Addressing training data sparsity and interpretability challenges in AI based cellular networks

Abstract

To meet the diverse and stringent communication requirements for emerging networks use cases, zero-touch arti cial intelligence (AI) based deep automation in cellular networks is envisioned. However, the full potential of AI in cellular networks remains hindered by two key challenges: (i) training data is not as freely available in cellular networks as in other fields where AI has made a profound impact and (ii) current AI models tend to have black box behavior making operators reluctant to entrust the operation of multibillion mission critical networks to a black box AI engine, which allow little insights and discovery of relationships between the configuration and optimization parameters and key performance indicators. This dissertation systematically addresses and proposes solutions to these two key problems faced by emerging networks.

A framework towards addressing the training data sparsity challenge in cellular networks is developed, that can assist network operators and researchers in choosing the optimal data enrichment technique for different network scenarios, based on the available information. The framework encompasses classical interpolation techniques, like inverse distance weighted and kriging to more advanced ML-based methods, like transfer learning and generative adversarial networks, several new techniques, such as matrix completion theory and leveraging different types of network geometries, and simulators and testbeds, among others. The proposed framework will lead to more accurate ML models, that rely on sufficient amount of representative training data. Moreover, solutions are proposed to address the data sparsity challenge specifically in Minimization of drive test (MDT) based automation approaches. MDT allows coverage to be estimated at the base station by exploiting measurement reports gathered by the user equipment without the need for drive tests. Thus, MDT is a key enabling feature for data and artificial intelligence driven autonomous operation and optimization in current and emerging cellular networks. However, to date, the utility of MDT feature remains thwarted by issues such as sparsity of user reports and user positioning inaccuracy. For the first time, this dissertation reveals the existence of an optimal bin width for coverage estimation in the presence of inaccurate user positioning, scarcity of user reports and quantization error. The presented framework can enable network operators to configure the bin size for given positioning accuracy and user density that results in the most accurate MDT based coverage estimation.

The lack of interpretability in AI-enabled networks is addressed by proposing a first of its kind novel neural network architecture leveraging analytical modeling, domain knowledge, big data and machine learning to turn black box machine learning models into more interpretable models. The proposed approach combines analytical modeling and domain knowledge to custom design machine learning models with the aim of moving towards interpretable machine learning models, that not only require a lesser training time, but can also deal with issues such as sparsity of training data and determination of model hyperparameters. The approach is tested using both simulated data and real data and results show that the proposed approach outperforms existing mathematical models, while also remaining interpretable when compared with black-box ML models. Thus, the proposed approach can be used to derive better mathematical models of complex systems. The findings from this dissertation can help solve the challenges in emerging AI-based cellular networks and thus aid in their design, operation and optimization.