ALS BiasedMF on LM1M

This page analyzes the hyperparameter tuning results for biased matrix factorization with ALS.

Parameter Search Space

Parameter	Type	Distribution	Values
embedding_size	Integer	LogUniform	4 ≤ \(x\) ≤ 512
regularization.user	Float	LogUniform	1e-05 ≤ \(x\) ≤ 1
regularization.item	Float	LogUniform	1e-05 ≤ \(x\) ≤ 1
damping.user	Float	LogUniform	1e-12 ≤ \(x\) ≤ 100
damping.item	Float	LogUniform	1e-12 ≤ \(x\) ≤ 100

Final Result

Searching selected the following configuration:

{
    'embedding_size': 463,
    'regularization': {'user': 0.794561339814651, 'item': 0.0016328354077841572},
    'damping': {'user': 9.986183348529642e-11, 'item': 3.1733506385207053e-09},
    'epochs': 6
}

With these metrics:

{
    'RBP': 0.0026159918135242014,
    'NDCG': 0.16254211766932172,
    'RecipRank': 0.019662310846581755,
    'RMSE': 0.8179058539897006,
    'TrainTask': '7901e7b7-5c59-4870-88ef-7caa2d74e031',
    'TrainTime': 35.217326331010554,
    'TrainCPU': 430.26249,
    'max_epochs': 30,
    'timestamp': 1743096093,
    'checkpoint_dir_name': None,
    'done': True,
    'training_iteration': 6,
    'trial_id': '34569_00042',
    'date': '2025-03-27_13-21-33',
    'time_this_iter_s': 5.317098140716553,
    'time_total_s': 36.84841513633728,
    'pid': 732477,
    'hostname': 'CCI-ws21',
    'node_ip': '10.248.127.152',
    'config': {
        'embedding_size': 463,
        'regularization': {'user': 0.794561339814651, 'item': 0.0016328354077841572},
        'damping': {'user': 9.986183348529642e-11, 'item': 3.1733506385207053e-09},
        'epochs': 6
    },
    'time_since_restore': 36.84841513633728,
    'iterations_since_restore': 6,
    'experiment_tag': '42_item=0.0000,user=0.0000,embedding_size=463,item=0.0016,user=0.7946'
}

Parameter Analysis

Embedding Size

The embedding size is the hyperparameter that most affects the model’s fundamental logic, so let’s look at performance as a fufnction of it:

Learning Parameters

User Reg.

User Damp.

Item Reg.

Item Damp.

Iteration Completion

How many iterations, on average, did we complete?

How did the metric progress in the best result?

How did the metric progress in the longest results?