The way in which we construct conventional machine studying fashions is to first prepare the fashions on a “coaching dataset” — sometimes a dataset of historic values — after which later we generate predictions on a brand new dataset, the “inference dataset.” If the columns of the coaching dataset and the inference dataset don’t match, your machine studying algorithm will often fail. That is primarily attributable to both lacking or new issue ranges within the inference dataset.
The primary downside: Lacking components
For the next examples, assume that you just used the dataset above to coach your machine studying mannequin. You one-hot encoded the dataset into dummy variables, and your absolutely reworked coaching information appears like under:
Now, let’s introduce the inference dataset, that is what you’d use for making predictions. Let’s say it’s given like under:
# Creating the inference_data DataFrame in Python
inference_data = pd.DataFrame({
'numerical_1': [11, 12, 13, 14, 15, 16, 17, 18],
'color_1_': ['black', 'blue', 'black', 'green',
'green', 'black', 'black', 'blue'],
'color_2_': ['orange', 'orange', 'black', 'orange',
'black', 'orange', 'orange', 'orange']
})
Utilizing a naive one-hot encoding technique like we used above (pd.get_dummies
)
# Changing categorical columns in inference_data to
# Dummy variables with integers
inference_data_dummies = pd.get_dummies(inference_data,
columns=['color_1_', 'color_2_']).astype(int)
This may rework your inference dataset in the identical method, and also you receive the dataset under:
Do you discover the issues? The primary downside is that the inference dataset is lacking the columns:
missing_colmns =['color_1__red', 'color_2__pink',
'color_2__blue', 'color_2__purple']
When you ran this in a mannequin educated with the “coaching dataset” it will often crash.
The second downside: New components
The opposite downside that may happen with one-hot encoding is that if your inference dataset contains new and unseen components. Think about once more the identical datasets as above. When you study intently, you see that the inference dataset now has a brand new column: color_2__orange.
That is the alternative downside as beforehand, and our inference dataset comprises new columns which our coaching dataset didn’t have. That is really a standard prevalence and may occur if one in every of your issue variables had modifications. For instance, if the colors above characterize colors of a automotive, and a automotive producer out of the blue began making orange vehicles, then this information won’t be obtainable within the coaching information, however may nonetheless present up within the inference information. On this case you want a strong method of coping with the difficulty.
One may argue, nicely why don’t you listing all of the columns within the reworked coaching dataset as columns that may be wanted to your inference dataset? The issue right here is that you just usually don’t know what issue ranges are within the coaching information upfront.
For instance, new ranges might be launched recurrently, which may make it tough to keep up. On prime of that comes the method of then matching your inference dataset with the coaching information, so that you would wish to examine all precise reworked column names that went into the coaching algorithm, after which match them with the reworked inference dataset. If any columns have been lacking you would wish to insert new columns with 0 values and in case you had additional columns, just like the color_2__orange
columns above, these would have to be deleted. This can be a reasonably cumbersome method of fixing the difficulty, and fortunately there are higher choices obtainable.
The answer to this downside is reasonably easy, nevertheless lots of the packages and libraries that try and streamline the method of making prediction fashions fail to implement it nicely. The important thing lies in having a operate or class that’s first fitted on the coaching information, after which use that very same occasion of the operate or class to remodel each the coaching dataset and the inference dataset. Under we discover how that is accomplished utilizing each Python and R.
In Python
Python is arguably one the most effective programming language to make use of for machine studying, largely attributable to its intensive community of builders and mature package deal libraries, and its ease of use, which promotes speedy improvement.
Relating to the problems associated to one-hot encoding we described above, they are often mitigated by utilizing the broadly obtainable and examined scikit-learn library, and extra particularly the sklearn.preprocessing.OneHotEncoder
class. So, let’s see how we will use that on our coaching and inference datasets to create a strong one-hot encoding.
from sklearn.preprocessing import OneHotEncoder# Initialize the encoder
enc = OneHotEncoder(handle_unknown='ignore')
# Outline columns to remodel
trans_columns = ['color_1_', 'color_2_']
# Match and rework the information
enc_data = enc.fit_transform(training_data[trans_columns])
# Get function names
feature_names = enc.get_feature_names_out(trans_columns)
# Convert to DataFrame
enc_df = pd.DataFrame(enc_data.toarray(),
columns=feature_names)
# Concatenate with the numerical information
final_df = pd.concat([training_data[['numerical_1']],
enc_df], axis=1)
This produces a ultimate DataFrame
of reworked values as proven under:
If we break down the code above, we see that step one is to initialize the an occasion of the encoder class. We use the choice handle_unknown='ignore'
in order that we keep away from points with unknow values for the columns once we use the encoder to remodel on our inference dataset.
After that, we mix a match and rework motion into one step with the fit_transform
technique. And eventually, we create a brand new information body from the encoded information and concatenate it with the remainder of the unique dataset.