UKNN Implicit

This page analyzes the hyperparameter tuning results for the ItemKNN scorer in implicit-feedback mode.

Parameter Search Space

Parameter Type Distribution Values Selected
max_nbrs Integer Uniform 2 ≤ \(x\) ≤ 50 48
min_nbrs Integer Uniform 1 ≤ \(x\) ≤ 5 1
min_sim Float LogUniform 1e-06 ≤ \(x\) ≤ 0.1 0.00138

Final Result

Searching selected the following configuration:

{'max_nbrs': 48, 'min_nbrs': 1, 'min_sim': 0.001379808386693685}

With these metrics:

{
    'RBP': 0.10932398176010141,
    'NDCG': 0.36032996267360856,
    'RecipRank': 0.3664214976179329,
    'TrainTask': 'b6a1ab6f-a108-470b-a8d6-208aae2d3f10',
    'TrainTime': 0.01956150800015166,
    'TrainCPU': 0.025275999999999632,
    'TestTask': '88af01f4-dca4-4e69-8ff7-802cdb4806b3',
    'TestTime': 56.197305957000026,
    'TestCPU': 57.993985,
    'timestamp': 1742697946,
    'checkpoint_dir_name': None,
    'done': True,
    'training_iteration': 1,
    'trial_id': '81f6e_00038',
    'date': '2025-03-22_22-45-46',
    'time_this_iter_s': 56.56378650665283,
    'time_total_s': 56.56378650665283,
    'pid': 192935,
    'hostname': 'CCI-ws21',
    'node_ip': '10.248.127.152',
    'config': {'max_nbrs': 48, 'min_nbrs': 1, 'min_sim': 0.001379808386693685},
    'time_since_restore': 56.56378650665283,
    'iterations_since_restore': 1,
    'experiment_tag': '38_max_nbrs=48,min_nbrs=1,min_sim=0.0014'
}

Parameter Behavior

Neighborhood Size

The neighborhood size is key argument to consider, let’s look at RMSE with repspect to that:

Filtering

There are two filtering parameters we also test — minimum neighbors and mininmum simialrity.