An Automatic Modeling Method of Kansei Evaluation from Product Data Using a CNN Model Expressing the Relationship Between Impressions and Physical Features

Abstract

In the field of Kansei engineering, the approach is often taken of Kansei evaluation modeling expressing the relationships between physical features and impression of an object. However, in the conventional modeling method, personnel and time costs are very high because multiple experiments and analyses are needed to high precision modeling. In contrast, study using machine learning has been conducted as a method of modeling the relationship between physical features and impressions of products. However, no studies have been reported that considering how the nature of an impression that there is evaluation vary from person to person. In this study, we work on automatically Kansei evaluation modeling using images and review-text data of products existing on the web. A convolutional neural network (CNN) is used for modeling, and variation in the impressions of each product are taken into consideration when learning. In the proposed method, we performed the following: (1) Extraction of the main impressions of target domain and calculation of values that express the strength of each impression from review-text data through text mining based on the previous study [1], (2) creation of a product image data set that uses the distribution of products’ impression scores as a training label and (3) construction of the CNN model using the created data set. We applied proposed method to wristwatches as the target domain and verified the estimation accuracy of constructed CNN model. As a result, a high positive correlation was confirmed between estimated impression score and impression scores that were calculated from review-text data. In addition, since present results exceeded the estimation accuracy of CNN model hasn’t learned distribution of impression scores, learning variations in the evaluation of peoples’ impressions were shown to be effective for improving estimation accuracy.