How Weak Supervision Powers Ant Group’s Real‑World AI Challenges

1. Introduction to Weak Supervision

Weak supervision addresses situations where fully labeled data are scarce or noisy. It can be categorized into three typical problems: incomplete supervision (few labeled samples, many unlabeled), inaccurate supervision (labels contain systematic noise), and inexact supervision (labels are coarse‑grained, e.g., multi‑instance learning).

2. Weak‑Supervision Scenarios at Ant Group

Two common business challenges motivate the use of weak supervision: (1) limited labeled data in a target risk‑assessment scenario, where cross‑domain data from higher‑risk users or historical customers can be leveraged; (2) high labeling cost or noisy rules in fraud detection, where expert knowledge, existing rules, or legacy models provide imperfect labels.

3. Modeling with Sample Scarcity

3.1 Cross‑Domain Causal Effect Estimation

The goal is to estimate the effect of an intervention (treatment) on an outcome when the target domain lacks treatment or outcome data. Two cases are considered: (a) no labeled data in the target domain, and (b) a few labeled samples in the target domain.

We propose a Direct Learning framework that first predicts pseudo‑effects using source‑domain control and treated models, then learns an effect model on these pseudo‑effects. To handle distribution shift, we apply density‑based re‑weighting via domain adaptation.

Unreliable pseudo‑effects are mitigated by estimating uncertainty with Monte‑Carlo dropout; samples with low uncertainty receive higher weights (reliable scoring).

3.2 Scenario with Few Labeled Target Samples

Beyond the two‑stage approach, we explore a neural‑network‑based method that predicts factual outcomes and adds a debiasing loss to align source and target distributions, improving generalization.

4. Modeling with Noisy Labels

In many target scenarios, multiple noisy label sources (expert tags, rule‑based tags, outdated models) are available. The objective is to build a robust model that leverages these sources.

Two classic approaches exist: (a) two‑stage methods that first aggregate noisy labels (e.g., voting) and then train; (b) joint estimation of source confusion matrices and model parameters. We extend these by exploiting the model’s self‑cognition ability.

4.1 Theoretical Insights

Models can identify instance‑wise label noise: samples with higher loss are likely mislabeled.

Models can also discern annotator‑wise quality, distinguishing high‑quality from low‑quality label sources.

4.2 Framework Design

The proposed architecture contains three key modules:

Self‑cognition : estimates the reliability of each sample’s label and produces a quality vector for each annotator.

Mutual‑denoising : uses reliable annotators to generate pseudo‑labels for other sources, weighting them by the learned quality scores (1‑w). This improves each source’s learning.

Selective Knowledge Distillation : distills the multi‑source model into a lightweight version for deployment, guided by the reliability scores.

5. Experimental Validation

Extensive experiments on synthetic and real‑world datasets show that the proposed method outperforms baselines, especially under strong distribution shift. Ablation studies confirm that adding the Reliable Scoring module consistently improves performance, while the Distribution Adaptation module alone may cause degradation without reliable scoring.

The work has been accepted at CIKM (titled “Treatment Effect Estimation across Domains”) and at ICML (titled “Self‑cognitive Denoising in the Presence of Multiple Noisy Label Sources”).

6. Practical Applications

Sample‑scarcity solutions enable cross‑domain risk modeling, such as staged marketing coupon allocation. Noisy‑label techniques support user‑profile inference where expert tags or legacy rules provide imperfect annotations.

Overall, the combination of cross‑domain causal learning, uncertainty‑aware pseudo‑effect estimation, and multi‑source noisy‑label denoising offers a versatile toolkit for real‑world weak‑supervision problems.