In this work, we propose a distributed learning framework that classifies users' transportation modes by leveraging heterogeneous networks (HetNets) architecture and employing a federated learning (FL) algorithm. The ultra-densification of small cells and the dynamic mobility of users impact performance by triggering unnecessary handovers. Therefore, information on user mobility allows the network to perform handover management more intelligently and efficiently. The proposed machine learning framework adopts distributed learning using a federated learning algorithm to detect transportation modes, including driving a car, riding a bicycle, walking, and running. In the proposed framework, local models are trained at small cells using user history information inherently distributed on the network side. A macro cell aggregates the local models to get a global model for classifying the transportation modes of users. Training the local models by small cells over user history information addresses critical FL issues, such as non-independent and identically distributed data and system heterogeneity. Simulation results demonstrate that the proposed framework achieves an accuracy of 98.85% in classifying transportation modes, utilizing input features extracted from user history information.

In this paper, we propose a power allocation algorithm to guarantee quality of experience (QoE) while considering the impact of imperfect channel state information (CSI) in multi-connectivity multiple-input multiple-output (MIMO) systems. Multi-connectivity is a promising solution for wireless technologies to fulfill ever-increasing demands for QoE via massive MIMO. However, the performance of MIMO systems mostly depends on the accuracy of the channel estimation algorithm that provides CSI. Under imperfect CSI scenarios, increasing transmit power to satisfy QoE for one user exacerbates interference with others, thus leading to a tradeoff between power consumption and user satisfaction. To guarantee QoE while minimizing transmission power under imperfect CSI, the effect of imperfect CSI is quantified, and a closed form signal-to-interference-plus-noise ratio (SINR) is derived. Then, a two-stage algorithm categorizes UEs into two groups: those that only need single connectivity and those that need dual connectivity (i.e., adaptive connectivity). After classification, transmission power is allocated to guarantee QoE via a power control strategy that minimizes the transmission power. By comparing performance with a single connectivity–based algorithm and fixed multi-connectivity–based algorithms, we show that our proposed algorithm not only satisfies all the UEs in the system but also consumes less transmission power than the benchmarks.