Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Lossy Compression of Cellular Network KPIs

Andrea Pimpinella, Fabio Palmese, Alessandro E. C. Redondi

TL;DR

The paper tackles the challenge of storing and analyzing massive cellular-network KPI time-series by adopting a task-centric rate–distortion framework. It compares PCM, DPCM, DCT, and KLT with uniform quantization to compress KPIs (downlink volume, PRB occupancy, active users) and evaluates the impact on downstream analytics, including core-network aggregation and Median Weekly Signature forecasting. Key findings show that 3–4 bits per sample at around 30 dB per-cell reconstruction enables 8–16× compression relative to 32-bit floats, and aggregation across thousands of cells preserves analytics quality, with negligible forecasting degradation when per-cell SNR exceeds ~30 dB. The results demonstrate that KPI compression can be both efficient and transparent to network analytics, guiding practical reporting strategies and motivating future work on spatially-aware and learning-based approaches.

Abstract

Network Key Performance Indicators (KPIs) are a fundamental component of mobile cellular network monitoring and optimization. Their massive volume, resulting from fine-grained measurements collected across many cells over long time horizons, poses significant challenges for storage, transport, and large-scale analysis. In this letter, we show that common cellular KPIs can be efficiently compressed using standard lossy compression schemes based on prediction, quantization, and entropy coding, achieving substantial reductions in reporting overhead. Focusing on traffic volume KPIs, we first characterize their intrinsic compressibility through a rate-distortion analysis, showing that signal-to-noise ratios around 30 dB can be achieved using only 3-4 bits per sample, corresponding to an 8-10x reduction with respect to 32-bit floating-point representations. We then assess the impact of KPI compression on representative downstream analytics tasks. Our results show that aggregation across cells mitigates quantization errors and that prediction accuracy is unaffected beyond a moderate reporting rate. These findings indicate that KPI compression is feasible and transparent to network-level analytics in cellular systems.

Lossy Compression of Cellular Network KPIs

TL;DR

The paper tackles the challenge of storing and analyzing massive cellular-network KPI time-series by adopting a task-centric rate–distortion framework. It compares PCM, DPCM, DCT, and KLT with uniform quantization to compress KPIs (downlink volume, PRB occupancy, active users) and evaluates the impact on downstream analytics, including core-network aggregation and Median Weekly Signature forecasting. Key findings show that 3–4 bits per sample at around 30 dB per-cell reconstruction enables 8–16× compression relative to 32-bit floats, and aggregation across thousands of cells preserves analytics quality, with negligible forecasting degradation when per-cell SNR exceeds ~30 dB. The results demonstrate that KPI compression can be both efficient and transparent to network analytics, guiding practical reporting strategies and motivating future work on spatially-aware and learning-based approaches.

Abstract

Network Key Performance Indicators (KPIs) are a fundamental component of mobile cellular network monitoring and optimization. Their massive volume, resulting from fine-grained measurements collected across many cells over long time horizons, poses significant challenges for storage, transport, and large-scale analysis. In this letter, we show that common cellular KPIs can be efficiently compressed using standard lossy compression schemes based on prediction, quantization, and entropy coding, achieving substantial reductions in reporting overhead. Focusing on traffic volume KPIs, we first characterize their intrinsic compressibility through a rate-distortion analysis, showing that signal-to-noise ratios around 30 dB can be achieved using only 3-4 bits per sample, corresponding to an 8-10x reduction with respect to 32-bit floating-point representations. We then assess the impact of KPI compression on representative downstream analytics tasks. Our results show that aggregation across cells mitigates quantization errors and that prediction accuracy is unaffected beyond a moderate reporting rate. These findings indicate that KPI compression is feasible and transparent to network-level analytics in cellular systems.
Paper Structure (12 sections, 3 equations, 3 figures)

This paper contains 12 sections, 3 equations, 3 figures.

Table of Contents

  1. Introduction
  2. Methodology
  3. Dataset
  4. KPI Time-Series Model
  5. Quantization Model
  6. Prediction and Transform-Domain Representations
  7. Rate--Distortion Metrics
  8. Rate--Distortion Analysis
  9. Impact on Downstream Analytics
  10. Aggregated KPI Accuracy
  11. Forecasting Performance
  12. Conclusions

Figures (3)

  • Figure 1: Rate--distortion characteristics for the three considered KPIs. SNR is reported as a function of the entropy-estimated rate for PCM, DPCM, DCT, and KLT.
  • Figure 2: Aggregate SNR as a function of the per-cell SNR for three representative KPIs: (a) PRB, (b) Active Users (RRC), and (c) Volume (VOL). Results are shown for different numbers of aggregated cells ($N=10,100,1000$). In all cases, aggregation by summation yields a substantial SNR gain that increases with $N$, reflecting the attenuation of uncorrelated quantization noise across cells.
  • Figure 3: Forecasting performance using the Median Weekly Signature (MWS) predictor. Average RMSE as a function of the per-cell SNR for three representative KPIs: (a) PRB occupancy, (b) number of active users (RRC), and (c) downlink traffic volume. Errors are reported in KPI-native units (percentage for PRB, number of users for RRC, and MB/hour for volume). Results are shown for original and KLT-quantized data. Across all KPIs, forecasting accuracy remains stable once a moderate cell-level SNR is achieved.