Research On Recommendation Algorithm Based On User Behavior Sequenc

Recently,sequential recommendation have attracted more and more interest as a research point of recommendation.Different from collaborative filtering-based and feature fusion-based recommendation,sequential recommendation take user’s behavioral sequence data as the research object to capture the evolution of user’s interest over time.With the development of deep learning,deep networks are widely used to model user’s behavior sequences.Compared with the traditional sequential recommendation based on factorization machine and Markov chain,which failed to model long sequence data,deep networks can take into account the modeling task of both long and short sequence data,and handle continuous interaction behavior data more effectively.Although existing deep learning-based sequential recommendation have made many breakthroughs,there are still two problems: 1)While existing recommendation have achieved great performance improvements by incorporating the modeling of temporal correlation of interactions in the sequential modeling framework,they only consider the time interval of interactions,making them limited in capturing the temporal dynamics of user preferences.2)Sequential recommendation based on deep networks rely on a large amount of training data,and the high cost of acquiring user behavior data leads to a serious data sparsity problem,which greatly limits deep recommendation models from reaching their full potential.In this paper,we conduct the following two studies to solve the above problems in sequential recommendation.1)To address the problem of inadequate extraction of temporal correlations in the existing sequence modeling framework,a new time-aware positional embedding method is designed to combine temporal information with positional embedding to learn the temporal correlations of items while modeling position correlations.Subsequently,the Time-aware Sequential Recommendation based on Dual-Tower Self-Attention(Ti DSA)is proposed based on the time-aware positional embedding.Ti DSA contains item-level and feature-level self-attention blocks,which analyze the process of user preference change over time from two perspectives,respectively,and achieve the unified modeling of time,item and feature.In addition,in the feature-level self-attention blocks,the self-attention weights are designed to be calculated from multiple perspectives to fully learn the correlation from item-item,item-feature and feature-feature.Finally,the final user preference representation is obtained by fusing the item-level and feature-level information,thereby providing users with reliable recommendation results.Adequate comparison experiments and ablation experiments are also done on four real recommendation datasets.The results show that Ti DSA outperforms many advanced baseline models.2)To address the problem of static sequence segmentation and fused multi-scale interest,We design a Contrast Learning based and multi-scale interest fusion for Session-based Recommendation(CL4SRec).Firstly,CL4 SRec separates the sequence data into multiple sessions based on the interaction time,and divides the global/current sessions by the start time of the latest session as the split line.Then,We design two independent encoders to model global and current interests of users respectively.Moreover,to capture the relevance and difference between the global-current interests,we introduce a contrast learning task to assist the training and enhance the correlation between the global-current interest representation vectors.Finally,we design an attention-based interest fusion network to achieve global-current interest fusion in an adaptive way.We have conducted extensive experiments on classical datasets from industry.The experimental results show that CL4 SRec outperforms many baseline models.In addition,the effectiveness and necessity of the components of CL4 SRec are further validated by ablation experiments.