Precision@10
15.1%
Relevant items / K
Hybrid model · 100 eligible test users
Measured, not made up
The intelligence, evaluated.
Actual model metrics. Chronological holdout. A transparent look at what’s working.
In-memory seed dataset · storage not connected
100 sample users · 100 products · 2,400 interactions. Synthetic data; these are not live customer statistics.
Recall@10
75.5%
Relevant items found / all relevant items
Hybrid model · 100 eligible test users
NDCG@10
41.2%
Rank-aware relevance / ideal ranking
Hybrid model · 100 eligible test users
Which model finds the better fit?
Computed comparison across the same time-based holdout
- Precision
- Recall
- NDCG
Time matters.
No peeking ahead.
Interaction data is sorted by timestamp. Models are trained only on the first 80%, then evaluated on the remaining 20%.
1,920 training events480 test events
Cutoff: February 27, 2025
Relevance: held-out wishlist, cart, purchase.
Excludes products already purchased in training.
Relevance: held-out wishlist, cart, purchase.
Excludes products already purchased in training.
Signals over time
Actual event counts in the selected dataset
- Interactions
- Purchases
| Model | Precision@10 | Recall@10 | NDCG@10 |
|---|---|---|---|
| Hybrid50 / 30 / 20 | 0.1510 | 0.7550 | 0.4117 |
| Collaborative | 0.1250 | 0.6250 | 0.3987 |
| Content-based | 0.1000 | 0.5000 | 0.2874 |
| Popularity | 0.1050 | 0.5250 | 0.4233 |
Educational results on synthetic seed data, not production accuracy. Metrics are recalculated when K or recorded activity changes.
Methodology