Abstract:
Understanding sophisticated user behavior-based feature interaction is crucial to optimizing
CTR for recommender systems. Despite having significant progress, the methods that are
used nowadays tend to have a strong bias towards low- or higher-order feature interactions
and involve a great deal of feature engineering. However, most of the feature engineering
methods are nontrivial and often requires rigorous feature engineering or exhaustive searching.
DNNs can learn feature interactions automatically; but they implicitly generate all of
those interactions and there is no control over the DNN about how it is generating all the
cross features thus resulting in many redundant crosses. The proposed model, incorporates
the strength of factorization machines for recommendation and applies PCA for learning
largest data variance to feed the important features to the deep model. Compared to the
Google’s newWide Deep design, we used the regular input for the factorizationmachines,
but to get rid of the redundant cross features we feed the deep model with major features
that impacts the prediction most. The results are very promising and they show a significant
increase in accuracy on the predictions for CTR compared to state of the artWide & Deep
and DeepFM model.
