--- title: Cluster Paths in Data Analysis url: https://www.emergentmind.com/topics/cluster-paths type: topic --- # Cluster Paths in Data Analysis Cluster paths are path-centered representations used to organize cluster structure in several distinct senses. In graph-based clustering, a path is a sequence of edges whose cumulative, minimax, or learned cost determines assignment to density peaks; in manifold clustering, it may be an angle-constrained graph path, a maximum-density path between mixture components, or a bottleneck path on a simplex graph; in convex clustering and local graph diffusion, it is a regularization path indexed by a control parameter; in temporal analysis, it is an inferred evolution path through a learned cluster-transition graph; and in neural-network interpretability, it is the sequence of cluster IDs traversed by an input across layers [1812.11850] [1802.07416] [2503.15582] [2501.15964] [1503.00322] [1512.03501] [2510.06541]. ## 1. Meanings of the term The term is not standardized across fields. In some literatures it denotes a path **used to define clusters**; in others, a path **of clusterings** generated as a parameter changes; in still others, a path **through clusters** that summarizes evolution or representation learning. | Setting | Path object | Representative formulation | |---|---|---| | Graph-based assignment | Node path $\Gamma$ from a super-source to a sample | Non-decreasing shortest-path cost [1812.11850] | | Intersecting manifolds | $\alpha$-constrained neighborhood-graph path | Turning-angle threshold [1802.07416] | | Density hierarchy | Continuous path between mixture centers | Maximum-density bottleneck path [2503.15582] | | Simplex geometry | Path on a weighted simplex graph | LAPD bottleneck metric [2507.10710] | | Regularization | Sequence $\{U^*(\lambda_t)\}$ or $\{x(\delta)\}$ | Convex clustering and seeded PageRank paths [2501.15964] [1503.00322] | | Neural interpretability | Layer-wise sequence $\gamma(x)=(c^{(1)},\dots,c^{(L)})$ | Activation-cluster path [2510.06541] | A temporal-data usage is explicit in ClusPath, where observations are partitioned into clusters $C_1,\dots,C_k$ while simultaneously learning a directed weighted graph $A=(a_{pq})$ that models transitions between successive phases under a “slow changing world” assumption [1512.03501]. A separate combinatorial-optimization usage studies clustering directly **on a path graph** with arbitrary pairwise correlation costs; in that setting the problem is polynomial-time solvable by dynamic programming, and the associated lifted multicut polytope for paths has a complete totally dual integral description [1706.06822]. This suggests a common design pattern: replacing purely local point-to-point similarity with an object that encodes continuity, bottlenecks, admissibility, or multiscale evolution. ## 2. Shortest-path formulations for clustering on graphs A canonical path-based clustering construction appears in the shortest-path method for joint human-machine analysis of complex datasets. Given a point cloud, a weighted graph is built with vertices $V=\{v_1,\dots,v_N\}$ and nonnegative edge costs $w_{ij}$. Local point-density and distance-to-higher-density are defined as $$ \rho_v = \sum_{u\in V}\mathbf{1}(\|x_u-x_v\|\le r), \qquad \delta_v = \min_{u:\rho_u>\rho_v}\|x_u-x_v\|, $$ and density peaks are selected as $$ P_\tau = \{v \mid \rho_v>\tau_\rho \wedge \delta_v>\tau_d\}. $$ A fictitious source $s$ is connected to each density peak with negligible cost $\epsilon$, and a single-source shortest-path problem is solved from $s$ to all nodes [1812.11850]. The crucial constraint is that the path-cost functional must satisfy Bellman’s optimality, i.e. it must be non-decreasing on path extension. The paper gives three standard forms: $$ c_{\mathrm{sum}}(\Gamma)=\sum_{(i,j)\in \Gamma} w_{ij}, \qquad c_{\max}(\Gamma)=\max_{(i,j)\in \Gamma} w_{ij}, $$ and a learned local-window cost $$ c_f(\Gamma)=\max_{(i,j)\in \Gamma} f(w_{i-k+1,i-k+2},\dots,w_{ij}). $$ Because these costs are non-decreasing under extension, Dijkstra-style search remains valid. This directly addresses a failure mode of local association rules: on the “dendrites + blob” benchmark, CDP mis-assigns tips of long thin projections and DBSCAN cannot separate two touching high-density blobs, whereas minimax-path clustering corrects both artifacts. The reported F1-score is CDP $\approx 0.85$, DBSCAN $\approx 0.80$, minimax-path $\approx 0.92$, SVM-enhanced $\approx 0.99$; the corresponding Jaccard scores are CDP $\approx 0.80$, DBSCAN $\approx 0.75$, minimax-path $\approx 0.88$, SVM-enhanced $\approx 0.98$ [1812.11850]. The same work extends path costs with a trainable path classifier. Features are extracted from the last $w$ edges of a candidate extension, including densities, gap sizes, and $$ \Delta \rho(\Gamma_w)=\max_{u,v\in \Gamma_w}|\rho_u-\rho_v|. $$ A binary SVM with RBF kernel is trained on admissible versus forbidden fragments, and the classifier score is converted into a pseudo-cost. In 4D centroid tracking of T-cells, CDP gives F1 $\approx 33\%$, minimax only gives F1 $\approx 37\%$, and the SVM path-classifier reaches F1 $\approx 92\%$ with Jaccard $\approx 0.90$ [1812.11850]. A related high-dimensional scheme first bins the data into occupied partitions, builds a nearest-neighbor partition graph, and computes shortest-path matrices such as $\Delta L$. The resulting path-length features support MAXPATHL clustering around local maxima and a Line-Of-Sight criterion based on convex-hull connectivity, corner-turn tests, and variance tests. The dominant cost is all-pairs shortest paths on the reduced partition graph, nominally $O(N_P^3)$, but the method is designed for the regime $N_P \ll N$ [1710.04886]. ## 3. Path admissibility, bottlenecks, and manifold separation For multi-manifold clustering, path admissibility can be imposed geometrically. In the angle-constrained method, a symmetric $q$-nearest-neighbor graph is built on $X=\{x_1,\dots,x_N\}\subset\mathbb{R}^D$, and a path $(i_0,\dots,i_m)$ is called $\alpha$-constrained when every turning angle satisfies $$ \theta_t \le \alpha, \qquad t=1,\dots,m-1. $$ Several landmarks are chosen at random, and for each landmark a modified Dijkstra algorithm marks which nodes are reachable by an $\alpha$-constrained path. Each sample then receives a binary signature $$ \varphi_i = [\phi_1(i),\dots,\phi_M(i)] \in \{0,1\}^M, $$ and hierarchical clustering with complete linkage is applied to these signatures [1802.07416]. Reported synthetic accuracies exceed $93\%$ on all eight benchmark cases, including TP $93.0\%$, TSI $98.2\%$, FS $98.1\%$, DSPR $97.5\%$, and RCC $99.4\%$. On COIL-20, clustering the objects $\{3,6,19\}$ yields $99\%$ accuracy and clustering all 20 objects yields $83.6\%$ [1802.07416]. A density-landscape variant appears in t-NEB. Let $p:\mathbb{R}^d\to\mathbb{R}_+$ be a smooth density estimated by a Gaussian or Student’s-$t$ mixture. For two cluster centers $\mu_i,\mu_j$, the maximum-density path is $$ \gamma^* = \arg\max_{\gamma\in\Gamma_{ij}} \min_{t\in[0,1]} p(\gamma(t)), $$ equivalently $$ \gamma^* = \arg\min_{\gamma\in\Gamma_{ij}} \max_{t\in[0,1]} E(\gamma(t)), \qquad E(x)=-\log p(x). $$ After overclustering, nudged elastic band optimization estimates the bottleneck density $\rho_{ij}$ on edges of a $k_{\mathrm{nbr}}$-NN graph, an MST is computed on weights $w_{ij}=-\rho_{ij}$, and bottom-up merging yields a dendrogram [2503.15582]. On 16D MNIST-Nd, NEB-based merging achieves $\mathrm{ARI}\approx 0.80$, whereas Euclidean and dip-statistic merging are both around $\mathrm{ARI}\approx 0.45$; stability across random EM seeds is pairwise ARI $>0.8$ in $d\le 32$ [2503.15582]. A simplex-graph construction generalizes bottleneck paths to intersecting manifolds. Candidate $d$-simplices are built from an annular graph, adjacent simplices share a $(d-1)$-face, and edge weights are defined from the dihedral angle. The largest-angle path distance is $$ \mathrm{LAPD}(\Delta_i,\Delta_j)= \min_{\Gamma:\Gamma_1=\Delta_i,\Gamma_L=\Delta_j} \max_t W_S(\Gamma_t,\Gamma_{t+1}). $$ Denoising removes simplices with large $\kappa$-nearest-neighbor LAPD, and single-linkage on the denoised simplex graph recovers manifold components [2507.10710]. The theory gives $$ \mathrm{wLAPD}\lesssim \frac{\tau}{e}, \qquad \mathrm{bLAPD}_{\mathrm{dns}}\gtrsim \Theta, $$ under the stated sampling, curvature, and noise conditions, and the scalable implementation reaches overall complexity $O(D\,n\log n)$ when $B,d,k=O(1)$. Empirically, LAPD is reported as robust to noise, curvature, and small intersection angle, with runtime for $n$ up to $20000$ remaining $5$–$100\,\mathrm{s}$ [2507.10710]. ## 4. Trajectory and flight-path clustering Path-based clustering is particularly natural for trajectory data. Song et al. use the U.S. Federal Aviation Administration’s Traffic Flow Management System dataset together with DV8 to categorize flight paths by agglomerative hierarchical clustering under two distance models: a spatial-based geographic distance model and a vector-based cosine similarity model [2310.00773]. For trajectories sampled to $M$ points, the geographic model averages great-circle distances, $$ D_{\mathrm{geo}}(T_i,T_j)=\frac{1}{M}\sum_{k=1}^M d(\phi_{i,k},\lambda_{i,k};\phi_{j,k},\lambda_{j,k}), $$ while the directional model computes $$ S_{ij}=\frac{1}{M-1}\sum_{k=1}^{M-1} \frac{v_{i,k}\cdot v_{j,k}}{\|v_{i,k}\|\|v_{j,k}\|}, \qquad D_{\cos}(T_i,T_j)=1-S_{ij}. $$ A silhouette-score cutoff selects the dendrogram cut, and point extraction uniformly subsamples trajectories for speed [2310.00773]. The reported behavior is phase-dependent. For enroute segments, the geographic-distance model yields silhouette scores in the $0.8$ range, including $0.84$ for CMH$\to$ATL. For CMH$\to$PHL arrival patterns, cosine-based clustering gives a silhouette of $0.61$ versus $0.46$ for geographic distance. On SFO$\to$PIT enroute data, geographic distance gives score $0.51$, while cosine similarity gives $0.39$ [2310.00773]. Point extraction with $N=1,2,4,8,16$ shows that for long-haul queries, $N$ up to $8$ or $16$ incurs negligible silhouette loss $(<0.01)$ while reducing runtime by $4\times$–$6\times$; for short-haul or highly curved approaches, smaller $N$ values preserve cluster quality. The practitioner recommendation is to segment flights into departure, enroute, and arrival portions; use cosine for departure/arrival and geographic for enroute; and expose a human-in-the-loop threshold override [2310.00773]. A related trajectory hierarchy appears in scan-path generation. HPCGen decomposes raw gaze data into a multi-level tree by hierarchical K-Means, reclustering shift vectors relative to parent centroids and computing a PCA model within each cluster. New scan paths are generated by traversing the tree top-down and sampling principal-component perturbations [2201.08354]. This is not a path-cost clustering algorithm, but it uses cluster hierarchies explicitly as a path-generating mechanism. ## 5. Regularization paths and clustering paths In convex clustering, the path is a sequence of optimization solutions indexed by a fusion parameter. PyClustrPath formulates convex clustering as $$ \min_{U\in\mathbb{R}^{n\times d}} \frac12\|X-U\|_F^2 + \lambda\sum_{i