Explicit Item KNN on ML100K

This page analyzes the hyperparameter tuning results for the Item KNN explicit-feedback rating predictor.

Parameter Search Space

Parameter	Type	Distribution	Values
max_nbrs	Integer	Uniform	2 ≤ \(x\) ≤ 50
min_nbrs	Integer	Uniform	1 ≤ \(x\) ≤ 5
min_sim	Float	LogUniform	1e-06 ≤ \(x\) ≤ 0.1

Final Result

Searching selected the following configuration:

{'max_nbrs': 16, 'min_nbrs': 1, 'min_sim': 0.0008994464114952931}

With these metrics:

{
    'RBP': 0.009393756061435993,
    'NDCG': 0.18756120317017017,
    'RecipRank': 0.04171114352752735,
    'RMSE': 0.8502323823394599,
    'TrainTask': '9b2da9c5-f167-4bfa-9718-7c444eec6972',
    'TrainTime': 0.03031399999963469,
    'TrainCPU': 0.030417999999999723,
    'TestTask': '7c606f9c-91b8-4cc8-96ba-8713ada5bb46',
    'TestTime': 0.9276122160008526,
    'TestCPU': 1.0888609999999999,
    'timestamp': 1743029692,
    'checkpoint_dir_name': None,
    'done': True,
    'training_iteration': 1,
    'trial_id': '46f0c_00051',
    'date': '2025-03-26_18-54-52',
    'time_this_iter_s': 1.1398043632507324,
    'time_total_s': 1.1398043632507324,
    'pid': 298907,
    'hostname': 'CCI-ws21',
    'node_ip': '10.248.127.152',
    'config': {'max_nbrs': 16, 'min_nbrs': 1, 'min_sim': 0.0008994464114952931},
    'time_since_restore': 1.1398043632507324,
    'iterations_since_restore': 1,
    'experiment_tag': '51_max_nbrs=16,min_nbrs=1,min_sim=0.0009'
}

Parameter Influence

Neighborhood Size

The max neighbors is one of the more influential properties. Note that this chart may be noisier than typical, because we are plotting the influence from random search, so the other parameters are simultaneously changing.

RMSE

RBP

Minimum Neighbors

We also examine the minimum neighbors, which reduces the model’s willingness to make low-confidence predictions.

Minimum Similarity

Finally, the minimum similarity filters out low-information relationships: