UNIFIED UNSUPERVISED DEEP LEARNING MODEL FOR PREDICTION OF TIME AND TOPIC-SPECIFIC INFLUENCERS FROM SOCIAL MEDIA
In Online Social Networks (OSNs), an efficient Influential User Prediction (IUP) is essential for different applications like sentiment analysis, online recommendation, etc. Among several prediction models, a Grey Wolf optimization with Graph Convolutional Neural Network (GW-GCNN) model can predict influencers by learning the latent vector representation of each netizen, including different centrality measures. But it models the netizen influence based on a fixed-size sub-network from the netizen’s social action log and topic distributions, as well as was highly reliant on the different topics. It cannot learn an entire huge corpus since merely a limited part of the information was annotated by Ground Truth (GT). Hence, this article proposes a unified unsupervised GW-GCNN with the Long Short-Term Memory (LSTM) model without GT supervision. It aims to design the netizen’s influence dynamics and discover the Influence Propagation (IPN) on multiple topics. The major contributions of this model are (i) measuring the time-aware and topic-related influences, (ii) modeling the IPN related to interval and topics using the Influence Attention-GCNN (IA-GCNN) that learns the netizen’s latent vector representation under multiple topics and (iii) extracting temporal influence and learning the Influence Scores (ISs) by using a matrix-adaptive LSTM that considers the unsupervised objective. Moreover, the learned ISs of every topic are summed and max-pooled over a period to get every netizen's IS for predicting influencers. At last, the extensive experiments reveal that the GW-GCNN-LSTM achieves 93.9%, 92.5% and 92.4% accuracy for Facebook, Weibo and Twitter datasets, respectively during training, whereas it attains 94.1%, 93.5% and 94% accuracy for Facebook, Weibo and Twitter datasets, respectively during testing compared to the KSGC, Multi-view Influence (MvInf), InfACom-GCN and GW-GCNN algorithms.
Online social networking, User influence, GW-GCNN, Temporal influence, LSTM, unsupervised learning