An Influence-based Approach for Root Cause Alarm Discovery in Telecom Networks

Keli Zhang, Marcus Kalander, Min Zhou, Xi Zhang, Junjian Ye

Abstract

Alarm root cause analysis is a significant component in the day-to-day telecommunication network maintenance, and it is critical for efficient and accurate fault localization and failure recovery. In practice, accurate and self-adjustable alarm root cause analysis is a great challenge due to network complexity and vast amounts of alarms. A popular approach for failure root cause identification is to construct a graph with approximate edges, commonly based on either event co-occurrences or conditional independence tests. However, considerable expert knowledge is typically required for edge pruning. We propose a novel data-driven framework for root cause alarm localization, combining both causal inference and network embedding techniques. In this framework, we design a hybrid causal graph learning method (HPCI), which combines Hawkes Process with Conditional Independence tests, as well as propose a novel Causal Propagation-Based Embedding algorithm (CPBE) to infer edge weights. We subsequently discover root cause alarms in a real-time data stream by applying an influence maximization algorithm on the weighted graph. We evaluate our method on artificial data and real-world telecom data, showing a significant improvement over the best baselines.

Introduction

In modern telecommunications, hardware failures rarely occur in isolation. Due to the deeply interrelated nature of network devices, a single fault often triggers a "flood" of secondary alarms cascading across the system. For network engineers, identifying the "root cause" alarm within these millions of events is a high-stakes challenge essential for rapid failure recovery. Traditional methods rely heavily on manual labor or rigid, expert-defined rules that struggle to scale with evolving network topologies.

This research presents a novel unsupervised framework that treats root cause discovery as an influence maximization problem. By modeling how alarms influence one another through a hybrid causal graph and representation learning, the proposed system identifies root causes in real-time data streams with high accuracy and efficiency.

System Architecture

The framework is divided into two primary phases: an offline process for graph construction and an online process for real-time prediction.

Figure 1: The dual-process architecture for root cause discovery, featuring offline graph construction and online influence ranking.

As illustrated in Figure 1, the offline process utilizes historical alarm data and network topology to construct an influence graph. This graph is then used by the online process to rank incoming alarms in real-time, identifying the most likely root cause candidates based on their propagation potential.

Methodology

The proposed solution consists of three core components: HPCI for graph structure learning, CPBE for edge weighting, and IRIE for root cause ranking.

Hybrid Causal Graph Learning (HPCI)

Constructing an accurate causal graph is difficult due to redundant edges and indirect causal links. The researchers developed HPCI (Hawkes Process-based Conditional Independence), a hybrid method that combines the temporal modeling capabilities of Hawkes Processes with the pruning power of Conditional Independence (CI) tests. While Hawkes Processes effectively capture how past events trigger future ones, they often over-report connections. CI tests are applied to ensure only direct causal links are preserved, resulting in a more robust and sparse causal map.

Causal Propagation-Based Embedding (CPBE)

Once the graph structure is established, determining the "strength" of causal relationships is critical. The Causal Propagation-Based Embedding (CPBE) algorithm uses representation learning to infer edge weights.

Figure 2: Causal context generation process. Alarm transactions are mapped onto the causal graph, and Depth-First Search (DFS) is used to extract causal sequences (contexts) for embedding.

As shown in Figure 2, the system generates "causal contexts" by traversing the causal graph using Depth-First Search (DFS) or random walks based on observed alarm transactions. These sequences are fed into a skip-gram model (similar to Word2Vec) to learn node embeddings. The vector similarity between nodes is then used to define the weights of the edges in the influence graph.

Influence Ranking (IRIE)

The final stage views the root cause as a "seed" node in an influence diffusion model. Using the IRIE algorithm, the system ranks alarms based on the total influence they exert over other observed alarms in a given transaction. This approach shifts the perspective from simple search to social network-style influence analysis, where the most "influential" alarm is identified as the root cause.

Experimental Evaluation

The framework was evaluated using both synthetic datasets and real-world telecom data. The real-world dataset comprised over 672,000 alarm records from a major carrier, with 6,000 manually labeled transactions for validation.

Key results include:

  • Graph Accuracy: HPCI achieved an F1-score of 0.612, significantly outperforming standard PC-algorithm variants (0.19–0.44).
  • Localization Accuracy: The system achieved a Top-1 accuracy of 61.8% and a Top-5 accuracy of 96.1% in identifying the correct root cause.
  • Efficiency: The online process is highly efficient, processing 6,000 transactions in approximately 5.5 minutes, making it suitable for live network monitoring.

Conclusion

This influence-based approach provides a powerful, data-driven alternative to traditional expert-rule systems. By combining causal inference with network embeddings, the framework adapts to network changes without manual intervention. Future work will focus on refining time-window sensitivity in preprocessing and expanding validated causal maps to cover a broader range of alarm types.

Summary updated

Abstract

Alarm root cause analysis is a significant component in the day-to-day telecommunication network maintenance, and it is critical for efficient and accurate fault localization and failure recovery. In practice, accurate and self-adjustable alarm root cause analysis is a great challenge due to network complexity and vast amounts of alarms. A popular approach for failure root cause identification is to construct a graph with approximate edges, commonly based on either event co-occurrences or conditional independence tests. However, considerable expert knowledge is typically required for edge pruning. We propose a novel data-driven framework for root cause alarm localization, combining both causal inference and network embedding techniques. In this framework, we design a hybrid causal graph learning method (HPCI), which combines Hawkes Process with Conditional Independence tests, as well as propose a novel Causal Propagation-Based Embedding algorithm (CPBE) to infer edge weights. We subsequently discover root cause alarms in a real-time data stream by applying an influence maximization algorithm on the weighted graph. We evaluate our method on artificial data and real-world telecom data, showing a significant improvement over the best baselines.
BibTeX
@inproceedings{zhang2020influence,
  author = {Keli Zhang and Marcus Kalander and Min Zhou and Xi Zhang and Junjian Ye},
  title = {{An Influence-based Approach for Root Cause Alarm Discovery in Telecom Networks}},
  booktitle = {ICSOC Workshops},
  year = {2020},
  url = {https://openreview.net/forum?id=Qf7uzRUWJKM},
}
Figure