Condensation Point Clustering

• Condensation point clustering is defined as the grouping of data around high-confidence representative points obtained via learned latent representations.
• The method combines property regression with geometric losses, utilizing attractive and repulsive potentials to accurately cluster and reconstruct objects.
• Its applications range from image and graph data processing to detector reconstruction in particle physics, highlighting its versatility and scalability.

Searching arXiv for recent and foundational papers on object condensation / condensation-point clustering. arxiv_search.query({"^^^^^^^^3search_query3 condensation\" OR ^^^^^^^^3all:\3 clustering\"","start":^^^^3search_query3^^^^,"max_results":^^^^3all:\3search_query3 arxiv_search.search({"query":"object condensation clustering","max_results":^^^^3all:\3search_query3^^^^}) arxiv_search({"query":"object condensation clustering","max_results":^^^^3all:\3search_query3^^^^}) {"query":"object condensation clustering","max_results":^^^^3all:\3search_query3^^^^} Condensation point clustering denotes a family of clustering and condensation formulations in which representative points emerge and act as anchors for grouping. In the machine-learning formulation introduced as object condensation, each vertex predicts object properties PRESERVED_PLACEHOLDER_^^^^3search_query3^^^^, a condensation score PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^, and coordinates PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^ in a learned clustering space; high-$\beta$ vertices act as condensation points, and nearby vertices are assigned to them by a simple thresholded procedure (&&&^^^^3search_query3^^^^&&&). Related uses of the same vocabulary appear in graph condensation, where class-wise cluster centroids become synthetic nodes (&&&^^^^3all:\3^^^^&&&), and in statistical mechanics or random CSPs, where condensation refers to regimes in which a small number of clusters or condensates dominate the state space (&&&^^^^3 OR all:\3^^^^&&&, Sly et al., 2023).

^^^^3all:\3^^^^. Object-condensation formalism

The canonical machine-learning formulation treats pixels, detector hits, graph vertices, or point-cloud points as the primitive carriers of both instance membership and object properties. For each vertex $i=1,\dots,N$, the ground truth comprises membership indicators $M_{ik}\in\{0,1\}$ over objects $k=1,\dots,K$, a background label $n_i\in\{0,1\}$, and target properties $t_i$. The network predicts per-vertex properties $p_i$, a condensation score PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3^^^^, and clustering-space coordinates PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^. From PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3^^^^ one defines a charge

PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^

with PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^4 a small hyperparameter (&&&^^^^3search_query3^^^^&&&).

The total loss is

PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^5

where PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^6 balances condensation against property regression. The property term uses the charge-weighted factor

PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^7

and

PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^8

The condensation-point assignment term enforces exactly one high-PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^9 representative per object and suppresses background activations:

PRESERVED_PLACEHOLDER_^^^^3 OR all:\3search_query3^^^^

where PRESERVED_PLACEHOLDER_^^^^3 OR all:\3all:\3^^^^ is the highest-charge vertex in object PRESERVED_PLACEHOLDER_^^^^3 OR all:\3 OR all:\3^^^^ (&&&^^^^3search_query3^^^^&&&).

The geometric part of the loss approximates each object by its highest-charge vertex. For a vertex PRESERVED_PLACEHOLDER_^^^^3 OR all:\33^^^^ belonging to object PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^4, the attractive potential is

PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^5

while for a vertex outside object PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^6 the repulsive potential is

PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^7

The resulting pull-push term is

PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^8

This construction makes the method independent of assumptions on object size, sorting or object density, and it generalises to non-image-like data structures such as graphs and point clouds (&&&^^^^3search_query3^^^^&&&).

^^^^3 OR all:\3^^^^. Inference, clustering mechanics, and implementation choices

After training, losses are no longer evaluated. Inference proceeds by collecting all vertices with PRESERVED_PLACEHOLDER_^^^^3 OR all:\3^^^^9 into a candidate list, sorting them by descending $\beta$^^^^3search_query3^^^^, and accepting a candidate as a final condensation point only if it lies farther than $\beta$^^^^3all:\3^^^^ from all previously accepted points in clustering space. The accepted set $\beta$^^^^3 OR all:\3^^^^ is then used for nearest-representative assignment: every vertex is assigned to its nearest $\beta$3 if $\beta$4, and the object properties are taken directly from the condensation point, i.e. $\beta$5. The paper gives example values $\beta$6 and $\beta$7 (&&&^^^^3search_query3^^^^&&&).

Several implementation choices are explicit. The lowest useful clustering-space dimension is $\beta$8 for symmetry breaking; the graph example uses $\beta$9. Raising $i=1,\dots,N$^^^^3search_query3^^^^ accentuates segmentation over property regression, with typical values $i=1,\dots,N$^^^^3all:\3^^^^. The background suppression weight $i=1,\dots,N$^^^^3 OR all:\3^^^^ is an order-one scalar, and no $i=1,\dots,N$3 is added in denominators because $i=1,\dots,N$4 is strictly forced to be $i=1,\dots,N$5 by the sigmoid activation. Batch size and learning-rate scheduling, including cyclic learning rates between $i=1,\dots,N$6 and $i=1,\dots,N$7, are described as stabilizing choices. In graph or point-cloud applications, any GNN layer such as GravNet can be used to produce $i=1,\dots,N$8, $i=1,\dots,N$9, and $M_{ik}\in\{0,1\}$^^^^3search_query3^^^^; for images, the reference implementation uses a small U-Net-style convnet predicting a grid of vertices each with $M_{ik}\in\{0,1\}$^^^^3all:\3^^^^ (&&&^^^^3search_query3^^^^&&&).

The method is positioned against both computer-vision and detector-specific baselines. For image-like data, each pixel is a vertex and the condensation loss merges pixels of the same object without anchors or NMS. For graphs and point clouds, graph edges need not be pre-built because $M_{ik}\in\{0,1\}$^^^^3 OR all:\3^^^^ is learned directly. Traditional clustering methods such as DBSCAN or HDBSCAN cluster in input space and require separate seeding or thresholding, whereas object condensation integrates seeding via $M_{ik}\in\{0,1\}$3 and clustering into a single end-to-end loss. Likewise, particle-flow algorithms use hard rules for seeding, clustering, and track-cluster linking; object condensation replaces seeding, clustering, track-cluster linking and thresholding by a single differentiable network and loss (&&&^^^^3search_query3^^^^&&&).

3. Detector reconstruction and empirical realizations

As introduced, object condensation was presented as a one-stage grid-free multi-object reconstruction method for physics detectors, graph data, and image data, with proof-of-concept applications to a simple object-classification task in images and to reconstruction of multiple particles from detector signals (&&&^^^^3search_query3^^^^&&&). The detector setting is especially natural because sparse calorimeter or tracker readouts are more naturally represented as irregular sets or graphs than as dense images.

A detailed later realization is the CLAS^^^^3all:\3 OR all:\3^^^^ calorimeter study, which applies object condensation clustering to neutron and photon reconstruction in a hodoscopic detector (&&&^^^^3all:\3all:\3^^^^&&&). In that formulation, each hit predicts a $M_{ik}\in\{0,1\}$4-D condensation coordinate $M_{ik}\in\{0,1\}$5 and a confidence score $M_{ik}\in\{0,1\}$6. For each true cluster $M_{ik}\in\{0,1\}$7, the representative hit is

$M_{ik}\in\{0,1\}$8

and the latent distance is $M_{ik}\in\{0,1\}$9. The loss splits into $k=1,\dots,K$^^^^3search_query3^^^^, with an attractive term that pulls hits toward their own condensation point, a repulsive term with margin $k=1,\dots,K$^^^^3all:\3^^^^ that separates different clusters, a “coward” term that forces at least one hit per true cluster to have high $k=1,\dots,K$^^^^3 OR all:\3^^^^, and a noise term that penalizes large $k=1,\dots,K$3 on padded or background hits (&&&^^^^3all:\3all:\3^^^^&&&).

The CLAS^^^^3all:\3 OR all:\3^^^^ inference procedure sorts hits by descending $k=1,\dots,K$4, repeatedly promotes the highest unassigned hit with $k=1,\dots,K$5 to a new seed, and collects all unassigned hits within latent distance $k=1,\dots,K$6 into the cluster; all remaining hits are labeled noise. The reported hyperparameters are $k=1,\dots,K$7 and $k=1,\dots,K$8. The model accepts up to $k=1,\dots,K$9 hits per event, each with $n_i\in\{0,1\}$^^^^3search_query3^^^^ features, constructs per-hit embeddings with BatchNorm and three fully connected layers, computes positional encodings with four GravNet blocks over a fixed tensor for all $n_i\in\{0,1\}$^^^^3all:\3^^^^ strips, and contextualizes the resulting tokens with a four-layer Transformer encoder before predicting $n_i\in\{0,1\}$^^^^3 OR all:\3^^^^ (&&&^^^^3all:\3all:\3^^^^&&&).

Evaluation is reported on one million simulated $n_i\in\{0,1\}$3 collision events. A reconstructed neutral cluster is defined as trustworthy when there exists exactly one true particle of that type within $n_i\in\{0,1\}$4 and $n_i\in\{0,1\}$5 of the reconstructed direction, and no other reconstructed cluster of the same type lies within that cone. Under this criterion, the fraction of reliable neutron clusters increases from $n_i\in\{0,1\}$6 to $n_i\in\{0,1\}$7, and the photon fraction increases from $n_i\in\{0,1\}$8 to $n_i\in\{0,1\}$9 (&&&^^^^3all:\3all:\3^^^^&&&). The study also states that it is the first application of AI clustering techniques for hodoscopic detectors (&&&^^^^3all:\3all:\3^^^^&&&).

4. Graph-condensation usage of condensation-point clustering

A distinct but related use of the term appears in GECC, a graph condensation method designed for large-scale and evolving graph data (&&&^^^^3all:\3^^^^&&&). GECC replaces gradient- or trajectory-matching approaches with a two-stage, clustering-based procedure. At each time step $t_i$^^^^3search_query3^^^^, it first applies an SGC-style propagation over the normalized adjacency,

$t_i$^^^^3all:\3^^^^

and forms propagated features

$t_i$^^^^3 OR all:\3^^^^

It then performs class-wise clustering on the propagated representations, representing class $t_i$3 by $t_i$4 synthetic nodes and producing a condensed graph $t_i$5 with node features $t_i$6 and identity adjacency (&&&^^^^3all:\3^^^^&&&).

The clustering stage uses an assignment matrix $t_i$7 and centroids $t_i$8 such that $t_i$9 approximates $p_i$^^^^3search_query3^^^^, with centroid update

$p_i$^^^^3all:\3^^^^

To control both representation distortion and parameter-matching error, GECC minimizes the balanced objective

$p_i$^^^^3 OR all:\3^^^^

where the second term penalizes deviations from ideal cluster sizes. Optimization is performed with hard or soft EM, and in evolving settings previous centroids are inherited through an incremental k-means++ scheme that augments $p_i$3 with new seeds from arriving data (&&&^^^^3all:\3^^^^&&&).

The paper provides theoretical support through three bounds. The training-stage bound shows that matching propagated features and model parameters suffices to match loss gradients; the test-stage bound implies that limiting the final parameter shift yields generalization guarantees on unseen data; and the third theorem gives

$p_i$4

so balanced cluster sizes directly tighten the parameter-distance bound (&&&^^^^3all:\3^^^^&&&). Complexity is reported as $p_i$5 for feature propagation and $p_i$6 per restart for hard or soft k-means, plus $p_i$7 for the balance term. The method is described as scaling linearly in $p_i$8 and $p_i$9 with PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3search_query3^^^^ (&&&^^^^3all:\3^^^^&&&).

Empirically, GECC is evaluated on transductive datasets including Citeseer, Cora, Pubmed, Ogbn-arxiv, and Ogbn-products, and inductive datasets including Flickr and Reddit. The abstract reports an around PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3all:\3^^^^ speedup on large datasets (&&&^^^^3all:\3^^^^&&&). The detailed Reddit example at reduction rate PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3 OR all:\3^^^^ gives PRESERVED_PLACEHOLDER_^^^^3all:\3search_query33^^^^ accuracy and PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3^^^^4 s condensation for GCond, PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3^^^^5 and PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3^^^^6 s for GEOM, and PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3^^^^7 and PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3^^^^8 s for GECC; removing feature propagation causes a PRESERVED_PLACEHOLDER_^^^^3all:\3search_query3^^^^9–PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3search_query3^^^^ absolute accuracy drop, omitting centroid reuse leads to PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3all:\3^^^^–PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3 OR all:\3^^^^ more k-means iterations at large PRESERVED_PLACEHOLDER_^^^^3all:\3all:\33^^^^, and selecting the best of PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^4 restarts by lowest PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^5 yields PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^6 on Citeseer relative to random-initialized k-means (&&&^^^^3all:\3^^^^&&&).

5. Condensation and clustering in statistical mechanics

Outside machine learning, condensation point clustering describes physical cluster formation under cooling, exclusion constraints, or size-dependent stationary weights. In the two-dimensional Lennard-Jones study under controlled exponential cooling, particles are considered bonded when their separation satisfies PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^7, and clusters are connected sets of such bonds. The temperature follows

PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^8

with PRESERVED_PLACEHOLDER_^^^^3all:\3all:\3^^^^9 and PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3search_query3^^^^. The final-equilibrium RMS displacement exhibits a broad maximum at PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3all:\3^^^^, the cluster-size distribution is well fit by a two-parameter Gamma form, and for PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3 OR all:\3^^^^ with PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\33^^^^ all particles coalesce into a single macroscopic cluster (&&&^^^^3 OR all:\33^^^^&&&).

In the driven pair exclusion process, condensation is mesoscopic rather than macroscopic. Above the threshold density

PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3^^^^4

the system develops multiple condensates with scaling

PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3^^^^5

The reported exponents are approximately PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3^^^^6 and PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3^^^^7. The drive induces spatial correlations among condensates, so a cluster of condensates appears; in two dimensions the cluster is anisotropic, with PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3^^^^8 and PRESERVED_PLACEHOLDER_^^^^3all:\3 OR all:\3^^^^9 (&&&^^^^3 OR all:\3^^^^4&&&).

A more recent rigorous treatment considers stochastic lattice gases with size-dependent stationary weights

PRESERVED_PLACEHOLDER_^^^^3all:\33search_query3^^^^

For PRESERVED_PLACEHOLDER_^^^^3all:\33all:\3^^^^, the system phase-separates into a bulk at the critical grand-canonical measure and a condensed phase containing the excess mass. In the mesoscopic regime, the characteristic condensate scale is

PRESERVED_PLACEHOLDER_^^^^3all:\33 OR all:\3^^^^

and the size-biased sampled cluster size obeys

PRESERVED_PLACEHOLDER_^^^^3all:\333^^^^

conditionally on being in the condensate, where PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^4. When PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^5, one instead obtains a single macroscopic condensate with PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^6 (&&&^^^^3 OR all:\3^^^^&&&).

A related one-dimensional hopping model with attractive logarithmic interactions has equilibrium gap distribution

PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^7

The critical coupling is PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^8, and in the scaling limit

PRESERVED_PLACEHOLDER_^^^^3all:\33^^^^9

the rescaled gap distribution for the gap containing a uniformly chosen point converges to

PRESERVED_PLACEHOLDER_^^^^3all:\3start3search_query3^^^^

The model thereby exhibits multiple correlated clusters and a self-similar spatial structure (&&&^^^^3 OR all:\3^^^^6&&&).

6. Distinctions across domains and broader theoretical interpretations

The same terminology covers formally different clustered objects. In learned object condensation, the clustered entity is a set of pixels, hits, or vertices organized around a high-confidence latent representative. In GECC, the clustered entity is a class-wise centroid that becomes a synthetic node in a condensed training graph. In statistical mechanics, the clustered entity is a particle aggregate or condensate, while in random CSPs it is a dominant solution cluster.

Domain	Representative papers	Clustered object
Object condensation	(&&&^^^^3search_query3^^^^&&&, &&&^^^^3all:\3all:\3^^^^&&&)	Vertices or hits around learned condensation points
Graph condensation	(&&&^^^^3all:\3^^^^&&&)	Propagated node embeddings around class-wise centroids
Physical or combinatorial condensation	(&&&^^^^3 OR all:\33^^^^&&&, &&&^^^^3 OR all:\3^^^^4&&&, &&&^^^^3 OR all:\3^^^^&&&, Sly et al., 2023, Billam et al., 2013)	Particle clusters, condensates, vortices, or solution clusters

This distinction is especially important in the two theoretical literatures that use “condensation” without a learned latent geometry. In decaying two-dimensional quantum turbulence, negative-temperature states exhibit both macroscopic vortex clustering and kinetic-energy condensation, termed an Onsager-Kraichnan condensate; the clustered fraction, maximum cluster charge, pair-sign correlations, and the large-scale excess in the incompressible kinetic-energy spectrum serve as order parameters (Billam et al., 2013). In random regular NAE-SAT, the condensation regime is the interval PRESERVED_PLACEHOLDER_^^^^3all:\3start3all:\3^^^^ in which PRESERVED_PLACEHOLDER_^^^^3all:\3start3 OR all:\3^^^^ clusters carry almost all solutions, and the local weak limit becomes explicitly non-Markovian because the solution measure is dominated by a small number of large clusters (Sly et al., 2023).

Across these settings, the common structural theme is concentration onto a reduced set of representatives: latent representatives in object condensation, centroids in graph condensation, and dominant physical or combinatorial clusters in condensed phases. The technical content, however, remains domain-specific. In detector reconstruction the decisive quantities are PRESERVED_PLACEHOLDER_^^^^3all:\3^^^^43, learned coordinates, and pull-push losses; in graph condensation they are propagated embeddings, balanced assignments, and centroid evolution; in statistical mechanics and CSPs they are cluster-size distributions, critical densities or temperatures, and the transition from many small clusters to a small number of dominant ones (&&&^^^^3search_query3^^^^&&&, &&&^^^^3all:\3^^^^&&&, &&&^^^^3 OR all:\3^^^^&&&, Sly et al., 2023).