Predictive Modelling for Sensitive Social Media Contents Using Entropy-FlowSort and Artificial Neural Networks Initialized by Large Language Models

Abstract

This work offers an integrated methodological framework that integrates the capabilities of large language models (LLMs), rules-based reasoning, multi-criteria sorting, and artificial neural networks (ANN) in developing a predictive model for classifying the intensity of sensitive social media contents. The current literature lacks a holistic consideration of multiple attributes in evaluating social media contents, and the proposed framework intends to bridge such a gap. Three actions constitute the development of the framework. First, LLMs (i.e., GPT4) evaluate the social media contents under a predefined set of attributes, leveraging the power of LLMs in content analytics. Second, rules-based reasoning and multi-criteria sorting (i.e., entropy-FlowSort) determine the categories of social media contents. Lastly, the two previous actions produced a complete dataset that can be used to train a predictive model using ANN to classify sensitive social media contents. With 1100 randomly extracted social media contents and the predefined categories of violations against community standards set by Facebook, the proposed integrated methodology produces an ANN-based classification model with 86.36% prediction accuracy. Comparative analysis using Decision Trees, k-nearest neighbors, Linear Discriminant Analysis, Random Forest, and Naive Bayes classification yields the highest performance of ANN. The predictive model can be used as a decision-support tool to design moderation actions on social media contents.