Sangwan Neeti, Bhatnagar Vishal
GGS Indraprastha University and Maharaja Surajmal Insitute of Technology, New Delhi, India.
NSUT East Campus (Formerly Ambedkar Institute of Advanced Communication Technologies and Research), New Delhi, India.
Sci Rep. 2025 Jan 20;15(1):2508. doi: 10.1038/s41598-025-86785-3.
As digital media grows, there is an increasing demand for engaging content that can captivate audiences. Along with that, the monetary conversion of those engaging videos is also increased. This leads to the way for more content-driven videos, which can generate revenue. YouTube is the most popular platform which shared the revenue from advertisement to video publisher. This paper focuses on the work of video popularity prediction of the YouTube data. The idea of mapping the video features into low-dimensional space to get the latent features is presented. This mapping is achieved by a novel multi-branch child-parent Long Short Term Memory (LSTM) network. These latent features train the fused Convolutional Neural Network (CNN) with LSTM to predict the popularity of unseen videos on the trained deep learning network. We compared our results against Linear Regression (LR), Support Vector Regression (SVR) and Fully Convolutional Networks (FCN) with LSTM. A significant improvement with a 50% reduction in MAE and a 0.61% increase in the coefficient of determination (R²) has been observed by the proposed Multi branch LSTM encoded features with a fused deep learning predictor (MLEF-DL predictor) when compared to existing methods.
随着数字媒体的发展，对能够吸引观众的引人入胜的内容的需求日益增加。与此同时，这些引人入胜的视频的货币转化率也在提高。这为更多能够产生收入的内容驱动型视频开辟了道路。YouTube是最受欢迎的平台，它将广告收入分享给视频发布者。本文重点研究YouTube数据的视频热度预测工作。提出了将视频特征映射到低维空间以获得潜在特征的想法。这种映射是通过一种新颖的多分支父子长短期记忆(LSTM)网络实现的。这些潜在特征训练融合了卷积神经网络(CNN)和LSTM的模型，以预测经过训练的深度学习网络上未见过的视频的热度。我们将我们的结果与线性回归(LR)、支持向量回归(SVR)以及带有LSTM的全卷积网络(FCN)进行了比较。与现有方法相比，所提出的具有融合深度学习预测器的多分支LSTM编码特征(MLEF-DL预测器)在平均绝对误差(MAE)上显著降低了50%，在决定系数(R²)上提高了0.61%。
