Algorithm Transparency Registry

Last updated June 25, 2026

This page describes the recommendation, matching, and scoring algorithms ZhiYin uses, in line with China's Provisions on the Administration of Algorithm-Generated Recommendations and the personal information protection requirements for automated decision-making. For each algorithm we explain its purpose, how it broadly works, the data it uses, and how long results are kept.

Scope of this disclosure

We list the algorithms that influence what content and candidates you see and the scores attached to them. AI outputs on ZhiYin are advisory only — every hiring decision is made by a human employer, and candidates always see the full job listing rather than an algorithm-filtered subset.

Job–candidate matching

Purpose: Suggest relevant job openings to candidates, and relevant candidates to employers, based on professional fit.
Basic operating principle: A deterministic, fixed-weight, two-sided formula. A candidate-fit score combines skills (40%, matched through a bilingual English–Chinese skill lexicon), experience (20%), verified evidence such as mock-interview and proctored-assessment results weighted by their published predictive validity (25%), and recent activity (15%). A job-appeal score combines salary fit (30%, compared only when candidate and job use the same currency), location (25%), employer quality — reply rate, offer-acceptance rate, and candidate satisfaction (25%) — salary transparency (10%), and job freshness (10%). The two sides are combined by harmonic mean, so a match that is good for only one side ranks low, and may be blended with multilingual text-embedding similarity (60% formula / 40% similarity). Recommendation feeds also apply aggregate-level balancing: jobs already flooded with applications are demoted up to 12% and under-exposed jobs boosted up to 6%. Every score includes a per-factor breakdown and the matched and missing skills. No large language model is used in scoring, and protected characteristics and hukou are never ranking factors.
Data inputs: From the candidate profile: skills, years of experience, verified evidence scores (mock interviews, proctored assessments, courses, resume analysis), recent activity, desired salary and currency, preferred locations, and remote preference. From the job posting: required skills, experience range, salary range, currency, location, salary transparency, and posting age. Aggregate employer-quality metrics (reply rate, offer-acceptance rate, candidate satisfaction) and aggregate application counts per job; optionally, text embeddings of the profile summary and job description.
Retention: Match scores are recomputed as profiles and jobs change and are not stored as a permanent ranking; underlying profile and job data follow the retention periods in our personal information list.

Candidate–job fit scoring

Purpose: Give employers an at-a-glance indication of how well a specific candidate fits a specific role, shown alongside the full profile.
Basic operating principle: A rule-based model compares the candidate's profile against the role's requirements across skill match, experience, education, location, and salary fit, then combines these into an overall score with short explanatory notes. The score informs, but does not make, the hiring decision.
Data inputs: Candidate skills, experience, education, location, and salary expectations; job requirements and compensation range.
Retention: Computed on demand for an employer viewing a candidate and not retained as a standalone profile attribute.

Talent readiness ranking

Purpose: Help candidates understand their overall job-readiness and surface candidates who have actively engaged with preparation and assessments.
Basic operating principle: A composite 0–100 score blends five signals with published weights: resume quality (20%), interview-preparation engagement (25%), completed courses and learning paths (15%), skill-assessment results (20%), and mock-interview averages (20%). It reflects activity on the platform and is recalculated as new data arrives.
Data inputs: Resume analysis results, interview-prep activity, course and learning-path completion, skill-assessment scores, and mock-interview scores.
Retention: Recalculated from your latest activity; component scores are derived from records kept per our personal information list and removed when those records are deleted.

AI resume analysis

Purpose: Help candidates improve their resume and help employers assess fit by extracting and rating skills, experience, and qualifications.
Basic operating principle: A large language model reads the resume text (and, when provided, a target job description), identifies skills and experience, rates them, and writes improvement suggestions. It produces guidance and ratings, not an automatic accept-or-reject outcome.
Data inputs: Resume text supplied by the candidate and, optionally, a target job description.
Retention: Analysis results are stored with your resume and removed when you delete the resume or your account.

Your right to opt out

You can choose not to be subject to personalized recommendation or automated decision-making. Opting out does not affect your account status or your ability to browse and apply to jobs manually.

Opt out of personalized recommendation. Turn off personalized recommendation so your profile is excluded from algorithmic matching and you browse jobs without algorithm-tailored suggestions.
Object to automated decision-making. Ask that decisions affecting you are not based solely on automated processing, and request human review where a decision has a significant effect on you.
Request an explanation. Ask us to explain in clear terms the main factors behind any score or recommendation that affects you.

Human oversight

AI scores and recommendations on ZhiYin are advisory. Employers make all hiring decisions independently and see full candidate profiles, not algorithm-filtered shortlists. We review algorithm outputs regularly and run periodic bias checks across our matching and scoring algorithms.

Exercising your rights

To opt out, object to automated decision-making, or request an explanation, contact our Data Protection Officer at privacy@talentbridge.com

Job–candidate matching

Purpose

Suggest relevant job openings to candidates, and relevant candidates to employers, based on professional fit.

Basic operating principle

A deterministic, fixed-weight, two-sided formula. A candidate-fit score combines skills (40%, matched through a bilingual English–Chinese skill lexicon), experience (20%), verified evidence such as mock-interview and proctored-assessment results weighted by their published predictive validity (25%), and recent activity (15%). A job-appeal score combines salary fit (30%, compared only when candidate and job use the same currency), location (25%), employer quality — reply rate, offer-acceptance rate, and candidate satisfaction (25%) — salary transparency (10%), and job freshness (10%). The two sides are combined by harmonic mean, so a match that is good for only one side ranks low, and may be blended with multilingual text-embedding similarity (60% formula / 40% similarity). Recommendation feeds also apply aggregate-level balancing: jobs already flooded with applications are demoted up to 12% and under-exposed jobs boosted up to 6%. Every score includes a per-factor breakdown and the matched and missing skills. No large language model is used in scoring, and protected characteristics and hukou are never ranking factors.

Data inputs

From the candidate profile: skills, years of experience, verified evidence scores (mock interviews, proctored assessments, courses, resume analysis), recent activity, desired salary and currency, preferred locations, and remote preference. From the job posting: required skills, experience range, salary range, currency, location, salary transparency, and posting age. Aggregate employer-quality metrics (reply rate, offer-acceptance rate, candidate satisfaction) and aggregate application counts per job; optionally, text embeddings of the profile summary and job description.

Retention

Match scores are recomputed as profiles and jobs change and are not stored as a permanent ranking; underlying profile and job data follow the retention periods in our personal information list.

Candidate–job fit scoring

Purpose

Give employers an at-a-glance indication of how well a specific candidate fits a specific role, shown alongside the full profile.

Basic operating principle

A rule-based model compares the candidate's profile against the role's requirements across skill match, experience, education, location, and salary fit, then combines these into an overall score with short explanatory notes. The score informs, but does not make, the hiring decision.

Data inputs

Candidate skills, experience, education, location, and salary expectations; job requirements and compensation range.

Retention

Computed on demand for an employer viewing a candidate and not retained as a standalone profile attribute.

Talent readiness ranking

Purpose

Help candidates understand their overall job-readiness and surface candidates who have actively engaged with preparation and assessments.

Basic operating principle

A composite 0–100 score blends five signals with published weights: resume quality (20%), interview-preparation engagement (25%), completed courses and learning paths (15%), skill-assessment results (20%), and mock-interview averages (20%). It reflects activity on the platform and is recalculated as new data arrives.

Data inputs

Resume analysis results, interview-prep activity, course and learning-path completion, skill-assessment scores, and mock-interview scores.

Retention

Recalculated from your latest activity; component scores are derived from records kept per our personal information list and removed when those records are deleted.

AI resume analysis

Purpose

Help candidates improve their resume and help employers assess fit by extracting and rating skills, experience, and qualifications.

Basic operating principle

A large language model reads the resume text (and, when provided, a target job description), identifies skills and experience, rates them, and writes improvement suggestions. It produces guidance and ratings, not an automatic accept-or-reject outcome.

Data inputs

Resume text supplied by the candidate and, optionally, a target job description.

Retention

Analysis results are stored with your resume and removed when you delete the resume or your account.

Your right to opt out

You can choose not to be subject to personalized recommendation or automated decision-making. Opting out does not affect your account status or your ability to browse and apply to jobs manually.

Opt out of personalized recommendation. Turn off personalized recommendation so your profile is excluded from algorithmic matching and you browse jobs without algorithm-tailored suggestions.

Object to automated decision-making. Ask that decisions affecting you are not based solely on automated processing, and request human review where a decision has a significant effect on you.

Request an explanation. Ask us to explain in clear terms the main factors behind any score or recommendation that affects you.