FM-based Recommendation Model for Short-video with Topic Distribution

Abstract With the popularity of mobile internet terminals, the speed of the network and With the popularization of mobile Internet terminals, the speed of network and the reduction of traffic tariff, people can watch videos through cell phones at any time. As short videos have the characteristics of short time and rich content, they can maximize people's need to watch videos in fragmented time. Therefore, short videos that combine filming techniques, music, stories and images. Short videos can meet users' content consumption needs in a more diversified way. Due to the widespread popularity of short videos, many short video platforms have been born. Many short video platforms have been created. The short video recommendation algorithm has become an important means of competition among platforms. How to ensure the accuracy of the recommendation algorithm and the real-time of the algorithm How to ensure the accuracy of the recommendation algorithm and the real-time of the algorithm have been the focus of research. In recent years, a great deal of progress has been me in the study of problems related to preference recommendation and recommendation. The most widely used method in industry today is the use of LR to learn click-through recommendation models. LR has the vantage of being simple and very easy to implement for massively real-time parallel processing, but linear models have a limited learning capability and do not capture the information carried by higher-order features , thus limiting the recommendation performance. Based on the above analysis, this paper proposes a FM-based short video preference recommendation model from the multi-topic nature of short videos. The main contributions of the model are: Topic-based segmentation of the original training set is performed using LDA, and each sub-training set generated by the segmentation is significantly smaller than the original training set, which reduces the computational complexity to a certain extent. Automated feature selection and linear transformation of features for training sets under different topics, reducing the dependence of manual feature engineering on time and labor in the baseline algorithm. By integrating the recommendation results of different topics, thus improving the recommendation accuracy. Experiments demonstrate that our model can effectively improve the recommendation performance.