def holdout_eval(recommender, train, test, at=10):
# train the recommender
logger.info('Recommender: {}'.format(recommender))
tic = dt.now()
logger.info('Training started')
print(train.sum())
recommender.fit(train)
logger.info('Training completed in {}'.format(dt.now() - tic))
# evaluate the ranking quality
roc_auc_, precision_, recall_, map_, mrr_, ndcg_ = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
n_eval = 0
nusers = train.shape[0]
for test_user in range(nusers):
user_profile = train[test_user]
relevant_items = test[test_user].indices
if len(relevant_items) > 0:
n_eval += 1
# this will rank **all** items
recommended_items = recommender.recommend(user_id=test_user,
exclude_seen=True)
# evaluate the recommendation list with ranking metrics ONLY
roc_auc_ += roc_auc(recommended_items, relevant_items)
precision_ += precision(recommended_items, relevant_items, at=at)
recall_ += recall(recommended_items, relevant_items, at=at)
map_ += map(recommended_items, relevant_items, at=at)
mrr_ += rr(recommended_items, relevant_items, at=at)
ndcg_ += ndcg(recommended_items,
relevant_items,
relevance=test[test_user].data,
at=at)
roc_auc_ /= n_eval
precision_ /= n_eval
recall_ /= n_eval
map_ /= n_eval
mrr_ /= n_eval
ndcg_ /= n_eval
return roc_auc_, precision_, recall_, map_, mrr_, ndcg_
recommended_items = recommender.recommend(user_id=test_user, exclude_seen=True)
if args.prediction_file:
# write the recommendation list to file, one user per line
# TODO: convert user and item indices back to their original ids
user_id = test_user
rec_list = recommended_items[:args.rec_length]
s = str(user_id) + ','
s += ','.join([str(x) for x in rec_list]) + '\n'
pfile.write(s)
# evaluate the recommendation list with ranking metrics ONLY
roc_auc_ += roc_auc(recommended_items, relevant_items)
precision_ += precision(recommended_items, relevant_items, at=at)
recall_ += recall(recommended_items, relevant_items, at=at)
map_ += map(recommended_items, relevant_items, at=at)
mrr_ += rr(recommended_items, relevant_items, at=at)
ndcg_ += ndcg(recommended_items, relevant_items, relevance=test[test_user].data, at=at)
roc_auc_ /= n_eval
precision_ /= n_eval
recall_ /= n_eval
map_ /= n_eval
mrr_ /= n_eval
ndcg_ /= n_eval
# close the prediction file
if args.prediction_file:
pfile.close()
logger.info('Recommendations written to {}'.format(args.prediction_file))
logger.info('Ranking quality')
for test_user in range(nusers):
user_profile = train[test_user]
relevant_items = test[test_user].indices
if len(relevant_items) > 0:
n_eval += 1
# this will rank **all** items
recommended_items = recommender.recommend(user_id=test_user,
exclude_seen=True)
# evaluate the recommendation list with ranking metrics ONLY
roc_auc_[nfold] += roc_auc(recommended_items, relevant_items)
precision_[nfold] += precision(recommended_items,
relevant_items,
at=at)
recall_[nfold] += recall(recommended_items, relevant_items, at=at)
map_[nfold] += map(recommended_items, relevant_items, at=at)
mrr_[nfold] += rr(recommended_items, relevant_items, at=at)
ndcg_[nfold] += ndcg(recommended_items,
relevant_items,
relevance=test[test_user].data,
at=at)
roc_auc_[nfold] /= n_eval
precision_[nfold] /= n_eval
recall_[nfold] /= n_eval
map_[nfold] /= n_eval
mrr_[nfold] /= n_eval
ndcg_[nfold] /= n_eval
nfold += 1
logger.info('Ranking quality')
