Improving Hot Topic Detection in Microblog by Metaheuristic Based Clustering and Hybrid Neural Network Classifier

Abstract

The most popular social media platform to exchange the ideas, opinions, daily activities in wide range is the twitter. Each short message in the twitter is called as tweet. Conventional and current work is around latest hot topic detection while hot topic detection using the past tweets is missing. However, processing the older posts helps the organization to analyze the most successful products in the past. Those successful products can be re-released by adding advanced features to generate more revenue and helps the health officers to take preventive measures or medications by processing the posts to detect the health problem suffered by majority of people in past. This work concentrates in identification of the most popular thing in past. This work comprises of three approaches. The Density Peak Clustering_Adaptive Cut-off distance (DPC_ACD) enhances density peak clustering (DPC) specifically using an adaptive cut-off distance (ACD) for identifying hot topics in historical tweets. This modification helps to cluster the tweets. The second approach, further increased the clustering results by combining Density Peak Clustering with Remora Optimization Algorithm (DPC_ROA) for deciding the optimal cut-off distance. This technique is more data-driven and minimize the need for human judgment. Optimization and automation in deciding the optimal cut-off distance is also possible with DPC_ROA. The aforementioned proposed methods do not handle the context of words in the tweet. Hence to automate the feature extraction and to understand the context of words in the tweet, the combination of Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) and modified bald eagle search (MBES) is proposed in the contribution-III. Bald Eagle Search (BES) excels in refining search spaces and adjusting it based on the best positions found during optimization. Moreover, BES is a gradient-based algorithm. The gradient-based algorithms use the gradient information to direct the search towards the best solution, resulting in faste