How GIN-Based Cohort Modeling Boosts Cold-Start CTR Prediction by 2%

Background

CTR prediction in JD advertising suffers for cold‑start users lacking interaction history, leading to random recommendations and low click‑through rate.

Motivation

Existing encoder‑decoder methods generate virtual behavior embeddings but (1) use simplistic encoding that ignores high‑order user‑item interactions, and (2) feed the virtual embeddings directly into the CTR model, causing biased representations.

Proposed Method: GINCM

GINCM (Graph Isomorphism Network‑based Cohort Modeling) consists of three cooperating modules.

1. GIN‑based Graph Encoder

Encodes each user’s bipartite interaction sub‑graph (users ↔ items, edges = click/browse) with a Graph Isomorphism Network (GIN). GIN captures higher‑order and non‑linear topological patterns better than GCN or GraphSAGE, producing a dense user behavior representation.

2. Joint Denoising Auto‑Encoder (JDAE)

Generates a virtual behavior embedding for a cold user from its attribute vector (e.g., gender, age, city, device). The encoder processes both attribute and behavior features; the decoder reconstructs each from the other, enforcing a bidirectional consistency loss:

Loss = ||Attr - Dec(Beh)||² + ||Beh - Dec(Attr)||²

This forces the model to learn a mapping “attribute → behavior” and its inverse, yielding richer virtual embeddings than raw attributes.

3. Cohort Modeling Network (CMN)

The virtual embedding is binarized into a short code using a Straight‑Through Estimator (STE). Each binary code indexes a cohort (group of similar users). Cohort embeddings are obtained by averaging the embeddings of all users sharing the same code, and the averaged cohort vector replaces the raw virtual embedding for the downstream CTR model, thereby correcting individual bias.

Experimental Setup

Two datasets were used:

Amazon Reviews benchmark (1.2 M samples).

JD advertising logs (30 B samples over 31 days; first 30 days for training, last day for testing).

Ablation studies removed each module in turn. Results (new‑user AUC) show:

Without GIN (Mean‑Pooling): AUC ↓ 0.7321 → high‑order interaction modeling is crucial.

Without JDAE (default vector): AUC ↓ 0.7112 → learned virtual behaviors outperform raw attributes.

Without CMN (direct virtual embedding): AUC ↓ 0.7145 → cohort correction yields more robust predictions.

Online Deployment

During data ingestion, user sub‑graphs are constructed and processed by GINCM; the resulting cohort embeddings are cached. At real‑time ranking, the cached vectors are read directly, avoiding graph reconstruction latency.

Production Impact

A/B testing on JD’s platform reported a 2.13 % lift in click‑through rate and a matching 2.13 % increase in revenue per mille (RPM).

https://dl.acm.org/doi/10.1145/3726302.3731936