| description abstract | Customer segmentation divides customers into groups with different characteristics and supports the design of customized products and tailored marketing strategies. Recent studies explore using online reviews as the data source and social network analysis as the fundamental technique for customer segmentation. These studies usually utilize the frequency of mentioned product attributes and/or customers' sentiments from online reviews in the segmentation process. However, few of them investigate the influence of different types of information (e.g., with or without sentiment, order information) on the segmentation performance. In addition, previous studies seldom consider and tackle the challenge of clustering high-dimensional data when online reviews contain customers' rich opinions towards multi-faceted attributes of a product. To fill these gaps, we propose a comprehensive framework for customer segmentation and need analysis based on sentiment network of online reviewers and graph embedding. The frequently mentioned product attributes and customers' sentiments are first extracted from online reviews. Then, a customer can be represented as a vector consisting of his/her sentiment polarities on each product attribute as well as rating and order information. After that, a social network of customers is established by examining the similarity of customer vectors. The network nodes are embedded into low-dimensional vectors, which can be further clustered into different groups, i.e., customer segments, and their respective needs can be analyzed by methods such as Importance–Performance Analysis. Our framework enables the construction and performance comparison of various types of networks, node compositions, and embedding methods. A case study employing the online reviews of a passenger vehicle in China's market is used to demonstrate the validity of the proposed framework. The results indicate that the customer segmentation generated by the sentiment network of online reviewers with Graph Autoencoder (GAE) embeddings performs better than other alternative models that do not utilize vector embeddings, fail to consider the sentiment information, or leverage bipartite network structures. Our framework provides more nuanced insights for designers to improve customers' satisfaction and increase the market competitiveness of their products. | |
