A3S: A General Active Clustering Method with Pairwise Constraints
Authors: Xun Deng, Junlong Liu, Han Zhong, Fuli Feng, Chen Shen, Xiangnan He, Jieping Ye, Zheng Wang
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In extensive experiments across diverse real-world datasets, A3S achieves desired results with significantly fewer human queries compared with existing methods. and 3. Experiments We organize the experiments as follows: we explain the experimental setup in Section 3.1; we compare A3S with stateof-the-art active clustering methods and present the detailed results in Section 3.2; then we compare the performance of A3S when applied to different clustering algorithms in Section 3.3; lastly, we explore the influence of components in A3S in Section 3.4 to 3.5. |
| Researcher Affiliation | Collaboration | Xun Deng 1 2 Junlong Liu 2 Han Zhong 3 Fuli Feng 1 Chen Shen 2 Xiangnan He 1 Jieping Ye 2 Zheng Wang 2 1University of Science and Technology of China 2Alibaba Group 3Peking University. Correspondence to: Zheng Wang <wz388779@alibaba-inc.com>, Fuli Feng <fulifeng93@gmail.com>. |
| Pseudocode | Yes | Algorithm 1 Adaptive Active Aggregation and Splitting Input: Data X, query limit Qmax, index q = 0, clustering algorithm A Initialization: Using A to generate initial clustering Ω for iter in 1 : L2 do... and Algorithm 2 Subcluster Partition Input: Cluster w; subcluster lists N = {} Sort samples in w in ascending order by their distance to the centroid for i in w do... and Algorithm 3 Fast Transitive Inference Input: State matrix S, new constraints (s, t). for i in (s, t) do... |
| Open Source Code | Yes | 1The code is available at https://github.com/xiangtanshi/A3S. |
| Open Datasets | Yes | Datasets. We sampled six datasets from four real-world image sources for the experiments: Market-1501 (Zheng et al., 2015), which comprises human body images from 1501 individuals. We use two subsets: MK20 (351 images from 20 people) and MK100 (1650 images from 100 people); Humbi (Yu et al., 2020), a large multiview image dataset focused on human expressions like faces, and we extracted a subset Humbi-Face containing 5600 face images from 100 different people; Handwritten (Dua et al., 2017), a collection containing 2000 samples of handwritten digits from 0 to 9 . We use the Fourier coefficient features in the experiments. (4) MS1M (Guo et al., 2016), a substantial benchmark dataset commonly used in face recognition tasks, and we sampled two large subsets: MS1M-10k, MS1M-100k. |
| Dataset Splits | No | The paper describes the datasets used for experiments but does not provide specific details on how these datasets were split into training, validation, and test sets, such as exact percentages or sample counts for each split. |
| Hardware Specification | Yes | Computing Resources We utilize a [Ge Force RTX 3090 Ti] for feature extraction using DNN models. For the implementation of baseline methods and A3S, we perform the experiments on a machine equipped with an Intel(R) Xeon(R) Platinum 8163 CPU @ 2.50GHz. |
| Software Dependencies | No | The paper mentions various algorithms and models used (e.g., 'isotonic regression', 'Fast Probabilistic Clustering (FPC)', 'K-means clustering', 'Arcface model'), but it does not provide specific version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | B.4. Hyperparameter Setting The implementation of A3S involves two hyperparameters: the threshold for density test and the number of neighbors considered in Fast Probabilistic Clustering. We report our choice of these two parameters in Table 4. In particular, the hyper-parameter τ is used to filter out the clusters with low density, and it is set to be slightly lower than the average density among all clusters. For the selection of τ, we first compute density using the formula in Eq. (5) for all clusters, then calculate the mean value as d, and set τ as d 0.1. The final results of A3S are not sensitive to this value, and perturbing it to d or d 0.05 has a negligible impact on the final clustering result. and Table 4. Hyperparameter setting of A3S. dataset MK20 MK100 Handwritten Humbi-Face MS1M-10k MS1M-100k τ 0.5 0.8 0.8 0.5 0.5 0.5 neighbors 50 50 50 50 50 50 |