How matching works
We first separate your positive research interests from recognized instructions, such as topics to exclude. The positive research text is compared with indexed academic evidence, including publications, eligible CC-BY/CC0 abstract terms, supervised work, research presentations, biographies and advertised PhD projects. TF-IDF gives more weight to distinctive words and adjacent word pairs; cosine similarity measures their weighted overlap.
Instructions and exclusions
For example, “Exclude civil engineering” makes civil engineering an exclusion, rather than an interest that earns a higher score. Matching evidence is removed before scoring. A professor can still qualify through other relevant work.
How search instructions are interpreted
The current parser recognizes separate English instruction clauses beginning with Exclude, Avoid, Do not include, Not interested in, Prefer or Prioritize. It does not understand every form of negation. Ambiguous wording, such as “unrelated applications of AI”, produces a warning rather than an assumed filter.
Exclusions use indexed words and adjacent word pairs, not a semantic topic classifier. A phrase that cannot be fully represented in the index is flagged as unenforced. “Structural or civil engineering” is expanded into two exclusions. Check the interpretation shown by the search.
Advanced search lets you replace the extracted exclusions. Clearing an edited exclusion field removes those filters; restoring exclusions from the description uses the original instructions again. Prefer and Prioritize clauses are recorded as discipline preferences and removed from the positive research text, but do not currently add a ranking bonus or change the selected schools.
Known faculty names and labeled identity or supervisor lines are stripped from the matching text to reduce name-based bias; your editable description is preserved. Each professor is scored using up to five distinct qualifying sources. Repeated pages from one presentation and duplicate copies of the same work do not count as independent evidence. Topic fingerprints can then refine the order of these evidence-qualified candidates.
Technical details and formula
For these nonnegative TF-IDF vectors, each source's cosine similarity ranges from 0 (no weighted overlap) to 1 (identical direction in the representation). This measures textual overlap, not identical research interests.
A source must pass a minimum cosine similarity of 0.025, shared-word checks and topic-specificity checks before it contributes. The source weights below are unchanged by instruction parsing.
Evidence score B = S1 + 0.35S2 + 0.15S3 + 0.05S4 + 0.05S5
S1 to S5 are the strongest qualifying source scores, ordered after source adjustments. Advertised PhD project evidence receives a 1.25 multiplier, publication evidence 1.0, and supervised-work evidence 0.8. Availability notes add no score. Project descriptions reflect the linked source at its capture date; confirm current availability with the professor. Missing sources contribute zero. These are fixed heuristic weights, not weights learned from student-supervisor outcomes.
Final research score = B × (1 + 0.5F)
F is a guarded similarity between your positive research text and the professor's cleaned topic fingerprints. Core topic terms form the main fingerprint; supporting terms contribute 15% of its similarity when present. At least two distinctive overlapping words and support in the leading evidence for detected application contexts are required for a bonus. Fingerprints cannot introduce a candidate who fails the evidence checks.
The 0.5 coefficient is a bonus strength, not a 50/50 blend: F = 0.4 increases B by 20%. If no fingerprint bonus qualifies, F = 0 and the evidence score is retained. A coverage safeguard also suppresses the bonus in close comparisons with a coauthor who shares leading evidence but lacks fingerprints. This reduces the effect of uneven keyword coverage.
The combined score can exceed 1; it is not a normalized cosine similarity. Cards display three decimal places, so small differences can be hidden by rounding. These fixed heuristic weights have not been calibrated against human relevance assessments. Neural embeddings are not used in the live ranking.
What does 0.145 mean?
A research match score of 0.145 combines the qualifying evidence score with any guarded fingerprint bonus. It does not mean 14.5% similarity, a 14.5% chance of a suitable supervisor, or a measure of the professor's qualifications.
Use it to compare results for the same research description and school selection. A score of 0.145 ranks above 0.110 for that search, but does not establish that the professor is a better supervisor. There is no validated cutoff for a good match, and scores should not be compared across different queries.