Multi-Epoch learning with Data Augmentation for Deep Click-Through Rate Prediction

TL;DR

This work tackles the persistent one-epoch overfitting observed in deep CTR models by decoupling the embedding and MLP components and introducing MEDA, a framework that performs multi-epoch training with data augmentation through varied embedding spaces. In non-continual MEDA, embeddings are reinitialized each epoch to break data dependency, while continual MEDA trains multiple independently initialized embedding spaces sequentially to augment data without excessive embedding-MLP coupling. Empirical results across public and business datasets show consistent AUC gains (0.8% to 4.6%) over single-epoch training, with online A/B testing confirming practical benefits in a production ad system. The approach demonstrates that pre-trained MLPs can adapt to new embedding spaces, enabling effective matching based on relative embedding relationships and offering significant data-augmentation advantages with potential privacy-aware extensions.

Abstract

This paper investigates the one-epoch overfitting phenomenon in Click-Through Rate (CTR) models, where performance notably declines at the start of the second epoch. Despite extensive research, the efficacy of multi-epoch training over the conventional one-epoch approach remains unclear. We identify the overfitting of the embedding layer, caused by high-dimensional data sparsity, as the primary issue. To address this, we introduce a novel and simple Multi-Epoch learning with Data Augmentation (MEDA) framework, suitable for both non-continual and continual learning scenarios, which can be seamlessly integrated into existing deep CTR models and may have potential applications to handle the "forgetting or overfitting" dilemma in the retraining and the well-known catastrophic forgetting problems. MEDA minimizes overfitting by reducing the dependency of the embedding layer on subsequent training data or the Multi-Layer Perceptron (MLP) layers, and achieves data augmentation through training the MLP with varied embedding spaces. Our findings confirm that pre-trained MLP layers can adapt to new embedding spaces, enhancing performance without overfitting. This adaptability underscores the MLP layers' role in learning a matching function focused on the relative relationships among embeddings rather than their absolute positions. To our knowledge, MEDA represents the first multi-epoch training strategy tailored for deep CTR prediction models. We conduct extensive experiments on several public and business datasets, and the effectiveness of data augmentation and superiority over conventional single-epoch training are fully demonstrated. Besides, MEDA has exhibited significant benefits in a real-world online advertising system.

Figures (9)

• Figure 1: Our proposed MEDA framework. For non-continual learning, MEDA reinitializes the embedding parameters at the onset of each training epoch; for continual learning, MEDA maintains multiple independently initialized embedding layers and for each dataset, trains each embedding layer once successively. The embedding layers can be selected based on requirements or costs.
• Figure 2: The test AUC curves of the Direct multi-epoch learning and our non-continual MEDA on the public datasets.
• Figure 3: The test AUC curves of various models trained with our non-continual MEDA on the public datasets.
• Figure 4: The training metric curves of training DNN on the Taobao dataset, with or without non-continual MEDA.
• Figure 5: The test metric curves of training DNN on the Taobao dataset, with or without non-continual MEDA.