Research On Video Recommendation Based On Attention Mechanism

In the Information Era,rapid developments of the Internet and multimedia technology have witnessed a dramatic increase in the number of online videos and users,but have also caused the problem of Information Overloading.As an effective means to alleviate the pressure of Information Overload,the video recommendation system enables users to quickly find the videos that interest them most from massive amounts of data on the internet.In other words,video recommendation system can greatly minimize the cost of time that our users spend on browsing the internet.Meanwhile,online video providers can provide accurate and personalized recommendations to users by exploring their preferences.In this way,customer loyalty is increased,and user experience is improved.Therefore,recommendation system is a mutual benefit system for both users and providers.Under this circumstance,this thesis proposes two attention-based recommendation algorithms,based on researches on video recommendation system.1)The first algorithm is ADeepFM,which can be used for movie recommendation.In this algorithm,we introduce attention mechanism so that the improved model can learn the output weight of the second-order combination feature adaptively.This new algorithm can effectively reduce voices due to invalid features.When calculating output weights,the second-order combination feature vectors and the target movie vectors are equally divided to form several different implicit semantic spaces.After computing the attention weights in different implicit semantic spaces,we can learn the relationship between two vectors at a fine granularity,which can effectively mitigate the problem of weight smoothing between vector elements.By predicting possible rating scores that a user may give to the target movies,we can recommend movies to the user according to their rankings.Results of off-line experiments show that compared with original DeepFM algorithm,ADeepFM algorithm achieves better performance in RMSE and MAE evaluation indexes,for it overcomes the problem of output weight.2)The second attention-based video recommendation algorithm is SAT2 Rec.In this algorithm,both “rough sorting” and “fine sorting” are adapted.In the stage of "rough sorting",the sequence of user's video viewing history is compared to the word sequence in natural language.While the continuous bag-of-words(CBOW)model is used to build up the user's history sequence,continuous and dense video feature vectors are extracted to avoid the impact of data sparseness.To generate video candidate set,we can calculate the similarity of the video vectors.In the stage of "fine sorting",we construct the user's interest model by applying attention mechanism.Attention mechanism can learn the relationship between the user's viewing history and the target videos,so as to find out the user's interest,further filter candidate videos based on videos similarity,and finally recommend those with high similarity.SAT2 Rec,the attention-based video recommendation algorithm,considers both video similarity and user interest distribution,and thus achieve higher results in evaluation indexes including recommendation accuracy,recall rate,AUC value.