Identifying successful search patterns for improved job recommendation

Job recommender systems help job seekers find a job by recommending vacancies the system believes relevant for them. This work presents the first research done on the impact of the temporal aspect of online job search on job recommendation. Our starting hypothesis is that job search is a process, during which job seekers can have evolving ideas of what vacancies are relevant for them. E.g. a specific job seeker might have no clue what jobs he is interested in at first, preferring a wider set of job recommendations that can inspire him/her. Over time, the job seeker will hopefully get an idea of the type/location/… of jobs (s)he is interested in, preferring a narrower set of job recommendations that fall within the range of that idea. Still, if a suitable job is not found over a longer period of time, the job seeker might be open to a wider range of alternatives. Intuitively there likely exist several of such search strategies. By modelling these job search strategies, we can use them to predict what the current stage of a job seeker’s job search process is, and adapt job recommendations accordingly. Similar temporal dynamics in user preferences have been studied in the context of online music recommendation: in one listening session someone might be open to discovering new music, whereas in another session (s)he could only be interested in known favorites, making novel recommendations a poor choice. For online music listening, this has shown to improve recommendation. Since the job search process is quite complex, and hard to compare to listening to online music, the scope of this research goes beyond predicting the novelty for a single session. The research goals of this work are threefold: First, we model the search process of job seekers, with focus on job diversity, novelty and reciprocity. Second, we investigate the link between search process and unemployment duration, to see which search strategies are more effective than others. Third, we propose changes to existing recommendation strategies that take into account the personalized search process of job seekers. We perform our experiments on data gathered via a job search engine provided by the Flemish public employment services during the period 2015 – 2016. Our data contains approximately 33,000,000 clicks covering 250,000 job seekers and 950,000 vacancies. We aggregate this data into sessions, storing for each session the diversity (job type, location, etc.), novelty, and reciprocity of vacancies looked at. Furthermore, we link demographic information and unemployment duration to the search behavior. We employ a two-step approach to model the existing job search processes. First, we perform a clustering on the sessions to define a state-space of session types (an example session-type could be “session during which vacancies are looked at that have a highly diverse job type, but in a very small location radius”). Secondly, we look at how job seekers transition between session-clusters over time. In order to model common transition paths, we use various techniques, such as hidden Markov models, sequence- and process mining techniques. Our results show that there is a clear difference in job search processes between individual job seekers, indicating there is much to be gained from tailoring recommender systems towards the job search process. We present a strategy on how to tune existing job recommender systems to best consider each job seeker’s personal evolution. We believe that this work is not only valuable w.r.t. improved job recommendation, but that its results can also be used to improve job seeker counseling, potentially leading to a more efficient job search experience and a reduced unemployment duration.status: publishe