Hull Clustering in Data Analysis

• Hull clustering is a method that uses convex and conical hulls to select extreme representatives, capturing boundary behavior effectively.
• It is applied in energy system optimization, image-set matching, conical anchor recovery, and improved k-means initialization to enhance both accuracy and efficiency.
• By focusing on boundary points and local hull decomposition, hull clustering offers a robust alternative to centroid-based methods for handling data extremes and latent structures.

Hull clustering denotes a class of clustering and representative-selection procedures organized around convex or conical hull geometry rather than around interior summaries such as centroids. In the literature covered here, the term is used in several closely related senses: selecting representative periods (RPs) as extreme points of a time-series dataset for energy system optimization, partitioning an image set into multiple local convex hulls for set-to-set matching, and recovering extremal rays that span a conical hull for subsequent clustering or latent-structure recovery (Neustroev et al., 29 Aug 2025, Chen et al., 2014, Zhou et al., 2014). Across these settings, the common premise is that boundary structure—vertices or rays—captures operational extremes, discriminative variation, or latent generators more directly than mean-like representatives.

1. Geometric scope and core definitions

In the representative-period setting, base-period vectors $x^1,\dots,x^T\in\mathbb{R}^p$ define a convex hull

$$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$

and the RP set $R\subset\{1,\dots,T\}$ is chosen so that $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$ (Neustroev et al., 29 Aug 2025). In image-set matching, an image set $S=\{I_1,\dots,I_n\}$ is modeled by its convex hull

$$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$

with set-to-set distance defined as $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$ (Chen et al., 2014). In conical-hull learning, the target is an anchor set $A$ such that every $x_i$ lies in

$$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$

which yields the minimum-conical-hull problem: minimize $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$0 subject to $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$1 (Zhou et al., 2014).

Context	Hull object	Main operation
Energy system optimization	Convex hull of base-period vectors	Select extreme-point RPs; fit convex or conic blends
Image-set matching	Global and local convex hulls of image sets	Partition sets into local hulls; match corresponding hulls
Divide-and-conquer learning	Conical hull generated by anchors	Recover extremal rays; cluster from anchor structure
K-means initialization	Convex hull of the dataset	Use hull points to choose initial centroids

These formulations share the use of hull membership, hull distance, or hull coverage as the organizing primitive. A plausible implication is that “hull clustering” is best understood not as a single standardized algorithm, but as a geometric design pattern in which clustering is driven by boundary points, local hull decomposition, or extremal generators.

2. Representative-period hull clustering in energy system optimization

For time-series aggregation in energy system optimization models (ESOMs), hull clustering is defined as selecting RPs as the extreme points, or vertices, of the dataset in feature space rather than as centroids or medoids (Neustroev et al., 29 Aug 2025). The stated rationale is twofold. First, binding constraints in ESOMs, including ramping, unit-commitment limits, and storage extremes, often coincide with data extremes. Second, any point in $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$2 can be reconstructed without error from its hull vertices, while centroids lie interior and may miss boundary behavior.

The paper introduces a greedy convex-hull clustering procedure adapted to RP selection. Starting from the point furthest from the data mean, the method repeatedly adds the point with maximum distance to the convex hull of the currently selected representatives. Distance to the convex hull is computed by projecting $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$3 onto $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$4 via projected gradient descent (PGD). Quickhull is also described as a divide-and-conquer method to find the convex hull of a set of points in $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$5, with average complexity $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$6 in 2D/3D and worst-case $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$7 (Neustroev et al., 29 Aug 2025).

After hull clustering yields RP data vectors $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$8, each base period $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$9 is approximated by a blend

$R\subset\{1,\dots,T\}$0

Three blend types are specified: convex blend, with $R\subset\{1,\dots,T\}$1 and $R\subset\{1,\dots,T\}$2; sub-unit conic, with $R\subset\{1,\dots,T\}$3 and $R\subset\{1,\dots,T\}$4; and general conic, with only $R\subset\{1,\dots,T\}$5 (Neustroev et al., 29 Aug 2025). The paper states that convex blends preserve affine combinations, sub-unit conic ensures inequality-constraint validity, and full conic allows arbitrary nonnegative scaling.

Integration into the ESOM is performed by replacing each base day’s time-dependent parameters with weighted sums over the RPs. In the single-stage ESOM structure, operational profiles satisfy

$R\subset\{1,\dots,T\}$6

and analogous substitutions are made for demand, renewable availability, and storage-related inputs. Regret is defined by comparing the full-resolution objective $R\subset\{1,\dots,T\}$7 with the cost $R\subset\{1,\dots,T\}$8 obtained when decisions from the reduced model are fixed and re-evaluated on full data:

$R\subset\{1,\dots,T\}$9

The reported computational costs are $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$0 for greedy hull clustering and $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$1 independent PGD problems, each $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$2, for weight fitting (Neustroev et al., 29 Aug 2025).

The empirical results are problem-specific. In a generation expansion planning case study based on European TYNDP 2022 data, with ramping but no storage, 5 RPs with conic hull and conic weights yield regret $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$3, whereas k-means with 5 RPs yields regret $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$4 and requires 80 RPs to reach similar regret; solve time is $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$5 s for the 5-RP conic hull model versus $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$6 s for 80-RP k-means. In a Power-to-X dispatch case study based on TYNDP 2024 data with seasonal storage and hydropower inflows, 10 RPs with convex-hull selection and convex weights yield regret $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$7 and runtime $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$8 s, while 10 RPs with k-medoids and Dirac weights yield regret $\mathrm{conv}\{x^r:r\in R\}\approx \mathcal{H}$9 at similar runtime (Neustroev et al., 29 Aug 2025). The paper’s recommendations are correspondingly differentiated: when ramping or extremes dominate, choose hull clustering; when interpolation or storage balances dominate, choose blended weights; the combined approach is described as robust across problem types.

3. Local convex hulls for image-set matching

In image-set recognition, hull clustering addresses a different failure mode: a single global convex or affine hull may generate strong artificial features and noisy combinations when image sets exhibit significant intra-class variations or noise (Chen et al., 2014). The proposed response is an adaptive multi-model construction in which each set is partitioned into several local convex hulls, thereby limiting interpolation to compact subsets rather than permitting combinations across the entire set.

The local hulls are extracted by Maximum Margin Clustering (MMC). Given a partition $S=\{I_1,\dots,I_n\}$0, each local hull is $S=\{I_1,\dots,I_n\}$1, and the objective is to maximize the minimum inter-hull distance. For the two-cluster case, this is stated to be exactly equivalent to a two-class hard-margin SVM problem with labels $S=\{I_1,\dots,I_n\}$2 and classifier $S=\{I_1,\dots,I_n\}$3 maximizing the margin $S=\{I_1,\dots,I_n\}$4. Allowing slack variables yields the standard MMC formulation of Xu et al. (2005). For $S=\{I_1,\dots,I_n\}$5, the same style extends by analogy with multiclass SVMs, and the implementation uses the convex-concave relaxations and cutting-plane solvers of Yu and Grauman (2009), alternating between fixing labels and training an SVM, then fixing the classifier and reassigning labels; convergence is reported to take a few dozen iterations (Chen et al., 2014).

Once the query set has been partitioned, Adaptive Reference Clustering (ARC) re-clusters each gallery set to resemble the query’s local hulls. For each gallery sample $S=\{I_1,\dots,I_n\}$6, ARC computes

$S=\{I_1,\dots,I_n\}$7

assigns the sample to the corresponding cluster, discards empty clusters, and forms gallery local hulls $S=\{I_1,\dots,I_n\}$8 (Chen et al., 2014). The set-to-set distance is then

$S=\{I_1,\dots,I_n\}$9

and classification selects the gallery class with smallest $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$0.

The comparative claim is explicit: single-hull methods can generate grossly interpolated points when extreme samples mix, while nearest-neighbour methods avoid interpolation but are highly sensitive to noise. The multi-convex-hull approach occupies an intermediate position by allowing limited interpolation only within compact clusters. Conventional multi-model approaches are also criticized for fixing clusters offline, which can yield poor cross-set correspondence; ARC instead re-clusters each gallery set on the fly, and noisy or rare query clusters without gallery support are effectively ignored (Chen et al., 2014).

Empirically, the adaptive multi-convex-hull method attains $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$1 accuracy on the Honda/UCSD face-video dataset, versus $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$2–$$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$3 for Sparse Approximated Nearest Points, Mutual Subspace Method, and Manifold Discriminant Analysis. On CMU-MoBo and ETH-80, it reduces error rates by $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$4–$$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$5 relative to single-hull baselines and by several percent relative to recent multi-patch and MANOVA-style alternatives (Chen et al., 2014). These results frame hull clustering here as a discriminative local modeling strategy rather than as an extreme-point selection problem.

4. Anchoring a conical hull and divide-and-conquer clustering

A third formulation centers on conical rather than convex hulls. In “Divide-and-Conquer Learning by Anchoring a Conical Hull,” the objective is to identify a small anchor set $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$6 such that every target point $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$7 lies in $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$8, equivalently $$H^{\mathrm{con}}=\bigl\{y=\sum_{k=1}^n w_k I_k \;\big|\; \sum_{k=1}^n w_k=1,\;0\le w_k\le 1\bigr\},$$9 under a separability assumption with $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$0 (Zhou et al., 2014). The paper states that this reduction applies to a broad class of machine learning problems commonly addressed by EM or sampling, including GMM, HMM, LDA, NMF, and subspace clustering.

The proposed Divide-and-Conquer Anchoring (DCA) algorithm projects the data into many very low-dimensional random subspaces, often 2D planes, solves the low-dimensional minimum-conical-hull subproblem there, and aggregates anchor frequencies across $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$1 trials. In 2D, the subproblem becomes purely angular: one computes angles of projected points relative to a coordinate axis, finds $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$2 and $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$3, and selects the candidate rays in $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$4 that enclose the angular range of $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$5 (Zhou et al., 2014). The resulting total complexity is

$$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$6

where $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$7 and the projection dimension is constant. The paper contrasts this with exact convex-hull computation in $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$8 dimensions, which can be exponential in $$D_h(H_1,H_2)=\min_{y_1\in H_1,\;y_2\in H_2}\|y_1-y_2\|_2$$9, and with standard EM, which costs $A$0 over $A$1 iterations.

The same framework provides a fast cone-membership subroutine. Rather than solving a small linear program to test whether $A$2, DCA repeatedly projects $A$3 to low dimension and checks whether the projected point lies between the min/max angles of the projected anchors. If it falls outside on any projection, it is declared not to lie in the cone (Zhou et al., 2014). This membership test then becomes part of a hull-based subspace clustering pipeline.

For subspace clustering, the paper imposes a block-diagonal separability on $A$4, with each cone using anchors $A$5. On each 2D projection, a 1D clustering such as mean-shift is run on the array of angles; the resulting cluster labels define an anchor–anchor co-occurrence graph $A$6, where $A$7 counts how often two anchors agreed in the same 2D cluster. Spectral clustering of $A$8 recovers the anchor blocks, and each $A$9 is assigned to the unique block whose anchor cone contains it (Zhou et al., 2014).

The reported results emphasize scale. On synthetic 4-cone data with $x_i$0, 10 anchors per cone, and $x_i$1 points each, DCA-SC achieves $x_i$2 anchor precision and recall with only $x_i$3 random 2D projections, with errors $x_i$4. On COIL-100, it reaches mutual information $x_i$5 versus $x_i$6 for SSC or LRR, but in $x_i$7 s versus $x_i$8 s. On a CMU-moCap sequence, it selects $x_i$9 anchor frames per motion, reconstructs in-cone frames with $$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$0-errors $$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$1, and clusters at $$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$2 s versus $$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$3 s for competing methods (Zhou et al., 2014). In this usage, hull clustering is principally an anchor-recovery and scalability mechanism.

5. Hull-derived initialization for k-means

A related but narrower use of hull geometry appears in Rahman et al.’s method for initial cluster center selection in k-means (Rahman et al., 2022). This work does not define hull clustering as a standalone clustering paradigm; instead, it uses a convex-hull algorithm to improve centroid initialization under the premise that the initial centroids should be well separated because final clusters are separated groups in feature space.

The procedure begins by extracting the outer points of the dataset using Quickhull. The first two centroids $$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$4 are then chosen as the pair of hull points with maximum Euclidean distance,

$$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$5

where

$$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$6

The remaining $$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$7 centroids are selected iteratively by maximizing cumulative distance to previously chosen centers,

$$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$8

while discarding the $$\mathrm{cone}(Y_A)=\Bigl\{\sum_{j\in A}\alpha_j y_j:\alpha_j\ge 0\Bigr\},$$9 nearest neighbors of each selected centroid to reduce the risk of choosing multiple centers from the same true cluster (Rahman et al., 2022). The paper specifies $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$00, or between $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$01 and $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$02 by empirical tuning.

The stated complexity is $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$03 for Quickhull, $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$04 for pairwise distances among hull points, and $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$05 for the remaining centroid passes, yielding total complexity $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$06 (Rahman et al., 2022). Standard Lloyd k-means is then run to convergence. Evaluation uses Error(%), the Cluster Center Proximity Index (CCPI), and the Rand Index.

The reported labeled-dataset results are specific. The proposed method yields $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$07 clustering error on Iris, $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$08 on Wine, $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$09 on Letter, and $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$10 on Ruspini. Computation times and CCPI are also given: for Iris, random k-means requires $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$11 ms with CCPI $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$12, while the proposed method requires $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$13 ms with CCPI $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$14; for Ruspini, the times are $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$15 ms and $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$16 ms with CCPI $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$17 and $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$18, respectively (Rahman et al., 2022). The note accompanying these results is that for $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$19, random k-means can be faster, but as $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$20 increases the hull-based initialization brings superior convergence speed and much lower CCPI.

Within a broader account of hull clustering, this method is best treated as a hull-derived initialization strategy. It shares the boundary-point emphasis of the other papers, but its objective is improved seeding for centroid-based clustering rather than hull-based representation or hull-based matching.

6. Comparative interpretation, limitations, and recurrent misconceptions

The surveyed literature assigns different technical roles to hulls. In ESOM reduction, hull vertices are used as representative periods because extremes are more likely to be constraint-binding and because boundary coverage matters for regret under reduced temporal resolution (Neustroev et al., 29 Aug 2025). In image-set recognition, local hulls are introduced to constrain interpolation and suppress artificial variations and noisy combinations (Chen et al., 2014). In DCA, conical hull anchors provide a separable latent structure that can be recovered efficiently through many low-dimensional subproblems (Zhou et al., 2014). In k-means initialization, hull points serve only to improve the spacing of initial centroids (Rahman et al., 2022).

One common misconception is that hull methods are simply substitutes for centroid-based clustering. The sources do not support that simplification. The ESOM paper explicitly contrasts extreme points with centroids or medoids, while the image-set paper distinguishes local hulls from both nearest-neighbour methods and single global hull models. Another misconception is that interpolation is uniformly either beneficial or harmful. The image-set work argues that global interpolation across an entire set can create unrealistic combinations, whereas the ESOM work shows that convex, sub-unit conic, and general conic blends can improve approximation accuracy when used to reconstruct base periods from representative periods under explicit constraints (Neustroev et al., 29 Aug 2025, Chen et al., 2014). This suggests that the effect of interpolation depends on whether it is global or local, unconstrained or problem-structured.

The main algorithmic trade-offs are also domain-specific. Hull extraction and hull-distance calculations may be expensive in high dimension, which motivates Quickhull only as a component in some settings and motivates low-dimensional random projections in DCA. Weight fitting by PGD adds overhead in blended RP methods as $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$21 grows, and the ESOM paper recommends $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$22–$$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$23 in practice as a balance between accuracy and runtime (Neustroev et al., 29 Aug 2025). By contrast, the image-set paper accepts the overhead of MMC and ARC to obtain better cross-set correspondence, and DCA explicitly replaces higher-dimensional hull or EM-style procedures with $$\mathcal{H}=\mathrm{conv}\{x^1,x^2,\dots,x^T\} =\Bigl\{\sum_{t=1}^T \lambda_t x^t \;\Big|\; \lambda_t\ge 0,\;\sum_{t=1}^T \lambda_t=1\Bigr\},$$24 very cheap 2D subproblems (Chen et al., 2014, Zhou et al., 2014).

Taken together, these works indicate that hull clustering is a boundary-aware alternative to interior-summary methods. Its most characteristic operations are selecting extreme representatives, decomposing a dataset into compact local hulls, or recovering anchors that span a conical hull. The practical value reported in the literature lies in lower regret for reduced ESOMs, stronger robustness in image-set matching, scalable anchor recovery for conical models, and improved k-means initialization when cluster separation matters.