AutoShape: Automatic Design of Click-Through Rate Prediction Models Using Shapley Value

Abstract

As a main part of automated machine learning, Neural Architecture Search (NAS) has been employed in exploring effective Click-Through Rate (CTR) prediction models of recommender systems in recent studies. However, these studies update the architecture parameters by different strategies, such as gradient-based methods or evolutionary algorithms, and may suffer from either interference problems or high time-consuming in the search stage. Besides, the blocks in the search space were usually homogeneous, which means they have the same candidate operations to choose. Such design will result in redundancy in the search space, because many structures are inherently invalid and just increase the complexity of searching. To address the above issues, we implement the three-level automatic design of CTR prediction models using NAS via Shapley value, named as AutoShape. For the search space, we divide it into three parts according to the characteristics of the CTR model. Each part comprises distinctive candidate operations and forms a cell as an individual processing level, which improves the stability of the searched models’ effect and reduces the computational complexity. For the search strategy, we leverage Shapley value, a metric derived from cooperative game theory, to estimate the contribution of the operations in each block and the connections between the blocks, which can find effective models and reduce the time cost. Furthermore, experiments on different benchmark datasets show that the structure obtained from the search performs better than several hand-crafted architectures and is more stable than another NAS-based algorithm.