Towards xApp Conflict Evaluation with Explainable Machine Learning and Causal Inference in O-RAN

This paper proposes a framework for managing conflicts between concurrent xApps in Open RAN by combining explainable machine learning and causal inference to identify conflicting control parameters, model their interactions via causal graphs, and quantify their impacts on network performance.

The Gist

Imagine a busy 5G cell tower as a high-tech kitchen run by a team of specialized chefs. In the old days, one head chef (the network operator) controlled everything. But with the new O-RAN system, the kitchen is open to many different third-party chefs (called xApps) who can jump in to help.

One chef might be an expert at speeding up orders (maximizing throughput), while another is a master at saving electricity (energy efficiency).

The Problem: The Kitchen Chaos

The trouble starts when these chefs don't talk to each other.

• Chef A turns up the stove heat (increases power) to cook faster.
• Chef B turns down the gas (decreases power) to save energy.
• Chef C rearranges the ingredients (changes bandwidth) to optimize the menu.

If they all act at the same time without coordination, they might accidentally burn the food (network crashes) or serve a cold, slow meal (poor performance). This is what the paper calls an "xApp Conflict."

Currently, there is no standard way to know who caused the mess or how bad the mess is. It's like trying to figure out why a cake collapsed when five people were baking it simultaneously, but you have no recipe and no cameras.

The Solution: The "Smart Detective" Framework

The authors propose a new system that acts like a super-smart detective to solve these kitchen fights. They use two main tools:

1. The "Why" Detector (Explainable Machine Learning)

First, they train a computer model to watch the kitchen and predict how the food turns out based on what the chefs do. But a normal computer model is like a black box; it just says "The cake failed."

This team uses a special tool called SHAP (think of it as a magnifying glass). Instead of just predicting the outcome, the magnifying glass highlights exactly which ingredient or action mattered most.

• Example: The magnifying glass reveals that "Turned up the heat" and "Added too much flour" were both pulling in opposite directions, causing the cake to fail.
• The Result: They can now draw a map (a graph) showing which chefs are fighting over which ingredients.

2. The "What If" Simulator (Causal Inference)

Knowing who fought isn't enough; you need to know how much damage they did. This is where Causal Inference comes in. It's like a time-travel simulator.

Instead of just saying "Heat and Flour are bad together," the system asks:

• "If Chef A had left the heat alone, but Chef B still added flour, would the cake have been okay?"
• "If we only changed the heat by 1%, how much would the speed of service change?"

They calculate two things:

• ATE (Average Effect): On average, across all days, how much does turning up the heat hurt the speed?
• CATE (Conditional Effect): How does it hurt specifically on a rainy Tuesday when the kitchen is already crowded? (This helps the manager make smarter, specific decisions rather than a "one-size-fits-all" rule).

The Big Picture

By combining these two tools, the system gives the network operator (the Head Chef) a clear report:

"Hey, the 'Speed Chef' and the 'Energy Chef' are fighting over the 'Power Knob.' Every time they fight, our service slows down by 5%. If we tell the Energy Chef to back off just a little bit, we can keep the speed high without wasting much power."

Why This Matters

This paper is the first to build a framework that doesn't just guess who is causing problems but proves the cause-and-effect relationship using real data. It moves network management from "guessing and hoping" to "knowing and fixing," ensuring that our 5G networks stay fast and reliable even when many different apps are trying to control them at once.

In short: They built a smart referee that watches the digital chefs, spots the arguments before they ruin the meal, and tells the manager exactly how to settle the dispute to keep the customers happy.

Technical Summary

Here is a detailed technical summary of the paper "Towards xApp Conflict Evaluation with Explainable Machine Learning and Causal Inference in O-RAN".

1. Problem Statement

The Open Radio Access Network (O-RAN) architecture enables flexible, multi-vendor 5G deployments by disaggregating base stations and allowing third-party applications (xApps) to run on the near Real-Time RAN Intelligent Controller (near-RT RIC). While xApps manage critical functions like traffic steering and energy efficiency, their concurrent operation often leads to conflicting control actions.

• The Core Issue: Multiple xApps may control different RAN Control Parameters (RCPs) that inadvertently influence the same Key Performance Indicator (KPI), or control the same RCP directly. These conflicts can degrade network performance (e.g., reduced throughput, increased latency).
• The Gap: Existing conflict management strategies (game theory, reinforcement learning, static prioritization) often rely on simplified assumptions (e.g., Gaussian distributions, known relationships) or require vendor cooperation that is unrealistic in commercial multi-vendor environments. There is a lack of a comprehensive framework to systematically identify conflicting RCPs and quantify their causal impact on network KPIs across dynamic network states.

2. Methodology

The authors propose a novel framework that integrates Explainable Machine Learning (ML) and Causal Inference to evaluate xApp conflicts. The methodology follows a four-step pipeline:

A. Data Collection & Regression Modeling

• Data: Observational data is collected from a network simulator (or digital twin), capturing RCP values (inputs) and KPIs (outputs) over time.
• Modeling: A regression model is trained to predict KPIs based on RCPs. The authors tested various models (Decision Trees, Random Forest, XGBoost, MLP, SVR) and found XGBoost to be the most accurate.

B. Conflict Identification via Explainable ML (SHAP)

• Technique: The authors use SHAP (SHapley Additive exPlana-tions) to interpret the trained regression model.
• Goal: To identify which RCPs jointly influence the same KPI.
• Logic: If multiple RCPs exhibit high SHAP values (feature importance) for a specific KPI, they are flagged as potential sources of conflict. This helps construct a Conflict Graph representing direct, indirect, and implicit conflicts.
  • Direct: Two xApps control the same RCP.
  • Indirect: Two xApps control different RCPs affecting the same KPI.
  • Implicit: One RCP influences another, which then affects a KPI.

C. Causal Inference & Impact Estimation

• Causal DAG Construction: Insights from the SHAP analysis are used to build a Causal Directed Acyclic Graph (DAG). This graph maps the causal relationships between RCPs (treatments), KPIs (outcomes), and confounding variables.
• Metrics:
  • Average Treatment Effect (ATE): Estimates the average causal impact of changing an RCP on a KPI across all network conditions.
  • Conditional Average Treatment Effect (CATE): Estimates the impact of an RCP on a KPI under specific network states. This addresses heterogeneity, revealing when a parameter adjustment is effective or harmful depending on the current network context.
• Tools: The framework utilizes the DoWhy and EconML libraries for causal estimation and refutation testing (e.g., placebo tests, random common cause tests) to ensure statistical robustness.

3. Key Contributions

1. Novel Framework: This is the first work to comprehensively integrate ML model explainability (SHAP) with causal inference for xApp conflict management in O-RAN.
2. Conflict Taxonomy & Detection: Provides a systematic method to detect indirect and implicit conflicts (which are harder to spot than direct conflicts) by analyzing feature importance on shared KPIs.
3. Quantitative Impact Analysis: Moves beyond qualitative detection to quantify the exact magnitude of conflict impacts using ATE and CATE, allowing operators to understand the "cost" of a specific control action.
4. Vendor Agnosticism: The approach relies on the underlying causal relationships between parameters and performance metrics, making it applicable regardless of the specific xApp vendor or algorithm used.

4. Experimental Results

The authors evaluated the framework using a MATLAB-based 5G network simulation (single gNB, one UE, video conferencing traffic).

• Regression Performance: XGBoost achieved the highest accuracy ($R^2 \approx 0.93-0.98$) in predicting KPIs (Throughput, Spectral Efficiency, BLER) from RCPs.
• Conflict Identification (SHAP):
  • Throughput: Identified Tx Power as the dominant RCP.
  • Spectral Efficiency: Identified both Bandwidth and Tx Power as significant influencers.
  • BLER: Identified Tx Power and PRBs as key factors.
• Causal Estimation (ATE & CATE):
  • ATE Results: A 1 dBm increase in Tx Power increased Spectral Efficiency by 0.00251, while a 1 MHz increase in Bandwidth decreased it by 0.00320.
  • Robustness: Refutation tests yielded high p-values (>0.05), confirming the causal estimates were not spurious.
  • CATE Insights: The effect of Tx Power on Spectral Efficiency varied across network states (CATE distribution centered around 0.001), demonstrating that a "one-size-fits-all" policy is insufficient. The framework successfully identified that the magnitude of conflict varies by network state.

5. Significance and Future Work

• Significance: The proposed framework empowers Mobile Network Operators (MNOs) to move from reactive conflict resolution to proactive, data-driven policy design. By understanding the causal impact of conflicting xApps, operators can define tolerance thresholds and dynamically adjust policies based on real-time network states (via CATE).
• Limitations: The current evaluation relies on a simplified simulation. Real-world O-RAN testbeds with high-dimensional, complex data are needed for further validation.
• Future Work: The authors plan to deploy this framework on the ns-O-RAN simulator with multiple xApp deployments to generate realistic, high-dimensional datasets and test scalability.

In conclusion, this paper provides a rigorous, mathematically grounded approach to solving the "conflict problem" in O-RAN, shifting the paradigm from heuristic conflict avoidance to precise causal impact evaluation.