Research On Time And Semantic-Aware Recommendation Method

This thesis is concerned with research on recommendation algorithm,a task of predicting the “rating” or “preference” that a user would give to a given item or object(e.g.,“hotel”,“video”,“goods”,“song”).The current state-of-the-art solutions generate a list of recommendation in one of two ways---through collaborative filtering or content-based filtering approach.Such approaches have been shown to be promising for different recommendation systems and are often combined as hybrid methods.However,each type of these methods has its own strengths and weaknesses.For example,collaborative filtering methods often suffer from three problems: cold start problems,scalability problems,and sparsity problems,while content-based methods are hampered by three problems: limited content analysis,over-specialization,and cold-start for new users.In order to tackle these limitations of these recommendation methods,this thesis investigates the time variation of user preferences and needs,the semantic information of the object content,and the user's special interest in the different component features of each object to improve the quality of the recommendation algorithm.Firstly,this thesis develop an effective scheme for time and semantic-aware hotel recommendations,and considers the user's specific needs for the hotel's multifaceted attributes(e.g.,cleanness,service),the time dynamics of the user's booking of hotels,and the semantic analysis of the review texts.These factors are of vital importance for hotel recommendations.Specifically,this thesis present a hotel recommendation method based on Poisson tensor decomposition to learn latent factor of user,hotel,time and features.Through mining the time variation of the user's reservation,the user-hotel's multifaceted preferences and the semantic information in user's review texts in the potential space,the method proposed in this thesis can more accurately predict the user's preference of the hotel.In addition,the model can be easily generalized to cold-start users,predicting the multifaceted preferences of users with only a few review texts,which in turn facilitates hotel recommendations for new users.Then,this thesis deeply explores the impact of time sequence information of user behavior on user preference learning,and considers the user's preference for different component features of each object.This thesis demonstrates that this consideration has a positive effect on predicting and recommending the next item that the user might prefer.Based on this,this thesis designs a time and semantic-aware collaborative neural networkframework based on attention mechanism Transformer to improve the quality of short video recommendations.Specifically,the method proposed in this thesis includes: short video encoder to capture the different importance of short video multi-modal features and then learn short video feature embedding representation;user preference decoder is used to distinguish user historical interaction records,modeling sequence behavior to learn the user's preference representation;the scoring decoder receives a new short video as input to predict user preferences or scores,thereby obtaining a ranking list of short videos and then making recommendation.This thesis demonstrates how attention mechanism networks can be combined with collaborative filtering to model user-video interaction sequences for short video recommendations with the time series and content semantic awareness.Extensive experiments have been conducted on several real-world datasets.In terms of hotel recommendation,this thesis crawls users' comments on the hotel from the TripAdvisor website to build user-hotel interaction datasets.In terms of short video recommendation,this thesis builds two short video datasets from short video platforms of Toffee and Tik-Tok,crawling video that the user prefer as user-video interaction.From the experimental results,the following conclusions can be drawn: First,the use of multi-modal information combinations can help improve the quality of user interest prediction and item recommendation.And considering the user's attention for different component features of the item helps to mine the potential preferences of users.Second,temporal dynamics and time sequence information play an important role in user behavior modeling.Third,content semantic information,such as text,video visual content,etc.,can improve the quality of item feature representation,and thus improve the recommendation effectiveness.