, I’d prefer to share a sensible variation of Uber’s Two-Tower Embedding (TTE) method for instances the place each user-related information and computing assets are restricted. The issue got here from a heavy-traffic discovery widget on the house display screen of a meals supply app. This widget reveals curated picks akin to Italian, Burgers, Sushi, or Wholesome. The picks are created from tags: every restaurant can have a number of tags, and every tile is actually a tag-defined slice of the catalog (with the addition of some handbook choosing). In different phrases, the candidate set is already recognized, so the actual drawback shouldn’t be retrieval however rating.
At the moment this widget was considerably underperforming compared to different widgets on a discovery (essential) display screen. The ultimate choice was ranked on normal recognition with out taking into consideration any customized alerts. What we found is that customers are reluctant to scroll and in the event that they don’t discover one thing attention-grabbing throughout the first 10 to 12 positions then they normally don’t convert. However the picks could be large generally, in some instances as much as 1500 eating places. On prime of {that a} single restaurant could possibly be chosen for various picks, which implies that for instance McDonald’s could be chosen for each Burgers and Ice Cream, however it’s clear that its recognition is simply legitimate for the primary choice, however the normal recognition sorting would put it on prime in each picks.
The product setup makes the issue even much less pleasant to static options akin to normal recognition sorting. These collections are dynamic and alter incessantly because of seasonal campaigns, operational wants, or new enterprise initiatives. Due to that, coaching a devoted mannequin for every particular person choice shouldn’t be lifelike. A helpful recommender has to generalize to new tag-based collections from day one.
Earlier than shifting to a two-tower-style resolution, we tried less complicated approaches akin to localized recognition rating on the city-district stage and multi-armed bandits. In our case, neither delivered a measurable uplift over a normal recognition kind. As part of our analysis initiative we tried to regulate Uber’s TTE for our case.
Two-Tower Embeddings Recap
A two-tower mannequin learns two encoders in parallel: one for the consumer facet and one for the restaurant facet. Every tower produces a vector in a shared latent house, and relevance is estimated from a similarity rating, normally a dot product. The operational benefit is decoupling: restaurant embeddings could be precomputed offline, whereas the consumer embedding is generated on-line at request time. This makes the method engaging for techniques that want quick scoring and reusable representations.
Uber’s write-up centered primarily on retrieval, however it additionally famous that the identical structure can function a ultimate rating layer when candidate technology is already dealt with elsewhere and latency should stay low. That second formulation was a lot nearer to our use case.
Our Strategy
We stored the two-tower construction however simplified probably the most resource-heavy elements. On the restaurant facet, we didn’t fine-tune a language mannequin contained in the recommender. As a substitute, we reused a TinyBERT mannequin that had already been fine-tuned for search within the app and handled it as a frozen semantic encoder. Its textual content embedding was mixed with express restaurant options akin to worth, rankings, and up to date efficiency alerts, plus a small trainable restaurant ID embedding, after which projected into the ultimate restaurant vector. This gave us semantic protection with out paying the complete price of end-to-end language-model coaching. For a POC or MVP, a small frozen sentence-transformer can be an affordable start line as properly.
We prevented studying a devoted user-ID embedding and as a substitute represented every consumer on the fly via their earlier interactions. The consumer vector was constructed from averaged embeddings of eating places the client had ordered from (Uber’s submit talked about this supply as properly, however the authors don’t specify the way it was used), along with consumer and session options. We additionally used views with out orders as a weak damaging sign. That mattered when order historical past was sparse or irrelevant to the present choice. If the mannequin couldn’t clearly infer what the consumer appreciated, it nonetheless helped to know which eating places had already been explored and rejected.
A very powerful modeling selection was filtering that historical past by the tag of the present choice. Averaging the entire order historical past created an excessive amount of noise. If a buyer largely ordered burgers after which opened an Ice Cream choice, a world common may pull the mannequin towards burger locations that occurred to promote desserts reasonably than towards the strongest ice cream candidates. By filtering previous interactions to matching tags earlier than averaging, we made the consumer illustration contextual as a substitute of world. In follow, this was the distinction between modeling long-term style and modeling present intent.
Lastly, we educated the mannequin on the session stage and used multi-task studying. The identical restaurant could possibly be optimistic in a single session and damaging in one other, relying on the consumer’s present intent. The rating head predicted click on, add-to-basket, and order collectively, with a easy funnel constraint: P(order) ≤ P(add-to-basket) ≤ P(click on). This made the mannequin much less static and improved rating high quality in contrast with optimizing a single goal in isolation.
Offline validation was additionally stricter than a random break up: analysis used out-of-time information and customers unseen throughout coaching, which made the setup nearer to manufacturing habits.
Outcomes
In accordance with A/B exams the ultimate system confirmed a statistically important uplift in conversion price. Simply as importantly, it was not tied to 1 widget. As a result of the mannequin scores a consumer–restaurant pair reasonably than a set checklist, it generalized naturally to new picks with out architectural adjustments since tags are a part of restaurant’s metadata and could be retrieved with out picks in thoughts.
That transferability made the mannequin helpful past the unique rating floor. We later reused it in Advertisements, the place its CTR-oriented output was utilized to particular person promoted eating places with optimistic outcomes. The identical illustration studying setup subsequently labored each for choice rating and for different recommendation-like placement issues contained in the app.
Additional Analysis
The obvious subsequent step is multimodality. Restaurant photographs, icons, and probably menu visuals could be added as further branches to the restaurant tower. That issues as a result of click on habits is strongly influenced by presentation. A pizza place inside a pizza choice might underperform if its essential picture doesn’t present pizza, whereas a finances restaurant can look premium purely due to its hero picture. Textual content and tabular options don’t seize that hole properly.
Key Takeaways:
- Two-Tower fashions can work even with restricted information. You don’t want Uber-scale infrastructure if candidate retrieval is already solved and the mannequin focuses solely on the rating stage.
- Reuse pretrained embeddings as a substitute of coaching from scratch. A frozen light-weight language mannequin (e.g., TinyBERT or a small sentence-transformer) can present robust semantic alerts with out costly fine-tuning.
- Averaging embeddings of beforehand ordered eating places works surprisingly properly when consumer historical past is sparse.
- Contextual filtering reduces noise and helps the mannequin seize the consumer’s present intent, not simply long-term style.
- Unfavorable alerts assist in sparse environments. Eating places that customers seen however didn’t order from present helpful info when optimistic alerts are restricted.
- Multi-task studying stabilizes rating. Predicting click on, add-to-basket, and order collectively with funnel constraints produces extra constant scores.
- Design for reuse. A mannequin that scores consumer–restaurant pairs reasonably than particular lists could be reused throughout product surfaces akin to picks, search rating, or advertisements.
