Breaking the Validation Trade-off in Topic Extraction: A Bi-Objective Metaheuristic Model for Labelling Short-Text Clusters and an Application on AirBnB Tokyo Reviews

Abstract

In service systems, online marketing is the leading B2C channel. Hence, service system reviews on online sites are beneficial feedback for service providers that can contribute to improved customer satisfaction. However, it is not straightforward to observe the most frequent topics among the customer reviews for human cognition. Topic modeling is developed to unearth the topics in a large-scale textual dataset by machine capability. In that study, a clustering & cluster labeling pipeline is used as a topic modeling approach. Firstly, customer reviews are retrieved from AirBnB for Tokyo, then reviews are clustered. Subsequently, each cluster is labeled by using a multiobjective metaheuristic (Non-dominated Sorting Genetic Algorithm - NSGA-II) on coherence and divergence objectives. Cluster labels in pseudo-weighted Pareto optimal solution are proposed as topics, to present the most discussed aspects of the service for both service providers and potential customers to interpret the past user experience along with the sentiment values for each topic.