Exploiting Cell Similarities in a Radio Access Network to Enhance Explainability for Autonomic Network Management Systems
A better-performing cell is useful as a guide only if the comparison makes sense. Inspect the patterns behind cell similarity, and see what a high correlation can leave out.
Why is this cell a useful comparison?
An autonomic network-management system may propose copying a configuration from a better-performing cell. That raises a prior question: are the two cells meaningfully similar in how they are used?
We examine time-series performance patterns across multiple counters and make that comparison inspectable. The Similarity Explainability Service also uses model explanations to show how individual counters relate to a target KPI prediction.
A matching pattern can hide a different operating level
Compare the target with two constructed cells across two counters. Cell A follows the target’s throughput pattern exactly, at a different level, but its modulation pattern runs in the opposite direction. Cell B is a close match on both counters.
Inspect the similarity, counter by counter
Target cellCell A
Twelve equal observation windows, arbitrary counter units. The vertical scale stays at 0–100 in every view. Colours and solid/dashed lines distinguish the two cells.
- Pearson correlation r
- 1.000
- Target mean
- 21.75
- Comparison mean
- 49.15
| Counter | Cell A | Cell B |
|---|---|---|
| Throughput | 1.000 | 0.990 |
| Modulation | −1.000 | 0.993 |
Cell A’s throughput correlation is exactly 1 because A = 1.8 × target + 10. Its modulation correlation is −1. Cell B follows both patterns closely. One matching counter is not enough to establish similarity across the operating environment.
Read the plotted values
| Window | Target | Comparison |
|---|---|---|
| 1 | 12 | 31.6 |
| 2 | 10 | 28.0 |
| 3 | 8 | 24.4 |
| 4 | 9 | 26.2 |
| 5 | 18 | 42.4 |
| 6 | 32 | 67.6 |
| 7 | 44 | 89.2 |
| 8 | 40 | 82.0 |
| 9 | 30 | 64.0 |
| 10 | 24 | 53.2 |
| 11 | 20 | 46.0 |
| 12 | 14 | 35.2 |
The paper uses Pearson correlations across 35 performance counters. The two counters and all values here are constructed to explain that measure; they are not extracted network traces or SHAP results.
What the service adds
The research used three weeks of performance data from 700 LTE FDD cells, together with a configuration snapshot. We selected 35 performance counters in 11 categories, including throughput, bearer usage, resource-block usage, modulation and sleep behaviour.
The Similarity Explainability Service stores pairwise, per-counter Pearson correlations. Category summaries help users inspect which aspects of operation make another cell a relevant comparison. Scalar averages provide additional context about operating levels.
A separate explanation step uses SHAP to examine each counter’s contribution to a model’s target-KPI prediction. The paper presents a directional summary alongside the similarity results. Correlation describes pattern agreement; SHAP explains the prediction model. The visual above calculates the former.
Similarity is specific to the measurements
Pearson correlation measures linear co-variation. It is unchanged by a positive rescaling and an offset, as Cell A illustrates. Equal correlations therefore do not establish equal traffic volume or capacity. A constant trace has undefined Pearson correlation.
Agreement across observed counters does not prove that copying a configuration will cause the same outcome. The relevance of the measurements, the target KPI and the observation period still matter. SHAP attributes a model’s predictions; it does not turn those attributions into causal effects. Mean absolute SHAP values describe magnitude, with direction requiring separate information.
The paper demonstrates similarity reports and explanatory views. It does not establish that these comparisons guarantee safe reconfiguration or improve every autonomic controller.
Making comparisons useful for action
PECDAFs develops the use of comparable cells to flag likely KPI degradation before a proposed configuration change. CAMINO brings intent and environmental context into the assessment.
These works form part of my autonomous network-management research. Their connections concern the practical use of evidence; they do not imply that each paper validates a single integrated deployment.
Paper and citation
J. Armstrong, S. Fallon, E. Fallon, “Exploiting Cell Similarities in a Radio Access Network to Enhance Explainability for Autonomic Network Management Systems,” 2023 12th International Conference on Control, Automation and Information Sciences (ICCAIS), pp. 775–780, 2023. DOI: 10.1109/ICCAIS59597.2023.10382371.