TCCD: Tree-guided Continuous Causal Discovery via Collaborative MCTS-Parameter Optimization

Jingjin Liu, Yingkai Xiao, Hankz Hankui Zhuo, Wushao WenIJCAI 2025 · DOI

TL;DR

This paper introduces TCCD, a novel framework for discovering causal relationships in complex, multi-type event sequences, like those found in telecommunication networks. TCCD uniquely combines Monte Carlo Tree Search (MCTS) for global exploration with continuous gradient optimization for local refinement, overcoming limitations of existing methods that struggle with large search spaces and local optima.

What Problem Does This Paper Address?

Accurately identifying causal relationships between events is crucial for understanding and managing complex systems, particularly in telecommunication networks where a single fault can trigger a cascade of alarms. These networks generate massive, multi-type event sequences that are often represented as Directed Acyclic Graphs (DAGs). However, existing causal discovery methods face significant challenges:

  • Explosive Search Space: The number of possible causal structures grows exponentially with the number of event types and network scale, making exhaustive search infeasible.
  • Limited Generalization: Dynamic network conditions and increasing event complexity make it hard for current models to accurately estimate spatio-temporal parameters.
  • Local Optima: Purely gradient-based methods can get stuck in suboptimal solutions due to their reliance on local information, failing to explore the full range of possibilities.
  • Non-IID Data: Many traditional methods assume independent and identically distributed (i.i.d.) data, which doesn't hold for event sequences influenced by network topology and temporal dependencies.

This paper aims to address these issues by proposing a method that can efficiently and accurately discover causal structures in large-scale, complex event sequences.

Methods

TCCD's core innovation lies in its collaborative optimization of causal structure search and parameter estimation. It achieves this by integrating two key components:

  • Monte Carlo Tree Search (MCTS): MCTS is used to explore the vast search space of possible causal DAGs. It balances exploration (trying new paths) and exploitation (following promising paths) to efficiently navigate the complex structure landscape.
  • Continuous Gradient Optimization: For each potential causal structure explored by MCTS, TCCD uses continuous gradient-based methods to optimize the underlying parameters (e.g., causal strengths, baseline intensities). This allows for fine-tuning and accurate estimation of the relationships.

Key novel aspects of TCCD include:

  • Tree-Guided Dynamic Parameter Learning: This strategy allows parameter estimation to guide the structure search and, conversely, structural updates to trigger efficient local parameter adjustments. This creates a hierarchical framework for global exploration and local optimization.
  • Influence-Driven Local Optimization: To improve efficiency, TCCD updates parameters only for the parts of the causal graph directly affected by a structural change. This localized approach significantly reduces computational complexity.
  • Hierarchical Potential Score (HPS): TCCD employs a unified scoring function that combines BIC scores, local likelihood, and global consistency. This score guides the MCTS search and helps in iteratively refining the causal structure.
  • Refined Likelihood Function: Building on prior work (S2GCSL), TCCD uses a log-likelihood function that accounts for event occurrences, historical sequences, and topological structure, transforming the discrete DAG search into a continuous optimization task.

Datasets

The paper evaluates TCCD on:

  • Real-World Metropolitan Cellular Network Alarm Data: Three datasets are used:
  • Wireless-18: Contains 18 alarm types and 55 topological network elements. Notably, this dataset lacks an underlying topological network structure, demonstrating TCCD's ability to infer causality even without explicit network information.
  • Micro-24: 24 alarm types and 439 topological network elements.
  • Micro-25: 25 alarm types and 474 topological network elements.
  • Synthetic Datasets: Generated using the gcatle API, simulating event sequences based on topological Hawkes processes.

Key Insights and Ideas

  • Synergy of MCTS and Gradient Optimization: The paper demonstrates that combining MCTS for global exploration with gradient-based methods for local refinement is a powerful strategy for tackling complex causal discovery problems. This hybrid approach effectively addresses the limitations of each method when used in isolation.
  • Importance of Spatio-Temporal Modeling: The results highlight that methods incorporating topological network structures and temporal decay (like TCCD) significantly outperform traditional point process-based approaches, especially in large-scale, complex networks.
  • Localized Optimization for Efficiency: The influence-driven local optimization mechanism is crucial for making the approach scalable. By focusing updates on affected parts of the graph, TCCD significantly reduces computational burden without sacrificing accuracy.
  • Hierarchical Scoring for Robustness: The HPS function, which balances different scoring criteria, provides a robust way to guide the MCTS search, ensuring that the discovery process considers various aspects of causal inference.
  • Effectiveness of Invalid Edge Pruning: The ablation study suggests that removing invalid edges (those that would create cycles) is an important step for improving precision and overall performance.

Key Results

  • Superior Performance on Real-World Data: TCCD consistently outperforms several state-of-the-art baselines (PC, GES, ADM4, PCMCI, Corl, THP, S2GCSL) across all three real-world datasets in terms of F1 score, Precision, and Recall.
  • Highest F1 Scores: On the Wireless-18 dataset, TCCD achieved an F1 score of 0.6719, with high Precision (0.7288) and Recall (0.6232), and the lowest Structural Hamming Distance (SHD) of 34.
  • Robustness to Network Scale: TCCD demonstrates strong stability and generalization ability, particularly in high-dimensional scenarios with a large number of event types.
  • Ablation Study Validation: The TCCD-MCTS variant (without full MCTS refinement) still outperformed the S2GCSL baseline, validating the core components of TCCD. Removing invalid edges (TCCD-remove) further boosted F1 scores, confirming its importance.
  • Accurate Causal Structure Reconstruction: Visualizations of causal graphs on synthetic datasets show TCCD's ability to accurately capture both dense and sparse causal relationships, even in complex scenarios.

Limitations and Future Work

  • Limitations:
    • While TCCD significantly improves efficiency, the computational complexity can still be a challenge for extremely large-scale networks.
  • The paper notes that TCCD-remove (without invalid edge pruning) still performs competitively, suggesting that while pruning helps, the core MCTS-optimization framework is robust.
  • Future Work:
    • Scaling MCTS: The authors suggest exploring logical action representations to scale MCTS to even larger search spaces.
  • Diverse Event Patterns: Adapting the framework to handle more diverse event patterns and temporal dynamics is a promising direction.
  • Real-time Applications: Further research could focus on optimizing TCCD for real-time causal discovery in dynamic environments.

Summary updated

BibTeX
@inproceedings{liu2025tccd,
  author = {Jingjin Liu and Yingkai Xiao and Hankz Hankui Zhuo and Wushao Wen},
  title = {{TCCD: Tree-guided Continuous Causal Discovery via Collaborative MCTS-Parameter Optimization}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/1030},
}
Figure