Personalised Outfit Recommendation via History-Aware Transformers (Amazon Science, WSDM 2025)
Overview
This forum entry indexes an Amazon Science publication:
- Title: Personalised outfit recommendation via history-aware transformers
- Venue: WSDM 2025 (ACM International Conference on Web Search and Data Mining)
- Source: https://www.amazon.science/publications/personalised-outfit-recommendation-via-history-aware-transformers
- Category: Recommender Engines, Verticals (fashion/e-commerce)
- Modelling item compatibility (style, color, category coordination), not just individual item relevance;
- Capturing user preference history over time;
- Producing coherent sets rather than ranked item lists.
- 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 (RecSys)
- 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)
- Outfit recommendation differs from standard item ranking: set coherence and compatibility matter as much as individual relevance.
- History-aware modelling lets the system condition on long-term user behavior rather than a single session.
- As with all industrial recommender research, offline metrics should be validated against online user-satisfaction signals before deployment conclusions are drawn.
The work tackles outfit recommendation in fashion: instead of recommending items one at a time, the system composes compatible sets of items into complete outfits, personalised to each user. The proposed approach uses a history-aware transformer, incorporating a user's historical interactions and context into the transformer-based modelling of outfit compatibility and personal preference.
Context within recommender systems research
Outfit recommendation is a challenging vertical of recommender systems because it requires:
Transformer architectures have become the dominant backbone for sequential recommendation, and this work applies a history-aware variant to the outfit-composition task. Specific architectural details, datasets, and experimental metrics are available in the original publication.
Note on this entry
This post is a catalog entry in a curated awesome-list of search/recommendation research. It does not reproduce the paper's quantitative results. Readers should consult the original Amazon Science page and the WSDM 2025 proceedings for the abstract, method details, and benchmark tables.