Research On Self-Attention Based Long-and Short-term Sequential Recommendation

The rapid development of the information age has brought convenience to people's lives but also brought information overload.Recommendation system has brought great convenience to consumers on various online platforms.With the outstanding performance of neural networks and attention mechanisms on sequential tasks such as natural language processing,they were quickly applied to recommendation systems.Neural network-related recommendation algorithms mainly obtain the user's overall preferences through multi-layer perceptrons,while the attention mechanism-related methods mainly use their characteristics that can capture the key points of the sequence to implement user sequence recommendation.These methods have achieved good results in implicit feedback recommendation,but with subsequent research,they does not capture the deeper features of user sequences.In order to extract deeper features of user sequences and improve sequence recommendations,a sequence recommendation method that uses neural networks and self-attention mechanisms to combine user long-and short-term features is proposed.On one hand,the user and item vectors are organized by vector combination and splicing,and the multi-layer neural network is used to learn the interaction between the user and the item,the global long-term characteristics of the user can be obtained based on all previous interaction information of the user.On the other,a sequence recommendation method based on the self-attention mechanism is used to learn the self-attention weights between sequence items through the most recent historical interaction information in the user interaction sequence,and a new self-attention calculation method is proposed to extract the user sequence A more fine-grained relationship between the dimensions,and then obtain the short-term sequence characteristics of the user based on these recent item information.Finally,a fusion strategy is used to comprehensively consider the two methods to recommend target users to achieve both long-term and short-term characteristics.For the proposed recommendation method,lots of experiments and comparation are performed on Amazon and Movie Lens datasets.The results show that under the same experimental environment,the Self-Attention based long-and short-term sequential recommendation methods to some extent surpass some of the current mainstream collaborative filtering recommendation and sequence recommendation algorithms.