Towards a better understanding of recommender system in the labor market.
The rise of information systems, and especially the internet, has given service providers the opportunity to increase the number of services they can offer their customers. E-commerce platforms, such as Amazon and AliExpress, can offer their customers millions of products because they do not need shelf space to present all their products. Entertainment providers, such as Youtube, Spotify and Netflix, have millions of songs and movies from which the user can choose. Also in the labor market, employment services and job-posting sites, such as LinkedIn, can now present hundreds of thousands of vacancies to job seekers. With so many open vacancies, it can be difficult for job seekers to find jobs they are interested in, provided that they even know what they are interested in. The general problem, in which the large amount of available information hinders users in their decision making, is called the information overload problem.
Recommender systems are developed as a solution for this information overload problem. They are the set of software tools and techniques that help users find items that are relevant to them, by predicting which items the user will be interested in. This thesis presents three studies on the properties of recommender systems that recommend vacancies to job seekers, or vice versa, recommend job seekers for vacancies.
The first study compares two ways in which job recommender systems can capture job seekers' interest in open vacancies: explicit user interest and implicit user interest. This study shows that there is a large discrepancy between the job types people explicitly state to be interested in (explicit interest), compared to the job types people show implicit interest in. It also shows that job recommendations generated using implicit feedback data resulted in a more diverse set of recommendations that better predict job seekers' job interest and received higher expert appreciation, compared to recommendations generated using explicit interest data.
The second study in this thesis zooms in on two topics. First, the fact that for job recommendation to be successful, the job seeker has to be interested in the recommended job, but that this interest also needs to be reciprocated by the job provider. Second, the comparison between the reciprocated interest of vacancies that the user consumes via search, and vacancies that the user consumes via recommendation. The key take-aways from this chapter are the following: Vacancies consumed by traditional recommender systems have much lower reciprocity than items consumed via traditional user search. This means that simply plugging in standard unidirectional recommender systems in a reciprocal recommendation context could have a negative impact on the consumer of the recommended items.
However, the reciprocity of recommended items can be increased beyond the reciprocity of user search, with a trade-off in predictive power, by varying the rating matrices used by the recommendation algorithms.
In the third study, we present the extension of a methodology that we originally developed to visually inspect the distribution of recommendation quality over all users. The methodology allows allows non-technical interpreters to answer the following questions about the relationship between a feature and a complex dataset: Is the feature distributed randomly, or is it systematically biased towards groupings in the data? If there is an observable bias, what features from the dataset drive this bias? Applied to recommender systems, this methodology can be used to study for which job seekers recommender systems work well, and for which they do not.