[Paper] MAC: A Conversion Rate Prediction Benchmark Featuring Labels Under Multiple Attribution Mechanisms

Source: arXiv - 2603.02184v1

Overview

The paper introduces MAC, the first publicly available conversion‑rate (CVR) benchmark that provides labels generated by multiple attribution mechanisms (e.g., last‑click, first‑click, time‑decay). By exposing the same user journey to several “views” of conversion, MAC enables multi‑attribution learning (MAL)—training models that can simultaneously reason about different business‑logic definitions of a conversion. The authors also release PyMAL, an open‑source library that bundles a suite of baseline MAL algorithms and utilities for reproducible research.

Key Contributions

• MAC dataset: a large‑scale, real‑world CVR benchmark containing conversion labels from several attribution schemes, filling a long‑standing data gap for MAL research.
• PyMAL library: a ready‑to‑use Python package (built on PyTorch) that implements a wide range of MAL baselines, evaluation metrics, and data loaders.
• Comprehensive empirical study: systematic experiments that uncover three practical insights about when and how MAL helps CVR prediction.
• Mixture of Asymmetric Experts (MoAE): a novel MAL architecture that outperforms the previous state‑of‑the‑art by explicitly separating multi‑attribution knowledge learning from main‑task inference.
• Open‑source release: both the benchmark and library are publicly hosted, encouraging community contributions and fair comparison.

Methodology

1. Data Construction – The authors start from an industrial advertising log (≈ hundreds of millions of ad impressions). For each user session they compute conversion labels under several common attribution rules (last‑click, first‑click, linear, time‑decay, etc.). The resulting dataset contains a shared feature set (user, ad, context) plus multiple binary targets, one per attribution method.
2. Multi‑Attribution Learning (MAL) – Instead of training separate models for each target, MAL treats the auxiliary attribution targets as auxiliary tasks in a multi‑task learning (MTL) framework. The main task is usually the business‑critical attribution (e.g., last‑click), while the others provide extra supervisory signals.
3. Baseline Suite – PyMAL implements several MTL strategies: hard parameter sharing, soft sharing (via attention or gating), and recent MAL‑specific designs (e.g., shared expert layers, task‑specific heads).
4. Proposed Architecture – MoAE
   • Asymmetric Experts: separate expert subnetworks specialize on each attribution label, allowing the model to capture attribution‑specific patterns.
   • Mixture Layer: a gating network learns to combine expert outputs differently for each task, ensuring the main task can selectively draw on the most relevant auxiliary knowledge.
   • Training Objective: a weighted sum of binary cross‑entropy losses, with dynamic weighting to avoid over‑emphasizing noisy auxiliary tasks.

All experiments are run on the MAC benchmark using identical data splits and hyper‑parameter search protocols to guarantee fair comparison.

Results & Findings

• Consistent gains for most attribution settings, especially when the user journey is long and noisy.
• Objective complexity matters: adding more auxiliary tasks helps when the main target is “hard” (e.g., last‑click) but can hurt for simpler targets like first‑click, highlighting the need for careful auxiliary selection.
• Design principles validated: (1) full multi‑attribution knowledge learning (dedicated experts) and (2) knowledge utilization (task‑aware mixing) are both essential.
• MoAE performance: outperforms the previous best MAL method by ≈ 2.3 % absolute AUC on the main task and shows better stability across attribution variants.

Practical Implications

• Better ROI estimation – Advertisers can train a single model that simultaneously predicts conversions under several attribution rules, reducing the overhead of maintaining multiple models.
• Improved bidding algorithms – Real‑time bidding systems can feed the richer multi‑attribution predictions into their utility functions, leading to more nuanced budget allocation (e.g., favoring users likely to convert under a long‑term attribution).
• Reduced data engineering effort – With MAC and PyMAL, data teams can quickly prototype MAL pipelines without building custom multi‑task scaffolding.
• Model interpretability – The asymmetric expert structure gives a natural way to inspect which attribution signals drive a prediction, aiding debugging and compliance reporting.
• Scalable deployment – MoAE’s modular experts can be pruned or quantized independently, allowing edge‑or‑cloud deployment strategies that match latency constraints.

Limitations & Future Work

• Domain specificity – MAC is derived from a single e‑commerce advertising platform; transferability to other domains (e.g., SaaS, gaming) remains to be validated.
• Label noise – Attribution labels are heuristic approximations of true causal impact; noisy auxiliaries can degrade performance, as seen with first‑click.
• Static weighting – The current loss‑weighting scheme is heuristic; more sophisticated meta‑learning or reinforcement‑learning approaches could adaptively balance tasks.
• Scalability of experts – As the number of attribution mechanisms grows, the expert pool may become unwieldy; future work could explore parameter‑efficient expert sharing or sparsity techniques.

The authors encourage the community to extend MAC with additional attribution schemes, test MAL on other downstream tasks (e.g., churn prediction), and explore richer architectural variants built on the PyMAL foundation.

Authors

• Jinqi Wu
• Sishuo Chen
• Zhangming Chan
• Yong Bai
• Lei Zhang
• Sheng Chen
• Chenghuan Hou
• Xiang‑Rong Sheng
• Han Zhu
• Jian Xu
• Bo Zheng
• Chaoyou Fu

Paper Information

• arXiv ID: 2603.02184v1
• Categories: cs.LG, cs.AI
• Published: March 2, 2026
• PDF: Download PDF