An Ensemble Learning Scheme for Indoor-Outdoor Classification Based on KPIs of LTE Network

Lei Zhang, Qin Ni, Menglin Zhai, Juan Moreno, César Briso-RodríguezIEEE Access 2019 · DOI

Abstract

Wireless Big Data has aroused extensive attention, as mass mobile devices have been developed and deployed for the upcoming 5G era. The context information of these devices is of importance for personalized services in a smart environment. Nevertheless, the constant change of scenes challenges to the network operator. In this paper, we propose an ensemble learning scheme for indoor-outdoor classification for a typical urban area, based on the cellular data captured in a commercial LTE network. The variables are extracted by network key performance indicators (KPIs) and radio propagation knowledge. Based on these main variables, the decision trees grow and split by the Gini index of sampled features. Then, all decision trees are assembled as weak learners to build the ensemble scheme, thus improving the discrimination ability. The self-validation results show the ensemble model achieves extreme accurate (with an out-of-bag error lower than 1%) classification for indoor and outdoor environments. Moreover, the prominent variables are selected based on the variable importance of in the initial training. The reconfigured model based on fewer variables and less weak learners also gains the highest accuracy and relative short compute time, compared with other classical machine learning methods.

Introduction

As we transition into the 5G era, "context awareness"—knowing exactly where a user is and what their environment looks like—is vital for personalized services, location-based systems, and network optimization. Traditionally, GPS is the go-to for location, but it has two fatal flaws for this specific task: it is notoriously unreliable indoors due to signal blockage from roofs and walls, and it is a massive drain on mobile battery life. Other approaches using smartphone sensors (like light or magnetic field sensors) raise privacy concerns and add processing complexity. This research proposes a "passive" way to determine a user's environment by using the data the network already generates: Measurement Reports (MRs) and Key Performance Indicators (KPIs).

Methodology and System Model

The proposed ensemble learning scheme utilizes Random Forest (Bagging) to classify environments. The core of the approach lies in "physics-informed" feature engineering, moving beyond raw data to incorporate domain-specific radio propagation knowledge.

System model illustrating signal interactions between outdoor/indoor eNodeBs and User Equipment (UE) within cellular cells.

Feature Engineering

  1. Feature Extraction: 20 variables were extracted based on radio propagation knowledge, including signal strength (RSRP) and quality (RSRQ) for the serving cell and six neighboring cells, as well as Timing Advance (TA) to estimate distance.
  2. The "Delta" Feature: A novel variable was introduced by calculating the difference between the measured RSRP and an estimated RSRP derived from the Cost 231 Hata propagation model. This helps the model identify signal attenuation caused by building penetration.
  3. Ensemble Construction: Multiple decision trees were grown using the Gini index for splitting. By aggregating these "weak learners," the model becomes robust against the noise and fluctuations inherent in radio environments.

Data Collection

The study utilized a real-world dataset captured from a commercial 4G FDD-LTE network in a typical urban environment in Shanghai, China. The dataset included 24,912 data samples filtered from raw measurement reports. Indoor data was collected from five different shopping malls, while outdoor data spanned a five-square-kilometer urban area.

Geographic distribution of eNodeBs and Measurement Reports (MRs) for indoor and outdoor environments in the urban study area.

Key Results and Insights

  • Extreme Accuracy: The initial 20-variable ensemble model achieved an accuracy of 99.5%, with an Out-of-Bag (OOB) error rate below 1%.
  • Efficiency Gains through Reconfiguration: Using "Variable Importance" (VI) metrics, the authors discovered that just 4 of the 20 variables (serving cell RSRP/RSRQ, first neighbor RSRP, and the propagation model delta) provided nearly the same predictive power. The reconfigured model maintained 99.3% accuracy while reducing computation time by 75% (from 19.5s to 4.8s).
  • Benchmarking: The ensemble approach significantly outperformed classical methods. For comparison, Support Vector Machines (SVM) achieved 97.6% accuracy but took nearly seven times longer to process (33.8s) than the optimized ensemble.
  • Handling "Missing" Data: In real-world networks, a phone might not detect all six neighboring cells. Ensemble trees naturally handle these null values better than distance-based algorithms like KNN or hyperplane-based ones like SVM.

Conclusion

This research demonstrates that combining radio propagation physics with ensemble learning creates a high-precision, computationally efficient framework for indoor-outdoor classification. Future work will explore the impact of high-speed movement (e.g., users in vehicles) and the integration of 5G-specific KPIs and beamforming data to provide even more granular context sensing.

Summary updated

Abstract

Wireless Big Data has aroused extensive attention, as mass mobile devices have been developed and deployed for the upcoming 5G era. The context information of these devices is of importance for personalized services in a smart environment. Nevertheless, the constant change of scenes challenges to the network operator. In this paper, we propose an ensemble learning scheme for indoor-outdoor classification for a typical urban area, based on the cellular data captured in a commercial LTE network. The variables are extracted by network key performance indicators (KPIs) and radio propagation knowledge. Based on these main variables, the decision trees grow and split by the Gini index of sampled features. Then, all decision trees are assembled as weak learners to build the ensemble scheme, thus improving the discrimination ability. The self-validation results show the ensemble model achieves extreme accurate (with an out-of-bag error lower than 1%) classification for indoor and outdoor environments. Moreover, the prominent variables are selected based on the variable importance of in the initial training. The reconfigured model based on fewer variables and less weak learners also gains the highest accuracy and relative short compute time, compared with other classical machine learning methods.
BibTeX
@inproceedings{zhang2019ensemble,
  author = {Lei Zhang and Qin Ni and Menglin Zhai and Juan Moreno and César Briso-Rodríguez},
  title = {{An Ensemble Learning Scheme for Indoor-Outdoor Classification Based on KPIs of LTE Network}},
  booktitle = {IEEE Access},
  year = {2019},
  doi = {10.1109/access.2019.2914451},
}
Figure