Sequential Recommendation Based On Deep Neural Network

To reduce information overload on the web,recommendation systems are widely deployed to perform personalized information filtering.The user-item interactions are sequential and constantly changing.Sequential recommendation considers these interactions as temporal-order sequences,which can effectively capture users’ recent references,thus improving the recommendation performance.Sequential recommendation has drawn a lot of attention due to the good performance it has shown in recent years.The application of deep neural networks in sequential recommendation has achieved many remarkable results in recent years.However,previous works mainly focus on itemitem temporal information of the sequence while ignoring the latent collaborative relations in user-item interactions.Moreover,when extracting the temporal information,the selfattention based method loses the position information in the interaction sequence.Some position encoding methods such as adding position embedding to the item embedding have been generated to alleviate this issue.However,the position embeddings are obtained by independent training,and there is a lack of mutual constraint between different position embeddings.Therefore,such methods allow modeling the absolute position but not the relative-distance relationship of items.This thesis studies the above issues in the following aspects.First,this thesis proposes a new method named Collaborative-Temporal modeling for Sequential Recommendation(CTSR)to learn both collaborative relations and temporal information.The proposed CTSR method introduces a graph convolutional network to learn the user-item collaborative relations while using self-attention to capture item-item temporal information.Second,this thesis extract a portion of item-item pairs that are most valuable and then feed these pairs as augmented information into adjacency matrix of the graph neural network.Besides,this thesis explicitly models relative position information to solve the problem that position embeddings capture the position of individual items but not the ordered relationship between individual item positions.Finally,this thesis apply the focal loss to reduce the loss contribution of the easy samples and increase the contribution of the hard samples,to train the model more effectively.The experimental results show that the method surpasses previous works on three real-world datasets,thus verifying the effectiveness of the proposed method.