基于网络表示学习的短视频流行度预测研究
Study of Short Video Popularity Prediction Based on Network Representation Learning
朱恒民 1徐凝 2魏静 2沈超2
作者信息
- 1. 南京邮电大学管理学院,南京 210003;江苏高校哲学社会科学重点研究基地—信息产业融合创新与应急管理研究中心,南京 210003
- 2. 南京邮电大学管理学院,南京 210003
- 折叠
摘要
预测短视频的流行度不仅有助于短视频平台高效地管理信息,还可以对社会舆情进行监控.针对已有工作仅考虑短视频多模态内容特征构建流行度预测模型这一现实情况,本文基于网络表示学习,提出融合短视频内容和网络结构特征的流行度预测模型.首先,基于爬取的抖音数据构建包含短视频、发布者和评论者节点,以及发布和评论关系的异质信息网络,将其映射为短视频和发布者两个同质网络,选择node2vec算法表征网络结构,作为网络模态;其次,提取短视频的多模态内容特征,采用低秩多视图子空间学习方法融合短视频内容和结构特征,作为流行度预测模型的输入;最后,构建短视频流行度预测的多层感知机回归模型,并进行对比和消融实验.结果表明,融合网络结构能够降低短视频流行度预测的误差,各模态对短视频流行度预测的影响程度依次为文本、网络、社交、音频和视觉模态.本文融合了短视频内容和网络结构特征,为基于特征工程的短视频流行度预测提供了新思路.
Abstract
Predicting the popularity of short videos not only helps short-video platforms with efficient information man-agement but also plays an important role in monitoring public opinion.Unlike existing studies that focus only on multi-modal content features of short videos,to construct a popularity prediction model,we propose a popularity prediction mod-el based on network representation learning,fusing content and network structural features.First,based on the dataset crawled in Douyin,a heterogeneous information network consisting of nodes was constructed,including short videos,pub-lishers,commenters,and edges.After mapping into two different homogeneous networks,namely,short-video and publish-er networks,node2vec was selected to represent the network structure in the embedding space as a network modality.Sec-ond,the multimodal content features of short videos were extracted and fused using low-rank multiview embedding learn-ing.Finally,a multilayer perceptron machine regression model was proposed for short-video popularity prediction.Com-parisons and ablation experiments were further conducted.The results show that fusing network structure features can re-duce the error of short-video popularity prediction.The degree of influence of the various modalities on short-video popu-larity prediction consisted of the textual,network,social,acoustic,and visual modalities,in decreasing order.Our method,which combines short-video content and network structure features,provides new ideas for short-video popularity predic-tion based on feature engineering.
关键词
短视频/网络表示学习/流行度预测/多层感知机Key words
short video/network representation learning/popularity prediction/multilayer perceptron引用本文复制引用
基金项目
国家自然科学基金项目(72374111)
江苏省研究生科研与实践创新计划项目(KYCX23_0933)
出版年
2024
NETL
NSTL
万方数据