Brand Advertising Value Modeling: From Instant CTR to Deep CVR and Incremental Uplift

Background : Alibaba Mama brand advertising provides high‑quality display resources to help advertisers improve brand awareness and reduce acquisition costs. Beyond guaranteeing exposure volume, advertisers increasingly care about the effectiveness of their placements. Value measurement enables the delivery of ads that match user preferences and allocates high‑value traffic to advertisers, benefiting both users and the platform.

Value Measurement System Construction : Starting from instant value (CTR), the system was extended to deep value (CVR) and incremental value (Uplift). CTR and CVR model the correlation between exposure and conversion, while Uplift captures the causal effect of an ad exposure on incremental conversion. To ensure homogeneity in Uplift modeling, a joint hash grouping of users and advertisers is applied, and special handling is used to mitigate exposure bias.

Deep Value Modeling (CVR) : A CVR model for out‑of‑Taobao brand ads was built. To address severe class imbalance, Focal Loss and Weighted Cross‑Entropy were employed. Delayed feedback was handled by decomposing the full‑cycle conversion probability into observable short‑term conversion probability and the probability of conversion within the observation window. The model uses a two‑stage training strategy and a multi‑expert architecture to improve learning and generalization. Multi‑task learning incorporates intermediate labels selected by conversion share. Online A/B tests showed a 4.0% lift in 7‑day delayed conversion.

Incremental Value Modeling (Uplift) : Uplift modeling relies on control/treatment groups. Because not all treatment requests are exposed, exposure rate is incorporated into the conversion relationship. A GAN‑based framework (GANLift) introduces a homogeneity constraint loss, encouraging the discriminator to be unable to distinguish generated samples from real control/treatment samples. The generator is supervised by the business conversion target, and the combined loss improves QINI by up to 12.5%.

Interaction Value Modeling : For the “Interactive City” scenario, a deep visit‑rate model was designed to filter out task‑incentive noise and focus on genuine user interest (second‑shop visit). Multi‑dimensional user sequence representations and UID embeddings were leveraged, yielding significant gains in second‑shop visit rate, deep interaction rate, and dwell time.

Model Technical Exploration : Feature representation and network architecture were the two main focus areas. High‑order feature interactions were enhanced by incorporating prior knowledge into embeddings, sample weighting, and adaptive network structures. References to DCN‑V2 and other works underline the importance of integrating domain knowledge into DNNs.

Multi‑Dimensional Sequence Representation : User behavior sequences were enriched with time, space, and depth dimensions. Each behavior’s embedding is weighted by a side‑information network that maps auxiliary signals (e.g., dwell time) to a weight, improving AUC by 0.13%.

User UID Representation : UID embeddings were added to capture user‑specific signals, boosting AUC by 0.24% but doubling model size. Hash‑bucketed UID compression reduced size while preserving most of the performance gain. Experiments showed larger improvements in high‑frequency interaction scenarios (e.g., Super Interactive City) than in low‑frequency ones (e.g., Super Storm).

Summary & Outlook : The work advances brand advertising value modeling by extending from CTR to CVR and Uplift, addressing delayed feedback, exposure bias, and homogeneity constraints. Future directions include deeper integration of prior knowledge across features, samples, network structures, and loss functions to accelerate convergence and further improve prediction accuracy.

References : [1] Y. Hou et al., “Conversion Prediction with Delayed Feedback: A Multi‑task Learning Approach,” ICDM 2021. [2] W. Ke et al., “Addressing Exposure Bias in Uplift Modeling for Large‑scale Online Advertising,” ICDM 2021. [3] Yoon et al., “GANITE: Estimation of Individualized Treatment Effects Using Generative Adversarial Nets,” ICLR 2018. [4] G. Zhou et al., “Deep Interest Network for Click‑Through Rate Prediction,” KDD 2018. [5] G. Zhou et al., “Deep Interest Evolution Network for Click‑Through Rate Prediction,” AAAI 2019. [6] R. Wang et al., “DCN V2: Improved Deep & Cross Network,” WWW 2021.