Deep Contextual Interest Network (DCIN) for CTR Prediction

Click‑through rate (CTR) prediction is a core task in online advertising and recommendation. Existing user‑behavior sequence models mainly focus on positive click signals and ignore the surrounding display context, which limits their effectiveness. To address this, the authors propose the Deep Contextual Interest Network (DCIN), which jointly models the clicked item and its surrounding items to capture context‑aware user interests.

DCIN consists of three key modules: (1) Position‑aware Context Aggregation Module (PCAM), which uses an attention mechanism to aggregate the representations of surrounding display items while incorporating absolute and relative position information; (2) Feedback‑Context Fusion Module (FCFM), a two‑layer MLP that non‑linearly fuses the click representation with its contextual representation; and (3) Interest Matching Module (IMM), which applies a DIN‑style attention to match the fused user interest with the target item.

In PCAM, each surrounding item is embedded, concatenated with its position embeddings, and weighted by an attention score that reflects both relevance to the click and positional bias. The weighted sum yields a context vector that is sensitive to the order of items on the screen.

FCFM takes the click embedding and the context vector as inputs, passes them through a two‑layer multilayer perceptron, and produces a fused representation that captures both positive and negative signals from the display context.

IMM then uses the fused representation together with the target item embedding in an attention network (similar to DIN) to generate the final user representation, which is fed into a binary classification layer for CTR estimation.

To meet strict latency requirements of industrial ad systems, the authors pre‑compute the context‑aware interest vectors offline because PCAM and FCFM are independent of the target item. This reduces online inference time by about 10 ms and allows DCIN to handle sequences 28 times longer than the previous state‑of‑the‑art RACP model with only ~1 ms additional latency.

Experiments are conducted on a proprietary 31‑day dataset containing billions of samples. Six baselines (Wide&Deep, DeepFM, DIN, DIEN, DFN, RACP) are compared using AUC and relative improvement (RelaImpr). DCIN achieves a 21.24 % RelaImpr over the best baseline and consistently outperforms all competitors in both offline and online A/B tests, delivering +1.5 % CTR and +1.5 % RPM gains.

Ablation studies confirm the importance of position information in PCAM and the fusion step in FCFM. Visualizations of the learned interest vectors further demonstrate DCIN’s ability to differentiate user interests under varying display contexts.

The paper concludes that incorporating display context and positional bias is essential for accurate interest modeling, and that DCIN provides a practical, high‑performance solution that has already been deployed at scale in Meituan’s advertising platform.