I'm trying to build autoencoder in keras in order to detect anomalies. However, most of the data is categorical and I have to encode it. When it comes to production, categorical features can take new values. So I thought of using entity embedding.
I've created an embedding layer for each categorical feature. In the code below, I concatenate all together and I add other Dense layers:
output = Concatenate()(output_embeddings) encoder = Dense(20, activation="relu", activity_regularizer=regularizers.l1(10e-5))(output) encoder= Dense(nbr_feautres, activation='relu')(encoder) decoder = Dense( ???? )(encoder) model = Model(inputs=input_models, outputs=decoder)
Is it possible to add a decoder layer and how can I implement this? Thanks.