What problem does this paper attempt to address?

The main problem that this paper attempts to solve is: how to automatically select the most appropriate model for the clustering problem in the absence of labels. Specifically, the author proposes a brand - new zero - shot model selection framework, which realizes the automatic clustering model recommendation for new datasets by pre - training on internal and external clustering validity indices (CVIs) from previous tasks. ### Core Problems of the Paper 1. **Challenges in Clustering Problems** - Clustering is a form of unsupervised learning, and the lack of labeled data makes it difficult to apply traditional AutoML techniques. - Traditional clustering methods rely on internal or external clustering validity indices (such as Calinski - Harabasz, Silhouette scores, etc.), but there is a lack of strong correlation among these metrics, and it becomes very difficult to optimize the external CVI for new datasets without labels. 2. **Limitations of Existing Methods** - Existing automated clustering methods usually only optimize the number of clusters or select algorithms through meta - learning, but most of these methods require labels or specific evaluation metrics. - Current AutoML systems mainly focus on supervised learning tasks, and there are relatively few automated solutions for clustering problems. ### Proposed Solutions The author proposes CLAMS (Clustering with Automated Machine Learning System) and its extended version CLAMS - OT (Zero - Shot Model Recommendation System Based on Optimal Transport) to solve the above problems: - **CLAMS**: An independent AutoML tool that can automatically select clustering algorithms and their hyper - parameter configurations. It supports internal and external CVIs and allows for complete pipeline selection. - **CLAMS - OT**: Utilizes the Optimal Transport Distance (OTD) to measure the similarity between different datasets, so as to recommend the most suitable clustering model without the need for labels. This method compares the new dataset with known datasets through pre - processing and transformation functions and selects the optimal clustering algorithm corresponding to the most similar dataset. ### Method Innovation Points - **Application of Optimal Transport Distance**: By introducing the low - rank Gromov - Wasserstein distance (GW - LR), a good balance between computational efficiency and accuracy can be achieved, thereby effectively measuring the similarity between different datasets. - **Zero - Shot Recommendation**: CLAMS - OT can quickly recommend a clustering model suitable for a new dataset without any labels, which is of great significance for practical application scenarios. ### Experimental Verification The author verifies the effectiveness of CLAMS - OT through experiments, and the results show that this method significantly outperforms existing clustering methods in multiple benchmark tests. In particular, CLAMS - OT shows higher performance on the Adjusted Mutual Information (AMI) metric. ### Summary This paper provides a novel and effective automated clustering solution by introducing optimal transport theory and zero - shot recommendation mechanisms, filling the gap in existing AutoML techniques in the field of unsupervised learning.

CLAMS: A System for Zero-Shot Model Selection for Clustering

A Survey on AutoML Methods and Systems for Clustering

AutoCl : A Visual Interactive System for Automatic Deep Learning Classifier Recommendation Based on Models Performance

Problem-oriented AutoML in Clustering

Clustering Indices based Automatic Classification Model Selection

Transferable Automl By Model Sharing Over Grouped Datasets

CLAMS: A Cluster Ambiguity Measure for Estimating Perceptual Variability in Visual Clustering

Selecting Optimal Trace Clustering Pipelines with AutoML

Automatic Parameter Selection for Non-Redundant Clustering

Point Cloud Classification via Deep Set Linearized Optimal Transport

An optimized intelligent open-source MLaaS framework for user-friendly clustering and anomaly detection

Exploring Zero-Shot Anomaly Detection with CLIP in Medical Imaging: Are We There Yet?

AutoCLIP: Auto-tuning Zero-Shot Classifiers for Vision-Language Models

STREAMLINE: A Simple, Transparent, End-To-End Automated Machine Learning Pipeline Facilitating Data Analysis and Algorithm Comparison

Benchmarking Automatic Machine Learning Frameworks

MLRR-ATV: A Robust Manifold Nonnegative LowRank Representation with Adaptive Total-Variation Regularization for scRNA-seq Data Clustering

Class label autoencoder for zero-shot learning

Fitting Multiple Machine Learning Models with Performance Based Clustering

Zero-Shot Recognition via Optimal Transport

AutoMMLab: Automatically Generating Deployable Models from Language Instructions for Computer Vision Tasks

AutoFCL: Automatically Tuning Fully Connected Layers for Handling Small Dataset