Talent science / analytics

Talent Analytics has emerged as a key area of competitive advantage in the modern economy. The progress made in Natural Language Processing, particularly with Deep Learning models since 2012, has been dramatic.

We have developed dozens of algorithms and approaches in the field of talent.

Please see custom solutions and case studies for more.


our CORE algorithm encodes the language of recruitment

A simplified model of high dimensional spaces generated through deep learning models applied to millions of data points, that encode the language of recruitment (not a terrible strain of bacteria)

A simplified model of high dimensional spaces generated through deep learning models applied to millions of data points, that encode the language of recruitment (not a terrible strain of bacteria)


Our algorithm enables you to create high quality, nuanced “job to candidate matching” at scale.


what our algorithm does and How YOU CAN TEST IT

Our algorithm:

  • Reads in the natural language of a candidate profile and a job description

  • Uses a model trained on hundreds of millions of data points to encode the natural language of the documents in a highly nuanced mathematical space

  • An artificial neural network (ANN) compares the quantitative encoding of the candidate with the quantitative encoding of the JD to determine the likelihood that the candidate is a fit for the role

  • The ANN provides a score -- which could range from 0 to 100 -- representing how well the candidate fits the role

  • This score can be used in countless ways, including to rank candidates, find similar candidates, recommend jobs, and automate decisions within digital platforms

 

Screenshot+2019-09-07+08.36.18.jpg
Screenshot+2019-09-07+08.36.18.jpg
 

Our algorithm does not use:

  • Any geographical matching, like a radial search. It's more efficient to pre-filter on your end rather than send all the data across the wire.

  • Hard filters on skills, like "the resume must include Java OR J2EE". Same reason as (1).

  • Level of education or educational institution. Our view is that the career background has sufficient signal without requiring the educational background. This is a debatable point. We have an alternative model which does contain this data, but in tests to date, it doesn't seem to improve accuracy.

  • A score explicitly based on seniority / years of experience, although this is implicitly taken into account by current job. If there is a hard filter (eg, the hiring manager will only interview someone with at least 3 years of industry experience), this should be applied before calling the API.


Ideal ways to test the API to demonstrate what it can do:

  • Pre-filter on your end for geography, hard skills, etc. You should be testing us for nuanced ranking.

  • Test the people with the highest score for a job -- they should be very good choices to put forward

  • Ideally a test should assess on a broad range of scores, because the model differentiates very well anywhere in the score range

  • We're most concerned with getting the natural-language fit spot on: e.g., are the top candidates in the correct vertical? if so, are they very specifically what the JD is asking for?

  • We like to describe our algorithm as 'nuanced, almost human-like'. See if you agree. Some of the results aren't obvious at first, but on closer inspection, make sense. It does make mistakes, but more often it seems to make surprisingly intuitive recommendations.

All models are wrong, but some are useful
— George Box