Using Twitter Data to Infer Personal Values of Japanese Consumers

Using Twitter Data to Infer Personal Values of Japanese Consumers Yinjun Hu Synergy Marketing, Inc. Dojima-Avanza 21F, 1-6-20 Dojima, Kita-ku, Osaka 530-0003, Japan. ko.inshun@syngery101.jp Yasuo Tanida Synergy Marketing, Inc. Dojima-Avanza 21F, 1-6-20 Dojima, Kita-ku, Osaka 530-0003, Japan. tanida.yasuo@synergy101.jp Abstract Our purpose is to use Twitter data to infer per- sonal values in marketing for Japanese con- sumers. In this paper, we reintroduce our per- sonal value system and apply the model for in- ferring personal values with tweets. To adapt the model to the rapid change of wording in tweets, we propose a dynamic model based on time-weighted frequency in this research. We evaluated the prediction results from our previous approach, newly proposed approach (the dynamic model), and other methods with 10-fold cross-validation. Our experiment re- sults show that personal values can be inferred from Twitter data, and our approach based on Bayesian network performs well with skewed training data. 1 Introduction In marketing science, personal values have been considered the central determinants of consumer be- havior. They are widely used for market segmenta- tion and behavioral prediction. VALS (Values And Lifestyles) is a personal value system for market segmentation,1 and (Wu, 2005) reconstructed it for Chinese consumers as China-Vals. Similarly, we developed a value system for Japanese consumers based on the AIO (Activities, Interests, and Opin- ions) (Plummer, 1974) rating statement during last two years. We gathered about 20,000 Japanese con- sumers’ personal values data via questionnaires, and filled the value system database with this data. 1VALS is developed by Mitchell, Arnold (May 1984) in the book Nine American Lifestyles: Who We Are and Where We’re Going. Although we can “talk to” consumers directly with a questionnaire approach, subjective biases may appear in answers, such as a response- bias (Peer and Gamliel, 2011) and a choice- supportive bias (Mather et al., 2000). On the other hand, a data mining approach like microblog anal- ysis, which is considered to have a sampling bias problem, as checked by (Mislove et al., 2011), can get objective “answers” for the “questions” without subjective biases. Hence, a microblog mining ap- proach, like mining tweets, may be a complement to the questionnaire approach. Moreover, (Chen et al., 2014) pointed out that word use may be influenced by values, and made an effort to analyze the associ- ations between personal values and their word uses in social media. In our research, we use Twitter for a data mining approach to infer personal values, be- cause we consider that users on Twitter tend to tweet their real intentions with limited impersonation. In our previous work, we proposed a model for extracting personal values from tweets with text mining technologies. This model demonstrated that we were able to predict consumer behavior with tweets and our value system. As a result, we applied it to marketing consulting and social contribution for trial-and-error. However, the wording on Twit- ter changes frequently, and the keywords used for inference in this model may be out of date. Hence, in this research, we …

Content truncated to 500 words