Building a vacancy recommender system | The case of Randstad
The right model for the job
tldr; Randstad built a vacancy recommender system to match the right candidate to the right job. BigData Republic helped Randstad to tackle two challenges: reciprocity and timely recommendations. The resulting recommender delivers an increase in application rate of 27%.
The business of matchmaking
Back when it all started, Randstad founder Frits Goldschmeding gave his first temporary worker a lift to her job on the back of his bike. He made sure that her new place of work was a good one and that she got there on time. Fifty years later, Randstad still puts all possible effort into matching the best candidate to the right client. The company has since become the global leader in the HR services industry.
Why a vacancy recommender?
In 2020, over 1.8 million candidates found a job through Randstad. Operating in such volumes means that it is nearly impossible to apply the same amount of personal attention to people as Mr. Goldschmeding did. That is why Randstad wanted to use data for personalization purposes. The idea of a vacancy recommender was born. But how do you go about building the right model for this particular job?
Challenge 1: Reciprocity in a recommender system
Building the vacancy recommender was the responsibility of the Smart Match Team at Randstad. The first challenge they faced was reciprocity. This particular recommender impacts two parties: the candidate and the client. It is in the interest of a candidate to find the vacancy that matches his preference, but a client needs to find a candidate that meets certain requirements. That is why the SmartMatch Team worked with BigData Republic to develop two models:
- The vacancy recommender: uses factors like experience, skills, and location to recommend vacancies to candidates. It also contains a submodel, called the preferred vacancy predictor, that uses the preferences of a candidate to predict which vacancies make the best match.
- The talent recommender: recommends candidates to recruiters based on the skills of the candidates.
Challenge 2: Building a recommender for timely products
The second challenge for the Smart Match Team applied to the method of building a recommender. A common approach is to use collaborative filtering, employed at for example Netflix. Their recommender system proposes movies to a user based on the preferences of similar users. In our case collaborative filtering on vacancies does not work, since they are usually filled once and then removed from the website. A hybrid approach was used to overcome this problem. A collaborative filtering model on job functions formed the base, with other features as optimizations, e.g. travel distance and salary.
How the vacancy recommender delivers value
Robbert van der Gugten - Machine Learning Engineer at BigData Republic: “In machine learning projects It is very important to start off with a simple model. First, prove that it has value. Then iterate to more complex models that have better performance. With this in mind, we started with an enhanced matrix factorization implementation for the preferred vacancy predictor. It was quickly integrated into the website where we gathered statistics. They showed that the model was performing well.
For the second model we had to take many features into account. A matrix factorization solution was inefficient and had issues with new candidates and new vacancies. We chose the Starspace model for the vacancy recommender, which creates embeddings for all candidates and vacancies. We recommend the vacancies that have the highest cosine similarity between a candidate embedding and all of the vacancy embeddings.``
We integrated the vacancy recommender with the talent recommender, using a weighted average of both scores. The combined efforts resulted in a 27% improvement in the application rate, compared to vacancies that candidates find by themselves.
Robbert van der Gugten - Machine Learning Engineer at BigData Republic