Research On Recommendation Algorithm Based On Stacked Auto-encoder

The era of big data has already come,and personalized recommendation has been applied to various fields of social life,as an effective means to solve the problem of information overload and it has a broader market prospects day by day.However,challenges and opportunities coexist.Especially as the capability of data collection continues to expand,the scope continues to expand,and the scale continues to increase.Traditional recommendation algorithms have a bottleneck in information fusion capability.In recent years,deep learning has achieved great success in the fields of image processing,natural language processing and speech recognition,and its excellent data processing capability has been widely recognized.In view of this,research of personalized recommendation algorithm based on deep learning has become a new hot spot.It has been shown that the stacked auto-encoder in the deep learning model can effectively integrate user and item side information,which is beneficial to alleviate the cold start and data sparsity of personalized recommendation.In this regard,this paper focuses on the recommendation algorithm based on stacked auto-encoder to further improve its Top-N recommendation ability,and to provide solutions for the temporal problem in user behavior.Firstly,for the Top-N recommendation problem,a Top-N recommendation model based on stacked denoising auto-encoder is proposed.Traditional recommendation metrics are designed to reduce the difference between predicted ratings and actual ratings.Top-N recommendation focuses on the fit between the user's actual choice and the recommended item list,and this is more closely with the needs of personalized recommendation applications.The model uses the decoding function of the auto-encoder to reconstruct the user's rating vector,and performs Top-N recommendation to the user according to the reconstructed rating vector.For the cold start and data sparseness problems,the model incorporates the user's side information.This paper also verifies the impact of missing rating value processing strategies on recommendation performance.Experiment shows that the model is superior to the existing model in Recall rate.Secondly,aiming at the temporal problem of user behavior,a loosely coupled recommendation framework that combines the temporal characteristics of user behavior is proposed.The temporal problem means that the user's interest often changes or fluctuates with time,thus the general judgment based on the user's historical behavior will damage the recommendation performance.In the traditional model,temporal characteristic processing strategy is closely coupled with the recommendation algorithm.If it is directly applied to the recommendation algorithm based on deep learning,the network structure of the model will become more and more complicated.In this regard,the framework for the first time deal with temporal problem from the perspective of data preprocessing,including temporal pre-processing,recommendation unit and recommendation synthesis three modules.Among them,the temporal pre-processing module is used to analyze the temporal characteristics of user behavior,and the "face" isolation strategy based on anomaly detection is designed for the phase characteristic of user behavior.The recommendation unit directly uses the existing recommendation algorithm and does not need to adjust.Experiment shows that the framework can enhance the application performance of the original recommendation algorithm by using the "face" isolation strategy.