Personality-Based Recommendations: Evidence from Amazon.com Panagiotis Adamopoulos Vilma Todri padamopo@stern.nyu.edu vtodri@stern.nyu.edu Department of Information, Operations, and Management Sciences Leonard N. Stern School of Business, New York University ABSTRACT between personality and preferences in multiple entertain- In this paper, we evaluate the accuracy of personality-based ment domains using explicit psychometric tests. There are recommendations using a real-world data set from Ama- several characteristics though that differentiate this study zon.com. We automatically infer the personality traits, needs, from the related work. For instance, apart from the Big and values of users based on unstructured user-generated Five model [9, 20] that the aforementioned studies employ, content in social media, rather than administering question- we also use the personality models of needs [12, 18] and val- naires or explicitly asking the users to self-report their char- ues [22]. Besides, rather than administering questionnaires acteristics. We find that personality characteristics signifi- or explicitly asking the users to self-report their characteris- cantly increase the performance of recommender systems, in tics as in previous studies in RSes, we automatically infer the general, while different personality models exhibit statisti- personality characteristics, needs, and values of users based cally significant differences in predictive performance. on unstructured user-generated content in social media. Categories and Subject Descriptors 3. PERSONALITY MODELS H.3.3 [Information Storage and Retrieval]: Information The personality traits [9, 20], needs [12, 18], and values Search and Retrieval - Information filtering [22] of the users in this study are automatically inferred Keywords based on a textual analysis of user-generated unstructured Recommendations; Personality traits; Big Five; Values; Needs data. In particular, for each user we analyzed the content of all the messages that there were publicly posted over time 1. INTRODUCTION on the social network of Twitter as well as the user-defined description of their accounts. From the messages of the users Personality traits have been found to influence various analyzed are excluded all the private messages between the aspects of individual behavior, including job performance users as well as non-English messages. In addition, we ex- [5], academic motivation [17], and romantic relationships cluded any messages that were not written by the specific [25]. Despite the initial promising evidence in various aca- target user each time (e.g., re-tweets) as those messages do demic fields and applications, including recommender sys- not correspond to the linguistic style of the specific user tems (RSes), personality traits are still not frequently used and hence might not reflect her/his personality. After the in predictive modeling, mainly because they usually require pre-processing of the corpus of user-generated content, there users to complete long questionnaires and hence they cannot were on average 26, 568 words per user; this number is much be easily applied at a large scale. In this study, we automat- higher than the typical number of words in other studies ically infer cognitive and social characteristics of users based (e.g., [13]) and can lead to more accurate results. The mes- on different personality models in psychology, including Big sages and the rest of the user-generated of each target user Five, Values, and Needs, and present a comparative analysis. are merged into a single “document” and the personality traits, intrinsic needs, and values of individuals are then de- 2. RELATED WORK rived using linguistic analytics. In particular, the tokens Tapping into the recent advances of data mining, various of the user-generated content -after some pre-processing of studies have successfully attempted to automatically derive the words, which includes removal of stop-words and non- personality traits from text based on the established rela- English words, stemming, and fuzzy matching- are matched tionship between word use and personality [11, 14, 27]. Ex- with the Linguistic Inquiry and Word Count (LIWC) psy- ploring the feasibility of deriving personality traits from so- cholinguistic dictionary, which has been developed over sev- cial media text, [19] demonstrated that computational mod- eral years and currently includes almost 4,500 words and els based on derived personality traits perform better than word-stems associated with one or more personality cate- models using self-reported traits. In addition, [7] found that gories [21], to compute relative scores in each dictionary predicted personality traits had the same effects as the traits category. Afterwards, based on [27], a weighted combina- measured by traditional personality questionnaires. tion is estimated based on the coefficient between category In RSes, the use of personality traits is a promising but scores and characteristics, using coefficients that were de- under-explored research direction. Among the most relevant rived by comparing personality scores obtained from sur- works, [10, 15] explicitly measure users’ personality based on veys with LIWC category scores from text [23, 27]. Simi- quizzes aiming at alleviating the cold-start problem. Using larly, user values are derived based on the same approach also questionnaires, [13] finds correlations between personal- [8] whereas for automatically inferring user needs a statisti- ity and movie preferences, while [6] studies the relationship cal model was employed based on ground-truth scores and a Copyright is held by the author(s). custom dictionary [26]; a publicly available implementation RecSys 2015 Poster Proceedings, September 16–20, 2015, Austria, Vienna. of the employed approach is available by [16]. . data from Amazon.com, we find that personality character- istics can increase the performance of RSes and we identify a specific model of personality that significantly outperforms the remaining models achieving promising performance. The main advantage of the employed approach is that automated methods for personality assessment are more ef- ficient and objective [11]. In particular, the traditional way of measuring personality, which requires people to complete long questionnaires, does not allow to obtain personality traits at a large scale for the population of interest [7]. Be- sides, user-generated content is more reflective of users’ ac- tual personalities, not “idealized” versions of themselves [4]. 6. REFERENCES [1] Adamopoulos, P. ConcertTweets: A Multi-Dimensional Data Set for Recommender Systems Research. [2] Adamopoulos, P. Beyond rating prediction accuracy: On new Figure 1: Predictive performance of different personality models. perspectives in recommender systems. In RecSys ’13 (2013). [3] Adamopoulos, P., and Tuzhilin, A. Estimating the value of 4. EXPERIMENTAL RESULTS multi-dimensional data sets in context-based recommender To empirically evaluate the employed approach, we build systems. In RecSys ’14 (2014). a factorization model incorporating the information of per- [4] Back, M., Stopfer, J., et al. Facebook profiles reflect actual personality, not self-idealization. Psychol. science (2010). sonality traits, needs, and values as well as item attributes. [5] Barrick, M., and Mount, M. The big five personality In particular, the user preferences are modeled as: dimensions and job performance: A meta-analysis. Pers. Psychol. (1991). y(x) = y(u; i; α1u , . . . , αm u ; α1i , . . . , αn i ) [6] Cantador, I., Fernández-Tobı́as, I., and Bellogı́n, A. Relating m n personality types with user preferences in multiple entertainment domains. In EMPIRE workshop (2013). X X = w0 + wu + wi + wj αju + wl αli + hvu , vi i [7] Chen, J., et al. Making use of derived personality: The case of j=1 l=1 social media ad targeting. In ICWSM (2015), AAAI. m X n X m X X n [8] Chen, J., Hsieh, G., et al. Understanding individuals’ personal values from social media word use. In CSCW (2014), ACM. + αju hvju , vi i + αli hvu , vli i + αju αli hvju , vli i, [9] Costa, P. T., and MacCrae, R. R. Revised NEO personality j=1 l=1 j=1 l=1 inventory (NEO PI-R) and NEO five-factor inventory (NEO FFI) manual. Psychological Assessment Resources, 1992. where the input vector x ∈ R|U |+|I|+m+n contains binary [10] Elahi, M., Braunhofer, M., et al. Personality-based active indicators for the user and item, the user attributes αu = learning for collaborative filtering recommender systems. In AI*IA 2013: Advances in Artificial Intelligence. Springer. (α1u , . . . , αm u ) capturing the personality characteristics of the [11] Fast, L., and Funder, D. Personality as manifest in word use: users, and item attributes αi = (α1i , . . . , αmi ) capturing the correlations with self-report, acquaintance report, and item categories, prices, etc.; the factorization of users vu , behavior. J. personality social psychology94, 2 (2008), 334. items vi , and attributes vju , vli is of dimensionality k. [12] Ford, K. Brands laid bare: Using market research for evidence-based brand management. John Wiley & Sons, 2005. Our data set was collected as in [1, 24] and contains [13] Golbeck, J., Robles, C., Edmondson, M., et al. Predicting 906, 277 purchases of 138, 536 distinct products on Ama- personality from twitter. In IEEE PASSAT Conference (2011). zon.com from 81, 475 users who shared their purchases on [14] Hirsh, J., and Peterson, J. Personality and language use in Twitter as well as the account information and the user- self-narratives. J. research personality43, 3 (2009), 524–527. [15] Hu, R., and Pu, P. Enhancing collaborative filtering systems generated content on the social network of Twitter for the with personality information. In RecSys (2011), ACM. same users. As our data set includes only implicit ratings, [16] International Business Machines Corporation. for each user we randomly select an equal number of non- https://github.com/watson-developer-cloud/ rated items (based on the frequency of ratings of each item) personality-insights-python, 2015. [17] Komarraju, M., and Karau, S. J. The relationship between the as negative examples in order to increase the accuracy of big five personality traits and academic motivation. Pers. our predictions. We use MCMC inference with Gibbs sam- individual differences39, 3 (2005), 557–567. pling to learn our factorization model. Moreover, we employ [18] Kotler, P., and Armstrong, G. Principles of Marketing 15th a holdout evaluation scheme with 80/20 random splits into Global Edition. Pearson, 2013. [19] Mairesse, F., and Walker, M. Words mark the nerds: training and test sets without filtering any ratings and we Computational models of personality recognition through evaluate each model in term of classification performance language. In CogSci (2006). based on accuracy. [20] Norman, W. Toward an adequate taxonomy of personality Figure 1 shows the experimental results. We see that per- attributes: Replicated factor structure in peer nomination personality ratings. The J. Abnorm. Soc. Psychol.66, 6 (1963). sonality characteristics increase the performance of RSes and [21] Pennebaker, J., Chung, C., Ireland, M., et al. The that different personality models can result in different pre- development and psychometric properties of LIWC2007, 2007. dictive accuracy. Interestingly, the under-explored personal- [22] Schwartz, S. H. Basic human values: Theory, measurement, ity models of needs [12, 18] and values [22] resulted in better and applications. Revue française de sociologie47, 4 (2006). [23] Tausczik, Y., and Pennebaker, J. The psychological meaning of predictive performance compared to the more popular model words: Liwc and computerized text analysis methods. J. of Big Five traits [9, 20]. We also see that combining the language social psychology29, 1 (2010), 24–54. attributes of the different personality models results in even [24] Todri, V., and Adamopoulos, P. Social commerce: An better performance and, hence, has the potential to further empirical examination of the antecedents and consequences of commerce in social network platforms. In ICIS (2014). increase the business value of recommendations [2, 3]. [25] Tupes, E., and Christal, R. Recurrent personality factors based on trait ratings. J. personality60, 2 (1992), 225–251. 5. CONCLUSIONS [26] Yang, H., and Li, Y. Identifying user needs from social media. In this study, we automatically infer the personality traits, Tech. rep., IBM Tech Report. goo. gl/2XB7NY, 2013. needs, and values of users based on unstructured user-generated [27] Yarkoni, T. Personality in 100,000 words: A large-scale analysis of personality and word use among bloggers. J. content in social media and build different RS models. Using research personality44, 3 (2010), 363–373.