When you browse the web, look for information or shop, you are surrounded by various recommendations. When viewing a product in an online store, you get recommendations for other products. When entering keywords into a search engine, you get a list of recommendations for websites that the engine assumes will best suit your needs. We are surrounded by recommender systems.
Although recommendations often seem intrusive and inappropriate, they are based on complex algorithms developed by numerous researchers.
The aims of recommender systems are often presented as “evil” (Stephen Baker in his book Numerati hints at conspiracy), but it is basically mainly an attempt to simplify the use of web technologies by interpreting the user’s wishes based on their previous behaviour.
Based on our experience, most of us suspect that recommender systems are far from perfect. The recommendations are often bad, sometimes even bizarre. This usually happens when the system does not have enough information about the user (the cold-start problem). Researchers are trying to better understand the user’s needs and are looking for the key parameters that say the most about the user. Recently, there has been an increase in research activity that, to improve the recommendations, uses the user’s personality and emotions.
The user’s personality is a characteristic that changes very little over time. Some are more extrovert than others. Some are more neurotic than others. Some are more open to diversity, others less.
A popular description model is the Big Five Model, which categorises the basic dimensions for describing someone’s personality as extraversion, agreeableness, conscientiousness, neuroticism and openness. Research has shown that people with different personalities have different desires [source], for example when it comes to music or film genres. Consequently, understanding the user’s personality can be used to improve recommendations [Tkalčič and Chen, publication expected in 2015]. The problem arises when we want to know the user’s personality. Usually, questionnaires with dozens of questions are used. However, such questionnaires are annoying and useless in practice.
Researchers from the University of Cambridge have developed a method for determining personality without annoying questionnaires [source]. Based on the user’s past activity on social networks (for example Facebook or Twitter), information other than their personality can be predicted, such as their gender, political orientation, IQ, religion, if they smoke and so on. For example, if someone liked the pages Leonard Cohen, Oscar Wilde or Leonardo da Vinci, this is a strong indicator of a personality that is open to new experiences. And conversely, people who have limited openness like the pages Nascar, ESPN2, Teen Mom 2 or I don’t read.
People’s emotions change rapidly, so it is difficult to capture them quickly enough to adapt the recommender system to the current emotion.
There are studies that try to identify the emotional state of a person by looking at their activities in social networks. This is illustrated by this study, which tries to find out from tweets whether their author is stressed or not.
The special characteristic of emotions is that they can be also artificially stimulated, which is called the induction of emotions. This can be utilised in recommender systems that are used to regulate emotions (e.g. recommending music for relaxation). But there is also a dark side, as social networks can also be used to manipulate people’s emotions.
This experiment was done by researchers at Facebook. One feature of Facebook is that its users can see some selected (recommended) content published by their friends. The reason for this is the amount of content, as friends create much more content than a user has time to view. A recommender system that filtered the content published by friends was adapted in such a way that one user group was recommended more content with negative emotions, while the other group was recommended more content with positive emotions. It turned out that users who received more negative emotions generated content that was more negative. A similar pattern was also observed in the second group, which created more content with positive emotions. This experiment showed that a recommender system designed to filter certain content can influence the emotional state of users.
It is important to know that our activity on the Internet leaves traces, which can be used to find out many things about us. Another important thing is the awareness that recommender systems can be used for manipulation that may be ethically questionable.
Author: Marko Tkalčič. Recommender systems, emotions, personality. Postdoctoral researcher at the Johannes Kepler University in Linz, assistant professor at the Faculty of Electrical Engineering, University of Ljubljana. He is interested in the use of psychological constructs in computer user modelling. You can find him on Twitter at @RecSysMare.
Note: This article was originally published in eSiNAPSA, online journal for scientists, experts and enthusiasts of neuroscience (year 2015, number 9).
Title photo via Wikimedia.
Translated by: Tina Goropečnik.