Recommendation Method by Fusing Interactive Sequences and Class Labels

EasyChair Preprint 1200

Abstract

The deep learning method introduces the recommendation system, which can solve the problem that the traditional recommendation system cannot capture the user's evolution preferences over time, but the existing sequence recommendation only considers the sequential similarity between the items and does not consider the content feature information between the items. This paper proposes a recommendation method by fusing interactive sequences and class labels based on the deep bidirectional LSTM model. The method is divided into two main parts: improved item embedding and deep bidirectional LSTM model preference learning. First, the item2vec model is used to embed the user interaction sequence into a low-dimensional space vector representation, and a category label vector is added for each embedded item vector. Secondly, the embedded item vector is input into the bidirectional LSTM model to learn the user's preference vector. Finally, the recommendation list is generated by the preference vector to calculate the most similar items in the embedded space. By experimenting on the real data set and comparing with the advanced methods, we proves that the method has improved the evaluation index recall rate and the Mean Reciprocal Rank compared with the comparison method, which better solves the problem of user interest evolution over time.

Keyphrases: Class Label, Deep Bidirectional LSTM, Interactive Sequences, deep learning

@booklet{EasyChair:1200,
  author    = {Chuanchuan Zhao and Jinguo You and Jiaman Ding},
  title     = {Recommendation Method by Fusing Interactive Sequences and Class Labels},
  howpublished = {EasyChair Preprint 1200},
  year      = {EasyChair, 2019}}