**SwitchTab: Switched Autoencoders Are Effective Tabular Learners** This paper introduces SwitchTab, a novel self-supervised learning framework designed to capture latent dependencies in tabular data. Traditional methods struggle with tabular data due to the lack of explicit spatial or semantic dependencies among samples. SwitchTab leverages an asymmetric encoder-decoder framework to decouple mutual and salient features, resulting in more representative embeddings. These embeddings improve decision boundaries and enhance performance in downstream tasks. Extensive experiments across various domains demonstrate superior performance in end-to-end prediction tasks with fine-tuning. Additionally, pre-trained salient embeddings can be used as plug-and-play features to boost the performance of traditional classification methods like Logistic Regression, XGBoost, and Random Forest. The framework also enhances explainability by visualizing decoupled mutual and salient features in the latent space. **Key Contributions:** 1. **SwitchTab Framework:** A novel self-supervised learning framework for tabular data that decouples salient and mutual embeddings. 2. **Competitive Performance:** Achieves competitive results across extensive datasets and benchmarks. 3. **Plug-and-Play Features:** Salient embeddings can be used to enhance traditional prediction models. 4. ** Explainability:** Visualizes decoupled mutual and salient features to improve model explainability. **Methodology:** - **Feature Corruption:** Enhances model performance through data augmentation. - **Self-supervised Learning:** Encodes and decouples feature vectors using projectors. - **Pre-training with Labels:** Introduces additional constraints for label prediction. - **Downstream Fine-tuning:** Integrates pre-trained encoder with linear layers for supervised tasks. **Experiments and Results:** - **Preliminary Information:** Details datasets, preprocessing, model architectures, and training settings. - **Performance Comparison:** SwitchTab outperforms mainstream deep learning and traditional models on various benchmarks. - **Plug-and-Play Embeddings:** Salient features significantly improve traditional models' performance. - **Visualization and Discussions:** t-SNE visualization confirms the effectiveness of decoupling mutual and salient features. - **Ablation Studies:** Analyzes the importance of the switching process, feature corruption rate, and computation cost. **Conclusion:** SwitchTab represents a significant step towards achieving more representative, explainable, and structured representations for tabular data, extending the success of representation learning from computer vision and natural language processing to tabular domains.**SwitchTab: Switched Autoencoders Are Effective Tabular Learners** This paper introduces SwitchTab, a novel self-supervised learning framework designed to capture latent dependencies in tabular data. Traditional methods struggle with tabular data due to the lack of explicit spatial or semantic dependencies among samples. SwitchTab leverages an asymmetric encoder-decoder framework to decouple mutual and salient features, resulting in more representative embeddings. These embeddings improve decision boundaries and enhance performance in downstream tasks. Extensive experiments across various domains demonstrate superior performance in end-to-end prediction tasks with fine-tuning. Additionally, pre-trained salient embeddings can be used as plug-and-play features to boost the performance of traditional classification methods like Logistic Regression, XGBoost, and Random Forest. The framework also enhances explainability by visualizing decoupled mutual and salient features in the latent space. **Key Contributions:** 1. **SwitchTab Framework:** A novel self-supervised learning framework for tabular data that decouples salient and mutual embeddings. 2. **Competitive Performance:** Achieves competitive results across extensive datasets and benchmarks. 3. **Plug-and-Play Features:** Salient embeddings can be used to enhance traditional prediction models. 4. ** Explainability:** Visualizes decoupled mutual and salient features to improve model explainability. **Methodology:** - **Feature Corruption:** Enhances model performance through data augmentation. - **Self-supervised Learning:** Encodes and decouples feature vectors using projectors. - **Pre-training with Labels:** Introduces additional constraints for label prediction. - **Downstream Fine-tuning:** Integrates pre-trained encoder with linear layers for supervised tasks. **Experiments and Results:** - **Preliminary Information:** Details datasets, preprocessing, model architectures, and training settings. - **Performance Comparison:** SwitchTab outperforms mainstream deep learning and traditional models on various benchmarks. - **Plug-and-Play Embeddings:** Salient features significantly improve traditional models' performance. - **Visualization and Discussions:** t-SNE visualization confirms the effectiveness of decoupling mutual and salient features. - **Ablation Studies:** Analyzes the importance of the switching process, feature corruption rate, and computation cost. **Conclusion:** SwitchTab represents a significant step towards achieving more representative, explainable, and structured representations for tabular data, extending the success of representation learning from computer vision and natural language processing to tabular domains.