Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
On Time Series Clustering with Graph Neural Networks
Authors: Jonas Berg Hansen, Andrea Cini, Filippo Maria Bianchi
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate the framework s qualitative performance through experiments on synthetic and real-world data, showing its effectiveness in different scenarios. We also provide a heuristic for model selection in unsupervised settings via a self-supervised forecasting loss. Code is available at https://github.com/NGMLGroup/Time-Series-Clustering-with-GNNs. In this section, we present and discuss the experimental results. Section 6.1 compares the proposed framework against several baselines and provides a qualitative cluster analysis on the synthetic datasets; Section 6.2 investigates the impact of using different approaches to build the adjacency matrix and provides a qualitative cluster analysis on the CER dataset; Section 6.3 reports ablation studies to assess the contribution of individual components; and finally, Section 6.4 evaluates the performance of a few variants of the proposed framework. |
| Researcher Affiliation | Academia | Jonas Berg Hansen EMAIL 1Department of Mathematics and Statistics, Ui T The Arctic University of Norway Andrea Cini EMAIL 2Swiss National Science Foundation Postdoc Fellow 3IDSIA USI-SUPSI, Università della Svizzera italiana Filippo Maria Bianchi EMAIL 1Department of Mathematics and Statistics, Ui T The Arctic University of Norway 4NORCE Norwegian Research Centre AS |
| Pseudocode | No | The paper describes methods in text and equations but does not include a clearly labeled pseudocode or algorithm block. |
| Open Source Code | Yes | Code is available at https://github.com/NGMLGroup/Time-Series-Clustering-with-GNNs |
| Open Datasets | Yes | Specifically, to test the effectiveness of the proposed STGNN model to perform clustering in a real-world setting, we consider the Commission for Energy Regulation (CER) dataset (Commission for Energy Regulation (CER), 2012). This dataset contains time series of energy load consumption from over 5,000 Irish homes and businesses, which were collected between 2009 and 2010. Each time series is associated with an electricity customer labeled as residential , SME (small/medium enterprise), or other . URL https://www.ucd.ie/ issda/data/commissionforenergyregulationcer/. |
| Dataset Splits | Yes | We split the datasets along the temporal dimension, where the first 80% of the time series was used for training and the next 10% for validation. The remaining 10% was kept aside for making qualitative plots. |
| Hardware Specification | Yes | RNN-based model requires a 35% longer training time on the same machine with an RTX 4090 GPU. |
| Software Dependencies | No | The paper mentions using Python and PyTorch implicitly due to the nature of the work (GNNs), and references libraries like TSFEL (Barandas et al., 2020) and Pytorch Geometric, but no specific version numbers for these or other core software dependencies are provided. For example, it mentions 'Pytorch Geometric1' with a footnote, but not a version. |
| Experiment Setup | Yes | In each layer of the model, the hidden size H is fixed at 16. The MLP components of the conditional block comprise a single hidden layer and an output layer, both with the same size H and Re LU activation functions. The TCN used as the temporal encoder consists of two temporal layers, with the dilation of the second layer set to 2. The Re LU function is used as the activation in each layer of the TCN. The hidden and output sizes are set to H, and the kernel size of the temporal convolution is set to 3. The TCN is also configured with a skip connection between the temporal layers. Finally, the temporal dimension is aggregated using a linear attention layer. The GNN encoder comprises two message passing layers, implemented as described in Eq. 4, with hidden and output sizes H. The temporal decoder is an MLP with a single hidden layer of size H and an output size TH = 1. In all experiments and configurations, the model was trained for 250 epochs with Adam using a multi-step learning rate scheduler reducing the learning rate by a factor of 0.5 every 50 epochs, with the initial learning rate set to 0.001. The optimizer was also configured with weight decay regularization (L2 penalty) set to 0.0001. The softmax temperature used when computing the cluster assignment matrix was initially set to 1.0 and then decreased linearly by 0.0099 per epoch down to a minimum of 0.01. Gradient clipping was applied during backpropagation with a clip value of 5. In all experiments, a grid search was conducted to find suitable values for the pooling loss coefficients where c1, c2 {0.1, 0.58, 1.06, 1.54, 2.02, 2.5} (six linearly spaced points between 0.1 and 2.5). |