AI4C3 – AI/ML Solutions for Communications
Thursday, 4 June 2026, 11:00-12:30, room Sala 2 (1st floor)
Session Chair: Harilaos Koumaras (NCSR Demokritos, GR)
Exploiting Cross-KPI Confidence Patterns to Expand Automation Scope in 5G Network Management
Joss Armstrong (LM Ericsson, Ireland); Sheila Fallon and Enda Fallon (Technological University of the Shannon, Ireland)
Autonomous network management relies on predictions of how key performance indicators will respond to configuration changes. When prediction intervals are trustworthy, the associated decisions can be automated. When they are not, the decision must be escalated for human review. Using conformal prediction at 95% coverage, high-confidence predictions are automated while low-confidence predictions are flagged. Across three European 5G production networks (500 to 12000 cells, 41.8 million escalated predictions), we find that 86–89% of flagged predictions did not actually need escalation: their errors fell within the conformal bound, meaning the prediction intervals were valid. This more than 6:1 over-escalation ratio overwhelms human reviewers. We show that the individual confidence signals that feed the combined score cannot, on their own, further separate bound violations from spurious escalations. However, cross-KPI confidence variance, the dispersion of confidence scores across co-escalated KPIs at the same cell and time, distinguishes two failure modes: uniform low confidence (cell-level weak peers, predictions typically accurate) versus uneven confidence (KPI-specific failure, with 2.7–2.8x higher bound violation rates). This pattern holds across performance KPI categories but can invert for connectivity KPIs, confirming a mechanistic interpretation. A gradient-boosted classifier exploiting cross-KPI features recovers 41–44% of the escalated band at 97% precision while missing fewer than 10% of bound violations, raising total automation scope from 70% toward 80%. A feature ablation across seven feature groups and three networks shows that feature engineering contributes far more than classifier choice, with KPI identity and category interaction features improving recovery by up to 87% over event-level features alone.
A Neuromorphic Approach to LDPC Decoding with Spiking Belief Propagation
Lucas Pelz (KTH Royal Institute of Technology, Sweden); Ahsan Javed Awan (Ericsson, Sweden); Mustafa Ozger (Aalborg University, Denmark & KTH Royal Institute of Technology, Sweden)
The rapid growth of mobile data traffic demands increasingly energy-efficient error correction solutions. Traditional decoding based on Belief Propagation (BP) often relies on the approximate Min-Sum (MS) algorithm for efficiency reasons, leading to performance degradation. While Machine Learning (ML) approaches like Neural Normalized Offset Min-Sum (NNOMS) mitigate this, they also introduce complexity inherent to artificial neural networks. To address this trade-off, this paper designs and evaluates a Spiking Normalized Offset Min-Sum (SNOMS), a Spiking Neural Network-based decoder compatible with Digital Neuromorphic Hardware (DNH) such as Intel's Loihi 2. SNOMS translates the parameter-learned decoding of NNOMS into the event-driven domain. Through simulation, SNOMS achieves performance competitive with its NNOMS baseline, incurring a maximum penalty of only 0.05 dB on Low-Density Parity-Check codes when measuring Block Error Rate. These results points us toward both more efficient trainable decoding algorithms and specialized, energy-efficient neuromorphic decoding hardware, building a foundation for future sustainable wireless communication.
Contextual Mobility: System Level Performance
Alperen Gündoğan, Panagiotis Botsinis, Sameh Eldessoki and Nikolai Bijovski Ribakov (Apple Technology Engineering, Germany); Christian Hofmann (Apple, Germany); Tarik Tabet (Apple Inc., USA)
This paper presents a contextual mobility framework for next-generation wireless networks that enhances quality of service by leveraging AI/ML capabilities at the user equipment (UE) during mobility. Traditional mobility procedures rely solely on signal measurements, which can lead to increased energy consumption, delayed reaction to abrupt signal changes, and unnecessary handovers. We propose a contextual mobility framework that utilizes AI/ML capabilities at the UE to predict mobility events based on user context, radio environment, and traffic patterns. The UE shares mobility contextual events with the network, enabling proactive decision-making and conditional configurations of the target cells. System-level simulations demonstrate improvements in key performance indicators, including reduced radio link failures, handover failures and improved throughput, particularly in challenging scenarios with coverage holes during mobility.
Clustered Latent Space Alignment for Multi-User Semantic Communications
Simone Gentile (Sapienza University of Rome, Italy); Danilo Menegatti (La Sapienza, Italy); Alessandro Giuseppi (University of Rome Sapienza, Italy); Antonio Pietrabissa (Università di Roma La Sapienza, Italy); Emilio Calvanese Strinati (CEA-LETI, France)
Semantic Communication has emerged as a promising paradigm to overcome the limitations of conventional bit-level transmission by focusing on the exchange of task-relevant information rather than raw data. However, when heterogeneous devices rely on independently trained neural encoders and decoders, latent space mismatches arise, generating semantic noise and degrading downstream task performance. This issue becomes particularly critical in multi-user scenarios, where efficiency and reduced computational complexity must be jointly addressed. A cluster-based latent space alignment framework for multi-user semantic communications is introduced. Users are grouped according to a similarity measure reflecting their communication requirements, and clusters are identified through a graph-based procedure that captures structural affinities among users. A shared semantic equalizer is then learned for each cluster, enabling efficient alignment between the access point and users while enforcing per-cluster power constraints. The resulting optimization problem is addressed through an alternating optimization strategy, iteratively updating cluster-level and user-specific equalizers. Simulation results on an image classification task with heterogeneous pre-trained models demonstrate that the proposed approach achieves high task accuracy compared with benchmarks, confirming the framework's effectiveness in balancing semantic alignment accuracy and cost efficiency in multi-user settings.
Measurement-Driven Probabilistic Prediction of Tail Latency in 5G Networks
Charbel Lahoud (Technical University Chemnitz, Germany); Hefdhallah Sakran (Chemnitz University of Technology, Germany); Christian Hofmann (TU Chemnitz, Germany); Chadi Lahoud (Boldyn Networks, Germany); Klaus Mößner (Chemnitz University of Technology, Germany)
Accurate prediction of end-to-end latency is essential for latency-sensitive 5G applications in autonomous driving and railway environments. In dynamic urban environments with non-line-of-sight (NLOS) conditions, handovers, and interference, tail latency events can significantly impact service quality, even with low average delays. This paper investigates whether passive layer-1 physical (PHY) key performance indicators (KPI), augmented by recent RTT history in hybrid models, contain sufficient predictive information to anticipate short-term RTT degradations. Using a large-scale, real-world 5G drive-test dataset comprising over 836,000 timestamped samples, we compare classical time-series methods with VMD-augmented deep learning models. Models are evaluated as probabilistic predictors of RTT exceeding application-relevant thresholds.
