NAOMIHIRE: AI CAN EXPLAIN ITS DECISIONS OR RECRUITING WITHOUT PAIN

NaomiHire team is glad to announce that we have successfully finished beta-testing of our AI engine “Naomi”. The main goal of the testing was to identify how Naomi explains its decisions regarding to: recruiting the best candidates for particular job; sourcing the best opportunity for particular candidate. Testing process was built on real job descriptions and real candidates. We use significant computer ontologies to build job descriptions and candidates profiles: locations (> 77K classes), soft skills ( > 300 classes), hard skills (>7900 classes), industries (260 classes), natural languages (> 800 classes). Below we explain data structures and the results of the testing process. Job (opportunity) descriptions Job or Opportunity description can be linked to the company or one of company projects. Opportunity profile consist of job details, job description, hard and soft skills requirements, additional information, salary, hiring strategy and hiring steps. Figure 1. – Hard skill requirements Skill requirements are designed in the way to replace BOOLEAN SEARCH  and help to avoid common mistakes made by recruiters used boolean search. As you can see in the fig.1 each skills requirement is a set of alternative skills with the same competence level. This approach allows us to build any search query with semantic like: {(“J2EE” OR “Java Web Start”) with level “EXPERT”} AND  {(“IntelliJ IDEA” OR “Eclipse”) with level “ADVANCED”} AND … The query shows that NaomiHire move the power of search to the new level because no other recruitment service gives you search by similar semantics. At the same time recruiter does not need to know how to build boolean query string. We do it behind the scene. Hiring strategy Figure 2. – Hiring strategy The next important feature is a hiring strategy. Why do we need that? Because as we know an importance of hard skills or soft skills depends on position level. Implemented hiring strategy let us to present any boolean search query in the semantic form like: [{(“J2EE” OR “Java Web Start”) with level “EXPERT”} AND  {(“IntelliJ IDEA” OR “Eclipse”) with level “ADVANCED”} with score “60”] AND [{(“Listening” OR “Attention”) with level “EXPERT”} AND  {(“Presentation skill” OR “Public speaking”) with level “ADVANCED”} with score “30”] AND … Incredible, isn’t it? Other blocks of information let recruiters to add other important filters and Naomi will return appropriate candidates exactly in accordance to search query relevance (shortlisting).  It is unreachable today for many technology companies. Candidate profiles Candidate profile is presented by the following blocks of information: personal info, skill set, education, work experience and vacancy search criteria (“Looking for”). Let investigate some of them in more details. Figure 3. – Candidate skill set Candidate skill set is a set of  hard skills, soft skills and natural languages. Hard skills may be actual and outdated and NaomiHire accounts it. Recruiter or candidate has to assign level of competence for each skill. Figure 4. – Candidate education history Candidate education history shows most valuable aspects of his/her education including dates, Alma-mater name, education level, location and hard skill used/learned in particular educational institution. Figure 5. – Candidate work experience Candidate work experience is represented by a set of records about his/her jobs in the present and in the past. Figure 6. – Candidate is looking for Candidate can also define types of companies, company maturity, career levels, possibility of remote job and expected salaries in different locations. Whole set of described search criteria from candidate profiles and job description lets Naomi provides best semantic matching comparing to all other candidate searching tools. Matching report NaomiHire uses Job and candidates profiles to perform semantic matching within a second and return short list of candidates relevant to particular vacancy. Figure 7. – Shortlist summary (matching report) Upper part of the shortlist summary report presents the aggregated information about most relevant candidates and shows figures about cost effectiveness, total relevance to job criteria, hard skill relevance, soft skill relevance, expected and evaluated by marked salary for each of potential candidates. Shortlisted candidates can be sorted by price efficiency, total relevance or other figures. Lowest part of report is filled by charts. We suppose that most useful could be a chart that shows  how candidates fit to vacancy budget (marked by red dotted lines). Figure 8. – Shortlist summary (matching report) Candidate matching report details For each shortlisted candidate Naomi builds a personal detailed report. Figure 9. – Candidate matching report In personal report HR expert can find the information how candidate matches to the vacancy. See some screenshots that shows all the advantages. The system reacts on any changes in job description or in candidate profile and in a few seconds returns new matching report. Testing result NaomiHire Team conducted a blind testing session with real companies and real candidates and have got accuracy 92,3% of best candidates selection for a specific job. We also have the same accuracy when we find the best job for the specific candidate. We also had another testing session when we asked recruiting agencies and companies to give us already closed jobs and list of candidates reviewed for those jobs. We didn’t know who was hired for those jobs. NaomiHire selected winners with 80% accuracy and TOP-2 candidates were selected with 95% preсision.

THE AMAZON LESSONS OR WHY NAOMIHIRE IS SAFE FOR COMPANIES

This week The Guardian published the article “Amazon ditched AI recruiting tool that favored men fro technical jobs”. The main point is that Amazon’s machine-learning specialists uncovered a big problem: their new recruiting engine did not like women. NaomiHire Team predicted such a problem and have designed own AI engine which doesn’t consider gender, nationality, age during the matching process. And there are no such filtering criteria when you create a job. Let see how our job description looks like: Figure 1. – Job details contains Any job profile in NaomiHire consists of the following data: title, company, job type, seniority level of position, years of experience in IT sphere, locations (where company looking for a candidate), hards skills requirements, soft skills requirements, required level of education, required languages and their levels, benefits in form of text, hiring strategy (was shown in the article “NaomiHire: AI can explain its decisions or recruiting without pain”  ), hiring process as a set of steps for candidate to be hired. Below you can see the data fields available for NaomiHire AI:   Figure 2. – Job description preview The same picture with the personal profile of the candidate. Figure 3. – Personal info about candidate Another important note for company: candidates are not able to see vacancy profiles before they will be matched by AI. It means that candidates can not tune their profile for a particular vacancy. The conclusions So, NaomiHire does not use gender, nationality, age in processing and helps companies avoid cheating through CV tuning because NaomiHire builds candidate rankings based on companies feedback after interview.

HOW DOES NAOMI AI WORK AND WHY NAOMI IS AI

Naomi AI is the heart of the NaomiHire Recruiting Career Service. The main scientific issue was how to evaluate a distance between any two concepts where a concept is anything that can be put into a taxonomy or ontology context. For the purpose/goal of the issue a distance (distance function) between two concepts means a strong mathematical metric. The answer to the main question has been given in the paper “MODEL OF COMPUTATIONS OVER CLASSIFICATIONS” But no reasons to think that a mind uses only one simple metric for evaluating distance in the issues of comparison two concepts. List of distance functions between two concepts were represented in “TRAINABLE MODEL OF THE CALCULUS OVER CLASSIFICATIONS” In the same paper a learnable model over classifications(ontologies) has been described and explained.   Using mathematical metrics is possible only if a context is a metric space. It required to define some rules how to create taxonomy/ontology in a way to guarantee that the result will be a metric space. The results are presented in the paper “CLASSIFICATION CALCULUS. THE CLASSIFICATION CORRECTNESS” During building ontologies for NaomiHire the co-founders validated results of each other in accordance with the rules “MEASURE OF DIFFERENCE BETWEEN CLASSIFICATIONS” The general approach used for matching a candidate and a job description was published in “THE CALCULUS OVER CLASSIFICATIONS. SELECTION OF PERSONNEL AS INTERPRETATION OF THE PROBLEM OF EXPERT SELECTION” Naomi AI uses a set of specific rules and applies them in accordance to feedforward neural network output for explaining its decisions.   The main goal of the testing was to identify how Naomi explains its decisions regarding to: a) recruiting the best candidates for a particular job; b) sourcing the best opportunity for a particular candidate. The testing process was built on real job descriptions and real candidates. We use significant computer ontologies to build job descriptions and candidates profiles: locations (> 77K classes), soft skills ( > 300 classes), hard skills (>8K classes), industries (260 classes), natural languages (> 800 classes). NaomiHire Team conducted a blind testing session with real companies and real candidates and have got accuracy 92,3% of best candidates selected for a specific job. We also have comparable accuracy (>90%) when we find the best job for the specific candidate. The details are available in our previous post “NAOMIHIRE: AI CAN EXPLAIN ITS DECISIONS OR RECRUITING WITHOUT PAIN”    The most often asked question is why NaomiHire team considers that it uses exactly AI but not machine learning. The answer is given on the header image that presents AI scope where following functionalities are provided by NaomiHire exclusively in the automation (machine) mode: Data filtering Data normalization and conceptualization Ensemble methods or models Insights value assessment and interpretation Decision variants and hypothesis producing Deduction or hypothesis pre-validation and Reflection It means that Naomi today corresponds to weak artificial intelligence and in the next decade will become a strong AI