The emergence of 5G networks has transformed the mobile ecosystem. Higher access speeds have driven explosive growth in mobile traffic demand. 5G is projected to handle 75 % of all mobile traffic by 2029, compared to just 26 % at the end of 2024. Traffic forecasting at both base station and city levels has become crucial for network operators. It enables optimization across multiple domains including network deployment planning [1] , and enhanced mobility management [2] .

The next-generation of mobile networks will increasingly rely on Artificial Intelligence (AI)/Machine Learning (ML) for effective network automation, resource orchestration and management. This translates into performing classification and regression tasks on time series data. Unfortunately, the existing AI/ML models are inherently complex and hard to interpret, which hinders their deployment in production networks. Further, the vast majority of the existing EXplainable Artificial Intelligence (XAI) techniques are either primarily conceived for computer vision and natural language processing and thus fail to provide useful insights.In this paper, we take the research on XAI for time series classification and regression tasks one step further proposing ChronoProf, a new tool that builds on legacy XAI techniques. By creating a linearized version of the original model for different observations, ChronoProf provides insights about the dynamic changes in the model decision-making process across observations and is agnostic to the influence of feature magnitude, which is a key limitation of legacy explainers. Thus, ChronoProf highlights the real influence of model parameters on the output. Our extensive evaluation with real-world mobile traffic traces shows that ChronoProf is able to measure the feature importance, especially in classification tasks where linearized explanations across observations show high consistency.

Interpreting time series predictions from advanced Machine Learning and Deep Learning (ML/DL) models is challenging, as these models often function as black boxes, limiting their applicability in critical domains. To address this, we leverage CHRONOPROF, an Explainable AI (XAI) technique specifically designed for time series data, built upon the SHAP framework. CHRONOPROF improves interpretability by deriving virtual weights from SHAP values, offering a linearized representation of complex model decisions while preserving temporal coherence. However, CHRONOPROF’s complexity can pose challenges for non-expert users. To mitigate this, we developed an interactive dashboard that simplifies interpretation by retrieving stored data and SHAP values to compute and visualize virtual weights along with other representations that are derived from them. This user-friendly interface enables users to explore model behavior across different models and datasets. Ultimately, this innovation facilitates the adoption of CHRONOPROF and fosters trust in AI-driven network operations.

Forecasting is increasingly considered a fundamental enabler for the management of next-generation mobile networks. While deep neural networks excel at short- and long-term forecasting, their complexity hinders interpretability, a crucial factor for production deployment. The existing EXplainable Artificial Intelligence (XAI) techniques, primarily designed for computer vision and natural language processing, struggle with time series data due to their lack of understanding of temporal characteristics of the input data. In this paper, we take the research on EXplainable Artificial Intelligence (XAI) for time series forecasting one step further by proposingAIChronoLens, a new tool that links legacy XAI explanations with the temporal properties of the input.AIChronoLensallows diving deep into the behavior of time series predictors and spotting, among other aspects, the hidden causes of forecast errors. We show thatAIChronoLens’s output can be utilized for meta-learning to predict when the original time series forecasting model makes errors and fix them in advance, thereby improving the accuracy of the predictors. Extensive evaluations with real-world mobile traffic traces pinpoint model behaviors that would not be possible to identify otherwise and show how model performance can be improved by 32 % upon re-training and by up to 39 % with meta-learning.

Next-generation mobile networks will increasingly rely on the ability to forecast traffic patterns for resource management. Usually, this translates into forecasting diverse objectives like traffic load, bandwidth, or channel spectrum utilization, measured over time. Among the other techniques, Long-Short Term Memory (LSTM) proved very successful for this task. Unfortunately, the inherent complexity of these models makes them hard to interpret and, thus, hampers their deployment in production networks. To make the problem worsen, EXplainable Artificial Intelligence (XAI) techniques, which are primarily conceived for computer vision and natural language processing, fail to provide useful insights: they are blind to the temporal characteristics of the input and only work well with highly rich semantic data like images or text. In this paper, we take the research on XAI for time series forecasting one step further proposing AIChronoLens, a new tool that links legacy XAI explanations with the temporal properties of the input. In such a way, AIChronoLens makes it possible to dive deep into the model behavior and spot, among other aspects, the hidden cause of errors. Extensive evaluations with real-world mobile traffic traces pinpoint model behaviors that would not be possible to spot otherwise and model performance can increase by 32%.