Using Standardscaler on 3D data

user_8907239 0
Viewed 3993

Im trying to scale a dataset with multiple features and time-series data using the scikit-learn standardscaler. At the moment I am creating a seperate scaler for every feature:

scale_feat1 = StandardScaler().fit(data[:,:,0])
scale_feat2 = StandardScaler().fit(data[:,:,1])
..

Is there a way to scale all features separately using one scaler? Also what is the easiest way to save a scaler for all features and apply it to a valdidation dataset?

Edit: Standardscaler only works on 2D Arrays, so the array would have to be flattened for scaling. In 2D Standardscaler creates a seperate mean and std-dev for every feature

scikit-learn python
2 Answers

Assuming that your data is shaped [num_instances, num_time_steps, num_features] what I would do is first reshape the data and then normalize the data.

import numpy as np
from sklearn.preprocessing import StandardScaler

scaler = StandardScaler()
num_instances, num_time_steps, num_features = train_data.shape
train_data = np.reshape(train_data, shape=(-1, num_features))
train_data = scaler.fit_transform(train_data)

This will reshape the data in a format where each feature is one column and it will normalize each feature separately. Afterwards, you can just return the data in the same shape before training.

train_data = np.reshape(train_data, shape=(num_instances, num_time_steps, num_features))

When it comes to the using the scaler on the validation set, the fit_transform method computes the mean and std on the train set and stores them in the object. Then, when you want to normalize the validation set you can do:

num_instances, num_time_steps, num_features = val_data.shape
val_data = np.reshape(val_data, shape=(-1, num_features))
val_data = scaler.transform(val_data)

And afterwards reshape the data in the shape that you need for training.

val_data = np.reshape(val_data, shape=(num_instances, num_time_steps, num_features))

This should do the trick for you.

Update:

As per @Medomatto comment, in the later numpy versions the correct way to reshape would be:

... = np.reshape(data, newshape=(...))
user_39870850

From the way you are doing it, it seems one could straightforwardly do

scale_feat = StandardScaler().fit(data[:,:,:])

to apply the scaling to all the features.

If you want the scaling to be applied on let's say first 5 features, you can do 

scale_feat = StandardScaler().fit(data[:,:,0:5])
user_49323160
Related