition weights of the user's rule-breaking probability of the lessee, the equipment asset devaluation rate and the industry risk conduction intensity are obtained, that is, the definition weights of the user's rule-breaking probability of the lessee, the equipment asset devaluation rate and the industry risk conduction intensity are obtained, for example, the weight of the user-defined rule-breaking probability of the lessee is 0.4, the weight of the equipment asset devaluation rate is 0.3 and the weight of the industry risk conduction intensity is 0.3.
In this embodiment, based on the user's three-party definition weights of the tenant breach probability, the equipment asset devaluation rate, the industry risk conduction intensity, the calculation of the tenant financing lease risk prediction value may use a weighted summation method, for example, the user defines the tenant breach probability weight as 0.4, the equipment asset devaluation rate weight as 0.3, and the industry risk conduction intensity weight as 0.3;
Assuming that the probability of the lessee default is 0.2, the equipment asset devaluation rate is 0.1, and the industry risk conduction intensity is 0.15;
Then lessee financing lease risk prediction = 0.4× 0.2+0.3×0.1+0.3×0.15=0.2 =0.3×0.1 =0.15' 0.08+0.03+0.045=0.155.
In this embodiment, the preset prediction threshold is a preset numerical standard for determining the risk level of financing lease of the lessee.
In this embodiment, the lessee-financing lease risk prediction value is a quantized value calculated by a model, which indicates that the lessee is at risk in the financing lease business.
In this embodiment, extracting a guarantee chain, a device mortgage state and an industry upstream and downstream enterprise associated with a lessee from a financing lease knowledge graph, and generating a visualized risk transmission sub graph refers to selecting information such as a guarantee relationship, a device mortgage condition and an industry upstream and downstream enterprise associated with the lessee from the overall financing lease knowledge graph, and displaying a path and a range of possible risk transmission in a graphical manner.
In this embodiment, the risk propagation subgraph is a local graph extracted from the entire knowledge graph specifically for risk propagation, and is used to analyze the risk propagation situation more focused.
In this embodiment, the risk impact weight of a node is a numerical value that measures the magnitude of the impact of each node in the risk propagation subgraph on risk propagation.
In this embodiment, the critical risk nodes are nodes that have an important role and a large impact on risk propagation in the risk propagation subgraph.
In this embodiment, the conductive path is a specific route where risks propagate between nodes.
In this embodiment, the critical risk path is a path of greater impact that is critical to risk propagation among the numerous conductive paths.
In the embodiment, the generation of the risk tracing report by combining the financing lease risk prediction value of the lessee integrates the financing lease risk prediction value of the lessee and the analysis result of risk propagation to form a report for tracing the source and the cause of the risk.
The technical scheme has the beneficial effects that the three-party weight defined by the user is obtained, so that the risk prediction result is more fit with the requirements and focus of attention of the user. And calculating a lessee financing lease risk prediction value, and providing a quantified index for risk assessment. And when the risk prediction value exceeds the threshold value, generating a visualized risk propagation sub-graph, and intuitively displaying the risk propagation condition. And calculating the risk influence weight, marking key risk nodes and conducting paths, and accurately identifying risk sources and key links. And a risk traceability report is generated, so that a detailed and clear basis is provided for risk management and decision making. The scheme improves individuation, visualization and traceability of financing lease risk prediction, and is beneficial to effectively coping with and managing risks.
Example 6:
based on embodiment 5, the financing lease risk prediction method based on the knowledge graph calculates the risk influence weight of each node in the risk propagation subgraph, marks the key risk nodes and the conduction paths based on the risk influence weights of all the nodes in the risk propagation subgraph, and forms the key risk path, including:
determining a risk influence weight initial value of all nodes in the risk propagation subgraph based on the total number of nodes in the risk propagation subgraph;
Acquiring a risk correlation value and a risk propagation strength of adjacent nodes in a risk propagation subgraph;
Performing iterative computation on risk influence weight initial values of all nodes in the risk propagation subgraph based on risk correlation values and risk propagation intensities of adjacent nodes in the risk propagation subgraph and a preset iteration formula until the difference between the risk influence weights obtained after the latest iteration process and the risk influence weights obtained after the last iteration process of all nodes in the risk propagation subgraph is smaller than a preset threshold value, marking all nodes in the risk propagation subgraph, wherein the risk influence weights obtained after the latest iteration process of all nodes are not smaller than the preset risk influence weight threshold value, as all key risk nodes;
Conducting path fitting is conducted based on all key risk nodes in the risk propagation subgraph, and a key risk path is formed.
In this embodiment, determining the initial value of the risk impact weight of all the nodes in the risk propagation sub-graph based on the total number of nodes in the risk propagation sub-graph means taking the inverse of the number of nodes in the risk propagation sub-graph as the risk impact weight value of each node at the beginning.
In this embodiment, the risk correlation value and the risk propagation strength of the neighboring nodes in the risk propagation subgraph refer to a measure of the magnitude of the risk correlation degree between the neighboring nodes in the risk propagation subgraph and the magnitude of the risk propagation strength between the neighboring nodes.
In this embodiment, the preset threshold is a preset numerical limit for determining or comparing the difference between risk impact weights obtained by the same node after adjacent iteration processes.
In this embodiment, the preset risk influence weight threshold is a preset weight numerical limit for determining the magnitude of the risk influence of the node.
The technical scheme has the beneficial effects that a starting point is provided for subsequent iterative computation by determining the initial value of the risk influence weight. And acquiring the risk correlation value and the propagation strength of the adjacent node, and providing important parameters for accurately calculating the risk influence weight. And the risk influence weight is continuously optimized by using iterative calculation, so that the accuracy and stability of the weight calculation are improved. The key risk nodes are screened through the preset threshold value and the weight threshold value, so that the nodes with important influence on risk propagation can be accurately identified. Conducting path fitting is conducted based on the key risk nodes, so that a key risk path is constructed, and key links of risk transmission can be clearly grasped. The method improves the accuracy and reliability of risk propagation analysis and provides powerful support for effective management of financing leasing risks.
Example 7:
Based on embodiment 6, a knowledge graph-based financing lease risk prediction method obtains a risk correlation value and a risk propagation strength of neighboring nodes in a risk propagation subgraph, including:
generating a risk related feature vector of each node based on business transaction data of each node in the risk propagation subgraph;
taking the similarity between the risk related feature vectors of the adjacent nodes in the risk propagation subgraph as a risk correlation value of the adjacent nodes;
and calculating the risk propagation intensity of the adjacent nodes based on the total transaction amount and the transaction frequency between the adjacent nodes in the risk propagation subgraph.
In this embodiment, the business transaction data of the nodes refers to various types of information and data of specific business transactions related to each node in the risk propagation subgraph.
In this embodiment, the risk related feature vector of each node is generated based on the business transaction data of each node in the risk propagation subgraph, and the risk related feature vector is converted into a vector form capable of characterizing the risk feature of the node by processing and analyzing the business transaction data of each node. Assume that in one risk propagation sub-graph, there is a node "enterprise A". The business transaction data comprises transaction amounts, transaction frequencies, time-based performance of transactions and the like of a plurality of suppliers in the past year. The data is processed, such as calculating an average value, standard deviation, change rate of transaction frequency, good rate of performance, etc. of the transaction amount. These processed values are then combined to form a vector, such as [ average transaction amount, standard deviation of transaction amount, rate of change of transaction frequency, performance goodness ], which is a risk-related feature vector of the node "enterprise a" for characterizing its features in terms of risk.
In this embodiment, the risk-related feature vector of the node is a vector for describing features of the node in terms of risk, and can reflect the risk condition of the node. For example, in a risk-spread subgraph of financing lease, a node represents a "certain manufacturing enterprise". The risk related feature vector may be [0.8,0.3,0.6,0.1], where the first element 0.8 may represent a higher liability ratio for the business, meaning a higher liability, the second element 0.3 may represent a reduced proportion of its recent market share, reflecting the business risk, the third element 0.6 may represent general stability of collaboration with the vendor, with supply chain risk, and the fourth element 0.1 may represent a lesser degree of adverse impact of recent policies on its industry. Such vectors can comprehensively describe the risk aspect of the node (i.e., the manufacturing enterprise).
In this embodiment, the similarity between risk related feature vectors of adjacent nodes is a measure of the similarity between risk related feature vectors of two adjacent nodes, and may be calculated by cosine similarity or euclidean distance between the vectors.
In the embodiment, the risk propagation intensity of the adjacent nodes is calculated based on the total transaction amount and the transaction frequency between the adjacent nodes in the risk propagation subgraph, namely, the intensity of risk propagation between the two adjacent nodes is determined according to the total transaction amount and the occurrence frequency of the transaction between the adjacent nodes, namely, the product of the total transaction amount between the two adjacent nodes and the total asset average value of the two adjacent nodes and the transaction frequency between the adjacent nodes is taken as the risk propagation intensity of the adjacent nodes.
The technical scheme has the beneficial effects that the risk related feature vector is generated, and a basis is provided for quantifying risk correlation and propagation strength. The similarity of the adjacent node risk related feature vectors is used as a risk correlation value, so that the degree of association of risks among nodes can be accurately reflected. The risk propagation strength is calculated through the transaction amount and the transaction frequency, and the possibility and influence of risk propagation are objectively evaluated from the service perspective. The key information of the adjacent nodes in the risk propagation subgraph can be acquired more scientifically and accurately, and powerful support is provided for subsequent risk analysis and evaluation. The method and the device improve the understanding and evaluation capability of the node relation in the risk propagation subgraph, and are helpful for constructing the key risk path more accurately.
Example 8:
based on embodiment 6, the method for predicting financing lease risk based on a knowledge graph presets an iteration formula, which comprises:
In the formula, RIWk+1 (i) is risk influence weight of an ith node in a risk propagation sub-graph after k+1 iteration, alpha is a risk propagation trend coefficient, n is the total number of nodes in the risk propagation sub-graph, relevance (j, i) is a risk correlation value of the jth node and the ith node in the risk propagation sub-graph, M (i) is a set of all nodes pointing to the ith node in the risk propagation sub-graph, RIWk (j) is risk influence weight of the ith node in the risk propagation sub-graph after k iteration, strength (j, i) is risk propagation Strength from the jth node to the ith node in the risk propagation sub-graph, O (j) is a set of all nodes pointing to the jth node in the risk propagation sub-graph, and Strength (j, k) is risk propagation Strength from the jth node to the kth node in the risk propagation sub-graph.
In this embodiment, the risk propagation trend coefficient is a self-defined influence factor, and has a value ranging from 0 to 1, and is used for adjusting the influence duty ratio transmitted from other nodes, namely balancing random risk propagation and risk propagation based on node relationships. It represents the probability of propagating depending on the actual relationship between nodes in the risk propagation process, and has no dimension.
The technical scheme has the beneficial effects that an explicit preset iteration formula is provided, and a standardized and quantifiable method is provided for calculating the risk influence weight. The formula considers a plurality of factors such as risk propagation trend coefficient, total number of nodes, risk correlation value among nodes, risk propagation strength and the like, so that the calculation result is more comprehensive and accurate. Through iterative computation, the change of the risk influence weight can be dynamically reflected, and the method is suitable for the dynamic property and complexity of risk propagation. The risk influence of the nodes is determined more accurately, so that the key risk nodes are marked more accurately and the key risk paths are constructed. The method improves the scientificity and accuracy of risk propagation analysis, and provides a more reliable basis for financing lease risk prediction and management.
Example 9:
Based on embodiment 6, the financing lease risk prediction method based on the knowledge graph performs conducting path fitting based on all key risk nodes in the risk propagation subgraph to form a key risk path, and includes:
fitting at least one conductive path based on edges between all critical risk nodes in the risk propagation subgraph;
When there is only one conduction path, then regard only one conduction path as the critical risk path;
When there is more than one conduction path, then the overall conduction probability for each conduction path is calculated based on the risk correlation values and the risk propagation strengths between all adjacent critical risk nodes in each conduction path, and the conduction path of the largest overall conduction path among all conduction paths is taken as the critical risk path.
In this embodiment, fitting at least one conductive path based on edges between all critical risk nodes in the risk propagation subgraph means deriving at least one possible risk propagation route from the connection relations between the critical risk nodes.
In this embodiment, the comprehensive conduction probability of each conduction path is calculated based on the risk correlation values and the risk propagation intensities between all the adjacent critical risk nodes in each conduction path, so that the comprehensive probability value of risk propagation on each conduction path is obtained by comprehensively considering the risk correlation degree and the strength of risk propagation between the adjacent critical risk nodes on each path. For example, the risk correlation value and the risk propagation intensity between adjacent key risk nodes can be weighted and summed by adopting preset weights to obtain a parameter between adjacent nodes, and then the parameters of all groups of adjacent nodes on each conducting path are sequentially summed to obtain the comprehensive conducting probability of each conducting path.
In this embodiment, the overall conduction probability of a conduction path is a value that represents the overall likelihood that a risk on a particular conduction path will propagate smoothly.
The technical scheme has the beneficial effect that various possible conduction paths among the key risk nodes can be comprehensively considered. When there is only one conductive path, it is determined directly, simplifying the process. When a plurality of conduction paths exist, the path with the highest probability is selected as the key risk path by calculating the comprehensive conduction probability, so that the accuracy and the rationality of path selection are improved. The method helps to determine the critical path of risk propagation more accurately, and provides a clear direction for risk management and control. The method improves scientificity and effectiveness of fitting the key risk paths, and enhances capacity of predicting and managing financing lease risks.
Example 10:
The invention provides a financing lease risk prediction system based on a knowledge graph, which is used for executing the financing lease risk prediction method based on the knowledge graph, which is described in any one of embodiments 1 to 9, and referring to FIG. 2, the system comprises the following steps:
The initial knowledge unit construction module is used for acquiring structured data, unstructured data and real-time transaction stream data from the financing lease business system and forming an initial knowledge unit set;
the knowledge graph construction module is used for identifying the entity and the semantic relation from the initial knowledge unit set, carrying out entity alignment and knowledge fusion, and constructing a dynamically updated financing leasing knowledge graph;
The composite risk feature vector generation module is used for capturing dynamic association features among entities in the financing lease knowledge graph, extracting a time sequence risk evolution mode of historical transaction data and generating a composite risk feature vector integrating static attributes and dynamic behaviors;
the model prediction evaluation module is used for inputting the composite risk feature vector into an integrated model fused by the multi-layer perceptron and XGBoost and outputting the default probability of the lessee, the equipment asset devaluation rate and the industry risk conduction intensity;
And the risk prediction module is used for generating a lessee financing lease risk prediction result based on the lessee default probability, the equipment asset devaluation rate and the industry risk conduction intensity.
The method has the beneficial effects that various data are acquired from the financing lease business system to form an initial knowledge unit set, structured data, unstructured data and real-time transaction stream data are integrated, an information barrier is broken, and a comprehensive data base is provided for risk prediction. The dynamic updated financing lease knowledge graph is constructed, an entity alignment algorithm and a knowledge fusion technology are introduced, a multi-hop association network covering multiple parties is established, the graph is updated dynamically through real-time data, the hidden risk propagation path is identified, the change of the service and new information can be reflected timely, and the timeliness and accuracy of risk assessment are improved. Capturing dynamic association features among entities, extracting risk evolution modes, generating a composite risk feature vector, and realizing feature coupling of static attributes and dynamic behaviors to more comprehensively and deeply describe risk features. The integrated model of the multi-layer perceptron and XGBoost fusion is adopted for prediction, the multi-layer perceptron learns high-dimensional nonlinear characteristics, XGBoost enhances analysis of sparse characteristics and time sequence trends, the advantages of the two models are fully exerted, the generalization capability of the model is improved, and the prediction accuracy and reliability are improved. And introducing an industry risk conduction intensity index, quantifying risk conduction probability based on an industry node topological structure in a knowledge graph, and realizing cross-level prediction from microscopic enterprise default to macroscopic industry risk. And a risk prediction result is generated based on various risk indexes, so that a more comprehensive and accurate reference basis is provided for financing lease decisions. The overall scheme solves the defects of the traditional method in three aspects of data dimension, entity association modeling and risk dynamic evolution through dynamic association mining of knowledge maps and multi-mode data fusion; meanwhile, the integrated model design gives consideration to prediction precision and interpretability, and provides panoramic early warning support from individual default to systematic risk for financing leasing business. The scientificity, the accuracy and the comprehensiveness of financing and leasing risk prediction are improved.
It will be apparent to those skilled in the art that various modifications and variations can be made to the present invention without departing from the spirit or scope of the invention. Thus, it is intended that the present invention also include such modifications and alterations insofar as they come within the scope of the appended claims or the equivalents thereof.