[CIKM 2025] MoAE: Breaking the Single-View Bottleneck in CVR Prediction with Multi-Attribution Learning
MAC: A Conversion Rate Prediction Benchmark Featuring Labels Under Multiple Attribution Mechanisms
This paper introduces the Multi-Attribution BenChmark (MAC), the first public conversion rate (CVR) prediction dataset featuring labels from four diverse attribution mechanisms. The authors propose the Mixture of Asymmetric Experts (MoAE) model, which achieves a new SOTA by effectively integrating multi-attribution knowledge into a main-task-centric architecture.
TL;DR
The accuracy of Conversion Rate (CVR) prediction is the lifeblood of online advertising. However, most models are "blind" to the full user journey because they only see labels from a single attribution rule (like last-click). This paper introduces MAC, the first large-scale public benchmark with multi-attribution labels, and MoAE, a model that achieves SOTA performance by learning from multiple attribution "perspectives" simultaneously.
Problem & Motivation: The "Single-View" Trap
In modern e-commerce, a user might click several ads before finally purchasing. Attribution mechanisms decide which click gets the credit. Currently, models are trained on a single mechanism, which is like trying to understand a 3D object from a 2D shadow—it lacks the causality captured by other rules (e.g., first-click, linear, or data-driven attribution).
The hurdle? No public data existed to study this "Multi-Attribution Learning" (MAL) until now.
Methodology: Two Principles for Better Learning
The authors argue that a successful MAL model must satisfy two design principles:
- Full Knowledge Acquisition: Use specialized experts to absorb different attribution logic.
- Task-Prioritized Utilization: Ensure the main business goal isn't "diluted" by auxiliary tasks.
The MoAE Architecture
MoAE (Mixture of Asymmetric Experts) implements these via:
- MoE Backbone: Separates "Shared Experts" (common conversion patterns) from "Attribution-specific Experts" (unique logic for Last-click vs. DDA).
- Asymmetric Transfer: Instead of a symmetric shared layer, knowledge flows from auxiliary towers into the primary task tower to maximize the main metric.
The MAC Benchmark & Experimental Insights
The authors released MAC, a dataset of 79 million clicks from Taobao, featuring four attribution views: Last-click, First-click, Linear, and Data-Driven Attribution (DDA).
Key Results
- MAL is Universal: Across all settings, MAL-based models beat single-attribution baselines.
- Complex Journeys Benefit Most: Users with longer conversion paths (more clicks) showed the highest GAUC gains (see Figure 5).
- MoAE Dominance: MoAE outperformed NATAL (the previous SOTA) across the board, proving that its structural design is superior to simple gradient surgery or loss weighting.
Critical Analysis & Conclusion
The value of this paper lies in the MAC benchmark and the PyMAL library. By providing the tools, the authors enable the community to move beyond the simplistic "last-click" assumption.
Limitations: The "First-click" target remains challenging. The authors found it notoriously noisy—adding too many auxiliary tasks actually hurt performance here. This suggests that the temporal gap in first-click attribution introduces confounding factors that current MDL/MAL architectures still struggle to resolve.
Future Outlook: The inclusion of Chain-of-Thought (CoT) reasoning for conversion path analysis could be the next frontier in understanding why certain touchpoints have higher attribution weights.
Takeaway: In advertising AI, more labels from different viewpoints are better than one perfect label. Architecture matters: don't just "share everything"; transfer knowledge asymmetrically to protect your main KPI.