In plain words
This paper introduces EASformer, an architecture for ranking apps for users. Past app interactions include what the user was doing and what they chose; a new candidate has context but no choice yet.
EASformer compares matching context and keeps past actions separate, using them alongside the user's current state to predict a response.
On one marketplace's offline test, EASformer ranked positive examples above negative ones better than the production model. Online tests also found higher eCPM in two scenarios during one evaluation period.
The research
In app ranking, can a model use past choices to assess a new candidate when past interactions contain both context and action, but the candidate has context without an observed action?
The paper asks whether making the comparable information explicit can support more meaningful history-to-candidate comparisons.
EASformer aligns comparable context fields and separates actions, then combines retrieved history with current user state to predict candidate responses.
Under the shared offline data split and input settings, EASformer scored 0.89411 AUC, a measure of how well a model ranks positive examples above negative ones. That was a 0.88% relative improvement over production HyFormer, supporting an advantage in this marketplace test.
The authors interpret the result as support for aligned, role-specific interaction. Ablations also favored the full design, while online tests found higher eCPM than HyFormer in two scenarios.
The evidence covers one anonymized app marketplace offline test and two online ranking scenarios during one evaluation period. The comparisons do not establish performance in other marketplaces, scenarios, or periods.
Analysis
Deep
- Original paper authors
- Bohao Li, Dinglun Cao, Pengcheng Huang, Yingda Tang, Zhiting Zhao
- Submitted to arXiv
- License
- arXiv non-exclusive