| dc.description.abstract | The emergence of 5G networks and the projected massive growth in mobile traffic demand necessitate accurate time series forecasting for efficient network operations. While AI-based time series forecasting models, particularly transformer-based ones, have shown significant advancements, their black-box nature hinders adoption in production networks due to a lack of interpretability. This opacity creates challenges in troubleshooting and vulnerabilities to adversarial attacks. Traditional explainable AI (XAI) techniques are often inadequate for time series analysis, failing to provide insights into the model's reasoning and the influence of input data characteristics, and often generating impractical explanation objects. To address these limitations, the research introduces two novel explainability techniques: AIChronoLens for univariate time series, which identifies crucial points by correlating SHAP values with Gramian Angular Fields, and ChronoProf for both classification and regression tasks on time series data, explaining feature importance using virtual weights derived from SHAP values. | es |
