Design of collaborative filtering recommendation algorithm combining time weight and reward and punishment factors
Panpan Yang · Academic Journal of Computing & Information Science · 2022
For distributed recommendation systems built on Spark and Flink big data platforms, when machine learning libraries are used for offline recommendation, Cartesian product calculation will be carried out, which consumes a large amount of memory and causes long algorithm execution time. In real-time recommendation, traditional recommendation algorithm model cannot dynamically sense the interest drift of users, resulting in poor recommendation results. To solve the above problems. This paper introduces the heapsort algorithm into the offline recommendation algorithm to solve the problem that ALS algorithm in MLlib will perform Cartesian product operation in model prediction, which consumes a lot of memory and takes a long time to execute the algorithm. The real-time recommendation algorithm introduces Ebbinghaus forgetting curve, and the fusion of timing weight and reward and punishment factors to dynamically perceive the drift of user interest and generate personalized TOP-N recommendation results. Experimental results show that the execution rate of the offline algorithm adopted in this paper is significantly improved when the RMSE index is basically unchanged. The accuracy rate and recall rate of the real-time recommendation algorithm are significantly improved, and the recommendation results are more in line with users' interests.