USER SESSION INTEREST PREDICTION MODEL BASED ON SEQUENCE FEATURE FUSION

Because of the time sequence features naturally existing in session, traditional session-based models focus on this feature but pay little attention to the mining of sequence-independent features. In view of this, this paper proposes a recommendation model which integrates two sequential features of sessions. The sequential dependent model used Bi-LSTM and attention mechanism to obtain the representation of user's homogeneous interest features. In the sequence-independent dependency model, multi-layer perceptron, residual network and attention mechanism were used to capture the heterogeneous interest representation. The user's next click item was predicted by fusing these two kinds of interest feature representation. Experiments show that compared with the suboptimal model, the precision is improved by 0.88 percentage points and the mean reciprocal rank is improved by the 0.41 percentage points on the Yoochoose dataset, the mean reciprocal rank is improved by 0.15 percentage points on the Diginetica dataset.