Advanced Clustering
- Publication Type:
- Thesis
- Issue Date:
- 2022
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Clustering is a classical technique in the field of data mining. It has played a key role in domains such as biology, medicine, business, and climatology, and is employed in nearly all scientific and social sciences. Despite the significance and pervasiveness of clustering and the plethora of existing algorithms, the current clustering methods suffer from a variety of drawbacks. For example, standard hierarchical clustering has an excessive computational overhead and requires some manually determined conditions. Partition clustering, such as K-means, demands that the number of clusters must either be known or estimated in advance and cannot detect non-convex clusters of varying size or density. Density clustering typically requires a suite of thresholds to be set in advance, such as cut-off distance. Model-based clustering generally relies on prior knowledge of many parameter settings, which is often very difficult to acquire in practice. Classic grid clustering also depends on many user-provided parameters, such as interval values to divide space and density thresholds.
On the other hand, in recent years, multi-view clustering has become a new research hotspot. Essentially, multi-view clustering arises from the combination of clustering problems and multi-view learning. Different from the various conventional single-view clustering methods mentioned above, as an extension of single-view clustering, multi-view clustering is used to handle multi-view data gathered from numerous feature collectors or collected from various sources in various domains. However, most current multi-view clustering approaches suffer from the following three problems: a) parameter tuning, b) significant computational cost, and c) difficulty in finding globally optimal view weights.
To solve the above problems, this thesis first proposes a brand-new efficient parameter-free autonomous clustering algorithm called Torque Clustering (TC). The proposed TC overcomes almost all the shortcomings in previous clustering methods. Furthermore, considering the good performance of the proposed TC, this thesis extends TC to two multi-view clustering algorithms, containing multi-view adjacency-constrained hierarchical clustering (MCHC) and particle swarm optimization (PSO)-based multi-view nearest neighbor clustering (PMNNC). MCHC tries to solve two problems in current multi-view clustering methods: a) parameter tuning and b) significant computational cost. PMNNC focuses on solving the third problem: c) difficulty in finding globally optimal view weights. Finally, we further apply the pseudo labels generated by TC to propose a new metric learning framework, named almost ultrametric learning using pseudo labels of torque clustering (AUMLTC), which can help other algorithms improve performance in a parameter-free and unsupervised manner.
Please use this identifier to cite or link to this item: