Heterogeneous Information Fusion based Topic Detection from Social Media Data

Due to the pervasive nature of social networking platforms, as well as the proliferation of user generated content, the internet has become a repository of unstructured multimedia data. The use of this huge data for user experience enhancement is still a problem, where topic detection is one of the solutions to solve this issue, not having been explored in the literature for this application. Videos with similar content or related to the same topic can be grouped together with the help of topic detection methods. In this paper, a framework for topic detection using web videos textual metadata has been developed. The key contribution in this paper is to leverage multimedia metadata to find web video topics using a two-step process . First, we used transformer-based model to perform topic modeling for identification of topics from the heterogeneous textual data of web videos. Second, topic-based video retrieval has been accomplished using a classification approach. Further, experiments are carried out on a publicly available dataset to assess the performance of the proposed method. The proposed work is compared to the state-of-the-art methods Discriminative Probabilistic Models (DPM), Event clustering based method (ECBM),Multi-Modality Based Method (MMBM), Side-Information Based Method (SIBM), and Similarity Cascades(SC), which shows that the proposed system outperforms others in terms of Precision, Recall, F-measure and Accuracy. The experimental results demonstrates the effectiveness of proposed method for topic detection.