Problem
Etsy sells sponsored listings next to organic search results, so an ad must match what the buyer wants right now. A query like "jacket" means different things to different people. The obstacle is data: in a one-hour session most users view fewer than ten listings, so a session looks too short to personalize from. Without personalization, ads are less relevant for buyers and sellers.
Example from the article: for a buyer who recently interacted with men's leather jackets, the ads shown for the query "jacket" differ with and without session personalization (Fig. 1 of the original).
Solution in brief
Encode the last hour of a user's actions into one vector with a reusable three-component module, the adSformer Diversifiable Personalization Module (ADPM): a transformer-style encoder, pretrained item embeddings and representations learned from scratch. Concatenate the vector to the input of the click-through rate (CTR) and post-click conversion rate (PCCVR) models.
Questions this case answers
- How to personalize a ranking model from a short, variable-length sequence of a user's recent actions when most sequences are very short
- How to combine learned, pretrained and on-the-fly representations of recent behavior into one user vector that any downstream ranker can consume
- How to personalize ad ranking for a user from the last hour of their actions?
- How to personalize CTR and post-click conversion models when most sessions contain fewer than ten viewed items?
Tools and methods
- ADPM — three-component module that turns a one-hour action sequence into a user vector
- adSformer block — transformer block with global max pooling instead of concatenated outputs
- AIR embeddings — item embeddings from metric learning with sampled in-batch softmax loss
- EfficientNetB0 — backbone for multitask image embeddings (taxonomy, color, material heads)
- Deep and cross network (DCN) — interaction layers that exploit the wider input vector
- TensorFlow Keras — the module is a Keras module imported into downstream models
Google or a one-time email code. No password.