Project Record

Explainable Recruitment Recommendation Engine

Summarize the Ionitsa project record titled "Explainable Recruitment Recommendation Engine" for a technical reader.
Cover the problem or research question, implementation or method, evidence or results, and limitations.
Separate facts stated on the page from your own assessment, note anything unclear or unverified, and avoid promotional language.

Primary source: https://ionitsa.com/projects/hr-recommendation-algo.md
Canonical page: https://ionitsa.com/projects/hr-recommendation-algo/

Interpretable candidate ranking that combines hard-skill fit, environment fit, experience, and constraints—no neural black box.

Executive Summary

Once candidates are sourced for a quant role, the question is not pass or fail—it is who should be interviewed first. I built an interpretable ranking engine that combines hard-skill fit, environment compatibility, experience relevance, and practical constraints into a single ordered shortlist. Every component of the score can be inspected and explained to recruiters.

We deliberately avoided neural recommendation systems. They can rank well but produce black-box results that are hard to defend in hiring decisions. Instead, the engine uses a weighted blend of explicit, auditable terms that compose outputs from the skill-profiling and environment-matching subsystems.

This demonstrates multi-objective scoring, explainable decision models, and recruitment optimization under real-world constraints.

Problem

We need a single ranking function that reflects (1) match to role requirements (skills, tools, domain), (2) behavioural/environment fit from platform signals, (3) experience level and relevance, and (4) optional factors like location or availability. Inputs are sparse and come from different sources. The engine must compose outputs from the hard-skill profiling and environment-matching subsystems into one ordered shortlist.

Model

Define component scores:

• $H$ — hard-skill fit (TF-IDF-weighted vector matching via cosine and F1-balanced scoring, with skill-graph smoothing from profession inference)
• $C$ — cultural/environment fit (compatibility between candidate preference vector and team environment)
• $E$ — experience relevance (tenure, domain alignment)
• $L$ — location or availability (binary or distance-based)

The overall recommendation score is a weighted blend:

with $\alpha + \beta + \gamma + \delta = 1$. In production we used $\alpha = 0.55$, $\beta = 0.25$, $\gamma = 0.15$, $\delta = 0.05$ as a starting point, tunable per role or team.

For candidate $i$ and role $r$:

Candidates are ordered by $R_1 \ge R_2 \ge \cdots \ge R_n$. Recruiters can drill into each term to see why a candidate ranked where they did.

Implementation

• Hard skills ($H$): TF-IDF-weighted skill vectors compared via cosine similarity and F1-style matching against role requirements; skill-graph partial credit for related tools.
• Environment ($C$): normalized fit from behavioural platform signals; down-weighted when data is sparse.
• Experience ($E$): tenure and relevance to role domain.
• Constraints ($L$): location match, availability, optional diversity or cap filters (e.g. max N per school).
• Ranking: sort by $R$; no neural network in the recommendation path—fully inspectable linear combination.
• Override: recruiters can adjust weights or exclude candidates when domain judgment conflicts with the score.

Trade-offs

Interpretability trades some predictive ceiling for trust and auditability. Culture signals depend on input consistency; hard-skill quality depends on CV completeness, TF-IDF corpus coverage, and graph coverage. Weights are heuristic and role-specific tuning is expected. The model transforms recruitment from binary filtering into an optimization problem: among available candidates, who maximizes the expected success function under explicit, explainable terms?

Related Work