基于注意力与神经网络的视频流行度预测模型

针对传统预测算法预测精度低及难以处理多变量的时序数据等缺点,文章提出一种采用双向长短期记忆(bi-directional long short-term memory, BiLSTM)网络和时间模式注意力(temporal pattern attention, TPA)机制相结合的视频流行度预测模型。双向长短期记忆网络从视频流行度时间序列的正向和反向提取时间特征,时间模式注意力机制从双向长短期记忆网络输出状态的深层特征提取时间模式,有利于视频流行度预测。真实视频数据的实验结果表明,与经典时序预测方法相比,TPA-BiLSTM模型能够有效地降低预测的误差,提高预测的准确性。

In view of the shortcomings of traditional prediction algorithms such as low prediction accuracy and difficulty in processing multivariate time series data, a video popularity prediction model using a combination of bi-directional long short-term memory (BiLSTM) and temporal pattern attention (TPA) is proposed. BiLSTM extracts temporal features from the forward and reverse of the video popularity time series, while TPA extracts temporal patterns from the deeper features of the output states of BiLSTM, which is beneficial for video popularity prediction. Experimental results on real video data show that the proposed TPA-BiLSTM combined model effectively reduces the prediction error and improves the prediction accuracy compared with classical time series prediction methods.