Research And Application Of Hybrid Recommendation Algorithm Based On Deep Learning

With the rapid development of the Internet,the problem of information overload has become increasingly serious.Recommender systems are an effective means to solve information overload and have attracted widespread attention in both academia and industry.Deep learning has become the mainstream of current recommendation algorithms due to its powerful feature learning capability,but the existing deep learning algorithms have low recall rate and not rich item candidate set.On the other hand,the current mainstream recommendation algorithms focus more on studying the explicit and implicit feature interactions between users and items,and less on considering the diversity of user interest expressions.To address this situation,a multi-strategy recall model and an explicit and implicit ranking model incorporating multi-headed attention mechanism are studied and constructed,and the performance of the recommendation system is improved.Finally,the improved methods are applied to the film recommendation system,and better experimental results are obtained.The thesis focuses on deep learning-based recommendation methods to improve the accuracy and efficiency of recommendation systems,and the main research includes the following three aspects.(1)A multi-strategy recall model is studied and constructed to address the traditional recall model with a single recall algorithm,which results in an incomplete set of candidate items and low recall rate.Three algorithms,improved DSSM’s two-tower recall,item similarity recall and heat recall,are used to generate recall candidate sets,where the improved DSSM’s two-tower recall model optimizes the parts of the hidden semantic extraction module and negative sample construction,and the improved item similarity recall optimizes the rating difference problem,and then the candidate sets obtained from the three recall strategies are fused using dynamic weighting method.The effectiveness of the model is demonstrated by improving the Recall metrics by 2.83%and 1.61% in the public datasets Movie Lens and Amazon Electronics,respectively.(2)To address the problems that current deep learning ranking algorithms focus more on studying the explicit and implicit feature interactions between users and items when constructing combinatorial features,often ignoring the correlation between features,an x Deep FM ranking model incorporating a multi-headed attention mechanism is studied and constructed.The improved model is able to adaptively learn user representations of different items,thus making the recommendation results more satisfying to the users’ needs.The AUC metrics of Movie Lens and Amazon Electronics in the public dataset improved by 1.06% and 1.33%,respectively,compared with the x Deep FM model,demonstrating the effectiveness of the model.(3)Based on the improved algorithm,we designed and implemented a personalized movie recommendation system.The system uses multi-strategy recall and x Deep FM with multi-headed attention mechanism as the core algorithm to make accurate personalized recommendations for users,and integrates the functions of movie hotness recommendation and rating ranking to achieve the effect of "a thousand people with a thousand faces".