Mosaic Cdsr Eval

---
name: mosaic-cdsr-eval
description: Evaluates a model's ability to perform next-item recommendation in cross-domain sequential settings by decomposing user intent into orthogonal preference components. It probes how well the model leverages shared and domain-specific signals across multiple item categories to predict future interactions. Use when the user wants to benchmark on Amazon Reviews (Movie–Book), Amazon Reviews (Movie–Music), Douban (Movie–Book), or asks about evaluating this task. Reports NDCG@10.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.10147
  bibtex_key: bahi2026mosaic
  confidence: high
---

# mosaic-cdsr-eval

> MOSAIC: Multi-Domain Orthogonal Session Adaptive Intent Capture for Prescient Recommendations — Bahi et al. (2026) (arXiv:2604.10147, 2026)

## What this evaluates

Evaluates a model's ability to perform next-item recommendation in cross-domain sequential settings by decomposing user intent into orthogonal preference components. It probes how well the model leverages shared and domain-specific signals across multiple item categories to predict future interactions.

## Datasets

- **Amazon Reviews (Movie–Book)** — total 2625327; splits: train (-1), val (-1), test (-1)
- **Amazon Reviews (Movie–Music)** — total 1138302; splits: train (-1), val (-1), test (-1)
- **Douban (Movie–Book)** — total 2070463; splits: train (-1), val (-1), test (-1)

## Metrics

- `HR@10` — range: [0, 1]
  - Hit Ratio at rank K: returns 1.0 if the ground-truth item appears in the top-K ranked list, else 0.0. Computed over the full item vocabulary.
- `NDCG@10` **(primary)** — range: [0, 1]
  - Normalized Discounted Cumulative Gain at rank K: (1/Z) * Σ_{i=1}^K (rel_i / log2(i+1)), where rel_i=1 if the item at rank i is the ground truth, else 0. Z is the ideal DCG (1.0 for a single relevant item). Rewards higher placements of the correct item.

## Input / output format

**Input**: Chronologically sorted user interaction history (per-domain and merged cross-domain sequences) up to timestep t-1.

**Output**: A ranked list of items from the entire item vocabulary, with the ground-truth next item expected at rank 1.

## Scoring recipe

```python
def compute_metrics(ranked_list, gold_item, K=10):
    rank = ranked_list.index(gold_item) + 1
    hr = 1.0 if rank <= K else 0.0
    ndcg = 0.0
    for i in range(1, K + 1):
        if i == rank:
            ndcg += 1.0 / math.log2(i + 1)
    ndcg /= 1.0  # IDCG for single relevant item
    return hr, ndcg
```

## Common pitfalls

- Using sampled negative items for ranking instead of the full item vocabulary, which introduces sampling bias and inflates metrics.
- Ignoring the leave-one-out split protocol where the last interaction is strictly held out for testing and the second-to-last for validation.
- Failing to report mean and standard deviation over five independent random seeds with paired t-test significance checks.

## Evidence (verbatim from paper)

> Following standard practice in the sequential recommendation literature, we adopt Hit Ratio at rank $K$ (HR@$K$) and Normalized Discounted Cumulative Gain at rank $K$ (NDCG@$K$), with $K\in{5,10,20}$. HR@$K$ measures whether the ground-truth item appears in the top-$K$ ranked list, while NDCG@$K$ additionally rewards higher placements. All metrics are computed in the full-ranking protocol, i.e., each test item is ranked against the entire item vocabulary, to avoid sampling bias.

## Citation

```bibtex
@misc{bahi2026mosaic,
  title={MOSAIC: Multi-Domain Orthogonal Session Adaptive Intent Capture for Prescient Recommendations},
  author={Bahi et al. (2026)},
  year={2026},
  note={arXiv:2604.10147}
}
```

- arXiv: 2604.10147