DL Models

Created: 2022-03-23 12:53

Pros:

  • they can deal with non linear transformations;
  • good for sequence modeling, next basket recommendation or session based recommendations (for example #click-logs);
  • representation learning Autoencoder Restricted Boltzmann machine(RBM);
  • easier to build hybrid systems

Cons:

  • interpretability -> black boxes;
  • they need a big amount of data;
  • hyperparameter tuning can take longer time;

Types:

  • Multilayer Perceptron (MLP) -> feed-forward neural network with multiple hidden layers, i.e. stacked layers of nonlinear transformations;
  • CNN can capture global and local features, good for grid-like data;
  • RNN for sequential data;
  • Autoencoder are unsupervised models that use bottleneck layer to learn salient features;
  • *Restricted Boltzmann machine(RBM)*are two layers models (visible and hidden layer);
  • Neural attentionthat use attention mechanism;
  • Neural Autoregressive Distribution Estimation (NADE) are unsupervised neural networks built on top of a autoregressive model and feedforward neural networks;
  • Adversarial Networks (AN) consist of a discriminator and a generator which compete against each other;
  • Deep Reinforcement Learninig (DRL) operate on a trial-and-error paradigm;
  • Transformer for sequential and session-based recommendation

Obviously a DL-based RecSys can use just one DL model or a combination of several DL models ( #deep_hybrid_models ).
Some combinations of models that have been proved to be effective are:

  • CNN + Autoencoder;
  • CNN + RNN;
  • RNN + Autoencoder;
  • Rnn + DRL

References

  1. https://arxiv.org/pdf/1707.07435.pdf
  2. https://medium.com/sciforce/deep-learning-based-recommender-systems-b61a5ddd5456
  3. https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9354169&tag=1

Code

  1. microsoft/recommenders
  2. Paddle
  3. DeepCTR
  4. prediction-flow

Tags

#dl #rnn #cnn #mlp #rbm #drl #nade #autoencoder #an #neural_attention #transformer