02-10: Sunghoon Kim
Chair: Paul Richard Messinger
Identifying Segment-Specific Key Drivers from Unstructured Online Service Review Data: Analytics Methodology and Application in Service
Customers are talking about, sharing, and complaining about their service experiences on online review sites. As such, large amounts of review data about service experiences (e.g., numerical star ratings, textual reviews in natural language) are publicly available, and these data are being updated every moment. These reviews include the direct but heterogeneous experiences of customers about services and/or products. As challenges for applying various statistical methods (e.g., segmentation analysis) to the review data, the data includes unstructured text reviews evaluating multiple service attributes and so we need to transform such qualitative, unstructured texts to quantified, structured format. In addition, the transformed data tend to be high-dimensional, correlated and sparse. Thus, there are needs for new analytics methodology to resolve such challenges.
We propose an integrated machine-learning algorithm to extensively apply a classic model-based segmentation method in marketing (e.g., latent class regression) to the unstructured online service review data. The proposed procedure extracts a quantified independent variables (IVs) matrix from unstructured textual reviews by developing a set of text analytics algorithms and then identifies segment-level key drivers by applying a suggested segmentation method with variable selection. Given the textual review data in service are typically highly dimensional, correlated, and sparse, we resolve this issue using simultaneous variable selection technique in the segmentation method. With the proposed method, firms or policy makers can focus on key drivers per each segment in their marketing activities (e.g., online banner advertising, search advertising) to improve resource allocation efficiency; this method will help them systematically keep track of periodic patterns of segment-level key drivers, which can be especially useful for service marketing.
We apply the proposed algorithm to two different kinds of unstructured review data: (1) reviews for rating 46,340 university professors, and (2) reviews for rating 10,876 restaurants in Arizona and Nevada states in the U.S. Through the two application studies, we show that the extracted IVs matrices are valid with several evidences. Next, we demonstrate that the proposed algorithm provides better predictive performance with out-of-sample data compared to two existing benchmark procedures (e.g., ordinal regression with a IVs matrix yielded by the deep-learning neural network approach, and collaborative filtering recommendation approach), which might help improve recommendation system. Finally, we discuss the key drivers across derived segments: E.g., “lecture quality,” “class materials,” “care by professors,” etc. across the three segments for the study 1, and “food quality,” “service quality,” “worth (money for value),” etc. across the four derived segments for the study 2.