Understanding the causal structure among the tags in marketing systems

The tagging system has become the basis of various systems, e.g., geo-social system, marketing system. Understanding the structure among the tags is one of the crucial tasks to the performance of various downstream marketing tasks, such as user behavior understanding, advertising, and recommendation. However, most of the existing methods mainly focus on the association among the tags, which usually results in false intervention suggestions, e.g., recommending a similar item after having bought one. To address this problem, we propose an Iterative Causal Structure Search (ICSS in short) algorithm for the high-dimensional social tags. In each iteration of the proposed approach, we first employ the constraint-based method to discover the skeleton of the causal structure and further employ the additive noise assumption to infer the edges whose directions are unknown in the previous stage. The proposed approach not only benefits from the good scalability of the constraint-based approach but also avoids the Markov equivalence class problem with the help of the additive noise assumption. We also theoretically show the correctness of the proposed algorithm. We test the ICSS and the baselines on both the simulated data and real-world data, further discover some interesting causal structures among the tags in a real-world marketing system.