Research On Personalized Recommendation Model And Key Technologies Based On Attention Mechanism

Recommendation system plays a more and more important role in modern Internet services,and has penetrated into all aspects of people's life.The research on recommendation model has high social value.The application prospect of model is broad,but its development still faces many challenges.For example,the traditional recommendation algorithms are shallow models,which cannot extract the high-dimensional hidden deep-level features between items from massive data.The embedded feature dimension extracted by the existing recommendation model based on deep learning is limited and cannot personalize the user interest features.Moreover,the structure of user interest feature vector extracted by the existing baseline model is single,which cannot fit the dynamic change process of user interest with time,and cannot avoid the interference of user interest drift.In addition,most of the existing conversational recommendation algorithms usually take the whole user behavior sequence as a whole,only use the attention mechanism in a specific space,and do not analyze the implicit characteristics of user behavior at the session level in a more fine-grained multi-dimensional space.In order to solve the above problems,thesis proposes a personalized recommendation model based on attention mechanism.the research on this topic is conducted with the following main work:(1)The research on the evolutionary recommendation model of deep interest based on attention mechanism.Aiming at the problems that the existing baseline recommendation model cannot personalize the characteristics of users' interests and cannot avoid the interference of users' interest drift,Thesis proposes a deep interest evolutionary recommendation model based on attention mechanism,and designs an interest extraction layer to locally activate some user behavior sequences,Fully excavate the implicit relationship between users and items,and realize the personalized expression of users' interest characteristics;The design of interest evolution layer uses the double-layer gate cycle structure with attention mechanism to extract the characteristics of users' long-term and short-term interests,and effectively capture the evolution process of users' interests.Finally,experiments on different comment databases verify the effectiveness of the recommendation model proposed.(2)The research on the multi head self-attention conversational deep neural network recommendation model.Most of the existing conversational recommendation algorithms do not fully consider the implicit characteristics of other dimensions in multiple different spaces,and have weak ability in personalized user interest modeling.Thesis proposes a head self-attention conversational deep neural network recommendation model,designs the session division layer,introduces the sequential location information between different behaviors in and between sessions,and excavates the multi-dimensional implicit relationship between users' behaviors in the session through the stacking of multiple head self-attention sublayers in the session interest interaction layer,Then,the evolution law of user interest between sessions is captured in the session interest activation layer combined with the context system.Finally,experiments on different session databases verify the effectiveness of the recommendation model proposed.