You must prepare your raw data using data transforms prior to fitting a machine learning model.
This is required to ensure that you best expose the structure of your predictive modeling problem to the learning algorithms.
Applying data transforms like scaling or encoding categorical variables is straightforward when all input variables are the same type. It can be challenging when you have a dataset with mixed types and you want to selectively apply data transforms to some, but not all, input features.
Thankfully, the scikit-learn Python machine learning library provides theColumnTransformer that allows you to selectively apply data transforms to different columns in your dataset.
In this tutorial, you will discover how to use the ColumnTransformer to selectively apply data transforms to columns in a dataset with mixed data types.
After completing this tutorial, you will know:
The challenge of using data transformations with datasets that have mixed data types.
How to define, fit, and use the ColumnTransformer to selectively apply data transforms to columns.
How to work through a real dataset with mixed data types and use the ColumnTransformer to apply different transforms to categorical and numerical data columns.
Kick-start your project with my new bookData Preparation for Machine Learning, includingstep-by-step tutorials and thePython source code files for all examples.
Let’s get started.
Update Dec/2020: Fixed small typo in API example.
Use the ColumnTransformer for Numerical and Categorical Data in Python Photo byKari, some rights reserved.
Tutorial Overview
This tutorial is divided into three parts; they are:
Challenge of Transforming Different Data Types
How to use the ColumnTransformer
Data Preparation for the Abalone Regression Dataset
Challenge of Transforming Different Data Types
It is important to prepare data prior to modeling.
This may involve replacing missing values, scaling numerical values, and one hot encoding categorical data.
Data transforms can be performed using the scikit-learn library; for example, theSimpleImputer class can be used to replace missing values, theMinMaxScaler class can be used to scale numerical values, and theOneHotEncoder can be used to encode categorical variables.
For example:
1
2
3
4
5
6
7
...
# prepare transform
scaler=MinMaxScaler()
# fit transform on training data
scaler.fit(train_X)
# transform training data
train_X=scaler.transform(train_X)
Sequences of different transforms can also be chained together using thePipeline, such as imputing missing values, then scaling numerical values.
It is very common to want to perform different data preparation techniques on different columns in your input data.
For example, you may want to impute missing numerical values with a median value, then scale the values and impute missing categorical values using the most frequent value and one hot encode the categories.
Traditionally, this would require you to separate the numerical and categorical data and then manually apply the transforms on those groups of features before combining the columns back together in order to fit and evaluate a model.
Now, you can use the ColumnTransformer to perform this operation for you.
Want to Get Started With Data Preparation?
Take my free 7-day email crash course now (with sample code).
Click to sign-up and also get a free PDF Ebook version of the course.
How to use the ColumnTransformer
TheColumnTransformer is a class in the scikit-learn Python machine learning library that allows you to selectively apply data preparation transforms.
For example, it allows you to apply a specific transform or sequence of transforms to just the numerical columns, and a separate sequence of transforms to just the categorical columns.
To use the ColumnTransformer, you must specify a list of transformers.
Each transformer is a three-element tuple that defines the name of the transformer, the transform to apply, and the column indices to apply it to. For example:
(Name, Object, Columns)
For example, the ColumnTransformer below applies a OneHotEncoder to columns 0 and 1.
The example below applies a SimpleImputer with median imputing for numerical columns 0 and 1, and SimpleImputer with most frequent imputing to categorical columns 2 and 3.
Any columns not specified in the list of “transformers” are dropped from the dataset by default; this can be changed by setting the “remainder” argument.
Settingremainder=’passthrough’ will mean that all columns not specified in the list of “transformers” will be passed through without transformation, instead of being dropped.
For example, if columns 0 and 1 were numerical and columns 2 and 3 were categorical and we wanted to just transform the categorical data and pass through the numerical columns unchanged, we could define the ColumnTransformer as follows:
A ColumnTransformer can also be used in a Pipeline to selectively prepare the columns of your dataset before fitting a model on the transformed data.
This is the most likely use case as it ensures that the transforms are performed automatically on the raw data when fitting the model and when making predictions, such as when evaluating the model on a test dataset via cross-validation or making predictions on new data in the future.
The dataset has 4,177 examples, 8 input variables, and the target variable is an integer.
A naive model can achieve a mean absolute error (MAE) of about 2.363 (std 0.092) by predicting the mean value, evaluated via 10-fold cross-validation.
We can model this as a regression predictive modeling problem with a support vector machine model (SVR).
Reviewing the data, you can see the first few rows as follows:
1
2
3
4
5
6
M,0.455,0.365,0.095,0.514,0.2245,0.101,0.15,15
M,0.35,0.265,0.09,0.2255,0.0995,0.0485,0.07,7
F,0.53,0.42,0.135,0.677,0.2565,0.1415,0.21,9
M,0.44,0.365,0.125,0.516,0.2155,0.114,0.155,10
I,0.33,0.255,0.08,0.205,0.0895,0.0395,0.055,7
...
We can see that the first column is categorical and the remainder of the columns are numerical.
We may want to one hot encode the first column and normalize the remaining numerical columns, and this can be achieved using the ColumnTransformer.
First, we need to load the dataset. We can load the dataset directly from the URL using theread_csv() Pandas function, then split the data into two data frames: one for input and one for the output.
The complete example of loading the dataset is listed below.
Note: if you have trouble loading the dataset from a URL, you can download the CSV file with the name ‘abalone.csv‘ and place it in the same directory as your Python file and change the call toread_csv() as follows:
1
2
...
dataframe=read_csv('abalone.csv',header=None)
Running the example, we can see that the dataset is loaded correctly and split into eight input columns and one target column.
1
(4177, 8) (4177,)
Next, we can use theselect_dtypes() function to select the column indexes that match different data types.
We are interested in a list of columns that are numerical columns marked as ‘float64‘ or ‘int64‘ in Pandas, and a list of categorical columns, marked as ‘object‘ or ‘bool‘ type in Pandas.
Finally, we can evaluate the model using 10-fold cross-validation and calculate the mean absolute error, averaged across all 10 evaluations of the pipeline.
1
2
3
4
5
6
7
8
9
...
# define the model cross-validation configuration
cv=KFold(n_splits=10,shuffle=True,random_state=1)
# evaluate the pipeline using cross validation and calculate MAE
Running the example evaluates the data preparation pipeline using 10-fold cross-validation.
Note: Yourresults may vary given the stochastic nature of the algorithm or evaluation procedure, or differences in numerical precision. Consider running the example a few times and compare the average outcome.
In this case, we achieve an average MAE of about 1.4, which is better than the baseline score of 2.3.
1
2
(4177, 8) (4177,)
MAE: 1.465 (0.047)
You now have a template for using the ColumnTransformer on a dataset with mixed data types that you can use and adapt for your own projects in the future.
Further Reading
This section provides more resources on the topic if you are looking to go deeper.
In this tutorial, you discovered how to use the ColumnTransformer to selectively apply data transforms to columns in datasets with mixed data types.
Specifically, you learned:
The challenge of using data transformations with datasets that have mixed data types.
How to define, fit, and use the ColumnTransformer to selectively apply data transforms to columns.
How to work through a real dataset with mixed data types and use the ColumnTransformer to apply different transforms to categorical and numerical data columns.
Do you have any questions? Ask your questions in the comments below and I will do my best to answer.
It providesself-study tutorials withfull working code on: Feature Selection,RFE,Data Cleaning,Data Transforms,Scaling,Dimensionality Reduction, and much more...
Bring Modern Data Preparation Techniques to Your Machine Learning Projects
Wow this function with the pipeline seems to be magical. I really wanted to know what pipelines can do(the power of pipelines). I have seen the documentation of pipeline(), it’s not as simple as you explain 🙂 I really likes your way of explanation. Have you written any blogs on introduction to pipelines which can start with simple and explain complicated pipelines. if so, please share the links.
if not, Can you please share some references where we can learn more about the sklearn pipeline?
I have followed your example , I am having a dataset with 3 columns as feature and 1 as label and all are categorical, no numeric values. I have printed X,y and seems they are correct but out is coming as MAE: nan (nan). Can you please suggest what wrong I am doing.
I was wondering how to get the full list of transformations that can be applied to ColumnTransformer and the best reference I found is below: https://scikit-learn.org/stable/modules/preprocessing.html# Do you know maybe some broader source for this topic?
Love this. Makes you realize how ColumnTransformer can make your life radically easy. But what I didn’t get is are the transformations applied in series ?
Cause I want to use an Imputer first and then normalize/onehot it.
Turns out you can\t :\ it throws error when dealing with NaN values, even though the imputer is the first transformation. Need to use Imputer first and then Minmax AGAIN
Actually the example I shared below produces 2 different results. The one without the pipeline gives a NaN found error. Perhaps the Pipeline is needed to execute in a sequence while ColumnTransformer doesn’t do that ?
Another issue I observed was while doing transformations of train and valid datasets. The resultant train dataset returned ascipy.sparse.csr.csr_matrix while the valid data just returned an ndarray.
I reduced thesparse_threshold but that results in a feature mismatch while predicting on the validation dataset.
Anyways I have bothered you enough already, I will figure it out somehow 🙂
In the example just after the paragraph “This is the most likely use case as it ensures that the transforms are performed automatically on the raw data when fitting the model and when making predictions, such as when evaluating the model on a test dataset via cross-validation or making predictions on new data in the future..” you say model.fit(train_X, train_y), but i think it should be: pipeline.fit(train_X, train_y)
Hi, Thanks for the excellent article. Had three clarifications:
1. So if I ever have to do more than just transform values in preprocessing (like impute & then transform), I need to put each individual combination in a Pipeline? These cannot directly go into the column transformer?
2. I assume we can do multiple splits of the dataframe too? Not just 2? Like if we want to LabelEncode some cat columns, ‘most_frequent’ impute & OneHotEncode other cat columns, impute ‘median’ for a few num columns and impute ‘mean’ for others followed by scaling them?
3. Lastly, if you want to drop certain columns – is it recommended to do so at first outside of any pipeline/transformer?
I was reading the entire syntax of Pipeline & ColumnTransformer. There is an explicit drop option in there too, where you can specify specific columns to be dropped. Might actually be better to use that, just to make it easier for validation/test/production data preprocessing. This way you can use the ‘pass_through’ option while dropping columns you want.
1. If I am using SMOTE and if i prefer using imblearn pipeline. Is there a way in imblearn pipeline to do a imputer or should we rely on sklearn pipeline
2. How to integrate imblearn pipeline and sklearn or is it advisable to do so?
Thanks much for the clarification but I would also like to know something in detail
What is the difference or similarity between the sklearn and imblearn pipeline then?
As far as I know – SMOTE can be performed only in imblearn pipeline it cannot be done in sklearn But how do other functionality work? Could you give an example to illustrate both of them (similarly and dissimilarity )- if you have any.
And also the possibility of integrating the above two pipelines (i.e) imblearn and sklearn? Is the above possible? Like say if I notice an imbalance in dataset so I have decided to go with imblearn pipeline – any means of adding up sklearn pipeline for Imputer or scaling after doing SMOTE in imblearn.
I was looking to apply a scaler and one hot encoding in place so that I don’t lose the index of my data frame . I came across your blog demonstrating the column transformer and, this was exactly what i needed. however there seems to be one drawback to this method, you lose the names of the features. Column transformer returns an sparse matrix scaled and encoded.
Is there a way to retrieve to column names and recreate a data frame? This would help when it comes to model interpretation
Let’s say we have a dataset with 5 features. We apply column_transformer only to 2nd a 3rd columns, and passthrough the rest of them. After applying the pipeline, the new order of columns will be: 2,3,1,4,5 (first the ones that have been transformed and then the rest of them) At the end, the dataset ends up having a different column order, why ? Is there any way to avoid this ?
Let’s say we have features 1,2,3,4 and 5 as part of a Pipeline Step 1 is MinMaxScaler for features 4 and 5 only (using column_transformer) Step 2 is RFE or RFECV Step 3 is a model
After step 1, new column order is 4,5,1,2,3 If step 2 (RFE) says that it kept only first two columns: which two columns were selected ? 4 and 5 or 1 and 2? I think 4 and 5, since at that point columns were re-ordered.
It happened to me and couldn’t believe RFE had seleted those features, but when I thought they were re-ordered, then it made sense.
I hope it is more clear now. Thanks a lot for your help
It does not matter what RFE selected, as you are evaluating the modeling pipeline. You are answering the question “what is the performance of this pipeline” not “what features would RFE select”. If you want to know the latter, you can study that in isolation.
You allow RFE to choose features based on training data, just like allow the model to fit internal parameters based on the training data. You don’t ask “what are the vales of the coefficients internal to the model” because it doesn’t really matter when you’re focused on model skill.
That all being said, you can access the RFE model in the pipeline or keep a reference to it and report the features that were selected from a single run.
Hi, one occasion where it matters is with LightGBM and categorical features. Using the sklearn API with LightGBM, the categorical features are specified as a parameter to .fit(). Since the DataFrame is casted to a numpy array during transformation (with for instance StandardScaler()), it is practical to specify categorical features with a list of int. Reordering of columns then makes for a “hard to find” bug.
model.fit(X_train,y_train,validation_data=(X_test,y_test),epochs=10,batch_size=64) ############# InvalidArgumentError: 2 root error(s) found. (0) Invalid argument: indices[55,1] = 170355 is not in [0, 5000) [[node sequential_3/embedding_3/embedding_lookup (defined at :9) ]] (1) Invalid argument: indices[55,1] = 170355 is not in [0, 5000) [[node sequential_3/embedding_3/embedding_lookup (defined at :9) ]] [[Adam/Adam/update/AssignSubVariableOp/_35]] 0 successful operations. 0 derived errors ignored. [Op:__inference_train_function_12567]
Errors may have originated from an input operation. Input Source operations connected to node sequential_3/embedding_3/embedding_lookup: sequential_3/embedding_3/embedding_lookup/11365 (defined at /usr/lib/python3.6/contextlib.py:81)
Input Source operations connected to node sequential_3/embedding_3/embedding_lookup: sequential_3/embedding_3/embedding_lookup/11365 (defined at /usr/lib/python3.6/contextlib.py:81)
Function call stack: train_function -> train_function
Great tutorial. Can we make Column Transformer part of param_grid ? I want to transform some columns using categorical encoding mechanisms, but there are several of interest: Target Encoding, Weights of Evidence encoding, catboost encoding, etc. Therefore, I want each encoding mechanism to be a hyper parameter inside grid search param_grid. Do you think its possible ?
I see ColumnTransformer() as a very powerful module to apply globally, but distinguishing transformation to every feature of dataset. I think it is useful not only inside the Pipeline steps but also as stand alone in order to get an inside analysis of features
I’m having a hard time applying columnTransform (or pipelines) to a group of: (i) one NLP-based feature (that require count vectorizer or tf-idf fit-transform); and (ii) other simpler features that require standard scaling – any chance you’ve ever faced a similar problem and know how to solve it?
The count vectorizer (or tf-idf) outputs a sparse matrix (while the rest remain as pandas series/df)… should I use numpy arrays for all of them?
Always a valuable piece of code to perform other experiments!
my experiments:
Is true that ColumnTransformer() API perform different transformation by column or feature (great!) and there is an argument remainder= that we can set to ‘passthrough’. Ok.
But my complain to Sklearn API is that transformations are performed consecutively, so if you decide e.g. to perform column 3 transformation before than 2 …so column 2 is replaced by column 3 transformed.
In addition to that, even if you decide to ‘passthrough’ a remainder column without transformation, between two alternate columns transformation, the ColumnTransformer() replace previous column with the following transformation and leave column (without transformation) at the end of columns serie …
This behaviour introduce a lot of confusion to the way we would expected to work the ColumnTransformer, and it make you wast your time with problems that can be avoided by a good API design :-(( unless would be some other arguments where you can specify them to avoid this missbehaviour columns alterations.
Anyway, as I said before, thank you to your piece of code you can foreseen this behaviour.
You manage to strike a nice balance between complexity and usability and the result is easy to read, easy to follow example that opens the door to a larger field to be explored at once’s own pace.
I’ve been researching high and low for an answer and haven’t been lucky yet. I’m training a pipeline that consists of data transformation pipelines for both numerical and categorical, added to a column transformer pipeline:
The issue that I’m facing is that I will fit_transform this data_pipeline to my training data and save this trained pipeline with joblib dump to use it for transforming with .transform() the validation data and also for inference in production. So far so good, however, when my training data is large-ish the saved file is very large and takes a long time to load (example: trained with 1.4M rows, saved pipeline is 49GB).
I’m wondering what I could be doing to make this saved trained data transformation pipeline lighter. Any tips?
Hi, thanks for writing these examples down. In the last line of your code example: I am not 100% sure, but is it OK to take
std(scores)
here?
This is because the values were previously set to absolute. The distances between the values can therefore be different due to the “zero border”. So if I had a value of -0.1 before and the mean is 0.1, then the distance = 0.2; but after I apply absolute(), then the value is now 0.1 and has a distance of 0 to 0.1. This affects the standard deviation, doesn’t it?
So you would have to take the standard deviation from the scores BEFORE applying absolute(), or am I wrong?
Hi , Im facing one error regarding fit_transform , here is my code :
import pandas as pd import numpy as np from sklearn.preprocessing import OneHotEncoder, StandardScaler, OrdinalEncoder from sklearn.compose import ColumnTransformer import category_encoders as ce
File ~\anaconda3\lib\site-packages\sklearn\utils\_set_output.py:142, in _wrap_method_output..wrapped(self, X, *args, **kwargs) 140 @wraps(f) 141 def wrapped(self, X, *args, **kwargs): –> 142 data_to_wrap = f(self, X, *args, **kwargs) 143 if isinstance(data_to_wrap, tuple): 144 # only wrap the first output for cross decomposition 145 return ( 146 _wrap_data_with_container(method, data_to_wrap[0], X, self), 147 *data_to_wrap[1:], 148 )
File ~\anaconda3\lib\site-packages\sklearn\compose\_column_transformer.py:727, in ColumnTransformer.fit_transform(self, X, y) 724 self._validate_column_callables(X) 725 self._validate_remainder(X) –> 727 result = self._fit_transform(X, y, _fit_transform_one) 729 if not result: 730 self._update_fitted_transformers([])
File ~\anaconda3\lib\site-packages\sklearn\compose\_column_transformer.py:658, in ColumnTransformer._fit_transform(self, X, y, func, fitted, column_as_strings) 652 transformers = list( 653 self._iter( 654 fitted=fitted, replace_strings=True, column_as_strings=column_as_strings 655 ) 656 ) 657 try: –> 658 return Parallel(n_jobs=self.n_jobs)( 659 delayed(func)( 660 transformer=clone(trans) if not fitted else trans, 661 X=_safe_indexing(X, column, axis=1), 662 y=y, 663 weight=weight, 664 message_clsname=”ColumnTransformer”, 665 message=self._log_message(name, idx, len(transformers)), 666 ) 667 for idx, (name, trans, column, weight) in enumerate(transformers, 1) 668 ) 669 except ValueError as e: 670 if “Expected 2D array, got 1D array instead” in str(e):
File ~\anaconda3\lib\site-packages\joblib\parallel.py:1051, in Parallel.__call__(self, iterable) 1048 if self.dispatch_one_batch(iterator): 1049 self._iterating = self._original_iterator is not None -> 1051 while self.dispatch_one_batch(iterator): 1052 pass 1054 if pre_dispatch == “all” or n_jobs == 1: 1055 # The iterable was consumed all at once by the above for loop. 1056 # No need to wait for async callbacks to trigger to 1057 # consumption.
File ~\anaconda3\lib\site-packages\joblib\parallel.py:782, in Parallel._dispatch(self, batch) 780 with self._lock: 781 job_idx = len(self._jobs) –> 782 job = self._backend.apply_async(batch, callback=cb) 783 # A job can complete so quickly than its callback is 784 # called before we get here, causing self._jobs to 785 # grow. To ensure correct results ordering, .insert is 786 # used (rather than .append) in the following line 787 self._jobs.insert(job_idx, job)
File ~\anaconda3\lib\site-packages\joblib\_parallel_backends.py:208, in SequentialBackend.apply_async(self, func, callback) 206 def apply_async(self, func, callback=None): 207 “””Schedule a func to be run””” –> 208 result = ImmediateResult(func) 209 if callback: 210 callback(result)
File ~\anaconda3\lib\site-packages\joblib\_parallel_backends.py:572, in ImmediateResult.__init__(self, batch) 569 def __init__(self, batch): 570 # Don’t delay the application, to avoid keeping the input 571 # arguments in memory –> 572 self.results = batch()
File ~\anaconda3\lib\site-packages\joblib\parallel.py:263, in BatchedCalls.__call__(self) 259 def __call__(self): 260 # Set the default nested backend to self._backend but do not set the 261 # change the default number of processes to -1 262 with parallel_backend(self._backend, n_jobs=self._n_jobs): –> 263 return [func(*args, **kwargs) 264 for func, args, kwargs in self.items]
File ~\anaconda3\lib\site-packages\joblib\parallel.py:263, in (.0) 259 def __call__(self): 260 # Set the default nested backend to self._backend but do not set the 261 # change the default number of processes to -1 262 with parallel_backend(self._backend, n_jobs=self._n_jobs): –> 263 return [func(*args, **kwargs) 264 for func, args, kwargs in self.items]
File ~\anaconda3\lib\site-packages\sklearn\utils\parallel.py:123, in _FuncWrapper.__call__(self, *args, **kwargs) 121 config = {} 122 with config_context(**config): –> 123 return self.function(*args, **kwargs)
File ~\anaconda3\lib\site-packages\sklearn\pipeline.py:893, in _fit_transform_one(transformer, X, y, weight, message_clsname, message, **fit_params) 891 with _print_elapsed_time(message_clsname, message): 892 if hasattr(transformer, “fit_transform”): –> 893 res = transformer.fit_transform(X, y, **fit_params) 894 else: 895 res = transformer.fit(X, y, **fit_params).transform(X)
File ~\anaconda3\lib\site-packages\sklearn\utils\_set_output.py:142, in _wrap_method_output..wrapped(self, X, *args, **kwargs) 140 @wraps(f) 141 def wrapped(self, X, *args, **kwargs): –> 142 data_to_wrap = f(self, X, *args, **kwargs) 143 if isinstance(data_to_wrap, tuple): 144 # only wrap the first output for cross decomposition 145 return ( 146 _wrap_data_with_container(method, data_to_wrap[0], X, self), 147 *data_to_wrap[1:], 148 )
File ~\anaconda3\lib\site-packages\category_encoders\utils.py:458, in SupervisedTransformerMixin.fit_transform(self, X, y, **fit_params) 451 “”” 452 Encoders that utilize the target must make sure that the training data are transformed with: 453 transform(X, y) 454 and not with: 455 transform(X) 456 “”” 457 if y is None: –> 458 raise TypeError(‘fit_transform() missing argument: ”y”’) 459 return self.fit(X, y, **fit_params).transform(X, y)
Hi neelanshuni…You may wish to try your implementation in Google Colab to determine if there may be an issue with your local Python environment. Otherwise, ensure that there are no issues that may have resulted from copy and paste of code.
