[Python] Why is my ALS model suggesting such a high alpha value?

I implemented a recommender system using Ben Fred’s implicit library.

Preprocessing steps I applied:

1. Removed all users with 200+ interactions per day (likely bots or employees).
2. Removed all users with fewer than 2 interactions.
3. Removed all products with fewer than 3 interactions.
4. Assigned different weights per interaction type:
   INTERACTION_WEIGHTS = {
   'purchase': 1.0,
   'initiate_checkout': 0.7,
   'add_to_cart': 0.5,
   'add_to_wishlist': 0.3,
   'content_view': 0.1
   }
5. Applied time-decay weighting so that recent interactions count more.
6. Used different decay scales for each interaction type.

Dataset stats:

• Number of users: 97,793
• Number of products: 32,943
• Number of non-zero interactions: 366,931
• Density: 0.00011390

Quantiles of weights:

• 0%: 0.0000
• 25%: 0.0000
• 50%: 0.0000
• 75%: 0.0027
• 100%: 13.5251

Training loss:

0.000003
...
0.000002


Interaction distribution:

Model configuration:


Default ALS hyperparams:

factors=100
regularization=0.01
alpha=1.0


My "tuned" model:

model = implicit.als.AlternatingLeastSquares(
factors=550,
regularization=0.01,
alpha=2.0,
iterations=25,
calculate_training_loss=True
)

The issue


When running hyperparameter tuning with my scoring method, the best performing alpha comes out around 140.0.

My concern:

• This seems extremely high compared to the default alpha=1.0.
• It makes me think something might be wrong with my preprocessing, weighting, or evaluation setup.

Evaluation metric: I use a weighted scoring function that combines several ranking metrics:

score = (
0.3 * precision_10 +
0.25 * precision_5 +
0.2 * auc_10 +
0.15 * map_10 +
0.1 * ndcg_10
)


Question:

• Is it normal for alpha to need to be set so high in an implicit ALS setup?
• Could this be an artifact of my interaction weighting or density?
• Or does it indicate a problem in how I’ve structured my dataset or evaluation pipeline?

Continue reading...