MATMCD

Causal discovery is an imperative foundation for decision-making across domains, such as smart health, AI for drug discovery and AIOps. Traditional statistical causal discovery methods, while well-established, predominantly rely on observational data and often overlook the semantic cues inherent in cause-and-effect relationships. The advent of Large Language Models has ushered in an affordable way of leveraging the semantic cues for knowledge-driven causal discovery, but the development of LLMs for causal discovery lags behind other areas, particularly in the exploration of multi-modal data.

To bridge the gap, we introduce MATMCD, a multi-agent system powered by tool-augmented LLMs. MATMCD has two key agents: a Data Augmentation agent that retrieves and processes modality-augmented data, and a Causal Constraint agent that integrates multi-modal data for knowledge-driven reasoning. The proposed design of the inner-workings ensures successful cooperation of the agents. Our empirical study across seven datasets suggests the significant potential of multi-modality enhanced causal discovery.

Hybrid methods for causal discovery typically prompt LLMs with the prior causal graph produced by some SCD algorithm, appended by some meta-data such as variable names and dataset titles as contexts. However, these inputs may fall short in fully activating the reasoning ability of LLMs. We were inspired by the observation that abundant semantic data from external sources, such as webs and logs, can serve as an additional modality to the observational data for improving prompts. On a variety of datasets, MATMCD reduces causal discovery errors (NHD) by up to 66.7% and improves root cause locating (MAP@10) by up to 83.3% over the best baselines.

To be comprehensive, we use five benchmark datasets covering both continuous variables and discrete variables: AutoMPG, which has five variables concerning city-cycle fuel consumption in miles per gallon; DWDClimate, which has six variables pertinent to observations from weather stations in Deutscher Wetterdienst; SachsProtein, which has eleven variables measuring the expression level of different proteins and phospholipids in human cells; Asia, which has eight variables relevant to lung disease diagnosis; and Child, which has twenty variables regarding congenital heart disease in newborn babies. By default we use GPT-4o mini with temperature 0.5 as the base LLM for all LLM-based and hybrid methods, and PC is used as the base SCD algorithm for all hybrid approaches.

Several observations follow from the table below. Methods involving LLMs generally outperform SCD algorithms in most cases except for the SachsProtein and Asia datasets, which pertain to biomedicine; this demonstrates the great potential of LLMs' commonsense knowledge acquired by pre-training for the task of causal discovery, but meanwhile draws our attention to their shortage of domain-specific knowledge. Hybrid approaches outperform LLM-based baselines in most cases, indicating that using SCD outputs as a prior for LLMs to reference has the benefits of complementing their causal effects related knowledge. Baseline methods that can leverage tools for retrieving external data, i.e. ReAct and LLM-KBCI-RA, sometimes outperform their counterparts, suggesting the potential of the augmented knowledge, but calling for a better way of retrieving and using external data.

The proposed MATMCD and MATMCD-RE achieved the best overall performance, considering that no baseline method consistently performs well across all datasets. In particular, they help alleviate the hallucination problem of other LLM baselines on the biomedical SachsProtein and Asia datasets via data augmentation. The results highlight the challenge of the existing LLM baselines in inferring causal effects solely by meta-data, i.e. node names and data titles, and validate the effectiveness of the proposed design of a multi-agent system for exploring multi-modality enhanced causal discovery. Finally, MATMCD-RE outperforms MATMCD on the Bayesian datasets Asia and Child, which could be a result of the better calibrated likelihoods by Top-K Guess reasoning for the probabilistic datasets.

We also investigate how MATMCD(-RE) refines the causal graphs generated by SCD algorithms by visualizing the true causal graph of the AutoMPG dataset, along with the causal graphs produced by PC, MATMCD and MATMCD-RE. Compared to PC, MATMCD(-RE) yields graphs that resemble the true causality better, with less wrongly added, missed, and reversed edges, suggesting the advantage of leveraging LLM agents' ability to impose multi-modal data empowered knowledge-driven constraints for screening erroneous edges. In contrast, PC alone relies solely on observational data thus is more error-prone. Moreover, LLM-provided knowledge offers precise guidance in establishing causal directions, resulting in fewer reversed edges on most of the datasets.

The ablation analysis uses three datasets, with MATMCD as our original model. We study the usefulness of the iterative search in DA-agent by replacing it with a single-round search template; we assess the effectiveness of the Knowledge LLM in CC-agent by removing it from CC-agent, which is equivalent to merging the Knowledge LLM and the Constraint LLM into a single LLM; we evaluate how MATMCD could generalize to different SCD algorithms by switching PC with ES and DirectLiNGAM; and we assess the impact of different LLMs on causal discovery.

We observe that iterative search is better than single-round search, as the latter may use biased query and have difficulty in producing comprehensive augmented data. Including a separate Knowledge LLM for explaining causal relationships is better than merging it with the Constraint LLM, suggesting their distinct roles in CC-agent. Using other SCD algorithms than PC generally degrades the performance, possibly because the constraint-based design of PC leads to better use of the constraints produced by LLMs; we also observe that, with MATMCD, the performance of ES and DirectLiNGAM were slightly enhanced compared to the counterparts without MATMCD. Intriguingly, the default GPT-4o mini is the most robust across different datasets, and GPT-4 is better than the open source LLMs. Compared to PC, all LLM variants improve performance, suggesting the effectiveness in leveraging the LLMs. Moreover, on the biological dataset SachsProtein, a larger model GPT-4 further boost the causal discovery performance, likely due to its better alignment with the specific domain.

Next, we evaluate MATMCD on AIOps datasets collected from Product Review (PR) and Cloud Computing (CC) microservice systems. The PR dataset has 216 variables, i.e. system pods, each associated with a multivariate time series containing 6 metrics, such as CPU and memory usage, of length 131,329. The CC dataset has 168 variables, each with a multivariate time series of 7 metrics and a length of 109,351. In both datasets, each variable also has a log recording its historical events. Through the log lookup tool, these logs can be leveraged as an additional data modality by MATMCD for enhanced causal discovery. The figure below demonstrates that the Summary LLM of DA-agent can effectively interpret and summarize the log data.

Since these datasets provide root causes of system failures identified by domain experts, but lack ground truth causal graphs, we use them to evaluate the root cause analysis performance based on the causal graphs produced by different methods. If running a random walk with restart on a causal graph can top-rank the root cause among all variables, it reflects the quality of the causal graph; therefore we use the widely adopted Mean Average Precision@K, with K set to 5 and 10, and Mean Reciprocal Rank, to assess the overall ranking performance on both datasets. By leveraging the log modality, MATMCD(-RE) significantly improves the accuracy of root cause locating on both datasets, with an 83.3% relative improvement over the best baseline on MAP@10. This highlights the potential of integrating prevalent log data in microservice systems for RCA and AIOps.