Predict CP2021 codes against static skill profiles

Scores each announcement against the profiles of the CP codes observed for its ESCO group, and returns the best one. No training data is consulted: the whole model is the object returned by build_cp_profiles().

Usage

predict_cp5_centroid(
  profiles,
  postings,
  skills,
  distance = "cosine",
  prior_weight = 1,
  dense_budget = 2e+08,
  explain = FALSE,
  verbose = TRUE
)

Arguments

profiles: A cp_profiles object from build_cp_profiles().
postings: A data.table of announcements to code. Needs general_id and the restrictor column the profiles were built on, plus idsector when the profiles carry a sector term. Every row gets exactly one prediction.
skills: A data.table with general_id and the profiles' skill column.
distance: Character naming the scoring rule, one of "multinomial_nb", "complement_nb", "bernoulli_nb", "hellinger", "cosine", "l2", "l1", "chisq", "jsd", "dice", "tfidf_cosine", "balassa_cosine".
prior_weight: Numeric multiplier on log n_c, the class-size prior (default 1). 0 removes the prior; large values reduce the rule to picking the largest class, which is the modal cascade.
dense_budget: Numeric cap on the elements of a dense query block, used to chunk large groups (default 2e8).
explain: Logical: also return the per-candidate score matrix for each announcement (default FALSE). Costly; intended for diagnosis.
verbose: Logical: print a one-line summary (default TRUE).

Value

A data.table with general_id, the profiles' target column, confidence and method. method is "centroid" where the model scored, "single_candidate" where the group offered only one code, "frequency" where the announcement carries no known skill and the class prior decided alone, and "no_match" where the group is absent from the profiles. When explain = TRUE the result carries a scores attribute.

Details

The rule is $$score(q, c) = \lambda \log n_c + s(q, \theta_c) + \mathrm{sector}$$ with $s$ the similarity implied by distance and $\theta_c$ the smoothed, group-shrunk profile of candidate $c$.

The shrinkage makes the model falsifiable. As shrink_beta grows, every $\theta_c$ in a group collapses onto that group's profile, the similarity term becomes constant across candidates, and the ranking reduces to $\lambda \log n_c$ — which is exactly the modal cascade. So a correct implementation must reproduce the modal baseline in that limit, and skillviz_workflow/run_cp5_centroid.R refuses to report any accuracy until it has checked that it does. Ties are broken on the smaller CP code, matching the cascade.

Announcements carrying no skill present in the profiles get "frequency" rather than "centroid": the prior decided them, not the skill vector. That matches predict_cp5_knn()'s convention, which keeps the two models' model-decided populations comparable.

Usage

Arguments

Value

Details

See also