Our work appeared in IEEE INFOCOM 2024 as:
- "AIChronoLens: Advancing Explainability for Time Series AI Forecasting in Mobile Networks": this is the paper describing the tool and will appear at the main event.
- "An In-Depth Analysis of Advanced Time Series Forecasting Models for the Open RAN": in this workshop paper, we benchmark different AI models (PatchTST, DLinear, and LSTM ) when they are utilized for single-step and multi-step predictions and discuss about integration of the tool in the Open RAN architecture.
- "Dissecting Advanced Time Series Forecasting Models with AIChronoLens": this is a poster that benchmarks different AI models (PatchTST, DLinear, and LSTM ) when they are utilized for single-step predictions.
🚩 Best main conference paper award on IEEE Infocom 2024
AIChronoLens is a tool for enhancing the explanation of AI models that are applicable to time series forecasting. AIChronoLens resolves the ambiguity of legacy explainable techniques in assigning the same relevance scores to highly diverse input sequences by exploring the Pearson correlation between relevance scores and an enriched expressiveness of the input sequence achieved with the Gramian Angular Field (GAF). This makes possible for AIChronoLens to dig deeper into the model operation, benchmark under equal conditions different AI models applied on the same dataset and spot the hidden causes of errors.
In our work we have used two datasets:
- EU Metropolitan Area (EUMA)
- RRC Connected Users (USERS-https://git2.networks.imdea.org/wng/madrid-lte-dataset)
The EUMA dataset contains aggregated traffic load from an European Metropolitan Area during 15 days taken in 2019. Disclaimer: we cannot release the EUMA dataset publicly.
The USERS dataset contains an estimation of RRC connected users over time. The base stations are located in two different areas of Madrid. BS 1-3 (zone I) are located within the diameter of the orbital highway M-30 circling all the central districts of Madrid municipality. BS 4-6 (zone II) are located in a suburbian area outside the orbital highway beforementioned.
The folder structure of the repository is as following:
- Manuscript_details contains the papers in .pdf format among the bibtex.bib files for referencing the papers.
- Figures contains the figures of the papers.
- Artifacts directory includes the code necessary to generate AIChronoLens results. It include the USERS dataset as an example along with the time series forecasting models PatchTST, DLinear, and LSTM used in "An In-Depth Analysis of Advanced Time Series Forecasting Models for the Open RAN" and "Dissecting Advanced Time Series Forecasting Models with AIChronoLens". While it currently includes these examples, the intention is to allow users to incorporate their own datasets and models.
- Artifacts/Main_conference_scripts directory contains the scripts used to generate the results that are specific for the main conference paper such as the clustering and sharpness score.
You can cite the papers as:
C. Fiandrino, E. Pérez Gómez, P. Fernández Pérez, H. Mohammadalizadeh, M. Fiore & J. Widmer (2024). AIChronoLens: Advancing Explainability for Time Series AI Forecasting in Mobile Networks. In IEEE International Conference on Computer Communications. [bibtex], [pdf], [dspace]
P. Fernández Pérez, C. Fiandrino, M. Fiore & J. Widmer (2024). An In-Depth Analysis of Advanced Time Series Forecasting Models for the Open RAN. In NG-OPERA@IEEE INFOCOM WORKSHOPS. [bibtex], [pdf], [dspace]
P. Fernández Pérez, C. Fiandrino, M. Fiore & J. Widmer (2024). Dissecting Advanced Time Series Forecasting Models with AIChronoLens. In IEEE INFOCOM POSTERS. [bibtex], [pdf], [dspace]
P. Fernández Pérez, C. Fiandrino, E. Pérez Gómez, H. Mohammadalizadeh, M. Fiore & J. Widmer (2025). AIChronoLens: AI/ML Explainability for Time Series Forecasting in Mobile Networks. In IEEE TRANSACTIONS AND MOBILE COMPUTING. [bibtex], [pdf], [dspace]
For any enquiry on the work, please reach out to Claudio Fiandrino (claudio.fiandrino@imdea.org)