Predicting Users’ Personality Based on Their ‘Liked’ Images on Instagram Alixe Lay Bruce Ferwerda University College London Jönköping University London Jönköping United Kingdom Sweden alixe.lay.17@ucl.ac.uk bruce.ferwerda@ju.se ABSTRACT With the development in technology and the increasing ubiquity implicitly, via observation of behavioural patterns [9]. Previous of social media services, it has created new opportunities to study process and reveals a person’s preferences to different personality from the digital traces individuals leave behind. The entertainment domains such as TV, music and books [7-8]. They large number of user-generated images on social media has can be assessed explicitly via psychometric questionnaire [6], or prompted renewed interests in understanding the psychological implicitly, via observation of behavioural patterns [9]. Previous factors driving production and consumption behaviours of visual studies have demonstrated that personality can be inferred from content. Instagram is currently the fastest growing photo-sharing individuals’ behaviours and user-generated content in digital social media platform, with more than 400 million active users environments, such as Facebook Likes [4], Twitter profiles [10], and nearly 100 million photos shared on the platform daily [1], the contents of personal websites [11], language used on and generates 1.2 billion likes each day [2]. The understanding of Facebook [12] and Twitter [13]. Although explicit questionnaires the appeal of visual content at an individual level is highly yield higher accuracy than methodologies inferring personality relevant to psychometric assessment, social media marketing and from user-generated content, they require more effort on interface personalisation. In this position paper, we address the participants’ part to complete [14]. Further, automatic personality need to explore the avenue of automatic personality assessment assessment from social media traces has the potential to allow using ‘liked’ images on Instagram. more efficient inquiry into personality at an unprecedented scale. KEYWORDS This position paper will outline a proposal for a study aimed at Personality, automatic personality recognition, Instagram, image predicting users’ personality based on their liked images on features Instagram. As automatic personality assessment within the domain of social media images is scarce, this research will be necessary to further our understanding of this field. The following 1 INTRODUCTION section will outline related work predicting personality using social media images, as well as the reason for choosing Instagram Recent advancements in technology and the increase in ubiquity as the SNS of interest. of digital services observing and recording human activities have opened up new opportunities for research into human behaviours [3]. With statistics showing that people are spending more of their time on the Internet on, or through, social networking services 2 RELATED WORK (SNS) [4], it is apparent that SNSs are data-rich avenues to study 2.1 Personality and Social Media Images personality and human behaviours. This allows researchers to base their predictions of individuals’ personality on digital records On SNSs, we are exposed to various images and videos on a daily of human behaviour [5]. basis. With the recent rise of photo-sharing SNS platforms (e.g., Instagram, Pinterest), photo-posting and sharing activities on Personality traits are the descriptions of people in terms of SNSs have increased vastly in popularity, making them a their relatively stable patterns of behaviour, thoughts and distinctive and fast-emerging phenomenon in digital environments emotions [6]. Personality influences human decision making [15]. Photographic data on social media are often connected with process and reveals a person’s preferences to different well-defined agents: the producers who create them, and the entertainment domains such as TV, music and books [7-8]. They consumers who consume them [16]. So, the creation of an image can be assessed explicitly via psychometric questionnaire [6], or is traceable from the first authorised post and the consumption of the same image can be inferred from various activities, one of ___________________________________________ them being to ‘like’ it [16]. © 2018. Copyright for the individual papers remains with the authors. Copying permitted for private and academic purposes. More recently, researchers have started to explore the links HUMANIZE '18, March 11, Tokyo, Japan. between personality and posted images on social media. Prior HUMANIZE’18, March 2018, Tokyo, Japan A. Lay & B. Ferwerda studies have been conducted to predict users’ personality using depressed also demonstrated a stronger preference to filter out all their Facebook profile images [17-18]. [18] analysed four families colours from their photos, and an aversion to artificially lightening of visual features and found interpretable patterns associated with photos, relative to their non-depressed counterparts. Importantly, the personality traits of the individuals who posted these images. these depressive signals are detectable in images posted on For instance, extraverted and agreeable individuals were found to Instagram even before the date of first diagnosis. Moreover, the have pictures with warm colours and many faces in their portraits, prediction model was more accurate than general practitioners at reflective of their tendency to socialise; whereas the images of correctly diagnosing depression, indicating that major those high on Neuroticism tended to be set indoor. When the psychological changes within individuals are transmitted in social performance of the classification approach was compared to the media use, and can be identified using computational methods. one obtained by human raters, this study showed that the former produced more accurate classifications than the latter for Extraversion and Neuroticism. Echoing previous psychological 2.3 Liked Images research [19-20], this study has demonstrated that Facebook As individuals engage in more ‘liking’ behaviours than posting profile pictures carry relevant information for classifying the behaviours [1-2], this warrants an investigation into personality personality traits of the individuals who post them. detection using ‘liked’ images. However, despite the contrast in both activities and the ubiquity of the ‘like’ or virtual endorsement function on various social media platforms [26], there has been 2.2 Instagram substantially less research attempting to infer personality from Instagram is an online, mobile phone photo-sharing, video-sharing virtually endorsed images on social media. [27] found that the and social network service (SNS) that enables its users to take features of images tagged as favourite on Flickr could be used to pictures and videos, and then share them on its own platform as predict both self-assessed and attributed personality traits. well as other social media platforms [21]. By recently outpacing However, they found covariation was high in attributed traits, but Twitter, YouTube, LinkedIn and Facebook in growth [22], not in self-reported traits. The authors explained that it is possible Instagram is currently the fastest growing social network site that when assessing their own traits, the participants used globally, with more than 400 million active users, nearly 100 information such as their personal history and life experiences million photos shared on the platform daily [1], and generates 1.2 [28], which is different or absent from their favourited pictures billion likes each day [2]. Despite the rapid rise of Instagram as [27]. It would be interesting to investigate whether the same one of the most popular social media platforms, there is limited findings will emerge on a different social media platform, academic research on this SNS compared to others, such as Instagram. Facebook and Twitter. As mentioned, posting an image is akin to production, while There have been two studies that have predicted personality liking an image can be considered as consumption of social media from posted images on Instagram. [23] found distinct features content, and hence both activities are fundamentally different in within Instagram photos (e.g., hues, brightness, saturation) that terms of purpose [16, 29]. [29] posits that the different uses are are related to personality traits, indicating that users with different driven by different motivations: people produce their own content personalities make their pictures look different. For instance, for self-expression and self-actualisation; consume the content for Openness to Experience was positively associated with the colour information and entertainment; and participate (by directly or green, low brightness, high saturation, cold colours and few faces; indirectly engaging with the content) for social interaction and individuals high on Conscientiousness tended to post images with community development. Previous psychological research has saturated and unsaturated colours; agreeable individuals were found that personality differences in posting behaviours [19-20], more likely to post images with few dark and bright areas; as well as virtual endorsement behaviours [30]. Within the Neuroticism was related to images with high brightness; personality computing literature, [31] looked at posted and Extraversion was linked with images of green and blue tones, low preferred images on Twitter and found that image posting and brightness, saturated and unsaturated colours [23]. In line with liking preferences using interpretable aesthetic and semantic previous research showing consistent links between Openness to features were associated with differences in personality. Further, Experience and aesthetic preferences [24], [23] also found combining the information from both posted and liked images Openness to Experience to be the trait with the most strongly leads to significant performance gain compared to individual significant correlations with image characteristics, followed by interactions, indicating that both posting and liking images allow Agreeableness and Conscientiousness. for more complete understanding of users’ personality. However, there has been no attempt to date comparing the predictive Another study investigated markers of depression within accuracy of posted and liked images on individuals’ personality Instagram photos posted by users [25]. The findings of the study on Instagram. With the growth of Instagram overtaking all the show that photos posted to Instagram by depressed individuals other SNSs [22], and the low generalisability of findings across were more likely to contain the colours blue and gray, to appear different social media platforms [26], it is necessary to study the darker, and to receive fewer likes. Instagram users who were links between personality traits and the images that users like and 2 Predicting Users’ Personality Based on Their ‘Liked’ Images on HUMANIZE’18, March 2018, Tokyo, Japan Instagram post on Instagram for the findings to be useful for designing We will devise a model to predict personality traits using effective advertising or personalisation strategies which are based features of participants’ posted images, and compare this model to specifically on Instagram activities. the one obtained from the analyses for RQ1 to determine whether the posted or liked images are more predictive of users’ personality traits. 3 RESEARCH PROPOSAL There has been no study to date which has attempted to predict RQ4 How do the images a user posts or likes differ? users’ personality based on their ‘liked’ images on Instagram. The correlations between personality traits and posted images will Further, it would also be interesting to look at the predictive be compared against those obtained from previous analyses of accuracy of posted and liked images on Instagram on users’ liked images in RQ2. personality, as they are considered as qualitatively different activities [16, 29]. With studies showing predominantly better RQ5 Is the computer-based or human-based personality accuracy of personality prediction using online behaviours [27; assessment using liked images of an individual more accurate? 40], we are also interested in comparing the accuracy of human- based and computer-based personality assessments using liked Human raters will be selected and asked to judge the liker’s Big images. In this position paper, we propose a research project Five personality traits on a 5-point Likert scale based on a random which aims to predict users’ personality based on the images that sample of 20 liked images from the collected data. They will first they ‘like’ on Instagram. be presented with the descriptions of each Big Five traits, and then asked to rate the images accordingly. As this is a labour intensive To assess participants’ self-reported personality, we choose to task, only 20 images will be used for each human rater. The focus on the Five-Factor Model (FFM), or “Big Five” as it is the accuracy results of human raters will then be compared to the most widely-accepted trait framework in the history of personality computer-based predictive model obtained in earlier analyses. psychology [32]. The FFM describes personality in terms of Extraversion, Agreeableness, Conscientiousness, Neuroticism and Openness to Experience [33]. In terms of the image features 4 IMPLICATIONS which will be used to predict personality, we will incorporate not Automatic personality assessment from liked images have only the standard colour- and content-based features, but also important implications for the field of personality and differential visual sentiment-based features. As suggested by [34], the psychology, as they can be used to measure psychological traits in standard methods used in studies of photographic data focus on a cheap, convenient and reliable manner. As this study will only identifying faces and features in the images, and is incapable to examine the prediction of personality from images, it may be a actually recognise the intention of the uploader, hence the social worthwhile avenue for future studies to explore the use of other value of the image. Hence, to bridge this affective gap, we will behavioural parameters within Instagram to assess personality, use visual sentiment-based features to form part of the features in such as written captions, comments, follower and following lists, the prediction model. A series of studies will be run, which will and profile descriptions. answer five research questions using the outlined approaches. The results of this study may contribute to the body of work RQ1 Can we predict users’ big five personality from the concerning personality-based personalisation [35]. For instance, characteristics of their ‘liked’ images? personality-based recommendation systems have been found to We will use computational methods to extract colour-based, increase users’ loyalty towards a system and lower their cognitive content-based and visual sentiment-based features from the effort in a more effective way, compared to systems without collected photographic data via an Instagram API. We will then personality information [36]. The adoption of personality devise a prediction model which can predict self-reported information into recommender systems may also have the personality scores from the characteristics of the images potential to lessen the cold start problem [37]. participants have ‘liked’. Further, the findings may also be of high relevance to social media marketing, particularly on Instagram. According to [38], RQ2 Which, if any, of the image features are indicative of liker’s personality? marketers are increasing their budget for social media marketing every year. With more brands competing for audience’s attention Correlational analyses will be used to identify the image features on social media, there is pressing need for more effective which are significantly correlated with liker’s personality traits. microtargeting strategies to increase the persuasive appeal to RQ3 Do the images a user posted or liked yield a higher marketing content. Importantly, a recent study found that when predictive accuracy over their personality? the content of persuasive appeals was matched to individuals’ psychological characteristics inferred from their Facebook Likes, 3 HUMANIZE’18, March 2018, Tokyo, Japan A. Lay & B. Ferwerda it resulted in up to 40% increase in clicks and up to 50% more [15] Kim, E., Lee, J. A., Sung, Y., & Choi, S. M. 2016. Predicting selfie-posting behavior on social networking sites: An extension of theory of planned purchases than when the content were mismatched or behavior. Computers in Human Behavior, 62, 116-123. unpersonalised [39]. It is possible that by understanding the links [16] Segalin, C. 2015. Social Signal Processing for Computational Aesthetics. [17] Celli, F., Bruni, E., & Lepri, B. (2014, November). Automatic personality and between characteristics of images individuals are consuming on interaction style recognition from facebook profile pictures. In Proceedings of Instagram and their personality, we may be able to further the 22nd ACM international conference on Multimedia (pp. 1101-1104). ACM. finetune the content of marketing content to increase its ability to [18] Segalin, C., Celli, F., Polonio, L., Kosinski, M., Stillwell, D., Sebe, N., ... & Lepri, B. 2017. What your Facebook Profile Picture Reveals about your persuade. Personality. Proceedings of the 25st ACM international conference on Multimedia. [19] McCain, J. L., Borg, Z. G., Rothenberg, A. H., Churillo, K. M., Weiler, P., & As discussed, automatic personality assessment may open up Campbell, W. K. 2016. Personality and selfies: Narcissism and the Dark Triad. new avenues for developing or elevating products and services. At Computers in Human Behavior, 64, 126-133. the same time, ethical challenges and privacy concerns may also [20] Qiu, L., Lu, J., Yang, S., Qu, W., & Zhu, T. 2015. What does your selfie say about you?. Computers in Human Behavior, 52, 443-449. arise from the capacity to identify individuals’ private [21] Frommer, D. 2010. Here's how to use Instagram. Business Insider, 11. psychological traits from their liked images. As the amount of [22] Chaffey, D. 2016. Global social media research summary 2016. Smart Insights, digital traces people leave behind grows in abundance, it becomes 8. increasingly difficult for individuals to control which of their [23] Ferwerda, B., Schedl, M., & Tkalcic, M. 2016. Using Instagram picture features to predict users’ personality. In International Conference on intimate attributes are being uncovered [4]. With exponential Multimedia Modeling (pp. 850-861). Springer, Cham. accumulation of digital behavioural records, continuous increase [24] McManus, I. C., & Furnham, A. 2006. Aesthetic activities and aesthetic attitudes: Influences of education, background and personality on interest and in pervasiveness and robustness of personality predictions, it is involvement in the arts. British Journal of Psychology, 97(4), 555-587. imperative that policymakers implement regulations on the uses as [25] Reece, A. G., & Danforth, C. M. 2017. Instagram photos reveal predictive markers of depression. EPJ Data Science, 6(1), 15. well as potential abuses of this kind of technology, in order to [26] Hayes, R. A., Carr, C. T., & Wohn, D. Y. 2016. One click, many meanings: ensure that the public is safeguarded from any potential harm that Interpreting paralinguistic digital affordances in social media. Journal of may incur. Broadcasting & Electronic Media, 60(1), 171-187. [27] Segalin, C., Perina, A., Cristani, M., & Vinciarelli, A. 2017. The pictures we like are our image: continuous mapping of favorite pictures into self-assessed and attributed personality traits. IEEE Transactions on Affective Computing, 8(2), 268-285. REFERENCES [28] Wright, A. G. 2014. Current directions in personality science and the potential [1] Instagram. 2016. Instagram statistics. https://instagram.com/press/ for advances through computing. IEEE Transactions on Affective Computing, [2] Sciberras, E. 2015. Social media statistics 2014. The latest overview of the 5(3), 292-296. social media world. http://socialmediabuzz.com/socialmedia-statistics-2014- [29] Shao, G. 2009. Understanding the appeal of user-generated media: a uses and latest-overview-social-media-world/. gratification perspective. Internet Research, 19(1), 7-25. [3] Kosinski, M. 2014. Measurement and prediction of individual and group [30] Lee, S. Y., Hansen, S. S., & Lee, J. K. 2016. What makes us click “like” on differences in the digital environment. Department of Psychology University of Facebook? Examining psychological, technological, and motivational factors Cambridge. on virtual endorsement. Computer Communications, 73, 332-341. [4] Alexa.com. 2017. Top Sites. https://www.alexa.com/topsites [31] Guntuku, S. C., Lin, W., Carpenter, J., Ng, W. K., Ungar, L. H., & Preoţiuc- [5] Kosinski, M., Stillwell, D., & Graepel, T. 2013. Private traits and attributes are Pietro, D. 2017. Studying personality through the content of posted and liked predictable from digital records of human behavior. Proceedings of the images on Twitter. In Proceedings of the 2017 ACM on Web Science National Academy of Sciences, 110(15), 5802-5805. Conference (pp. 223-227). ACM. [6] McCrae, R. R., & Costa, P. T. 2003. Personality in adulthood: A five-factor [32] Funder, D. C. 2001. Accuracy in personality judgment: Research and theory theory perspective. Guilford Press. concerning an obvious question. In B. W. Roberts & R. Hogan (Eds.), Decade [7] Cantador, I., Fernández-Tobías, I., & Bellogín, A. 2013. Relating personality of behavior. Personality psychology in the workplace (pp. 121-140). types with user preferences in multiple entertainment domains. In CEUR [33] Costa, P. T., & McCrae, R. R. 1992. Four ways five factors are basic. Workshop Proceedings. Shlomo Berkovsky. Personality and individual differences, 13(6), 653-665. [8] Rentfrow, P. J., & Gosling, S. D. 2003. The do re mi's of everyday life: the [34] Bechmann, A. 2017. Keeping it Real: From Faces and Features to Social structure and personality correlates of music preferences. Journal of personality Values in Deep Learning Algorithms on Social Media Images. In Proceedings and social psychology, 84(6), 1236. of the 50th Hawaii International Conference on System Sciences. [9] Fast, L. A., & Funder, D. C. 2008. Personality as manifest in word use: [35] Nunes, M.A.S. and Hu, R. 2012. Personality-based recommender systems: an correlations with self-report, acquaintance report, and behavior. Journal of overview. In Proceedings of the sixth ACM conference on Recommender personality and social psychology, 94(2), 334. systems (pp. 5-6). ACM. [10] Quercia, D., Kosinski, M., Stillwell, D., & Crowcroft, J. 2011. Our twitter [36] Hu, R., & Pu, P. 2011. Enhancing collaborative filtering systems with profiles, our selves: Predicting personality with twitter. In Privacy, Security, personality information. In Proceedings of the fifth ACM conference on Risk and Trust (PASSAT) and 2011 IEEE Third Inernational Conference on Recommender systems (pp. 197-204). ACM. Social Computing (SocialCom), 2011 IEEE Third International Conference on [37] Tkalcic, M., Kunaver, M., Košir, A., & Tasic, J. 2011. Addressing the new user (pp. 180-185). IEEE. problem with a personality based user similarity measure. In First International [11] Marcus, B., Machilek, F., & Schütz, A. 2006. Personality in cyberspace: Workshop on Decision Making and Recommendation Acceptance Issues in personal Web sites as media for personality expressions and impressions. Recommender Systems (DEMRA 2011) (p. 106). Journal of personality and social psychology, 90(6), 1014. [38] 4C. 2017. The State of Social Advertising. http://www.4cinsights.com/wp- [12] Farnadi, G., Zoghbi, S., Moens, M. F., & De Cock, M. 2013. Recognising content/uploads/2017/04/4C_TheStateOfSocialAdvertising_2017Q1.pdf personality traits using Facebook status updates. In Proceedings of the [39] Matz, S. C., Kosinski, M., Nave, G., & Stillwell, D. J. 2017. Psychological workshop on computational personality recognition (WCPR13) at the 7th targeting as an effective approach to digital mass persuasion. Proceedings of international AAAI conference on weblogs and social media (ICWSM13). the National Academy of Sciences, 201710966. AAAI. [40] Youyou, W., Kosinski, M., & Stillwell, D. 2015. Computer-based personality [13] Sumner, C., Byers, A., Boochever, R., & Park, G. J. 2012. Predicting dark triad judgments are more accurate than those made by humans. Proceedings of the personality traits from twitter usage and a linguistic analysis of tweets. In National Academy of Sciences, 112(4), 1036-1040. Machine learning and applications (icmla), 2012 11th international conference on (Vol. 2, pp. 386-393). IEEE. [14] Cantador, I., & Fernández-Tobías, I. 2014. On the exploitation of user personality in recommender systems. In CEUR Workshop Proceedings. Mouzhi Ge. 4