A model for representing and a method for feature extraction from streaming mobile network data

Abstract

The article addresses the problem of constructing a consistent feature representation of heterogeneous streaming data generated by mobile network components. The relevance of the study is determined by the asynchronous nature of data streams, different sampling rates, structural heterogeneity, missing observations, and varying data quality, all of which complicate their direct use in machine learning models. The aim of the study is to develop a model for representing streaming information and a method for feature space construction that preserve the temporal and structural characteristics of individual data sources and provide an integrated representation of the mobile network state. The proposed model describes network components, their data streams, individual time scales, dimensionality of primary observations, data quality estimates, and inter-component relationships without requiring all streams to be preliminarily resampled to a common sampling frequency. The developed method includes synchronization of data within common time windows, statistical aggregation, consideration of data availability, quality, and freshness, extraction of temporal and event-based features, and integration of local features into a fixed-dimensional network state vector. A distinctive feature of the method is the explicit preservation of information about observation availability, which prevents missing data from being interpreted as actual zero values of network indicators. The experimental study was conducted by comparing seven variants of feature representation using the same classification algorithm. The results showed an increase in macro-F1 from 0.781 for the simplest representation to 0.928 for the full integrated representation. The robustness analysis under incomplete data demonstrated lower performance degradation for the proposed representation compared with the use of the latest available values and resampling with interpolation. The obtained results confirm the effectiveness of explicitly accounting for data heterogeneity, temporal dynamics, quality, and availability when preparing streaming mobile network data for machine learning tasks.