Table 4.
Comparison of the NDCG at 10 value and overall model ranking for NetflixPrize.
| Carousel layout | ||||||
|---|---|---|---|---|---|---|
| Individual | TopPop | TopPop | ||||
| ItemKNN CF | ||||||
| NDCG | Rank | NDCG | Δ Rank | NDCG | Δ Rank | |
| TopPopular | 0.0799 | 13 | 0.0678 | - | 0.1261 | - |
| ItemKNN CF | 0.2060 | 10 | 0.1335 | 0 | 0.1261 | - |
| GlobalEffects | 0.0159 | 14 | 0.0733 | 0 | 0.1282 | 0 |
| UserKNN CF | 0.2581 | 3 | 0.1516 | 0 | 0.1573 | 0 |
| P3α | 0.1810 | 11 | 0.1194 | 0 | 0.1478 | 1 |
| RP3β | 0.2209 | 7 | 0.1345 | -2 | 0.1422 | -4 |
| IALS | 0.2380 | 5 | 0.1497 | 0 | 0.1566 | 1 |
| MF BPR | 0.1656 | 12 | 0.1172 | 0 | 0.1426 | 1 |
| MF FunkSVD | 0.2077 | 9 | 0.1446 | 2 | 0.1639 | 7 |
| PureSVD | 0.2508 | 4 | 0.1515 | 0 | 0.1552 | -2 |
| NMF | 0.2192 | 8 | 0.1434 | 0 | 0.1546 | 1 |
| EASER | 0.2619 | 2 | 0.1520 | 0 | 0.1534 | -6 |
| SLIM ElasticNet | 0.2913 | 1 | 0.1662 | 0 | 0.1669 | 0 |
| SLIM BPR | 0.2353 | 6 | 0.1467 | 0 | 0.1561 | 1 |
Each model is evaluated both individually (single-carousel) and as the last recommendation list in a multi-carousel interface of increasing complexity. The NDCG is computed with the single-list discount (concatenating all carousel lists). Higher ranks indicate better recommendation quality. The rank of models that are already used as carousels is removed. ΔRank is the difference between the rank when evaluated individually and the rank when evaluated in the corresponding carousel layout, e.g., a negative ΔRank indicates the model is in a worse ranking position.