Job interview 2.0: how a recruitment algorithm is born

And the great thing is that 'hybrid' selection (i.e. conducted by machines and humans side by side) works very well, at least according to a study conducted by Emil Palikot, Ali Ansari and Ada Aka of Stanford University.

The three researchers tried to use the most advanced artificial intelligence-based screening systems and compare them with more traditional automated systems.

The result? Thecandidates selected by AI algorithms turned out to be the best: in 53% of the cases they were hired by human HR managers after the face-to-face interview (against 28% of those selected by traditional methods).

Furthermore, the Stanford trio demonstrated how the 'conversational' approach of AI, with video interviews, is much more effective with candidates who are young or at the beginning of their careers.

Even the Harvard Business Review confirms how the use of specialised software (Workday, Oracle HCM and Greenhouse for example) is the norm in business to score candidates, finding the best ones then with the last human check.

But how is a personnel selection algorithm designed? How does it work? What does it consider in its evaluations? What are its secrets?

We turned these questions over to Rinaldo Festa, CTO of Cosmico, an Italian job matching start-up that, instead of buying the traditional 'generalist' software on the market (such as HireVue, Pymetrics or iCIMS), has created a particularly sophisticated proprietary one.

"First of all, it has to be said that selecting professionals is very challenging,' Festa clarifies: talent is a multidimensional concept, encompassing not only prior experience and declared skills, but also intrinsic abilities, growth potential, cultural compatibility and proactivity in learning.

In Cosmico, information is collected from various sources: resumes, responses to questionnaires on skills, experience and industry, data on behaviour and interaction with training or professional development platforms.

A critical aspect is managing self-declared information in resumes or multiple-choice tests: this may in fact be subjective or incomplete.

"In order to obtain a more reliable assessment, we need to integrate this data with more objective sources such as interaction with a platform (time spent on training content, completion of tests, participation in mentorship sessions), which provides objective indicators of interest, proactivity and knowledge acquisition," Festa emphasises, "but also with the formulation of more specific and scenario-based questions.

"The main technical challenge is then converting all this information - free text (role descriptions in the CV), multiple-choice answers (tool competence level), counters (years of experience) and behavioural data (videos viewed, quizzes completed) - into numbers that have mathematical meaning," Festa continues.

For text, advanced Natural Language Processing (NLP) techniques and the use of "embeddings" (vectors pre-trained on large amounts of text) are crucial for capturing the semantic meaning of words and phrases. The structured and behavioural data are in turn transformed into numerical values or coded categories.

"All these numerical characteristics are eventually combined (often concatenated) to form a single vector that represents the complete profile of the talent: the process is called "Vector Transformation Technology" and allows qualitative data to be converted into quantifiable vector representations, preserving important nuances".

Once talent profiles have become vectors, their comparison and analysis become simple mathematical operations.

"We can easily understand how similar two profiles are and especially which is the best match between position and professional (through the TRS Matching Algorithm), creating an objective ranking of the most suitable candidates for that specific role".

However, human intervention remains fundamental.

"The algorithm provides a powerful tool for analysis and pre-selection, generating precise and objective measurements. But the final decision, especially in the more qualitative interview and assessment phases (represented in the Evaluation Framework as 'HR team re-evaluation' phases), benefits greatly from the insight, experience and sensitivity of HR professionals," Festa clarifies.

The vector approach, in short, 'only' serves to facilitate and improve human decision-making. But it should never replace it.