🤖 machine learning

SCOPE: Sequential Causal Optimization of Process Interventions

This paper introduces SCOPE, a Prescriptive Process Monitoring approach that utilizes backward induction and causal learners to optimize sequential interventions from observational data, demonstrating superior performance over existing methods in aligning interventions to improve key performance indicators.

Original authors: Jakob De Moor, Hans Weytjens, Johannes De Smedt, Jochen De Weerdt

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jakob De Moor, Hans Weytjens, Johannes De Smedt, Jochen De Weerdt

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "One-Step" Trap

Imagine you are a manager at a bank, and you want to maximize profit from loan applications. You have a few levers you can pull to help a customer:

Speed up the process (Priority Procedure) vs. Keep it standard (Standard Procedure).
Set the interest rate (High, Medium, or Low).

Most existing computer systems that try to help managers make these decisions suffer from "tunnel vision." They look at each decision in isolation.

The Analogy:
Think of it like playing a video game where you have to make three choices in a row to win.

Bad AI: Looks only at the current screen. It thinks, "If I jump now, I get 10 points." So it jumps. It doesn’t care that jumping now makes you miss a secret door later that would have given you 1,000 points.
The Reality: In business, decisions are linked. If you choose the "Standard Procedure" (cheaper for the bank), the customer might get bored and quit. If they quit, the interest rate doesn't matter anymore. If you choose "Priority" (expensive), the customer stays, but now you must set a high interest rate to make back the money you spent on the speed.

Current methods either:

Only look at one decision at a time (ignoring the future).
Try to simulate the whole process using rough guesses (like a blurry map), which often leads them astray because the simulation isn't perfectly accurate.

The Solution: SCOPE (The "Backward-Planning" Navigator)

The authors created a new method called SCOPE. Instead of guessing or looking only at the next step, SCOPE plans from the end of the process back to the beginning.

The Analogy: The GPS Navigator
Imagine you are driving to a destination.

Standard GPS: Tells you, "Turn left in 500 feet." It doesn't really care why, it just reacts to where you are.
SCOPE: Works like a smart navigator that knows your final destination. It starts at the destination and works backward:
- "To arrive on time, I need to be on Highway 10 by 2:00 PM."
- "To be on Highway 10 by 2:00 PM, I need to take Exit 4 at 1:45 PM."
- "To take Exit 4 at 1:45 PM, I need to leave the house at 1:30 PM."

SCOPE does this for business processes. It asks: "What is the best interest rate to set at the very end to maximize profit?" Then it moves one step back: "Given that best interest rate, what was the best procedure to choose earlier?" It repeats this until it reaches the start of the process.

How It Works: "Causal Learning" (Finding the True Cause)

Business data is messy. Just because two things happened together doesn't mean one caused the other.

Example: Rich customers often get low interest rates. But did the low rate make them rich? No. They were already rich.

If a computer just looks at history, it might think, "Oh, low interest rates are associated with rich people, so I should give low rates to everyone to make them rich!" That’s wrong.

SCOPE uses Causal Learning. This is like a detective that tries to separate "correlation" from "causation." It asks: "If I change only the interest rate, keeping everything else the same, how does the profit change?" This allows SCOPE to figure out the true effect of its interventions, even if those interventions were rarely used in the past.

Why SCOPE is Better

The researchers tested SCOPE against other methods using simulated bank data. They found that:

It sees the big picture: Because it plans backward, it aligns all decisions. It doesn’t pick a "good" first step that leads to a "bad" second step.
It doesn’t need a fake simulation: Other methods try to build a fake version of the bank process (a simulation) to practice on. If the fake version is wrong, the advice is wrong (this is called the "reality gap"). SCOPE learns directly from the real historical data, so it doesn’t suffer from this "blurry map" problem.
It handles complexity: Even when there are many decision points (like 6 different steps in a process), SCOPE keeps getting better, while other methods get confused.

The Trade-off

SCOPE is more complex to build than simple methods. It requires calculating the value of every step from the end to the start. However, the paper shows that despite this complexity, it trains faster than some competing methods and gives much better results in terms of profit (or cost savings).

Summary

Old Way: Make the best decision for right now, ignoring how it affects the future. Or, use a rough simulation that might be inaccurate.
SCOPE Way: Start at the final goal, work backward to figure out the best sequence of moves, and use "causal detective work" to understand what actually changes the outcome.
Result: Better business outcomes because the decisions are coordinated and based on real cause-and-effect, not just historical patterns.

Technical Summary: SCOPE – Sequential Causal Optimization of Process Interventions

Problem Statement
Prescriptive Process Monitoring (PresPM) aims to recommend interventions during running business processes to optimize Key Performance Indicators (KPIs). A critical limitation in existing PresPM literature is the handling of sequential interventions. In realistic settings, interventions are rarely isolated; organizations must align sequences of interventions to jointly steer case outcomes. Current approaches fall short in three primary ways:

Single-Intervention Focus: Many methods address only a single intervention scenario, failing to coordinate multiple decisions across a case’s lifecycle.
Independent Optimization: Some methods handle multiple interventions but optimize each decision point in isolation (e.g., maximizing immediate remaining processing time), ignoring how earlier actions affect the efficacy of later ones. This can lead to suboptimal end-to-end outcomes.
Reality Gap via Approximation: Methods that do address sequential dependencies often rely on Reinforcement Learning (RL) agents trained on process approximations, such as Markov Decision Processes (MDPs) or data augmentation. These approximations may misspecify the true process dynamics, introducing bias and a "reality gap" that degrades performance in practice.

Methodology
The authors propose SCOPE (Sequential Causal Optimization of Process Interventions), a PresPM approach that learns aligned sequential intervention policies directly from observational event logs without requiring process simulators or data augmentation.

Theoretical Foundation: SCOPE is grounded in the framework of Dynamic Treatment Regimes (DTR) and utilizes backward induction to identify the optimal policy. It assumes sequential ignorability, the Stable Unit Treatment Value Assumption (SUTVA), and positivity.
Backward Induction with Regret-Based Value Function: The method iterates backward from the final decision point ( $K$ ) to the first ($1$). At each step, it estimates the Q-function (expected outcome given a prefix and action) and the Value function (maximum expected outcome from that prefix onward). To mitigate error propagation and model misspecification—common in heterogeneous business process prefixes—SCOPE employs a regret-based formulation of the value function. This formulation estimates the relative difference (regret) between the optimal action and the historically observed action, rather than absolute values.
Causal Learners: To estimate the Q-functions from observational data, SCOPE integrates causal learners. The paper evaluates three model-agnostic learners compatible with various base models (e.g., LSTMs, XGBoost):
- S-learner: Uses a single model taking the prefix and action as input.
- T-learner: Builds separate models for each possible action.
- RA-learner: A two-step approach that first estimates Q-values and then constructs pseudo-outcomes to explicitly estimate causal effects, aiming for better stability in causal effect estimation.
Algorithm: During training, SCOPE recursively computes estimated values for each case, carrying forward values for cases that skip certain decision points due to variable control flow. During inference, it moves forward, selecting the optimal action at each decision point based on the trained models.

Key Contributions

SCOPE Framework: The introduction of a PresPM method that combines causal learners with backward induction to derive causally grounded, aligned sequential intervention policies. It operates directly on observational data, avoiding the need for process-specific simulators or log augmentation.
Empirical Evaluation and Benchmark: The authors provide a comprehensive evaluation on an existing synthetic dataset (SimBank) and a novel semi-synthetic dataset (SimBPIC17) based on the real-life BPIC17_W event log. This new benchmark is released publicly to support future research in sequential PresPM.

Results
Experiments demonstrate that SCOPE consistently outperforms state-of-the-art baselines, including:

SEP-S: A method optimizing interventions independently at each decision point (adapted from de Leoni et al.).
KMeans-Q: An RL-based method using process approximation via KMeans clustering and Q-learning (adapted from Branchi et al.).

Key findings include:

Superiority over Independent Optimization: SCOPE significantly outperforms SEP-S across various confounding levels and training sizes, highlighting the benefit of aligning interventions.
Robustness to Confounding: SCOPE maintains strong performance even under high levels of confounding (up to 0.99), demonstrating its ability to identify effective actions rarely observed in historical data.
Scaling with Decision Points: On the SimBPIC17 dataset, SCOPE’s advantage over baselines increases as the number of decision points grows (from 2 to 6), as it better captures interdependent decisions.
Learner and Model Stability: The S-learner and RA-learner generally achieved the strongest performance, while the T-learner showed degradation under high confounding. XGBoost performed best as a base model.
Efficiency: Despite higher theoretical training complexity, SCOPE trained significantly faster than KMeans-Q in practice and matched SEP in inference complexity.

Significance and Claims
The paper claims that SCOPE addresses a critical gap in PresPM by enabling the optimization of sequential interventions without the biases introduced by process approximations. By leveraging causal learning and backward induction, SCOPE allows organizations to move from descriptive/predictive monitoring to active, aligned steering of process executions. The authors emphasize that their approach is robust to the heterogeneous and confounded nature of real-world event logs. The release of the SimBPIC17 benchmark is presented as a reusable resource to facilitate further research in this domain. The authors acknowledge limitations regarding the SUTVA assumption (no interference between cases) and sequential ignorability, noting these are common constraints in the field, but argue that their method remains effective under realistic business process conditions.