The problem
Screening is a ranking problem run under constraint: many applicants, little time, noisy signals, and reviewers who disagree with each other and with themselves a week later. Decisions that matter a great deal to the person on the other side get made against criteria nobody has stated. Three questions follow. Which attributes do reviewers actually weight, as opposed to the ones job descriptions claim? Once employability is written down as a binary outcome over measurable features, can it be predicted well enough to be worth anything? And what accuracy do you give up by insisting on a model whose reasoning a rejected candidate could be shown?
Asking both sides
Two survey instruments ran in parallel, one for candidates (12 responses) and one for hiring managers and team leads (10, drawn from large technology firms and startups). Both sides were surveyed deliberately, because disagreement between them is itself a result rather than noise to be averaged away.
Both groups put GPA first, at 43.5% of the importance weight among managers and 41.7% among candidates. Below that they part company. Managers give recommender scores 21.7% of their weight where candidates give 8.3%; candidates give coding assessments 37.5% where managers give 8.7%. Candidates, in other words, over-weight the signal they can grind and under-weight the one that requires a relationship.
Encoding it as a simulation
The elicitation fixed a six-feature description of a candidate: GPA, years of experience, coding assessment score, Leetcode progress, relevant skills score, and resume ATS score. Those six then parameterised a synthetic dataset of 1,000 observations. Simulating was a deliberate choice rather than a fallback. Because the data-generating process is fixed by construction, estimated coefficients can be checked against the true ones, and a pipeline that cannot recover known structure has no business being pointed at unknown structure. Simulating also turns class balance and feature correlation into knobs you can set, and it means no real applicant records are handled at all.
What the models did
Five classifiers spanning the interpretability range were compared. Logistic regression led at baseline, at 98.33% accuracy and 0.9988 ROC AUC, with the decision tree weakest at 85.33%. That ordering should not be read as a finding about estimators. It is what you would expect when the generator is close to linear in the log-odds, so it says more about the simulation than about logistic regression, and its real use is as a control confirming the pipeline recovers what was encoded. Principal component analysis plus hyperparameter tuning then moved the SVM from 89.00% to 96.67% accuracy and from 0.9344 to 0.9950 AUC. Of those two gains the AUC one tells you more, since it holds across thresholds instead of depending silently on class balance.
GPA carries the most weight in the manager survey, in the candidate survey, and in the fitted coefficients. So the criterion practitioners say they use is the one the encoded process rewards, and the pipeline picks that agreement up.
What it is for, and what it is not
The framework is the contribution here, not the accuracy. A tuned SVM scoring 96.67% on favourable synthetic data is unsurprising. What is worth keeping is the shape of the pipeline: elicit the criteria practitioners use, encode them as a generative process, then check whether standard estimators recover them. That works anywhere a selection judgement is implicit and contested, and hiring is only one instance.
The limits are worth stating as plainly as the results. At n = 12 and n = 10 the surveys support elicitation and the justification of a feature set; they do not support inference. Independent normal features are an idealisation, and real applicant attributes are skewed, censored, correlated and partly categorical, so the reported accuracy is an upper bound obtained under conditions chosen to be kind. “Employable” is a binary stand-in for a contested judgement and inherits whatever the elicitation encoded. No disparate-impact or subgroup-calibration analysis was run either, and that gap matters most precisely because the tuned model performs well. It is also why the paper argues interpretability and fairness belong in the design phase of a selection system. Added afterwards, they arrive too late to shape anything.