CoordinateAscent

This is a python implementation of the Coordinate Ascent algorithm (Metzler and Croft, 2007).

The structure of the code follows closely to the scikit-learn style, but still there are some differences in the estimator/metrics API (e.g. fit() method takes three arguments X, y, and qid rather than just two).

Four ranking metrics are implemented: P@k, AP, DCG@k, and nDCG@k (in both trec_eval and Burges et al. versions).

Dependencies

• numpy
• scikit-learn

Usage

The following code will run Coordinate Ascent for 2 random restarts and 25 iterations for line search in each dimension, optimizing for NDCG@10.

from coordinate_ascent import CoordinateAscent
from metrics import NDCGScorer

scorer = NDCGScorer(k=10, idcg_cache={})
model = CoordinateAscent(n_restarts=2, max_iter=25, verbose=True, scorer=scorer).fit(X, y, qid)
pred = model.predict(X_test, qid_test)
print scorer(y_test, pred, qid_test).mean()

Note that, when setting idcg_cache to some dict object, NDCGScorer will cache the ideal DCG score (wrt the qid) to speed up training. Be sure to use separate scorers if there is an overlap between training/test qids.

See test.py for more advanced examples.

References

Metzler and Croft. Linear feature-based models for information retrieval. Information Retrieval, 10(3): 257–274, 2007.