Graph Clustering Based Size Varying Rules for Lifelong Topic Modeling

Abstract

Lifelong learning topic models identify the hidden concepts discussed in the collection of documents. The concepts are represented as topics having groups of ordered words based on their relevance to the topic. Lifelong learning models have an automatic learning mechanism which allows continuous learning without external support. In the process, the model gets more knowledgeable with experience as it learns from the past in the form of rules. It is carries rules to the future and utilize them when a similar scenario arises. The existing lifelong learning topic models heavily rely on statistical measures to learn rules that leads to two limitations. The rules are evaluated for fixed number of words while ignoring the natural arrangement of words within the documents. Moreover, the rules have arbitrary orientation that causes repeated patterns of transferring the impact of a rule into a topic during the early iterations of the inference technique. In this research work, we introduce complex networks analysis for learning rules which addresses both of the limitations discussed. The rules are obtained through hierarchical clustering of the complex network that have different number of words within a rule and have directed orientation. The proposed approach improves the utilization of rules for improved quality of topics at higher performance with unidirectional rules on the standard lifelong learning dataset.