| The progress and development of the times have led to an increasing focus on the enrichment of the spiritual world.As one of the major entertainment industries,cinema has enriched people’s cultural life.In the context of rapid economic development,the film industry has also flourished.However,as the number of films has increased dramatically,it has become increasingly challenging for users to locate the films they are interested in quickly in the vast amount of information available.The emergence of recommender systems,an information filtering technique designed to provide users with personalized recommendations based on their historical behavior,personal preferences and other relevant factors,offers a new way of solving the problem of information overload.Currently,recommendation systems have a mature theoretical foundation and are widely used.However,the traditional single recommendation algorithm is not sufficient to cope with rich business scenarios.To address this problem,build a rich and easy-to-use movie recommendation system by combining the advantages of multiple recommendation algorithms in a hybrid recommendation approach,so as to overcome the shortcomings of a single recommendation algorithm and improve the coverage and variety of recommendations.Based on such a premise,the following work has been accomplished.1.DeepAFM is innovatively proposed to address the problem that DeepFM cannot distinguish the importance between features,which leads to weak model performance,and is applied as one of the recommendation methods in the movie hybrid recommendation system.The model uses the parallel combination structure of Wide&Deep to take into account the representation of both low-order and high-order feature combinations.The model learns the second-order cross-feature vector weights by introducing an attention network,and uses the top-k algorithm to introduce local correlations to optimize the model to sparse attention and measure the contribution of cross-features to the prediction results.The introduction of the attention mechanism saves arithmetic power and improves the model interpretability and accuracy.When the model is applied to a movie recommendation system,it is trained with an improved inverse propensity score,which balances the recommendations of popular and popular movies and weakens the influence of long-tail effect.The model is compared with six other baseline algorithms on the Criteo and Avazu datasets,and the results demonstrate that DeepAFM has higher AUC and ACC and lower Loss values,effectively improving the prediction accuracy.2.The overall design of the movie recommendation system is determined by analyzing user requirements.To meet the user’s needs for basic business functions,and by investigating the principles and characteristics of current mainstream recommendation algorithms in industry,the system’s hybrid recommendation strategy is designed comprehensively for different user needs.The system includes two recommendation modules,offline and real-time,with four recommendation methods.Offline recommendation includes statisticalbased recommendation,ALS collaborative filtering recommendation and DeepAFM model-based recommendation.Real-time recommendation is achieved by improving the collaborative filtering algorithm and introducing enhancement and weakening factors to strengthen the influence of users’ recent rating behaviors on the recommendation results,so as to achieve the effect of real-time recommendation result update.3.Java Web and Spark big data computing engine and other tools are used to build an easy-to-use,stable,hybrid recommendation system for movies that includes multiple recommendation methods.Through pseudodistributed deployment and testing,the system is found to be robust in basic functionality,effective in offline recommendations,timely in real-time recommendation updates,easy to use interface,timely in response,and stable and scalable overall. |
