[1M3-GS-13] AI application: Social data and prediction

[1M3-GS-13-02] User preference prediction of packaging designs using deep learning

〇Bohui Xia¹, Hideki Sakamoto², Xueting Wang¹, Toshihiko Yamasaki¹ (1. The University of Tokyo, 2. PLUG Inc.)

Keywords:Deep learning, Package design

Packaging design of merchandise has significant effects on consumers' behaviors. For example, some people might buy and try a new product because they think its packaging looks good. On the contrary, others might dislike it because of the design. Determining packaging design needs a lot of work, but if we could predict its users' preference in advance, it would reduce the cost. In this work, we propose a deep learning based method with ensemble learning to predict user preference of packaging design. We have collected results of surveys asking people to evaluate several packages and we use them as dataset. As a result, we can achieve 0.652 correlation coefficient between the ground truth user preference scores and the predicted values, which is quite high considering that such evaluation is highly subjective. Besides, in order to visualize which part of design affects preference, we show feature maps in a similar way to the class activation map (CAM) and conduct a qualitative analysis.