Published May 21, 2020 by Scott Hirsch
This is the second of five posts about machine learning at TalentMarketplace. Take a quick read of the introduction blog here.
Moe has a Bachelor's in Accounting and an Associate of Computer Science degree: while going through the latter, Moe joined the Applied Quantitative Methods program, a statistical and machine learning program that takes a select number of students, mostly with Masters and PhD degrees, who have expertise and aptitude in quantitative sciences.
When connected with TalentMarketplace through AQM, Moe was given the task of automating candidate recommendations to employers. Upon starting on this project, Moe noticed that, despite numerous similar candidate resumes, many employees were not viewing these similar candidate profiles.
To counter this problem, Moe set to work and began the creation of what would become the Candidate Resume Recommendation Engine. In short, this engine would show employers similar candidates to their previous searches, hires, and views. For the hiring decision makers, this means they can more easily find the right candidates, and build their hiring pipeline. For TalentMarketplace, this means more of the right views, thus more interview requests, and a stronger business overall.
One challenge with showing similar resumes is that resumes are inherently not standard. For example, some people say “dashboard”, while others say “reporting”. After some trial and error, Moe used Word Mover’s Distance, a measure of difference between vectors that represent words/documents in N-dimensional space, which can understand the similarity and dissimilarity between words such as “dashboard” and “reporting”, and suggest resumes to employers accordingly. More specifically, Word Mover's Distance (WMD) is an algorithm used to compute semantic similarities between collections of word vectors (known as "bag of words") by applying word2vec (an algorithm to encode text as vectors) to individual words. As a result, this approach enabled Moe to create the most efficient recommendation machine, allowing employers to find their ideal candidate in a short amount of time.
Semantic similarity between words is represented as the distance between vectors, so that a similarity ranking can be created through calculating the distance between resumes (in this example, documents).
 Kusner Matt J., Sun Yu, Kolkin Nicholas I., Weinberger Kilian Q. From Word Embeedings To Document Distance. 2015. http://proceedings.mlr.press/v37/kusnerb15.pdf
Moe, a Programmer with AQM and nearing the end of his second Bachelor’s degree, developed TalentMarketplace’s Candidate Resume Recommendation Engine. In Moe’s upcoming blog post, we explore how this recommendation engine enables employers to see lookalike candidates immediately upon viewing a candidate profile, saving them time. This model was made possible through the use of Word Mover's Distance to rank similar resumes and show the top three.
The number of applications that this one solution had within the recruitment space was particularly interesting to this project. As many of my mentors have noted, the true applications of data science lie within a number of algorithms that are applied at different parts of the value chain. Moe’s Candidate Resume Recommendation Engine is a 3 in 1 - helping employers find candidates when they’re actively looking or when they’re just browsing, and helps candidates better showcase their existing skills using the best keywords.