Research On Sequential Recommendation Algorithm Based On Self-supervised Learning And User Dynamic Preference

The sequence recommendation algorithm has better performance than the traditional recommendation algorithm.The sequence recommendation algorithm sorts the items through the timestamp information in the user’s historical behavior sequence and focuses on sequence pattern mining to predict the next item that the user may be interested in.However,with the passage of time,the existing sequence recommendation algorithms are usually affected by sparse data and noisy data,and have problems such as low robustness,insufficient recommendation accuracy,high time complexity,and poor interpretability.The sequence recommendation algorithm often uses the recurrent neural network as the core tool to capture the user’s dynamic preferences.The recurrent neural network can more accurately extract the global characteristics of the sequence.However,at present,the recurrent neural network has the problem that the time dependence monotonically changes with the position in the sequence.That is,for the recurrent neural network,the item or hidden state closest to the current time is always more important than the previous one,this has a great negative impact on the extraction of accurate local features of the sequence.In addition,the attention mechanism is also widely used in the sequence recommendation algorithm at present,but the traditional attention mechanism will distribute attention to all the contexts of the sequence,resulting in the concentration of attention.When the user interaction sequence is long,the main time complexity of the algorithm will be largely consumed in the attention mechanism module,these drawbacks will lead to the reduction of the recommendation accuracy of the recommendation algorithm and affect the overall operation efficiency of the algorithm.Therefore,this thesis first proposes a sequence recommendation algorithm(DGSR)based on different granularity feature extraction,and introducing text CNN to model local sequence features with different granularity between the output hidden states of bidirectional gated recurrent unit(Bi GRU),in order to alleviate the problem of insufficient local feature extraction of the sequence.In addition,considering the disadvantages of the traditional attention mechanism,the sparse attention mechanism is used to better focus attention on the key information between the output hidden states of the bidirectional gated recurrent unit,in order to mine more accurate global features of the sequence and alleviate the problem of low overall efficiency of the algorithm.This thesis compares with the existing excellent algorithms on the public datasets Movie Lens-1M and Steam,and the experimental results prove the effectiveness of the proposed algorithm.Aiming at the problems of low robustness,vulnerability to data sparsity and noise data and insufficient modeling of user dynamic preferences in existing sequence recommendation algorithms,this thesis designs a sequence recommendation algorithm based on self-supervised learning and user dynamic preferences based on the use of the preference learning layer in DGSR algorithm to model user long-term preferences.Specifically,two different data augmentation methods are used to construct effective self-supervised signals for the original user interaction sequence.Secondly,the contrastive self-supervised learning framework is used to conduct explicit selfsupervised learning on users’ long-term and short-term preference learning to optimize the parameters in the corresponding model.In this thesis,multi-task joint modeling is used to minimize the difference between positive and negative pairs of sequence views and maximize the difference between negative pairs of sequence views as pretext tasks,and the sequence recommendation task optimized by using the log-likelihood loss function is used as the main task.In addition,aiming at the problem that the existing self-attention mechanism cannot model the relative position relationship of items in the sequence,the disentangled attention mechanism in the field of natural language processing is introduced into the user’s short-term preference learning process to fully capture the relative position information of items in the user’s short-term sequence.Finally,on the basis of considering users’ long-term and short-term preferences,an effective fusion strategy is proposed to dynamically fuse users’ long-term and short-term preferences to improve the accuracy of recommendations.The proposed sequential recommendation algorithm based on self-supervised learning and user dynamic preference was compared with existing excellent algorithms on the public data sets Movie Lens-1M,Amazon-Beauty and Steam.Experimental results demonstrate the effectiveness of the proposed algorithm on the evaluation indexes HR@N and NDCG@N.At the same time,the ablation experiment proves that the sequence recommendation algorithm proposed in this paper effectively improves the recommendation accuracy of the current recommendation algorithm,and alleviates the existing problems of sparse data and poor robustness of the current sequence recommendation algorithm.