Restless Bandit Blog

Your best
candidates are
hiding in plain sight

Restless Bandit’s Search Algorithms

Posted by Steve Goodman on May 8, 2017 4:53:00 PM
Find me on:

Why are Restless Bandit’s search algorithms results best in class? After all, don’t most Applicant Tracking Systems (ATS) provide native search functionality?

This is a question we get all the time.

Restless Bandit is a technology search company through and through. Our core competency is search—or, in the parlance of Silicon Valley, information retrieval—and we chose to apply our search expertise to talent acquisition.

Started in 2014 by ex LinkedIn and Google search team members, Restless Bandit’s team has over 10 (collectively close to 30) years of experience in the discipline of information retrieval.

Applicant Tracking Systems are excellent solutions for workflow. Solutions such as Taleo, iCIMS, and others provide functionality to move candidates through a pipeline, similar to the way Salesforce moves deals through stages of sales qualification. The professional and academic qualifications to design and develop workflow applications versus information retrieval search engines are fundamentally different.

Informational retrieval, made famous by Google in the late 90s, is a discipline requiring specific skills developed through years of academically rigorous PhD programs and peer-reviewed published research. Restless Bandit, while just a three-year-old startup, has three technology PhDs on staff, and its engineering team of 17 professionals (as of March 2017) has 14 members exclusively focused on data science and information retrieval. It’s the cornerstone of the business.

Applicant Tracking Systems, for the most part, integrate third party search systems into their platform. The companies that build applicant tracking systems have neither a core discipline in search nor are they even focused on it. In most cases they use open source software search systems such as Solr or Lucene, and then modify these off-the-shelf generic systems slightly for their specific application.

While Restless Bandit’s search was purpose-built for recruiting, Solr and Lucene are not. These other search platforms were developed as keyword matching systems—and they’re excellent at that. Keyword matching, however, is the simplest form of search, and yields results far inferior to the latest methods in information retrieval.

Standardizing on generic search would be like standardizing on Google’s 1998 search versus Google’s vastly superior 2017 information retrieval models. Most would agree that today’s Google results are almost magical in their quality, and a quantum leap above the earlier iterations.

For example, a Boolean search in a native ATS for “marketing HubSpot” skills will provide an exact match of these skills. The results will be without context. With Restless Bandit you’ll get inferred skills back for marketing and HubSpot, such as Marketo and Eloqua (these are similar skillsets to HubSpot.) If a person doesn’t have specific HubSpot skills but they do have Marketo skills, then their skills are likely transferable to HubSpot and they’re qualified for the role.

This is just one simplified example. These inferences can cover tremendously complex requirements, and the only way to build them is to evaluate millions of resumes and build statistical models. Restless Bandit’s search builds a model around each resume, and evaluates tenure in job, time between jobs, title ascension over a period of years, skills ascension, career changes, quality of education and training, and whether someone previously worked at a competitor company. These are just a few of the over 100 features Restless Bandit’s information retrieval algorithms evaluate.

While the Restless Bandit website enumerates the work that went into development of our talent acquisition search engine, it’s worth repeating how we developed them:

  • We aggregated over 30,000,000 resumes and over 120,000,000 job descriptions
  • We trained the algorithms for several years via explicit human feedback
  • We developed the models
  • The algorithms automatically get better over time (machine learning), as more feedback is gathered. Most of this feedback happens via access to customer data

If you’re the technical type and want more detail, please check out our 30-page whitepaper explaining the exact methods we use to formulate our results. Request a link from our team, we’re more than happy to share.

Happy searching, your best candidates are hiding in plain sight!


Topics: machine learning, AI, artificial intelligence, big data, algorithms, search, boolean, ATS

Subscribe to email updates

Recent posts