Few-Shot Learning, Data Augmentation, and Semi‑Supervised Methods for Improving Safety and Governance Models at Didi

Didi, a large ride‑hailing platform, faces a critical challenge in safety and governance: the lack of sufficient, accurately labeled samples limits model performance. To address this, the team explores few‑shot learning techniques and builds a systematic solution covering data, model, and algorithmic strategies.

Related Work – Few‑shot learning is categorized into three approaches: (1) Data‑centric methods that use prior knowledge for data augmentation, (2) Model‑centric methods that reduce the hypothesis space via prior knowledge, and (3) Algorithmic methods that improve parameter search strategies.

Data Augmentation – For image data, classic transformations such as flipping, rotation, and scaling are used. For text, techniques include synonym replacement, random insertion, deletion, and swapping. The team also inserts domain‑specific “key phrases” (e.g., driver‑resistance sentences) into negative samples to create positive examples. Augmentation ratios of 3:3:3:1 (unchanged, add‑character, delete‑character, concatenate) are applied to generate diverse training data.

Semi‑Supervised Learning (Self‑Training) – The pipeline consists of: (1) training an initial model on the limited labeled set, (2) predicting pseudo‑labels on unlabeled data and selecting high‑confidence samples, (3) iteratively adding these pseudo‑labeled samples to the training set and retraining. This method is applied to expense‑complaint driver‑responsibility detection, achieving a >20% reduction in log‑loss compared with a baseline XGBoost model.

Model – Multi‑Task Learning – To mitigate extreme class imbalance in intercept scenarios, a multi‑task architecture shares embedding and hidden layers between a high‑volume “complaint” auxiliary task and the low‑volume “fact” task. This reduces the parameter search space for the primary task and improves ROC‑AUC and PR‑AUC, especially at low impact‑rate operating points.

Applications – The techniques are validated on several safety‑related use cases: (1) sexual‑harassment order detection, where key‑phrase insertion boosts TextCNN performance; (2) expense‑complaint driver responsibility, where self‑training lowers log‑loss; (3) route‑intercept, where multi‑task learning outperforms an online XGBoost baseline in ROC and PR curves.

Conclusion – By combining data augmentation, semi‑supervised learning, and multi‑task learning, Didi significantly reduces labeling costs while improving model effectiveness in sparse‑sample safety scenarios, providing a practical reference for other organizations facing similar data scarcity challenges.