结合内容特征提取和弹幕文本的短视频推荐模型构建及仿真

Construction and Simulation of short video recommendation model combining content feature extraction and bullet screen text

针对现有短视频推荐准确率不高的问题,提出一种融合视频内容与弹幕文本分析的短视频推荐方法。该推荐方法中首先采用LDA模型提取弹幕文本主题,然后提取出短视频内容特征,再根据短视频的高光时刻生成候选推荐列表;在候选推荐列表生成基础上,运用RNN对用户的长期兴趣进行建模,使用门控单元GRU处理短期兴趣,最终提出了一种结合长短期的短视频推荐模型,对用户进行推荐。结果证明,深度学习方法可最大化地获取短视频的内容特征,提高处理效率;结合用户长短期兴趣的短视频推荐模型在准确率、召回率以及MRR平均倒数排名等评价指标上优于其他目前使用广泛的推荐模型。由此说明,提出的推荐方法,可以更好地运用在实际推荐中。

Aiming at the problems of low accuracy and low efficiency in the existing short video recommendation algorithms,a short video recommendation method combining video content and bullet screen text analysis is proposed.This recommendation method uses LDA model to analyze the bullet screen text and determine its theme,then extracts the content features of short video through deep learning method and CNN analysis,and finally generates a candidate recommendation list according to the highlight time of short video.In this method,users’long-term and short-term interests are combined to generate a short video recommendation model combining users’long-term and short-term interests.After the comparative experiment,the experimental results show that the short video recommendation model combined with users’long-term and short-term interests is superior to other mainstream short video recommendation methods in accuracy,recall rate,MRR average reciprocal ranking and other evaluation indicators.This shows that the recommendation method proposed in this paper has better accuracy and execution efficiency.