Research And Implementation Of Dual Sequential Recommendation Model Baseed On Neural Network

In the era of information explosion,the ability of users to receive information is seriously "overloaded",so that it is difficult to filter out the expected information from the huge amount of information.In order to alleviate information overload and meet the needs of users,the recommendation system came into being.In the real world,user interactions are usually not independent of each other,but occur in dependent sequences,which promotes research in the field of sequence recommendation.Based on the sequential interaction of users,sequential recommendation mining user preferences hidden in the interaction behavior and providing personalized recommendation.Although the research on sequential recommendation has achieved certain results,there are still some shortcomings.On the one hand,users only interact with some of the items in the item set,resulting in a sparse dataset.Traditional models that use collaborative features to alleviate the problem of data sparseness cannot capture high-level collaborative features.On the other hand,most of the existing models assume that the attributes of items are static and only analyze the temporal dynamics from the user’s perspective.On the last aspect,due to the uncertainty of user behavior,there may be a problem that interdependent items are not adjacent to each other in the interaction sequence.Based on the above problems,this paper studies two recommendation algorithms in turn.First,the paper proposes Dual Sequential Recommendation with Collaborative Relations(DSRCR for short).This model aims to enrich embedded representations with higher-order collaboration relations and simultaneously model user and item dynamics.The DSRCR model uses a multi-layer graph attention network to model high-level collaboration relationships,and uses the user(item)representation with high-order features as the input of sequence modeling.In order to capture the dynamic characteristics of users and items at the same time,two sequences are constructed and trained in an interactive manner to obtain the dynamic preferences of users and the feature drift of commodities.The effectiveness of DSRCR was verified by comparing different datasets with baselines.Later,the paper proposed a skip behavior information modeling model based on reinforcement learning(Skip-GRU for short).This model aims to capture the skip behavior dependencies in the interaction sequence.Skip-GRU uses a recurrent neural network as the basic model,introduces a policy gradient agent to select the best state from the most recently obtained state,and transfers the selected state to the downstream state through a recursive transfer function.Meanwhile,the state selected by the agent at the previous moment is used to supplement the state selected by the agent.Based on the same dataset,the performance of the Skip-GRU model in evaluation indicators is better than the state-of-the-art baselines and DSRCR.Finally,the personalized paper recommendation system is designed and implemented by analyzing the requirements,sorting out the recommendation process,integratin the proposed dual-sequence collaborative relationship model and skip-behavior modeling model.