I have a dataset that contains text columns. I have used tf-idf to convert those text columns to numerical columns. I want to reduce the dimension of the dataset since tf-idf creates a multitude of new features/columns.

I am thinking of using autoencoder to reduce the dimension of the data by using the output of the encoded layer, concatenate those output to my dataframe and use it as a training set for the random forest.

My questions are: Do the above steps sound reasonable? In particular, should I train the autoencoder on the whole dataset and then use the output of the encoded layer (after training) as the new columns to feed random forest? [See code below]

I am using Keras so after defining the autoencoder model, I fit the whole dataset and then also predict on the whole dataset using only the encoded layer. And then using those predictions as my updated data.

Here is the code I am using:

def autoencoder(df):
    #return the lower dimensional data to feed directly into train_test_split
    df_copy = df.copy()
    input_dim = Input(shape = (df_copy.shape[1], ))

    encoded = Dense(int(df_copy.shape[1]/2), activation = 'relu')(input_dim)

    decoded = Dense(df_copy.shape[1], activation = 'sigmoid')(encoded)

    autoencoder = Model(input = input_dim, output = decoded)

    autoencoder.compile(optimizer = 'adadelta', loss = 'mean_squared_error')
    autoencoder.fit(df_copy.values, df_copy.values, nb_epoch = 10, batch_size = int(df_copy.shape[1]/3))

    encoder = Model(input = input_dim, output = encoded)
    encoded_input = Input(shape = (int(df_copy.shape[1]/2), ))
    encoded_out = encoder.predict(df_copy.values) #Note how I am training on the same data I am predicting on

    return pd.DataFrame(encoded_out)

I am new to autoencoders so any suggestions and help on how to use autoencoders in this context will be appreciated.


I will repost my answer on Quora here, about the "should I train the autoencoder on the whole dataset" part of the questions.

TL;DR: NO, you always use ONLY the training set for training, even for unsupervised learning with autoencoders!

What is the goal of using a test set ? It is meant to represent data that you model (including both the autoencoder and the random forest) have NEVER seen before.

Autoencoders are unsupervised learning models, but it does not mean that they cannot overfit. They can learn latent codes that do not generalize. Imagine an AE that encodes every data point x1, x2, … as a real number (one-dimensional code):

x1 -> 0.0001, x2 -> 0.0002, etc.

Do you think this autoencoder will generalize well ? No, it just remembered the entire training set and did not learn useful representations. That is why we use regularization in autoencoders, in very different forms : undercompleteness, sparsity, denoising, variational autoencoders (VAE)…

Conclusion, if you include the test set to train your autoencoder, it may in certain cases overfit, your performance estimates will be biased and you lose all the purpose of a test set.

(don’t hesitate to ask if you want me to detail some aspects of my answer or have more questions on this matter)

Concerning the architecture of you AE:

  1. using a sigmoid output activation only makes sense if you data takes values between 0 and 1, as a sigmoid function squashes the output into the [0, 1] range. So either you normalize your data to the [0, 1] range, or you use a linear output activation.
  2. With a sigmoid output activation, you should use a binary cross-entropy loss, which is designed for binary variables following Bernoulli distributions. With a linear output, use the mean squared error loss, which is designed for real-valued output (following Gaussian distributions).
  3. Maybe Adam optimizer will converge faster that Adadelta. Try for yourself.
| improve this answer | |
  • $\begingroup$ Thanks Florent for answering. After training the autoencoder, we use the output of the encoder as new features, right? $\endgroup$ – user9933484 Dec 2 '19 at 16:00
  • $\begingroup$ Exactly. I'm not an expert, but on text data, you may also consider a simple SVD on the TF-IDF matrix (if it's not too large), or word/sentence embeddings from generic pre-trained models such as word2vec or some variants of transformers (BERT among many others) $\endgroup$ – Florent F Dec 2 '19 at 16:44

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.