Analisis de Componentes Principales
1. Relación entre represión e integración o rechazo del actor no reconocido Sobre el recorte tempo-espacial efectuado en nuestro
The most important advantage of passive psychological assessment using machine learning is convenience for both participants and administrators. It allows a test to be pushed to millions of people quickly, and it requires no time commitment from the participants. Researchers who want to record an individual’s personality for a study can only request access to their social networking data, rather than asking the individual to complete long questionnaires, and in practice, an employer can easily request access to social networking data of the job seekers and infer their personality instead of requiring them to participate in self-report tests, which itself is prone to self-enhancement bias.
6.5 Summary
In this chapter, I used the methodology introduced in chapters 4 and 5 to construct a proof- of-concept machine learning system capable of accurate personality predictions. The pre- dicted personality scores were assessed in terms of correlations with self-reports, internal consistency, external validity and discriminant validity and were found to be very similar to self-reports in terms of their psychometric properties.
6.5 Summary 87
While this chapter mostly focused on investigating psychometric properties of the outcomes of the machine learning system, chapter 7 investigates why the predictions are the way they are. This includes analysis of predictability of various traits, predictability item-level data, and how personality scores relate to predictability of users on Facebook.
Chapter 7
Analysis of Predictability
In chapters 4 and 5, I introduced a dimensionality reduction method and predictive learning models that are capable of accurate prediction of personality of users from social networking data. In chapter 6, I investigated the psychometric properties of the outcomes of the ma- chine learning system when used to predict five-factor personality in terms of reliability and validity.
Five factor personality is very effective in predicting behaviour and important life outcomes [106, 121], such as happiness, spirituality, physical health, peer, family and romantic relation- ships, occupational choice and performance, political and ideological values, and criminality. This is partly why this research started, to examine whether it is possible to go from footprints of behaviour (online in the work of this thesis) to accurate assessment of personality.
Being able to predict personality scores presents a unique opportunity to investigate if the five-factor personality traits can predict a new outcome that has not been studied before, which is online predictability.
In section 7.1, I examine the level of predictability for each of the personality traits and com- pare them to the literature. In this section I also investigate why neuroticism is the most difficult trait to predict, as observed in chapter 6, and draw a new study which further inves- tigates a potential answer. In section 7.2, I perform new studies to investigate the per-item perfectibility for each of the 100-items of MyPersonality’s personality test, from the IPIP in- ventory and discuss the results. In section 7.3, I investigate what role demographics and per- sonality play in explaining how predictable a specific user is and perform relevant studies. Finally, section 7.4 summarises the findings.
90 Analysis of Predictability
7.1 Predictability of the five-factor model
When it comes to predicting behaviour and important life outcomes, different traits are bet- ter or worse at predicting different things. Conscientiousness is best able to predict aca- demic [104] and work performance [12]. My own analysis in section 6.3.3 showed that both ML-rated and self-report neuroticism are best at predicting depression and life satisfaction. Overall, conscientiousness is the most predictable trait from behaviour. This brings about the expectation that the two traits should be easiest to predict here, as well. However, this was not observed in chapter 6. Table 7.1 shows the correlations between the ML-rated and self-report personality traits, as discussed in chapter 6.
Trait Self-ML agreement
Openness .77
Conscientiousness .68
Extraversion .71
Agreeableness .72
Neuroticism .58
Table 7.1 Correlation of ML-rated and self-report five-factor personality model
Openness is the easiest trait to predict and neuroticism is the most difficult trait to predict. This is not a method-dependent limitation, as this pattern exists in all predictions made in chapter 5 with different model decision. Table 7.2 shows the correlation between predicted and self-report personality traits in literature.
Study O C E A N
Kosinski et al. [86] .43 .29 .40 .30 .30
Youyou et al. [152] .65 .52 .55 .56 .49
Park et al. [109] .41 .26 .36 .41 .39
Golbeck et al. [50] .65 .59 .55 .48 .53
Table 7.2 Correlation of ML-rated and self-report five-factor personality models in literature. O, Openness; C, Conscientiousness; E, Extraversion; A, Agreeableness; N, Neuroticism;
7.1 Predictability of the five-factor model 91
Across all studies, openness is the easiest trait to predict, which is inline with the results of this thesis as well. As explained in earlier chapters, this is most likely the result of the way Facebook is used as a medium for accessing news, politics and an online forum to access and discuss topics of arts. Kosinski et al. [86] reported conscientiousness as the most difficult trait to predict, with agreeableness and neuroticism being only slightly easier. Kosinski’s study is also based on MyPersonality and used a lot of the same users as this thesis, but with a narrower inclusion criteria, as described in chapter 3. Similarly, another study which used MyPersonality is Youyou et al. [152]. It also reports neuroticism as the most difficult trait to predict.
Park et al. [109] uses MyPersonality, but instead of Facebook Likes, they use the user’s status updates to train their models. They found conscientiousness to be the most difficult trait to predict. Golbeck’s study also primarily uses language features on social media and it finds agreeableness to be most difficult to predict.
There is no other literature focusing on the predictability of personality traits from Facebook data. However, having access to personality scores and predicted personality, I can investigate if personality traits themselves correlate with predictability of users. This is what I investigate in the next study in section 7.1.1.