[BELLE HISTOIRE] Using graphs and deep learning to make recruiters' lives easier

Easy Partner is a recruitment agency specializing in the tech sector. Its role is to help its customers identify and attract the best profiles, particularly in development, engineering or data science. And to achieve this, the company relies on innovation. In 2021, it intends to launch a research project to assist its recruiters on a daily basis.

" At the time, I was looking for a job in data science, having had to give up a CIFRE thesis a year earlier because of covid," recalls AI researcher Éric Behar. " After applying to the firm, I got in touch with the founder of Easy Partner, who told me about his innovation project and indicated that my profile might match this search. However, he still hadn't decided how to go about it, so I suggested that he set up a CIFRE thesis. With the company convinced, all that remained was to identify the right academic partner. In the end, it was the Distributed Services, Architectures, Modeling, Validation and Network Administration (SAMOVAR) laboratory at Télécom SudParis, part of the TSN Carnot Institute, with Amel Bouzeghoub, professor at Télécom SudParis, Julien Romero, lecturer at Télécom SudParis, and Katarzyna Wegrzyn Wolska, deputy director of the Efrei Research Lab, co-directing the thesis.

Helping recruiters pre-select candidates

The aim of this approach was to tackle one of the difficulties faced by recruiters: the pre-selection stage. In other words, for a given vacancy, how do you quickly find an initial list of relevant candidates? While professional networks now offer access to a colossal amount of information, this task is like looking for a needle in a haystack. " Our ambition was never to replace the recruiter, nor to eliminate human interaction, which is absolutely vital in recruitment," insists Éric Behar. " Rather, it was to help recruiters process a volume of data that would be impossible for a human being to analyze, and to carry out an initial filtering of candidates, by studying job offers and profiles in detail. "

Not, of course, a new idea. A number of automatic recommendation systems already exist. Nevertheless, developing a reliable solution remains a challenge, not least because of the sheer volume of data required. " As a result, the tools available today are often deemed disappointing by recruiters, compared to the promises made," notes Éric Behar. Conversely, a player like LinkedIn has a sufficient database and the appropriate technological resources. But its objectives are not those of a recruitment agency: instead, the professional network seeks to maximize the number of clicks from its users and sell paid subscriptions.

A graph to represent the tech recruitment market

As part of his CIFRE thesis, Éric Behar developed a new recommendation system for recruiters. The process is broken down into several stages, starting with a representation of the recruitment market in the form of a graph, made up of several nodes linked by edges. " Graphs are the subject of a great deal of research work, because they are a tool that represents complex information well," emphasizes the researcher. " Here, each node corresponds to a candidate, a job offer, a skill or a location. And there's an edge if there's a link between two entities. For example, a candidate who knows how to code in Python and lives in Paris will be linked to these two nodes. The same applies to a job offer requiring this programming language and located in Paris. This representation enables us to position a candidate globally on the tech recruitment market. And all thanks to the database provided by Easy Partner.

However, this preliminary work came up against a number of difficulties. Indeed, information such as skills comes from those declared by candidates or filled in by recruiters. In reality, two different terms can be attached to the same expertise. For example, one cloud engineer may mention "Microsoft Azure" on his CV, while another may mention "AWS": these are nevertheless two similar profiles. What's more, it's not always easy for a recruiter to keep track of changes in skills, or even job names. " For example, I used to be considered a ' data engineer', but this term has been replaced by ' AI engineer'", says Éric Behar. " The profession hasn't changed, it's only the popularity of AI that has induced this semantic evolution. "

Therefore, in order to have a graph that exhaustively represented the tech recruitment market, it was necessary to read between the lines, and even behind the lines. This is why the researcher also relied on two public databases: ESCO(European Skills, Competences and Occupations) and Wikidata. The former lists and categorizes skills, competencies, certifications and occupations relevant to the labor market in the European Union. The latter, with its vast open-access knowledge base, completes the information, particularly on proprietary technologies.

A deep learning model that learns from graphs and real cases

Then, based on this faithful representation of the tech recruitment market, how can we establish a match between job offers and candidates? To do this, Éric Behar used a " graph neural network ", a deep learning model based on an artificial neural network designed to operate on graphs. His aim: to reproduce the matching work carried out by a recruiter, between an offer and candidates. " Our model learns to establish links based on the proximities it observes between different nodes," explains the researcher. " For example, it can see, thanks to our graph enrichment work, that the terms "Azure" and "AWS" are close to the skill "cloud service". In this way, he understands that two a priori different profiles can correspond to a job offer requiring cloud skills. "

The training of the AI model is also based on past recommendations, actually made by Easy Partner recruiters. These are applications that have been shortlisted for a vacancy, but which have not necessarily led to recruitment. In fact, the system we have developed is designed to facilitate only the first screening stage, not to intervene at the moment of the customer's final decision. In addition, in order to refine its understanding of the matching process, the neural network was also fed with applications that were not selected by recruiters for a given vacancy.

How can time be taken into account in the recommendation system?

In this way, the model learned to make recommendations in both directions: relevant candidates for a job offer and offers matching a candidate. However, this approach had its limits, not least because it failed to take temporality into account. " Candidates don't stay on the market indefinitely, and job offers are eventually withdrawn from the market," notes Éric Behar. " So how can we inject a notion of temporality into our representation, while retaining its semantic richness?

Answering this question meant taking into account the specificities of the graph: in this case, time only had to be applied to certain types of node. Others, on the other hand, were not affected by this imperative: for example, the location "Paris" is not associated with a limited duration. It was therefore not possible to apply an ordinary temporal graph model. " We then opted for another approach, creating a new type of node, representing a recommendation, in which time-related information is stored," reveals Éric Behar. " Thus, candidates and job offers are no longer linked directly by an edge: matching now takes the form of a triplet associating a candidate, an offer and a recommendation, to which a time is associated. "

Towards greater explicability and use in healthcare?

The recommendation system developed was awarded the Best Paper Award at the international reference conference Web Intelligence and Intelligent Agent Technology (WI-IAT) 2024. In addition, learnings from the research work have enabled the implementation of a tool to help identify potential candidates, which is now used by the company's recruiters. "It's a real source of pride to see my work being of concrete use to those who have always been at the heart of this research project," enthuses Éric Behar.

The researcher is also considering other applications for his work, such as in the medical sector. This would involve linking patients, pathologies and treatments within a heterogeneous graph representation, a context in which temporality plays a key role. However, there are still many avenues to explore around recruitment, such as that of explicability. In this way, the system can justify the recommendations it makes, to boost user confidence and encourage adoption.

" Nevertheless, it's important to remember that the use of such a tool raises ethical issues," warns Éric Behar. " As a researcher, my role is to propose methods for designing reliable systems and implementing them transparently. But should we do it? The answer to this question is not easy, since it involves so many essential societal issues.