conceptinterest-graph-recs
On-device recommender
Notes
On-device recommender
One-line summary: Rec serving that runs inference (and sometimes training) on the user’s device with locally stored preference data, versus a fully cloud-hosted ranker.
The insight
The WWW 2024 ODRS tutorial treats traditional recommenders as cloud-trained and cloud-served, then names on-device systems as a response to energy, network, and privacy costs. RecGPT-Mobile (Taobao) is a split: local intent from a behavior store, remote item retrieval. That split is the closest fetched pattern to “user-owned graph + external catalog.” Neither source is YouTube or a hand-drawn graph. From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app.
Evidence
- From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app: Cloud RS — “the cloud server trains the RS model with all the user data it hosts and pushes recommendation results to users’ personal devices upon request.”
- From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app: ODRS — “inference can be efficiently performed on resource-constrained user devices with locally stored user data.”
- From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app: Industry instances listed: “Taobao’s mobile service, Google’s TensorFlow Lite Recommendation API, real-time short video recommendation on Kuaishou, and the built-in recommendation engine in the Brave Web Browser.”
- From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app: RecGPT-Mobile — “an Item Retrieval Module executes a search based on the generated intent query, and returns a set of interested items back to the client side.”
- From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app: Related work they cite: EdgeRec “ranking models directly on mobile devices”; DIR “re-ranking framework on mobile devices.”
Contradictions / tensions
- On-device ranking needs a local candidate cache. RecGPT-Mobile keeps catalog retrieval remote because the item set is huge. YouTube’s catalog is similarly remote and quota-bound (youtube-data-access-constraints).
- Federated recommendation is a subset of on-device training, not the whole ODRS paradigm (ODRS tutorial).
Open questions
Related
Referenced by