Personalised Outfit Recommendation via History-Aware Transformers (Amazon Science, WSDM 2025)
This entry indexes an Amazon Science publication presented at WSDM 2025: *Personalised Outfit Recommendation via History-Aware Transformers*.
- Source: Amazon Science publication page
- Venue: WSDM 2025
- Category: Recommender Engines / Verticals
- Classic pipeline: candidate retrieval → ranking → re-ranking → presentation, with recall covering the catalogue, ranking providing discrimination, and generation handling presentation.
- LLM-era additions: retrieval-augmented generation (RAG) and agentic search, where the number and policy of retrieval steps themselves become learnable decisions, and inference budget becomes a first-class constraint.
- In fashion recommendation specifically, core tensions include sparse user behaviour, enormous item catalogues, and multi-objective business goals; LLMs offer semantic priors and cold-start ability, but online inference cost and hallucination risk require careful system design.
- Metrics: nDCG@10, MRR, Recall@k, Hit@k, human preference judgements, task success rate, latency and token cost.
- Baselines: popularity heuristics, matrix factorisation, sequential transformers (e.g., BERT4Rec-style models), and non-personalised compatibility models.
- Ablations: contribution of history length, item representation, and outfit-compatibility modelling.
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers
- Augmenting Netflix Search with In-Session Adapted Recommendations
- Bridging Language and Items for Retrieval and Recommendation
- Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024
- DiffKG: Knowledge Graph Diffusion Model for Recommendation, WSDM 2024
- Original publication: Personalised outfit recommendation via history-aware transformers — Amazon Science, WSDM 2025.
Problem and Motivation
The work targets personalised outfit recommendation: recommending a compatible set of fashion items (an outfit) rather than isolated products. The key challenge is incorporating a user's history — past purchases, browsed items, and interactions — so that recommendations are consistent with the user's taste, existing wardrobe, and body/size constraints, while maintaining compatibility among the recommended items.
The proposed approach uses history-aware transformer architectures, following the broader industry trend of applying sequential transformer models to recommendation, where user behaviour sequences are encoded and used to condition generation or ranking of item sets.
Context Within Modern Rec Systems
The forum post situates the paper in the wider evolution of search and recommendation:
Typical Evaluation Setup
For papers in this area, evaluation generally involves:
> Note: Exact numerical results should be verified against the original PDF on the Amazon Science page; this post is a metadata-based summary, not a reproduction of the paper's tables.
Engineering Checklist for Similar Systems
| Concern | Question | Suggestion | |---------|----------|-----------| | Data | Does training/indexing data contain PII? How are versions managed? | Partitioned indexes, anonymisation, rollback-capable embedding versions | | Latency | What is the p99 budget? How many retrieval steps? | Cascaded retrieval with early stopping, caching, async re-ranking | | Quality | Do offline gains translate to online CTR/satisfaction? | Interleaving experiments, human audits of samples | | Safety | Does open retrieval introduce poisoning or bias? | Source whitelisting, adversarial detection, output filtering | | Cost | Per-query token/GPU cost? | Small-model routing, distillation, hybrid sparse+dense retrieval |
Key Takeaways for Search / Rec / Personalisation
1. Architecture: cascaded retrieval + ranking + generation remains mainstream; agentic paradigms make retrieval strategy itself learnable. 2. Data: high-quality instruction data and click/session logs matter as much as model architecture; synthetic data requires leakage and distribution-shift controls. 3. Evaluation: the gap between offline metrics and online satisfaction is widening; LLM-as-judge should be cross-validated with human evaluation. 4. Product: latency, cost, explainability, and safety are hard industrial constraints, not just academic benchmarks.