Introducing "Fujitsu Causal AI" #1 Causal Action Optimization Technology - fltech

Hello. We are Takagi, Okajima, Koyanagi, and Ogawa from the Artificial Intelligence Research Laboratory.

Fujitsu has developed "Fujitsu Causal AI" to support data-driven decision-making in enterprises. This technology analyzes causal relationships from corporate data and uses them to recommend actions that are most effective and have no negative impact, based on a wide range of information.

Starting in July 2025, an updated version of this technology has been made available as part of the AI service Fujitsu Kozuchi (R&D) lineup.

Over the course of three articles, starting with this one, we will introduce the appeal and technical background of "Fujitsu Causal AI."

Fujitsu Causal AI was created to solve the following challenges that conventional action recommendation technologies faced:

Difficulty in considering negative impacts (side effects) of actions.
Inability to simultaneously consider multiple causal relationships, potentially leading to sub-optimal action proposals.
Reliance on a single dataset, making it susceptible to data volume and bias, leading to limited action proposals.

To address these challenges, Fujitsu Causal AI consists of the following three core technologies:

Causal Action Optimization Technology
Integrated Causal Discovery Technology
Knowledge-Guided Causal Discovery Technology

We hope this series of articles will provide hints for solving your challenges. At the end of the article, we will also guide you on how to try out this technology.

(1) Causal Action Optimization Technology (This Article)
This technology rapidly analyzes causal relationships between phenomena from numerical data and explains those causal relationships with graphs and natural language sentences. Furthermore, based on the results of these causal relationships, it is a technology that recommends the most effective actions with no negative impact.

(2) Case Study of Causal Action Optimization Technology (Article scheduled around November 1421)
We will introduce a case study of our pilot project that utilized this technology for parameter design in the beer brewing process.

(3) Integrated Causal Discovery Technology & Knowledge-Guided Causal Discovery Technology (Article scheduled around November 28)
The Integrated Causal Discovery Technology analyzes integrated causal relationships from multiple datasets across different domains, enabling explanations of broad relationships that could not be explained by observed datasets from a single domain. Additionally, the Knowledge-Guided Causal Discovery Technology leverages past causal relationship graphs as prior knowledge during causal discovery, achieving highly reliable analysis even with small datasets.

Why is it necessary to capture causal relationships?

In data analysis, correlation only tells us things like "they increase or decrease together" or "they move at the same time." However, what businesses truly want to know is the causality: "Which indicators should be acted upon to achieve the desired results?"

For example, there's a high correlation between "ice cream sales" and "drowning incidents," but no direct mechanism connecting the two. The true cause is a third variable, "hot days," which affects both. Similarly, the correlation graph between "country-specific chocolate consumption and the number of Nobel laureates" is famous, but distributing chocolate won't increase awards. To avoid being misled by such spurious correlations and to derive correct actions, causal analysis, which can simulate the effect of intervention ("If X changes, how will Y change?"), is indispensable.

What can Causal Action Optimization do?

The goal of causal analysis is to dismantle "cause and effect" into a diagram and translate it into concrete actions for achieving desired results. "Fujitsu Causal AI" offers two values through its causal action optimization:

1) Causal Discovery: First, it can reveal a graph structure (causal graph) where causes and effects are connected by arrows from observed data.

2) Action Proposals: Using the obtained causal graph, it proposes how much to intervene in which variables under objectives and constraints such as "increase 'sales' by 10% without increasing 'overtime hours'."

Examples:

Manufacturing: Given desired beer properties (sensory indicators like pH, bitterness IBU) and factory settings (fermentation temperature, hop addition amount), it analyzes the causal relationship between them and the finished product, proposing optimal manufacturing settings.
Organizational Engagement: Automatically proposes the optimal combination of HR actions (training hours, weighting of performance evaluation, welfare benefits...) to improve employee satisfaction while preventing a decrease in productivity.

What are the challenges of existing technologies?

To achieve these, two major hurdles must be overcome:

Scalability barrier: Representative causal discovery algorithms (PC, FCI, DirectLiNGAM, etc.) experienced an exponential increase in computation when variables exceeded a few hundred, making them unable to handle thousands of variables common in real-world IoT sensor data or HR data.
Side effects of actions: Even if the causal structure is known, there were few adequate libraries for simultaneously optimizing multiple objectives, such as "increase A while decreasing B." Unexpected tradeoffs often occurred, such as "10% improvement in A led to a 20% deterioration in B."

How is Causal Action Optimization achieved?

Fujitsu addresses these challenges by centering its solution on its proprietary algorithm, Layered LiNGAM, and optimization technology, which were presented at the international conference ECML-PKDD 2024. fltech - Fujitsu Research Technical Blog

High-speed Causal Discovery: "Layered LiNGAM" dramatically speeds up the estimation of causal order, which was traditionally a computational bottleneck, by processing variables in a layered structure. In experiments, it achieved over 100 times faster performance compared to representative algorithms like DirectLiNGAM, enabling analysis of thousands of variables in real data within a practical timeframe.
Optimization under complex constraints: By applying constrained optimization to the derived causal graph, it can propose realistic actions that simultaneously consider complex business requirements (constraints), such as "maximize KPI B while maintaining KPI A" or "total cost within ... yen."

Here, we introduce the basic idea of LayeredLiNGAM. Typically, causal discovery is performed in two stages: (1) estimation of the causal order (topological order) and (2) estimation of edges according to that order. For (1), we use the following lemma:

If we create simple regression residuals $r_i^{(j)}=x_i-\hat\beta_{ij}x_j$ for each [tex j], then the following holds:

$x_j\ \text{is exogenous}\ \Longleftrightarrow\ x_j\ \text{is independent of all}\ {r_i^{(j)}}_{i\neq j}$

In other words, the causal order is determined by sequentially extracting $x_j$ with the highest "degree of independence." Exogeneity identification requires independence tests between variables and residuals. If the computational cost of one test is $C$ and the number of variables is d, the computational cost of order estimation reaches $\mathcal{O}(C\,d^{3})$ . This process has been a bottleneck. Therefore, while DirectLiNGAM extracts exogenous variables one by one, Layered LiNGAM simultaneously extracts exogenous sets (layers). By introducing a layered order $\mathcal{LO}$ , layers satisfying

$x_i=\sum_{j:\ \mathcal{LO}(j) \lt \mathcal{LO}(i)} b_{ij}x_j + e_i$

are successively removed from the top. Since the residual system $\mathbf{r}^J$ obtained by removing the exogenous set $X_J$ through multiple regression also satisfies LiNGAM (a set-wise version of the lemma), this process is recursively applied to obtain the layered order. As a result, the number of loops is reduced from "number of variables $d$ " to "number of layers $T\le d$ ", and the computational cost for order estimation is generally reduced to approximately

$\mathcal{O}\left(C\,T\,d^{2}\right)\quad(C:\ \text{computational cost of one independence measure})$

This is particularly effective for large-scale problems with fewer layers.

This ingenuity has led to a significant acceleration compared to existing technologies. For more details, please refer to the tech blog article that explains the details. fltech - Fujitsu Research Technical Blog

How will LLMs (Large Language Models) change analysis?

Furthermore, by incorporating LLM-powered support technology, we have made the analysis process smoother and easier to understand. Specifically, it supports users in the following three ways:

Prior Knowledge Setting Support: Prevents omissions in "prior knowledge" that users set before causal discovery.
Causal Graph Interpretation Support: Helps users interpret multiple causal graphs generated as a result of the analysis. It also assists in understanding the relationships of variables of interest within the causal graph.
Recommended Action Confirmation Support: Provides an easy-to-understand explanation of why the recommended action is effective by confirming its consistency with the causal graph.

This figure shows the overall flow of the technologies and features, including causal discovery and causal action optimization. Let's look at each technology in detail.

Prior Knowledge Suggestion Technology

To improve analysis accuracy, setting prior knowledge such as "A can cause B, but not vice versa" is crucial. This technology supports this task by suggesting candidate prior knowledge.

In causal discovery for this technology, the following four types of prior knowledge can be set:

Causal paths are set when one variable influences another (e.g., "advertising cost -> sales" specifies that advertising cost influences sales).
No direct causal relations are set when there is no causal relationship (e.g., "height -> academic performance" specifies that no direct causal relationship from height to academic performance should be derived).
Origin variables are set when they are not influenced by other variables (e.g., "birthplace, age" specifies that no causal relationships where these are influenced should be derived).
Sink variables are set when they do not influence other variables (e.g., "customer satisfaction" specifies that no causal relationships where customer satisfaction influences others should be derived).

This technology uses an LLM to generate and present candidate prior knowledge from the uploaded dataset's variable names, together with supporting evidence and confidence level. As a result, users can set up prior knowledge easily while reducing omissions.

Causal Graph and Action Explanation Technology

To facilitate the interpretation of complex causal graphs obtained from analysis results and to enhance the understanding of recommended actions, the LLM provides a function to generate explanations in natural language.

Causal Graph Explanation

Summary Graph Generation: From multiple causal graphs derived from the analysis, a "summary graph" is generated by extracting representative and important causal structures. The summary graph is optimized by setting rules such as the frequency of causal relationships between items and ensuring relationships are not cyclical. This concisely summarizes the relationships between items contained in multiple causal graphs, providing a clear visualization of the overall causal relationships between variables.
Enumeration and Explanation of Important Paths: From the summary graph, particularly noteworthy important causal chains (paths) are extracted, and the LLM explains "why such a relationship exists" in an easy-to-understand way, even for those without specialized knowledge. For example, specific explanations like "If resources are abundant, employees can work without excessive strain, which positively impacts employee performance" become possible. This explanation allows users to intuitively grasp the structure and impact of key causal relationships without having to interpret the entire complex causal graph.

Action Explanation

The LLM correlates the recommended actions with important paths on the causal graph and specifically explains "why this action is recommended" and "on which causal relationships its effectiveness is expected." For instance, an explanation such as "This action aims to improve performance by increasing overtime hours. However, the causal graph also suggests that increased overtime hours can reduce work-life balance" can be obtained. This enables users to easily evaluate the validity of actions and make decisions with confidence.

Explanation Generation for Action Content and Causal Graph Relationship

Would you like to try Fujitsu Causal AI?

The causal discovery and action optimization features we introduced in this article are readily available via API and GUI through Fujitsu's AI service "Fujitsu Kozuchi."

No programming is required. Simply upload a CSV file to visualize causal graphs from your data and obtain a list of candidate actions.

Use Case	Input	Output proposed by Kozuchi
Beer Development	Data on taste chart (bitterness, sweetness, aroma, etc.) and factory settings (fermentation temperature, hop addition amount, etc.), and target values for the taste chart	Entire manufacturing recipe, including fermentation temperature/time, malt blending ratio, hop addition timing, etc.
Employee Engagement	Current survey results and target scores	Specific HR actions such as training content, peer bonus amounts, frequency of 1-on-1s with supervisors.

Without programming, you can visualize causal graphs and obtain a list of candidate actions simply by uploading CSV files. Please try to instantly verify "What if...?" with your own data.

Paper Presentation (ECML-PKDD 2024):
Presented "LayeredLiNGAM," which speeds up LINGAM, a representative model for statistical causal discovery.
▶ Technical Blog Explanation
▶ Paper (Springer)

Application Examples (Materials Informatics Special Feature):
▶ #9: What causes changes in material properties? Causal Discovery AI answers!
▶ #10: Application of Causal Discovery AI to Semiconductor Device Design

Press Releases:
▶ Update on "Fujitsu Kozuchi" Causal Discovery Technology (2025/03/06)
▶ Launch of "Fujitsu Kozuchi" AI Service to Accelerate DX in a Wide Range of Business Systems (2023/05/17)
▶ Developed AI Technology to Incorporate Human Judgments and Hypotheses to Discover Highly Accurate Causal Relationships (2020/12/17)

fltech - Technology Blog of Fujitsu Research

A technology blog where Fujitsu researchers talk about a variety of topics

Introducing "Fujitsu Causal AI" (Total of 3 Parts) #1 Causal Action Optimization Technology