Bias Predictor

This page analyzes the behavior and hyperparameter search for the Bias scoring model on ML20M.

Parameter Search Space

Parameter	Type	Distribution	Values	Selected
damping.user	Float	LogUniform	0.1 ≤ \(x\) ≤ 100	51.5
damping.item	Float	LogUniform	0.1 ≤ \(x\) ≤ 100	9.58

Final Result

Searching selected the following configuration:

{'damping': {'user': 51.5088418568768, 'item': 9.581280544381151}}

With these metrics:

{
    'RBP': 0.09675027736979092,
    'NDCG': 0.3639622118919963,
    'RecipRank': 0.27040002494943277,
    'RMSE': 0.7985689660364931,
    'TrainTask': 'f6e5c93f-0409-4e40-ba33-6bcb681132ca',
    'TrainTime': 3.536836585997662,
    'TrainCPU': 3.5487679999999995,
    'TestTask': 'ae7ba272-533f-405e-beb0-2a6bb10f3234',
    'TestTime': 8.976197374002368,
    'TestCPU': 8.990714999999998,
    'timestamp': 1746055279,
    'checkpoint_dir_name': None,
    'done': True,
    'training_iteration': 1,
    'trial_id': 'a92ee974',
    'date': '2025-04-30_19-21-19',
    'time_this_iter_s': 16.861109972000122,
    'time_total_s': 16.861109972000122,
    'pid': 270632,
    'hostname': 'CCI-ws21',
    'node_ip': '10.248.127.152',
    'config': {'damping': {'user': 51.5088418568768, 'item': 9.581280544381151}},
    'time_since_restore': 16.861109972000122,
    'iterations_since_restore': 1,
    'experiment_tag': '100_item=9.5813,user=51.5088'
}

Metric Response

How does RMSE change with each setting independently?

User Damping

Item Damping

Best Configurations

Since this is an explicit-feedback rating prediction model, our primary search criteria is RMSE. The configuration with the best RMSE is:

	config.damping.user	config.damping.item	RBP	RMSE
Method
Random	51.508842	9.581281	0.09675	0.798569

If we instead searched for RBP, we would select:

	config.damping.user	config.damping.item	RBP	RMSE
Method
Random	33.241207	99.639145	0.098751	0.803312

Search Geometry

What is the geometry of the search space?