EXTR: Click-Through Rate Prediction with Externalities in E-Commerce Sponsored Search

Click‑through rate (CTR) prediction is vital for e‑commerce platforms. Besides the intrinsic quality of an ad, the surrounding items—both organic results and other ads, termed *external items*—significantly influence the target ad’s CTR. Existing CTR models usually ignore these externalities.

Two challenges arise: (1) *Diversity externalities*: a target ad may appear in any slot, each surrounded by a different set of external items; (2) *Incomplete externalities*: during the prediction stage the exact external ads are unknown. Inspired by the parallelism of Transformers, the paper proposes **EXternality TRansformer (EXTR)**, which treats every possible placement of the target ad as a query and the corresponding external items as key/value pairs, thus modeling diverse externalities simultaneously. To address incomplete externalities, a **Potential Allocation Generator (PAG)** module is introduced to infer plausible arrangements of external ads.

EXTR comprises two sub‑modules: a *Context Interaction Module* that encodes interactions among external items via stacked self‑attention layers, and an *Externality Extraction Module* that employs heterogeneous attention, using the target ad and slot encodings as queries and the context‑interacted vectors as key/value. A personalized externality term incorporates user embeddings via a dot‑product operator.

PAG generates ordering codes for external ads by weighting natural‑result rankings, and is trained with an auxiliary KL‑divergence loss using real slot positions as supervision.

Experiments on a week‑long Taobao search dataset (covering thousands of product categories) compare EXTR with strong baselines such as CAN, MLP, PNN, Wide&Deep, DCN, xDeepFM and a vanilla Transformer. EXTR consistently achieves higher AUC, COPC and lower LogLoss. Ablation studies confirm the contribution of each component (personalized externalities, external ad inclusion, PAG). Attention‑map visualizations illustrate how different consumer groups are affected by surrounding items.

The proposed framework is deployed in Alibaba’s advertising system, demonstrating practical efficiency and effectiveness. Future directions include handling data bias, cold‑start for new users, and improving model interpretability.