Bringing Causal Intelligence to IBM Prescriptive Process Mining with SAX4BPM

How causal insights make process recommendations faster, smarter, and more explainable.

Enhancing Prescriptive Process Mining with Causal Discovery

IBM Prescriptive Process Mining (PPM) is designed to identify the key drivers that influence process performance and provide data-driven recommendations for improvement.
Today, PPM already employs advanced statistical and machine learning algorithms to detect influencers — activities and factors that appear to impact KPIs such as process duration or cost.

However, while these influencer models are effective in identifying patterns, they often consider all activities in the process during analysis.
This can lead to:

• High computational overhead, especially for large process models,
• Reduced clarity due to the inclusion of weak or indirect influencers, and
• Limited explainability in understanding how exactly one activity impacts another.

To address these challenges, we introduced Causal Discovery into the PPM pipeline using SAX4BPM — an open-source Python framework built for business process explainability.

SAX4BPM combines Symbolic Aggregate approXimation (SAX) and causal inference techniques (such as the LiNGAM algorithm) to identify the true cause-and-effect dependencies between process activities.
By isolating only the causally relevant activities, PPM can now perform faster, more accurate, and more interpretable optimization for duration-type KPIs.

What is SAX4BPM?

SAX4BPM is an open-source Python framework designed for:

• Process model discovery
• Causal execution dependency analysis

SAX4BPM builds causal-execution models from process logs, discovering how activities actually influence each other. At the heart of this is SAX — Symbolic Aggregate approXimation — which transforms noisy numeric data into symbolic sequences that preserve trend patterns.

Why Causal Discovery Matters in PPM

In IBM Prescriptive Process Mining (PPM), performance optimization involves analyzing duration-type KPIs (like cycle time or waiting time) across all activities.
But not every activity actually affects the KPI.

Without causal filtering:

• High computation cost – Every activity is evaluated, even those with no causal effect.
• Noisy recommendations – Irrelevant activities can appear as false influencers.
• Limited interpretability – It becomes difficult to explain why certain activities seem to influence KPIs.

With causal discovery:

We only consider causally relevant activities, i.e., those that have a direct or indirect influence on KPI outcomes.

This causal filtering improves:

• Higher accuracy — The system focuses only on activities that truly affect performance.
• Better performance — Fewer irrelevant nodes reduce computational overhead.
• Improved explainability — The resulting recommendations are easier to trace and justify.

How SAX Works

SAX (Symbolic Aggregate approXimation) is a time-series transformation technique that turns numeric signals into symbolic patterns.

Example transformation:

A sequence like A → B → C → A captures an increasing-then-decreasing trend.

This symbolic form smooths out small fluctuations, making causal discovery robust to noise and missing data.

How Causal Discovery Happens in SAX4BPM

After the event log is preprocessed and transformed into symbolic time-series data using the SAX technique, SAX4BPM applies a causal inference algorithm called LiNGAM — Linear Non-Gaussian Acyclic Model — to uncover cause-and-effect relationships between process activities.

What LiNGAM Does

LiNGAM tries to determine which activities influence others by analyzing how their variations (e.g., duration or timing changes) are related across multiple process instances.
Unlike correlation-based methods, it infers directional influence — i.e., whether “Activity A causes Activity B” or vice versa.

Building the Causal Graph

LiNGAM computes causal strength values between every pair of activities.
The output is a Directed Acyclic Graph (DAG) that visually represents how activities influence one another.

In this DAG:

• Nodes represent activities.
• Arrows (edges) represent causal influence — e.g., “A → B” means Activity A affects Activity B
• Edge Weight → Strength of influence (0.0 – 1.0)

Understanding the Causal Graph (DAG)

A Directed Acyclic Graph (DAG) represents the cause-and-effect relationships discovered among activities.

• Root Nodes : Activities that start causal influence (drivers) - eg: Request Created
• Intermediate Nodes: Both influenced and influencing (chain links) - eg: Service closure with BO Responsibility
• Leaf Nodes: Outcome activities - eg: Request completed with Account Closure, Evaluating Request(NO registered letter)
• Independent Nodes: Unrelated, optional, or isolated activities - eg: Network Adjustment requested

So, PPM preprocessing layer will exclude leaf nodes (KPIs) and independent ones when calculating recommendations for duration type KPIs.

Integrating Causal Discovery into IBM Prescriptive Process Mining

In IBM Prescriptive Process Mining (PPM), the goal is to identify the activities that most significantly influence key performance indicators (KPIs) — such as lead time and waiting time — and then generate actionable recommendations for improvement.

With the integration of Causal Discovery using SAX4BPM, this process becomes more intelligent and targeted.

Here’s how it works in the PPM pipeline:

1. The causal discovery module runs on historical event logs, transforming raw process data into a causal structure.
2. Causally connected activities are identified — those that have a direct or indirect impact on KPI outcomes.
3. Independent or weakly connected activities are automatically excluded from further analysis

This integration enables PPM to focus its optimization algorithms only on the most relevant parts of the process model — those that genuinely drive performance outcomes.

The result:

• Reduced computational load
• Cleaner, more interpretable causal models
• Sharper and more accurate optimization insights

In essence, Causal Discovery acts as a precision filter for PPM — ensuring the system focuses only on what truly matters.

Business Impact

By combining SAX4BPM’s causal modeling with the PPM prescriptive engine, organizations can:

•  Prioritize the activities that truly influence KPIs, avoiding unnecessary analysis.
•  Improve computation speed and accuracy through model simplification.
•  Enhance explainability, delivering “Explainable Prescriptions” — recommendations that clearly show why and how an action improves performance.

This integration elevates PPM from being purely analytical to being causally intelligent.

Conclusion

Integrating Causal Discovery with SAX4BPM into IBM Prescriptive Process Mining marks a significant step toward more intelligent and explainable process optimization.
By uncovering the true cause–effect relationships between activities, PPM can now focus on what genuinely drives performance outcomes — not just what appears correlated.

This causal layer brings measurable improvements in accuracy, scalability, and explainability, enabling faster computations and more reliable recommendations.