Dynamic Sparse Learning (DSL) is a novel learning paradigm designed to enhance the efficiency of recommendation systems by reducing both training and inference costs while maintaining performance. The paper introduces DSL, which dynamically adjusts the sparsity of model weights during training, allowing the model to learn and retain important parameters while pruning redundant ones. This approach ensures a consistent and minimal parameter budget throughout the entire learning process, enabling end-to-end efficiency from training to inference. DSL is model-agnostic and can be applied to various recommendation models. It trains a lightweight sparse model from scratch, periodically evaluating and dynamically adjusting the significance of each weight and the sparsity distribution of the model. This method avoids the need for pre-training or complex architecture search, which are common limitations in existing solutions. DSL achieves this by maintaining a fixed sparsity ratio throughout training, ensuring that both training and inference complexities are minimized. The paper evaluates DSL on diverse recommendation models and benchmark datasets, demonstrating its effectiveness in reducing training and inference costs while maintaining comparable performance. Experimental results show that DSL can significantly reduce computational costs and memory usage, with performance that is competitive with full models. DSL's dynamic exploration mechanism allows it to discover optimal model architectures, leading to better performance and efficiency. DSL addresses the limitations of existing methods such as knowledge distillation, automated machine learning, and model pruning by providing a flexible and efficient learning framework. It dynamically adjusts the sparsity of model weights, enabling the model to adapt to different data distributions and maintain performance. The paper also provides ablation studies and visualizations to support the effectiveness of DSL, showing that it can achieve better "performance-sparsity" trade-offs compared to other methods. Overall, DSL offers a promising solution for efficient recommendation systems by combining dynamic sparsity adjustment with end-to-end training, leading to significant reductions in computational costs and improved model efficiency.Dynamic Sparse Learning (DSL) is a novel learning paradigm designed to enhance the efficiency of recommendation systems by reducing both training and inference costs while maintaining performance. The paper introduces DSL, which dynamically adjusts the sparsity of model weights during training, allowing the model to learn and retain important parameters while pruning redundant ones. This approach ensures a consistent and minimal parameter budget throughout the entire learning process, enabling end-to-end efficiency from training to inference. DSL is model-agnostic and can be applied to various recommendation models. It trains a lightweight sparse model from scratch, periodically evaluating and dynamically adjusting the significance of each weight and the sparsity distribution of the model. This method avoids the need for pre-training or complex architecture search, which are common limitations in existing solutions. DSL achieves this by maintaining a fixed sparsity ratio throughout training, ensuring that both training and inference complexities are minimized. The paper evaluates DSL on diverse recommendation models and benchmark datasets, demonstrating its effectiveness in reducing training and inference costs while maintaining comparable performance. Experimental results show that DSL can significantly reduce computational costs and memory usage, with performance that is competitive with full models. DSL's dynamic exploration mechanism allows it to discover optimal model architectures, leading to better performance and efficiency. DSL addresses the limitations of existing methods such as knowledge distillation, automated machine learning, and model pruning by providing a flexible and efficient learning framework. It dynamically adjusts the sparsity of model weights, enabling the model to adapt to different data distributions and maintain performance. The paper also provides ablation studies and visualizations to support the effectiveness of DSL, showing that it can achieve better "performance-sparsity" trade-offs compared to other methods. Overall, DSL offers a promising solution for efficient recommendation systems by combining dynamic sparsity adjustment with end-to-end training, leading to significant reductions in computational costs and improved model efficiency.