Abstract: Concerning the issue of traditional prediction model, this paper proposes a Deep Belief Networks（DBNs）based approach to predict the popularity of online videos. By modeling users’ attention with the information extracted from social network and search engine, it studies its feature selection method and the optimization of DBN’s parameters. Based on the data collected from one large-scale online video service provider, it implements experiments to evaluate the proposed approaches. The results show that DBN-based prediction obtains the highest performance up to 79.47%（for domestic TV drama）and 65. 33%（for foreign TV drama）. The prediction strategy can help provide decision-making information for risk assessment, publicity affair, and investment of online videos.

Key words: deep learning, online video service, popularity prediction, deep belief network, restricted Boltzmann machine

摘要： 针对在线视频热度预测研究中分类及预测效果欠佳，规则化较多和较缺乏实践检验等问题，通过对实际在线视频服务系统所采集的海量数据研究，提出一种基于深度信念网络（Deep Belief Networks，DBNs）的视频热度预测方法。首先，结合社交网络的关注度和视频关键词的搜索热度，对影响因子进行了建模和量化处理；其次，根据输入和输出变量确定了DBNs各层网络的结构，优化了网络参数和预测模型；最后，通过在线视频服务商的数据对深度信念网络进行训练，并多次交叉实验对比分析，结果表明基于DBNs方法在视频热度预测上准确率最高79.47%（国内视频）、65.33%（国外视频），可以为在线视频上映前的投资、宣传以及风险评估提供较全面可靠的参考决策。

关键词: 深度学习, 在线视频服务, 热度预测, 深度信念网络, 受限玻尔兹曼机