Research On Sequential Recommendation Algorithm Based On Convolutional Neural Network

The problem of information overload is becoming increasingly critical as the Internet advances at dizzying speed.The recommendation system appears at the correct time to shorten the time it takes for users to find things they are interested in and to increase the income of the Internet platform.As an important sub-domain of recommendation systems,sequential recommendation mines user interest by modeling the sequential nature of user behaviors.At present,sequential recommendation has made some progress,but still faces some challenges.For example,one-way modeling of user-item interaction limits the predictive power of the model.The model cannot mine dependencies beyond the length of subsequence since the user behavior sequence is trained in segments.Existing sequential recommendation algorithms merely model user behavior sequences and do not exploit user historical information efficiently.This thesis conducts research on the above-mentioned problems,and the main research contents include:1.A CNN-based sequential recommendation algorithm with bidirectional encoding was presented to address the problem that the sequential recommendation could not capture the bidirectional dependencies of the user-item interaction in one-way modeling.The item context features are extracted using the non-causal convolution encoder and fused with the item embedding vector,after which the causal convolution decoder is used to forecast the user’s next likely interaction.The results of the experiments demonstrate that mining the bidirectional dependencies of items can improve the recommendation algorithm performance.2.A CNN-based sequential recommendation algorithm with memory module is proposed to address the problem that the sequential recommendation divides user behavior sequences of different lengths into equal-length subsequences in the preprocessing stage,resulting in the loss of associations between subsequences and the model’s inability to capture dependencies beyond the length of the subsequences.During the convolution operation,the hidden state of the previous subsequence cached in the memory module is read as the padding content,thereby establishing the connection between the subsequences.Compared with the default zero padding,this algorithm can mine dependencies beyond the length of a subsequence,which helps to capture the user’s long-term interest.The results of the experiment suggest that the algorithm is effective.3.A CNN-based sequential recommendation algorithm combined with user information is presented to address the problem that the existing sequential recommendation algorithm just models the user behavior and does not fully utilize the user history information.The user features are extracted based on feature crosses,and the item features are extracted by convolutional neural network.By integrating the user features with the item features can predict the user’s next possible interaction.Experimental results show that the fusion of user information can enhance the expression ability of the model and improve the recommendation effect of the algorithm.