Neural network-based item recommendation method for memory-aware gated factorization machine

Abstract

The present invention provides a neural network-based item recommendation method for a memory-aware gating factorization machine. The method is implemented using an item recommendation model. The item recommendation model includes four parts: an input layer; feature extraction with a gated filter unit; memory perception Feature extraction and score prediction layers. The present invention closely combines the memory network and the collaborative filtering method, so that the performance of the model is greatly improved; inspired by the memory network, a memory matrix is ​​used for each user to record its historical interaction items, and the historical records read from the memory matrix The neural network is mapped to the feature representation of the user's recent preferred item to correct the feature vector of the current item; in addition, the present invention designs a gating unit to filter the auxiliary information to prevent the model from over-fitting.

Description

technical field [0001] The invention relates to the field of item recommendation, in particular to an item recommendation method based on a neural network-based memory-aware gating factorization machine. Background technique [0002] The recommendation system has been widely used in many fields. Collaborative filtering is one of the most widely used methods in the recommendation system. This method believes that users are more interested in items that are similar to the items they have interacted with in their history. Matrix factorization, the most popular collaborative filtering technique, is based on the assumption that there is a linear relationship between users and items. This assumption limits its performance, since in the real world, such relationships are often complex. In addition, there is a factorization machine, which is equivalent to matrix decomposition without the fusion of user and item auxiliary information. With more auxiliary information, the factorizati...

AI Technical Summary

Problems solved by technology

[0003] The existing item recommendation methods have the following deficiencies: 1) The current recommendation algorithm based on factorization machine does not treat the input features differently

The importance of different features in the input features is different, and the interaction between different features should also be different, but the existing factorization machine-based models all treat the input features equally, and the learned features cannot Accurately represent users or items; 2) In the real world, users' current preferences are greatly affected by their historical interactions. Many existing methods perform well, but they usually combine all historical interactions of a specific user to map to A fixed latent vector to predict the user's next likely interest item

This method does not treat all the user's historical interaction items differently, which weakens the recommendation performance of the model, because the influence of the user's historical interaction items on the user's current preference is not equally important; The accuracy of the results is very helpful, and many existing methods cannot effectively and automatically capture the features in these auxiliary information

Images