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 Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Shape-Informed Clustering of Multi-Dimensional Functional Data via Deep Functional Autoencoders
Authors: Samuel V. Singh, Shirley Coyle, Mimi Zhang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Section 5 presents FAEclust against state-of-the-art methods, and Section 6 discusses implications and future directions. We applied all eight FD clustering algorithms to nine one-dimensional and eight multi-dimensional functional datasets from the UEA & UCR Time Series Classification Repository. Clustering performance, evaluated using the Adjusted Mutual Information (AMI), is summarized in Table 1. For manifold-valued FD, we investigate five types of manifolds: Hypersphere, Hyperbolic, Swiss roll, Lorenz, and Pendulum. Details on the simulated FD are given in Appendix F.2. For each simulation scenario, we apply each model on 100 simulated datasets and report the mean and standard deviation of AMI (Table 2) and ARI (Table 7). |
| Researcher Affiliation | Academia | Samuel Singh School of Computer Science and Statistics Trinity College Dublin Shirley Coyle School of Electronic Engineering Dublin City University Mimi Zhang School of Computer Science and Statistics Trinity College Dublin |
| Pseudocode | Yes | A high-level outline of the algorithm is provided below. Input: latent representations X and similarity measures {s(yi,yj) : 1 i < j n}. while number of clusters > 2 do Compute the next breakpoint λK where a pair of centroids merge. Update the clustering by merging the corresponding pair of clusters. end while Output: breakpoints {λn 1, . . . , λ2} and corresponding hierarchical clustering. |
| Open Source Code | Yes | 1https://github.com/samuelveersingh/FAEclust |
| Open Datasets | Yes | When the FD are in a Euclidean space, we applied all eight FD clustering algorithms to nine one-dimensional and eight multi-dimensional functional datasets from the UEA & UCR Time Series Classification Repository. The real datasets are sourced from the UCR Time Series Archive, while the simulated datasets can be reproduced using the provided scripts. Detailed experimental settings are available in the Appendix and the associated Python package. |
| Dataset Splits | No | For all baseline methods, we run each method with the number of clusters ranging from 2 to 10 (extended to 20 for the Fungi dataset), ensuring that the true number of clusters is included in this range. We then report the best clustering performance based on ARI and AMI. In contrast, FAEclust determines the number of clusters in a fully data-driven manner using the silhouette score within each forward phase of the joint training and clustering framework. |
| Hardware Specification | Yes | In all the experiments above, dynamic time warping computations were performed on two 2.90 GHz CPUs (each with 16 cores), and the resulting distance matrices were used to train the FAE network on an NVIDIA Quadro RTX 5000 GPU with 16 GB of VRAM. |
| Software Dependencies | No | In our implementation, we utilized the Optuna package to perform the hyperparameter tuning task. In all experiments, both real and simulated, we fixed the FAEclust architecture to a seven-layer structure: one functional layer in the encoder, three functional layers in the decoder, and a three-layer MLP in between. The third layer in FAEclust serves as the bottleneck, producing the latent embedding. The maximum number of nodes allowed per layer was constrained to {64, 32, 16, 32, 64, 64, 64}. |
| Experiment Setup | Yes | All hyperparameters epochs, lr, batch_size, tau, layers, lambda_e, and lambda_d are optimized via Bayesian optimization over a predefined search space. The objective function in the Bayesian optimization problem is still the integrated objective function L = Lr + λw Lw + λc Lc. In our implementation, we utilized the Optuna package to perform the hyperparameter tuning task. In all experiments, both real and simulated, we fixed the FAEclust architecture to a seven-layer structure: one functional layer in the encoder, three functional layers in the decoder, and a three-layer MLP in between. The third layer in FAEclust serves as the bottleneck, producing the latent embedding. The maximum number of nodes allowed per layer was constrained to {64, 32, 16, 32, 64, 64, 64}. |