Tools for candidates
Solutions for companies
Get the Torre Messenger app
Stay on top of your messages and notifications in Torre.
Torre’s Job-Matching Model
Learn how Torre’s algorithms work
1.
How do we suggest and rank candidates
1.1.
Scoring
Our algorithm evaluates and assigns scores to every existing candidate or job listing within our database. The scoring process is driven by the job description, which acts as a query generator. We then compare the attributes of a person’s genome to see how well the candidate is able to perform the requirements of the jobs. Analyzing this data we come up with a score range, the more data we have the smaller and more accurate the range.
1.2.
Filtering
We use the scores and some other features to make sure we only show candidates who are able to do the job. To be able to perform a job you need to pass the minimum requirements of the job. Often we use the score previously calculated and use a threshold that’s calculated by analyzing millions of previous matches.
1.3.
Ranking
The final stage involves ranking the candidates in descending order of suitability, with the "highest potential match" candidates taking the top positions and gradually descending to "lower potential matches." The accumulated scores and overall compatibility with the job description determine this order. For ranking, we use more features than for filtering and scoring because there are some aspects of the job that are not necessarily required but make a difference between a good candidate and an amazing candidate.
2.
Certainty calculations
Our approach to candidate evaluation recognizes that many factors influence a candidate's ability to excel in a job, and sometimes, certain pieces of information may be missing. To address this, we use the concept of 'uncertainty' into our matching algorithm.
This concept comes into play when a candidate's profile lacks information about specific requirements for a job. Rather than making assumptions, our algorithm assesses the likelihood of whether the candidate possesses the unlisted skills or qualifications. Instead of presenting a definitive single score, we offer a range - a lower and upper bound - that reflects our prediction's possible variance regarding the candidate's capability to meet the job requirements. Therefore, when a candidate that applied for a job opportunity has not provided enough information for Torre to evaluate if they are a match or not, we label them as "Match to be determined".
This method offers a more nuanced and realistic view of a candidate's potential fit, acknowledging that profiles may not always capture every skill or experience a candidate possesses. By providing this range, we offer employers a more comprehensive understanding while respecting the inherent uncertainties in the candidate evaluation process.
3.
Scoring/filtering features
3.1. Skills proficiency
The skill-proficiency feature’s purpose is to validate if a user is able to perform at the required level of the job considering its skills and proficiencies on their profile.
3.2. Language fluency validation
The language-fluency feature’s purpose is to validate if a user is able to perform at the required level of the job considering languages and fluencies on their profile.
3.3. Salary validation
The salary-matching feature’s purpose is to ensure that a user's salary expectations align with the salary range offered for a job, based on the salary expectations of the candidate and the proposed salary on the job.
3.4. Organization size validation
The company-size validation feature's purpose is to ensure that a user's preference for company size aligns with the size of the company offering the job, based on the preferences of the applicant and the company's profile.
4.
Ranking-only features
4.1. Connection strength
Talent seekers tend to hire people from their network before looking at any other options for hiring. This network is often accessed through WoM or LinkedIn, but the network is a lot larger than that outreach usually goes. With this factor, we rank all candidates based on their multi-degree connectivity to members of the job post.
4.2. Behavioral similarity
Indicating the user's behavioral characteristics, linked to the recommendation algorithm. The purpose of this feature is to measure how well the job seeker compares to the job in terms of behavioral traits. This feature combines over 40 different individual factors and combines them into one score.
4.3. Years of experience
The skill experience ranker algorithm's purpose is to evaluate how many years of total experience the user has with all skills of the opportunity. In this case, we don’t only check if the job seeker can do the job like in the skill-proficiency factor, but we compare every user in how much relevant experience they have in each skill.
4.4. Self-awareness
The purpose of the self-awareness factor is to validate if candidates understand what they have to offer in their profile and if the genome is providing value to the job posts they applied to.
4.5. Recommendation weight
The skill recommendation-weight ranker algorithm's purpose is to evaluate the reliability of the genome and how much others have liked working with the genome.
4.6. Genome completion
The genome completion ranker ranks candidates based on how much information is on their genomes. Candidates that have more complete genomes will be more accurately ranked because we have more information to validate their ranking. This ranker ensures we penalize emptier genomes.
4.7. Screening questions
Screening questions are a way to add requirements on a job post that are not part of the quantification of the normal job post requirements yet. With these questions, you are able to dig deeper into candidate's motivation for this specific job post for example.
5.
Total ranking/scoring formula
Taking into account the comprehensive range of factors outlined earlier, we have meticulously trained a machine-learning algorithm using an extensive dataset encompassing millions of applications, mutual matches, invitations, hires, and disqualifications. This model is expertly optimized to prioritize the most promising candidates at the top of the list, significantly reducing the time required for candidate review. Importantly, we continuously refine and optimize the model to maintain its efficacy and accuracy.
Currently, we are utilizing a random forest model, renowned for its robustness and effectiveness in complex scenarios like candidate ranking. The strengths of this model in our context are manifold:
- Complex pattern recognition: Random forests excel at detecting intricate and non-linear relationships among variables. In terms of candidate ranking, this capability allows the model to evaluate how various elements of a candidate's profile synergistically affect their suitability for a position.
- Resistance to overfitting: Due to its ensemble approach, which amalgamates multiple decision trees, a random forest model is less prone to overfitting. This trait is vital in recruitment, ensuring consistent performance across diverse and new candidate profiles.
- Insightful feature analysis: The model offers insightful analyses on which aspects of a candidate's profile (like particular skills or experiences) are most predictive of their fit for a role, aiding recruiters in making more informed decisions.
- Versatility with data types: Random forests adeptly handle both numerical and categorical data, a common scenario in candidate profiles, thereby providing a holistic assessment.
- Scalability with large datasets: This model type remains efficient even with large-scale data, a usual occurrence in recruitment with extensive candidate pools.
- Interpretability and ease of use: Despite its complexity, the random forest model is relatively straightforward to interpret, an important feature for ensuring transparency and trust in recruitment processes.
6.
Open features for ranking
7.
Privacy
8.
The Team
We, the authors, acknowledge the potential for bias in this model due to the composition of the development team. Our team has consisted of two females, 16 males, and one or more researchers who self-identify as LGTBQ. The average age is 31. Our nationalities are represented by Brazil, Colombia, Ecuador, Egypt, Mexico, the Netherlands and the United States. Our professional backgrounds span computer science, machine learning, mathematics, statistics, finance, music, art, and self-taught coding. This demographic makeup may influence our perspectives, interpretations, and modeling of our algorithms. While we have strived for objectivity in our development, readers are encouraged to consider this potential bias when evaluating our approach and technology.
Fair artificial intelligence.
We’re committed to it.
