Career Fit Score
A transparent weighted similarity between a person's trait profile and an occupation's typical-incumbent profile, anchored in O*NET evidence. Four channels, disclosed weights, full audit trail. No black-box embedding.
What the score is, in one sentence
For each of the 2,536 careers in the knowledge graph, the Career Fit Score is the calibrated percentile of similarity between a person's assessed profile and that career's typical-incumbent profile, combined across four evidence channels with disclosed weights.
The four channels
Channel 1 — Interests (35%).RIASEC themes from the user's RIASEC assessmentare matched against O*NET's six interest profiles per occupation. RIASEC was the original person–environment-fit framework (Holland, 1997) and interest-congruence meta-analyses (Nye et al., 2017) consistently find it predicts vocational persistence and satisfaction better than other channels. This is why interests carries the heaviest default weight.
Channel 2 — Skills (30%).The user's self-rated and assessed skill profile is matched against the occupation's skill requirements from the skills sub-graph. Skills with O*NET importance ≥ 3.0 contribute proportionally to their importance score. Skills admitted via posting-evidence contribute proportionally to their empirical frequency. The user's skill rating is normalized to the population.
Channel 3 — Traits (20%).Big Five trait scores from the user's Big Five assessmentare matched against O*NET Work Styles (which are factor-aligned with Big Five). Conscientiousness has the highest weight within this channel because the meta-analytic evidence (Barrick & Mount, 1991) is clearest there; Neuroticism has reduced weight because raw level is less predictive than reactivity under specific job stressors.
Channel 4 — Cognitive (15%).The cognitive-reasoning composite is matched against O*NET's ability requirements (verbal, quantitative, spatial, working memory). Job-complexity adjusts the weight inside this channel: cognitive carries more weight for complex roles where general mental ability predicts performance at ρ ≈ 0.50, less for routine roles.
How the channels combine
Each channel produces a normalized similarity score in [0, 1]. Channel scores are combined as a weighted sum using the disclosed weights, then calibrated to a 0–100 percentile against the full universe of 2,536 careers using the user's own distribution. The calibration step is what makes a "73" comparable across users: it means "this career sits at the 73rd percentile of fit for yourelative to all 2,536 careers we know about."
The headline number is the calibrated percentile. Every career page additionally shows the channel-level contributions, so a user can see whether a high score was driven by interest congruence, skill match, trait fit, or cognitive alignment. Two careers at 73 can be reached by very different evidence paths, and that information is preserved.
Why a weighted similarity, not a learned ranker
The most fashionable approach today would be to train a learned embedding model on labeled career–person pairs and let a neural network do the ranking. We deliberately do not.
The reason is auditability. A counsellor explaining a recommendation to a teenager, a coordinator presenting cohort results to a school district, a coach defending a career suggestion to a paying client — each of them needs to be able to point at the specific evidence that produced a specific number. A learned ranker cannot do that, and asking it to produce post-hoc explanations is known to produce plausible but misleading rationales. A weighted similarity is mechanically explainable: every point of the score has a channel, every channel has a public source, every source has a citable dataset behind it.
The trade-off is real. A well-tuned learned ranker might extract a few additional percentage points of predictive accuracy from interaction effects we are not modeling. We have judged that gain less valuable than the cost of becoming a black box.
What the score does and does not predict
It does predictpopulation-level fit between a person's assessed profile and an occupation's typical-incumbent profile. High scores correlate with the outcomes the underlying literature predicts — interest congruence with vocational persistence and satisfaction (Nye et al., 2017), trait fit with longer-run engagement, cognitive fit with performance on complex tasks (Schmidt & Hunter, 1998).
It does not predictindividual outcomes. It does not say this person will succeed in this job. It does not control for the market timing of the role, the quality of the specific employer, the person's motivation, geographic constraints, or the dozen other factors that determine whether a career goes well in practice. We treat the score as a structured starting point for a conversation — with a counsellor, a coach, a parent, the user's own future — not as a forecast.
How institutional deployments can adjust
In institutional deploymentsthe coordinator can override the default channel weights for their cohort — for example, raising the Skills weight in a workforce-development programme where placement is the explicit goal, or raising the Interests weight in a high-school career-orientation programme where long-run fit matters more than immediate placement. Overrides are logged in the cohort's audit trail.
Limitations
The weighted-similarity approach inherits two known limitations from the underlying literature. First, RIASEC's six-theme structure has been challenged as imposing a discrete geometry on what is really a continuous interest space (Tracey & Rounds, 1996). We use RIASEC because the prediction literature is robust, but acknowledge the structural critique. Second, all four channels assume the user's self-report and assessed scores are honest and unbiased; we apply standard psychometric consistency checks but cannot fully eliminate response bias.
Citations
- Barrick, M. R., & Mount, M. K. (1991). The Big Five personality dimensions and job performance: A meta-analysis. Personnel Psychology, 44(1), 1–26. link
- Schmidt, F. L., & Hunter, J. E. (1998). The validity and utility of selection methods in personnel psychology. Psychological Bulletin, 124(2), 262–274. link
- Holland, J. L. (1997). Making vocational choices: A theory of vocational personalities and work environments (3rd ed.). Psychological Assessment Resources. link
- Nye, C. D., Su, R., Rounds, J., & Drasgow, F. (2017). Interest congruence and performance: Revisiting recent meta-analytic findings. Journal of Vocational Behavior, 98, 138–151. link
- Tracey, T. J. G., & Rounds, J. (1996). The arbitrary nature of Holland's RIASEC types: A concentric-circles structure. Journal of Counseling Psychology, 43(4), 431–439. link
Four signal channels, each weighted and combined transparently. (1) Big Five trait profile (Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism), weighted against O*NET Work Styles which are factor-aligned with Big Five. (2) RIASEC interest profile (Holland Codes), weighted against O*NET Interests — direct one-to-one mapping. (3) Cognitive reasoning composite, weighted against O*NET Abilities (verbal, quantitative, spatial). (4) Self-reported and assessed skill profile, weighted against the skills sub-graph of the knowledge graph. Each channel contributes a normalized similarity score; the four are combined into a single 0–100 Career Fit Score with disclosed channel-level breakdowns.
Transparency and auditability. A learned ranking model (gradient boosting, neural similarity, learned embeddings) can produce excellent results, but it produces them as a number you cannot explain. A weighted similarity model can: every percentage of the Fit Score has a source channel, every channel has an evidence base, every evidence base traces to a public dataset. For a product that families and counsellors rely on for major decisions, we judge auditability more valuable than the marginal accuracy gain of a black-box ranker.
The default weights are Interests 35%, Skills 30%, Traits 20%, Cognitive 15%. Two reasons for that order. First, interest-congruence meta-analyses (Nye et al., 2017) show RIASEC congruence is one of the strongest predictors of vocational persistence and satisfaction — the outcomes career advisors care most about. Second, skills weight matters more for placement decisions (employers screen on skills) while interests weight matters more for long-run fit. We disclose the weights on every result page and let coordinators in B2B deployments override them for their cohort.
The score is a calibrated percentile within the universe of 2,536 careers, not an absolute probability of success. A 73 means: combining your interests, skills, traits, and cognitive profile, this career sits at the 73rd percentile of fit relative to the full career universe. Two careers can both score 73 by getting there via different channels — one through interest congruence with low skill match, the other through high skill match with weaker interest fit. The channel breakdown matters more than the headline number.
Two reasons. First, our user is a person choosing or changing a career, not a recruiter staffing a requisition. We optimize for long-run fit, not next-week placement. Second, posting matching is well-served by other tools (LinkedIn, Indeed, ATS engines). We integrate posting evidence at the structural level — to detect emerging skills the knowledge graph should know about — but we do not match individuals to specific job listings.
Carefully. The Big Five and Cognitive channels are anchored in literature that predicts job performance at corrected validity coefficients of ρ ≈ 0.30–0.50 depending on job complexity (Schmidt & Hunter, 1998; Barrick & Mount, 1991). The RIASEC channel predicts persistence and satisfaction more strongly than performance per se. A high Career Fit Score is a meaningful signal of population-level compatibility. It is not a forecast that a specific person will succeed in a specific job — that depends on motivation, market timing, and many other factors the model does not see.
Yes. Every career page shows the channel-level contribution: how much of your Fit Score came from interests, skills, traits, and cognitive. For each channel, you can drill in to the specific O*NET Work Style, RIASEC theme, or skill that drove the contribution. This is the audit trail that the weighted-similarity model affords.