Active Learning Method for Pairwise Comparison Data

• Published in: 2023 IEEE 5th International Conference on Cybernetics, Cognition and Machine Learning Applications (ICCCMLA) pp. 310 - 315
• Main Authors: Yamagiwa, Ayako, Goto, Masayuki
• Format: Conference Proceeding
• Language: English
• Published: IEEE 07.10.2023
• Subjects: Active Learning; Companies; Data models; DNN; Learning systems; Machine learning; Pairwise Comparison; Product Image; Product Position Map; Semantics; Uncertainty
• DOI: 10.1109/ICCCMLA58983.2023.10346880

Companies need to create a product position map and correctly grasp the current situation. In the recent increase in the number of e-commerce sites, the influence of product images on customers' willingness to purchase is huge. In other words, if we can construct a product map based on customers' impressions of product images, we can expect to make the product lineup on e-commerce sites more efficient and attractive. Semantic Differential (SD) method and Multi-dimensional scaling (MDS) have been conventionally used to create product maps. However, these methods are unsuitable for analyzing a target with a large number of products, such as EC sites, because the number of product evaluation data required for the analysis increases depending on the number of products. Therefore, the authors proposed a method to construct a product image map of product images with high accuracy from a relatively small number of pairwise comparison data. Specifically, a model that estimates the pairwise comparison data by subjects is machine-learned. The learned model is used to supplement the missing pairwise comparison data to estimate the evaluation values of product images for a specific axis. The values are used to construct a map. The method of selecting a small number of data used to train the model can quickly identify the overall trend. Therefore, this study proposes an adequate data selection criterion to improve the accuracy of the map based on the concept of active learning. Experiments show that, in addition to uncertainty considerations, it is adequate to have a uniform amount of pairwise comparison data for each subject in the problem of complementing pairwise comparison data between subjects, as in the case of this study.