Excerpt
Quantified Personality – Automatic Personality Analysis from Online and Mobile Usage Data
Steffen Schumacher
11/2017
Algorithms are instructions for the stepwise execution of a method. Social and cultural scientists but tend to broaden the meaning of this notion and use it as an umbrella notion for digital automatization in general. But computer programs contain non-algorithmic command syntax, also. Furthermore, algorithms may develop and change during implementation and use which makes talking about “the” algorithm being always identical with itself often difficult or impossible. A comprehension of the notion of algorithm too distant from that of computer science hinders the comprehensibility of social and cultural scientific analyses by computer scientists. On the other hand, these sciences shouldn’t confine their usage of this notion to that of the latter to be still able to deal with the phenomenon from a different perspective. (cf. Dourish 2016)
Automatic personality analysis doesn’t use data gathered by questionnaires administered to respondents, any more, but uses usage data which are generated by default and in different contexts, respectively. This is the big novelty of this field of investigation which led to the two articles published by Kosinski and Stillwell in the „Proceedings of the National Academy of Sciences of the United States of America“ in 2013 and 2015 being the most influential articles ever published in the “Proceedings” according to their Altmetric Score. These two articles dealt with the analysis of the personality of Facebook users using their Facebook likes.
Since automated personality analysis is based on Big Data - for the development of its methods as well as their application - it is subject to the three paradoxes of Big Data described by Richards and King (2013): 1. The transparency paradox which follows from collecting more and more data to be able to make the world more transparent, but collecting them in more and more invisible and opaque ways. The data are saved on servers in unknown places and are analysed by methods and algorithms which are not easy to scrutinize. This holds in general, i.e. also for data and methods of automatic personality analysis.
2. The identity paradox: voluminous online usage data are analysed to reveal the behaviour and preferences – and in the future also the personality – of users, but the results are applied to influence and modify these very behaviours and preferences. Like in the case of filter bubbles generated by automated news feeds which send users only those informations which already conform to their preferences. The aim of automated personality analysis, too, is to be able to choose content and layout of advertising and political messages this way that they appeal to the users’ personality. Also, commercial websites and services and the behaviour of virtual agents should be able to be tailored to the individual users alike. Is there a personality bubble afoot which exceeds the filter bubble and renders the digital environment of the users accomodated to their personality to the maximum?
3. The third Big Data paradox which is linked to the first two is the power paradox which consists in the fact that, on the one hand side, Big Data should help everybody to gain informations about the world and to communicate with other users, but, on the other hand side, the very usage traces of the users doing this are stored and can be used by big organizations as well as agencies which have the right to access them to gain informations about the users and control them. This risk is particularly high in the case of automated personality analysis, since it yields intimate knowledge about the individuals’ personality and, thus, can uncover their weak points and ways to influence them.
Automatic personality analysis is a very recent field of investigation. The first summary article of it was published in 2014 (Vinciarelli 2014). According to it, the number of articles which have the notion “personality” in their title and have been published in journals of the two biggest international organizations of computer science and computer scientists, ACM and IEEE[1], jumped up in 2010. All articles found by me and dealt with in the following sections have been published in 2011 and afterwards, almost all of them in journals of the two professional organizations. By far most of the authors have a degree in computer science, followed by psychology and psychometrics.
In all articles, the model of personality used is the Big Five personality model, a psychological personality model which assumes the personality of every human being consisting of five dimensions which can be developed stronger or weaker. These five dimensions are called personality traits. They are openness, conscientiousness, extraversion, agreeableness and neuroticism. [2] That it is a psychological personality model means that values and attitudes which belong to the personality of individuals, too, aren’t investigated. [3] The Big Five model is the most widespread and accepted one in current psychology and possesses significant relations to external variables like job satisfaction, occupational position and achievement as well as choice of spouse and friends (Golbeck et al. 2011a).
In a study of employees of the IBM Almaden Research Center (Gou et al. 2014), two further models of personality have been used for analysis: the „basic human values“ of Shalom H. Schwartz – self-transcendence, conservation, self-enhancement, openness to change and hedonism – as well as a model which is developed by the authors themselves and based on Maslow’s hierarchy of needs and on marketing literature. It is called „fundamental needs“ and includes ideals, harmony, closeness, self-expression, excitement and curiosity.
It has to be said that the Big Five model has been critized in particular in regard to its intra-individual stability over time (Neuman 2016: 13), i.e. in regard to its test-retest-reliability.
The first kind of data which has been used to extract the personality of users are smartphone usage data or mobile phone usage data in general (mobile usage data).
There are a number of articles which describe studies with mobile phone users. The latter fill in a personality questionnaire, and either their smartphone usage data are recorded by an app – call logs, SMS logs, app logs, Bluetooth scans of the environment and profile logs, i.e. logs of the adjustments of their smartphone profile, as well as heart rate, GPS position and derived walking speed, speed-up, duration of sleep, light intensity, pressure[4] - or mobile usage data are used which are stored by their mobile provider, i.e. metadata. Between personality and usage data, a mathematical connection is searched for by methods like regression and correlational analysis. This is the general procedure for the development of algorithms and programs for personality analysis: they are always based on formulae generated by connecting personality data from questionnaires to other kind of data which are later used to extract the personality of their authoring users whithout them having filled in a personality questionnaire.
In a study from 2011 (Chittaranjan et al. 2011), a large number of usage log data serve as independent variables which have been stored from 83 test persons during a period of eight months by a specialised software: use of the Office software, internet, video, audio, maps, Youtube, calendar functions, camera, chat functionality, SMS and games, number of outgoing calls, duration of outgoing calls (in average and in total), number of incoming calls, duration of incoming calls (in average and in total), number of emails sent and received, number of SMS sent and received, word length in emails sent and received (in average and the median value), number of contacts, number of telephone numbers called and having called, the same for SMS contacts etc. Big Five traits were used. A regression analysis with all these variables as independent variables reached an increase in accuracy of personality prediction of 17 to 40 percent for the five personality traits compared with prediction by chance (25 percent in average, extraversion and neuroticism were predicted best). This means a correct classification of the Big Five personality of 69 to 76 percent of the individuals. The interpretation of the relations found isn’t easy, partly they are directly plausibel, partly not. E.g. it’s plausibel that extraverted persons got more calls and these calls lasted longer. But there didn’t exist the same relation for outgoing calls. Another example: the authors explain their finding that one and the same Bluetooth ids were found more often in the environment of more emotionally stable (i.e. less neurotic) as well as more agreeable persons by speculating about persons possessing these personality traits staying a longer time in one and the same place than other persons.
In two studies from 2011 and 2013 (Oliveira et al. 2011, Montjoye et al. 2013), the data used are exclusively from call detail records (CDR) of mobile providers, i.e. from metadata of calls, SMS and MMS of mobile phone users. These are e.g. the number of calls, SMS and MMS, outgoing as well as incoming, the duration of calls, the interval between receipt of text messages and reply to them and between telephone calls with one and the same contact (phone number) and the ratio between the number of calls and text messages and the number of contacts. So, it’s not necessary to record the smartphone usage behaviour as described above, any more. These two studies used the Big Five model, too. The mean increase in accuracy of prediction of the Big Five traits against prediction by chance is higher than in the study described above: 39 percent (of variance) and 42 percent (concerning the classification of individuals into the three categories low, middle, high per trait), respectively. While in Oliveira et al. 2011 extraversion and neuroticism could be predicted best, like in Chittaranjan et al. 2011, it were openness and conscientiousness in the study of Montjoye et al. from 2013. After the additional inclusion of properties of the personal network of the individuals like number of contacts, network density and number of frequent contacts, openness was predicted best in the former study and neuroticism worst which pattern can be often found in studies described below using data from social networks.
I haven’t found studies using smartphone usage data after the year 2013. In an article from 2016 (Guo 2016), considerations are at least made about which smartphone usage data could be appropriate for the analysis of the users’ personality, but they aren’t tested empirically: physical data like heart rate and GPS position as well as users’ preferences being identified through their browsing histories.
The second kind of data used in studies of automatic personality analysis are Facebook likes.
These data have been successfully used and made famous by Michael Kosinski, a psychologist and psychological programmer, and David Stillwell, a psychologist and psychometrician, as already mentioned above. One reason for their success was their way of formulating the degrees of accuracy of prediction reached: they equated them with the accuracy of judgement of acquaintances, friends and spouses. Both authors were employed at the centre of psychometrics of the University of Cambridge in 2013. In 2015, Kosinski had changed to Stanford University. They used the Big Five personality model, too.
In the first of their two pertinent articles (Kosinski et al. 2013), they predict not only the Big Five traits of individuals, but also personal properties like sexual orientation, sex, age, race and political affiliation from Facebook likes, the second article deals only with the Big Five traits (Youyou et al. 2015). In both articles, they use data from tens of thousands of people which have filled in a Big Five personality questionnaire in a Facebook app called myPersonality (which had been developed by Stillwell) and, at the same time, have agreed to the usage of their personality as well as Facebook profile data for scientific purposes. In total, 7.5 million Facebook user have filled in at least one of the personality questionnaires from the app which included 20 to 100 questions (Schwartz et al. 2013: 6) during the period between 2007 and 2012 when the app was online. Other psychometric tests could be made in it, too. Their data are used in the article from 2015 for external validation of the personality measures. To motivate participants to make tests from the app they could view their results afterwards. Facebook likes had been publicly accessible by API from 2009 to April 30th 2015 (Anonymous 2016), so, they respresented a publicly available stock of data during these years.
In their article from 2015, the authors work with data of 70.000 users which all had filled in the 100 item Big Five questionnaire. More than 17.000 of them had been rated by a Facebook friend in regard to their personality by means of a 10 item Big Five questionnaire. And 14.000 of them had been rated by a second Facebook friend, so that the agreement between both judgements could be measured, in these cases. By means of a regression analysis, the relation between Facebook likes and Big Five traits were computed and, after this, it was tested how good the personality traits of those participants could be predicted which hadn’t been used to compute the formula, but had been spared for testing. These were a tenth of the sample. The accuracy of the prediction of a person’s personality increased plausibly with the number of Facebook likes included. The authors obtained thresholds of accuracy of prediction from a meta-analysis of published study results, namely the accuracy of colleagues, cohabitants, friends and partners in assessing one’s personality by a questionnaire. This accuracy was expressed in Pearson’s correlation coefficients. The result of their regression analysis was that with ten Facebook likes included in the analysis one’s personality – cast as the mean of the Big Five traits – could be assessed better than by a colleague, with 70 Facebook likes better than by a friend or cohabitant, with 150 likes better than by a family member and with 300 likes better than by a spouse. [5] Furthermore, the congruence between judgements of different judges were higher in the case of algorithms than of human beings. In the former case, different judges were two independently computed algorithms each based on one randomly chosen half of the likes used. Finally, personality assessment by algorithms also correlated better with human behaviour and other aspects having been extracted from the participants’ Facebook profiles and from the other psychometric tests of the app: study subject, size of the personal network on Facebook, substance use, health, activities on Facebook, attitudes like party preference, general values and life satisfaction as well as depression. The authors conclude that Facebook likes were a better predictor of human personality than the social cognitive abilities of human beings. The latter would be influenced by factors external to the task in question. Computers could, thus, assess humans’ personality better than other humans. On the other hand side, the authors give in that personality was limited to the Big Five traits, here, and that there could possibly be other, e.g. more subtle, personality traits which weren’t measured by the Big Five questionnaire and could be only recognized by fellow human beings, so far.
So, they could formulate the results of their study of Facebook likes in a very handy way: to assess someone better than a friend or cohabitant when useing 70 likes, better than a family member when using 150 likes and even better than a spouse when using 300 likes. Naturally, this shows the data drivenness of their approach which is adopted by many Big Data projects: research doesn’t follow from a theory nor aims at it, but investigates a stock of data on the basis of rather basic assumptions or simply because of their availability. This research doesn’t in practice deal so much with users’ personality, but with attributes like creditworthiness, purchasing power, likelihood of the cancellation of credits and health forecast (Christl 2014).
For the practical purpose of data driven research, it suffices if the formula found works. There isn’t a need for explanation why it works. In my opinion, this implies that it can’t be predicted if the independent variables’ power of prediction changes over time, e.g. decreases, if their value or their overall quality largely change. This means that the formulae used for prediction must be monitored and updated regularly since it can’t be known with sincerity that the relations found still exist and haven’t changed, respectively.
One general assumption about the reason for the ability of Facebook likes to predict users’ personality is made by Youyou, Kosinski and Stillwell (2015):
“Why are Likes diagnostic of personality? Exploring the Likes most predictive of a given trait shows that they represent activities, attitudes, and preferences highly aligned with the Big Five theory. For example, participants with high openness to experience tend to like Salvador Dalí, meditation, or TED talks; participants with high extraversion tend to like partying, Snookie (reality show star), or dancing.” (Youyou et al. 2015: 2)
It’s plausible that likes have something to do with users’ attitudes, preferences and activities, because they are the result of an evaluation by them. Though, a theory proper is missing, here. And they don’t build one from the results of their study, also.
I remains to be said that their results relate to the mean value of the five personality traits as already mentioned above. This traits are but predicted with differing accuracy: openness best, conscientiousness and neuroticism worst. Reasons for these differences aren’t given by the authors. So, they stick to this very unspecific explanation.
An additional question is if their method is still useful today since Facebook likes aren’t freely accessible any more.
Digression: The role of automatic personality analysis based on Facebook likes in the US election campaigns from 2016
After Trump’s election victory, the US enterprise Cambridge Analytica claimed to have made this victory possible by the psychographic analysis of the electorate which had taken place also on the basis of Facebook likes. By this, Trump’s campaign staff had been enabled to address voters in a more effective way (Confessore et al. 2017). This enterprise seems to have recruted some employees from the University of Cambridge and also chosen his name in connection with it. Of course, such a name also carries the good reputation of this institution. According to press reports, Michael Kosinski blames the enterprise to have stolen the method of personality analysis from Facebook likes, facilitated perhaps by the contact to members of the Centre for Psychometrics of the University for which he and Stillwell worked and work, respectively (Grassegger et al. 2016, Hartlmaier et al. 2017: 35f.). Meanwhile, Cambridge Analytica had to abandon his grandiose claims. It had worked only on a part of Trump’s campaign and had done only conventional statistical, not psychographic, voter analysis in it (Beuth 2017). But it’s a matter of fact that it is making an effort to buy as much data allowing a psychographic analysis about as many US citizens as possible. E.g. to get the permission for the usage of Facebook likes which weren’t publicly accessible at the time of the US election campaign, any more, it used a quiz game app and some other smaller Big Five test apps which request access to the individuals’ Facebook likes if they want to do the quiz and tests (Grassegger et al. 2016).
[...]
[1] The Association for Computing Machinery (ACM) is the biggest international organization of computer science and the Institute for Electrical and Electronics Engineers (IEEE) is the biggest international professional organization of engineers from the area of electrical engineering and communications technology.
[2] The model is often called OCEAN model according to the acronym built from the notions for the five dimensions in this order.
[3] But they are investigated by the techniques of sentiment analysis and opinion mining aiming at individuals’ positive and negative attitudes to objects (Liu/Zhang 2013). These attitudes can be rather short-time and ephemeral, e.g. concerning products and services, which one wouldn’t conceive of as part of the individuals’ personality.
[4] Imaginable would be also die usage of data which could be transmitted to the smartphone by wearable devices like smart watches and fitness wristbands.
[5] Openness was the Big Five trait which could be predicted best. From the other four traits which were close together in this respect neuroticism was predicted worst.