#### DMCA

## AFFINITY PROPAGATION: CLUSTERING DATA BY PASSING MESSAGES (2009)

### Cached

### Download Links

Citations: | 8 - 0 self |

### Citations

11970 | Maximum likelihood from incomplete data via the em algorithm
- Dempster, Laird, et al.
- 1977
(Show Context)
Citation Context ...e Figure 2.1(B–C)), learning mixing weights on class priors (instead of assuming ∀k : πk = 1 ), and to account for cluster assignment uncertainty by using the ExpectationK Maximization (EM) algorithm =-=[23]-=- and representing it with a simple distribution Q(z) = ∏ N i=1 ∏ K . Cluster assignments can be determined by minimizing the Kullback-Leibler divergence [18] between Q(z) and P(z|x), D (Q(z)‖P(z|x)) =... |

8904 |
Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ...computationallyefficient rearrangement of these sums and products. 2.6.2 Sum-Product Algorithm A notable formulation of the sum-product algorithm is Judea Pearl’s use of it as “belief propagation” in =-=[84]-=- for marginalizing variables in Bayesian networks. The algorithm is a series of rules—framed as passing messages between factor graph nodes—that organize and automate the application of the distributi... |

5157 | Optimization by simulated annealing
- Kirkpatrick, Gelatt, et al.
- 1983
(Show Context)
Citation Context ...], which corresponds to finding a K = 2 clustering solution whose search space is only O(N 2 ). There are many approximate techniques for finding general K-way graph cuts, such as simulated annealing =-=[13, 62, 75]-=-, Gibbs sampling [41], and iterated conditional modes [6], but more recent techniques such as using expansion moves and swap moves [9] have shown greatly improved performance. For example, α-expansion... |

5126 | Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...ng a K = 2 clustering solution whose search space is only O(N 2 ). There are many approximate techniques for finding general K-way graph cuts, such as simulated annealing [13, 62, 75], Gibbs sampling =-=[41]-=-, and iterated conditional modes [6], but more recent techniques such as using expansion moves and swap moves [9] have shown greatly improved performance. For example, α-expansion moves involve iterat... |

3626 | Equation of state calculations by fast computing machines
- Metropolis, Rosenbluth, et al.
- 1953
(Show Context)
Citation Context ...], which corresponds to finding a K = 2 clustering solution whose search space is only O(N 2 ). There are many approximate techniques for finding general K-way graph cuts, such as simulated annealing =-=[13, 62, 75]-=-, Gibbs sampling [41], and iterated conditional modes [6], but more recent techniques such as using expansion moves and swap moves [9] have shown greatly improved performance. For example, α-expansion... |

3053 | Some methods for classification and analysis of multivariate observations. In:
- MacQueen
- 1967
(Show Context)
Citation Context ...i, k)}, and the number of exemplars to find (K) or a real-valued exemplar cost to balance against similarities. A simple and fast algorithm for finding clustering solutions is the k-medoids algorithm =-=[70]-=-, which begins by randomly selecting a set of K data points as initial exemplars and then refines these in alternating steps as shown in Figure 1.2. The algorithm monotonically maximizes the sum of si... |

2120 | Fast approximate energy minimization via graph cuts
- Boykov, Veksler, et al.
- 2001
(Show Context)
Citation Context ...ng general K-way graph cuts, such as simulated annealing [13, 62, 75], Gibbs sampling [41], and iterated conditional modes [6], but more recent techniques such as using expansion moves and swap moves =-=[9]-=- have shown greatly improved performance. For example, α-expansion moves involve iteratively solving binary subproblems constructed by choosing one class label and lumping the remainder in the other c... |

1776 | Near shannon limit errorcorrecting coding and decoding: Turbo-codes,”
- Berrou, Glavieux, et al.
- 1993
(Show Context)
Citation Context ...y propagation achieves outstanding results by employing “loopy belief propagation” techniques (see Section 2.6) that have previously been used to approach Shannon’s limit in error-correcting decoding =-=[5]-=- and solve random satisfiability problems with an order-ofmagnitude increase in size [76]. The objective function it maximizes is the net similarity, S, which is the sum of the similarities of non-exe... |

1708 | On Spectral clustering: analysis and an algorithm
- Ng, Jordan, et al.
- 2001
(Show Context)
Citation Context ...re satisfactory results.CHAPTER 4. BENCHMARKING AFFINITY PROPAGATION 72 tive similarity) between points to partition the dataset into clusters. For these simulations, a spectral clustering algorithm =-=[80]-=- (based on [74, 104]) is used. Briefly, the K largest eigenvectors are computed (using MATLAB’s svds command) for an N ×N normalized distance matrix with diagonal elements set to zero and stacked into... |

1147 |
Hierarchical grouping to optimize an objective function.
- Ward
- 1963
(Show Context)
Citation Context ...n in Fig. 4.185 . Hierarchical Agglomerative Clustering Hierarchical agglomerative clustering is shown alongside CPLEX and affinity propagation in Figure 4.19. Hierarchical agglomerative clustering ( =-=[61, 93, 103]-=-) involves creating a linkage structure (easily visualized as a dendrogram) for a dataset containing a series of nested clusters, beginning with N clusters and ending with one large cluster. At each o... |

1106 | Flows in Networks,
- Ford, Fulkerson
- 1962
(Show Context)
Citation Context ...e approaches to exemplar-based clustering borrow from techniques employed for minimizing the sum of cut weights while partitioning graphs (graph cuts) or its dual formulation, maximizing network flow =-=[31, 32]-=-. The optimal two-way (binary) graph-cut can be found in polynomial time [43], which corresponds to finding a K = 2 clustering solution whose search space is only O(N 2 ). There are many approximate t... |

1047 | What energy functions can be minimized via graph cuts
- Kolmogorov, Zabih
- 2004
(Show Context)
Citation Context ...ves involve finding swaps of two class labels that improve the objective similar to the vertex substitution heuristic described in Section 2.5. A useful overview of these formulations can be found in =-=[64]-=-. 2.5 The Facility Location Problem Facility location is an important area of operations research. Simply stated, it is concerned with finding facility locations to be matched with subsets of customer... |

1006 | Texture synthesis by non-parametric sampling,‖ in
- Efros, Leung
- 1999
(Show Context)
Citation Context ...e a clustering of input features as output or require it as a preprocessing step for subsequent analysis. Exemplars have been used with success in a variety of vision tasks, including image synthesis =-=[27, 101]-=-, super-resolution [33, 92], image and video completion [52, 105], and combined tracking and object detection [40, 97]. The use of exemplars is attractive for several reasons. A relatively small numbe... |

860 | A new polynomial-time algorithm for linear programming.
- Karmarkar
- 1984
(Show Context)
Citation Context ... BACKGROUND 19 can be included ( ∑ N k=1 bkk = K) if the net similarity, S, is being minimized and not just the data-point similarity, Sdata. A common approach is to solve a linear program relaxation =-=[55, 57]-=- where ∀i, j : ˆ bij ∈ R and 0 ≤ ˆ bij ≤ 1 or implemented in optimization software packages such as CPLEX [19]. If the resulting solution is non-integer, stochastic rounding techniques or heuristics [... |

696 | Clustering by passing messages between data points,”
- Frey, Dueck
- 2007
(Show Context)
Citation Context ...ergence for a coherent solution. The simplicity and effectiveness of these update equations have made it the standard incarnation of affinity propagation since its initial 2007 publication in Science =-=[38]-=-. All experiments in Chapters 4–5 use this form of the algorithm. It is interesting to note that the greedy k-medoids clustering algorithm can be rewritten to use responsibilities and thus more closel... |

676 | Loopy belief propagation for approximate inference: an empirical study”,
- Murphy, Weiss, et al.
- 1999
(Show Context)
Citation Context ... results from the area of coding theory [5] were shown to be a special case of belief propagation in this loopy case [73]. Additional factor graphs (involving applications such as medical diagnostics =-=[79]-=- and phase-unwrapping [63]) were investigated at the time; positive results gave additional credence to the idea that pseudo-marginals provided useful enough approximations that argmax xn q(xn) = argm... |

528 |
Numerical Taxonomy. The principles and Practice of Numerical Taxonomy Classification.
- Sneath, Sokal
- 1973
(Show Context)
Citation Context ...n in Fig. 4.185 . Hierarchical Agglomerative Clustering Hierarchical agglomerative clustering is shown alongside CPLEX and affinity propagation in Figure 4.19. Hierarchical agglomerative clustering ( =-=[61, 93, 103]-=-) involves creating a linkage structure (easily visualized as a dendrogram) for a dataset containing a series of nested clusters, beginning with N clusters and ending with one large cluster. At each o... |

501 |
A thermodynamical approach to the traveling salesman problem.
- Cerny
- 1985
(Show Context)
Citation Context ...], which corresponds to finding a K = 2 clustering solution whose search space is only O(N 2 ). There are many approximate techniques for finding general K-way graph cuts, such as simulated annealing =-=[13, 62, 75]-=-, Gibbs sampling [41], and iterated conditional modes [6], but more recent techniques such as using expansion moves and swap moves [9] have shown greatly improved performance. For example, α-expansion... |

443 | Clustering with Bregman Divergences.
- Banerjee, Merugu, et al.
- 2005
(Show Context)
Citation Context ... the problem can be reformulated to maximizing ∑N i=1 log [ ] ∑N −βdϕ(xi,xj) j=1 πje where mixture component densities are located at each data 1 N point. Here dϕ(xi, xj) must be a Bregman divergence =-=[3]-=- (e.g. dϕ(xi, xj) = −s(i, j)), and so the latter likelihood is convex whose global optimum, subject to N∑ πj = 1, can be found in polynomial time. The β parameter is used to control the sharpness of t... |

433 |
Maximal flow through a network,”
- Ford, Fulkerson
- 1956
(Show Context)
Citation Context ...e approaches to exemplar-based clustering borrow from techniques employed for minimizing the sum of cut weights while partitioning graphs (graph cuts) or its dual formulation, maximizing network flow =-=[31, 32]-=-. The optimal two-way (binary) graph-cut can be found in polynomial time [43], which corresponds to finding a K = 2 clustering solution whose search space is only O(N 2 ). There are many approximate t... |

404 | Turbo decoding as an instance of Pearl’s “belief propagation” algorithm.”
- McEliece, MacKay, et al.
- 1998
(Show Context)
Citation Context ...des of the network. (p. 195)CHAPTER 2. BACKGROUND 28 Previous impressive empirical results from the area of coding theory [5] were shown to be a special case of belief propagation in this loopy case =-=[73]-=-. Additional factor graphs (involving applications such as medical diagnostics [79] and phase-unwrapping [63]) were investigated at the time; positive results gave additional credence to the idea that... |

380 | Segmentation using eigenvectors: A unifying view.
- Weiss
- 1999
(Show Context)
Citation Context ... results.CHAPTER 4. BENCHMARKING AFFINITY PROPAGATION 72 tive similarity) between points to partition the dataset into clusters. For these simulations, a spectral clustering algorithm [80] (based on =-=[74, 104]-=-) is used. Briefly, the K largest eigenvectors are computed (using MATLAB’s svds command) for an N ×N normalized distance matrix with diagonal elements set to zero and stacked into an N ×K matrix whos... |

374 |
Graph Clustering by Flow Simulation.
- Dongen
- 2000
(Show Context)
Citation Context ...nsistent with other experiments [82], the convex clustering algorithm seems to have poor performance in practice 6 . j=1 Markov Clustering Algorithm (van Dongen, 2000) The Markov clustering algorithm =-=[99]-=- is a graph-based clustering algorithm based on simulation of stochastic flow in graphs; results of the algorithm applied to the Olivetti faces are shown alongside CPLEX and affinity propagation in Fi... |

359 | The Generalized Distributive Law,”
- Aji, McEliece
- 2000
(Show Context)
Citation Context ...lier. It should be noted that for singly-connected graphs the factorization in equation (2.20) is exact and the algorithm converges to the exact marginals. 2.6.4 Max-Product Algorithm As described in =-=[1]-=-, the idea behind the sum-product algorithm and factor graphs can be applied to any commutative semiring6 . In many cases it is more desirable or efficient to use the maxproduct algorithm, whose messa... |

295 |
An algorithm for integer solutions to linear programs,"
- Gomory
- 1963
(Show Context)
Citation Context ...ontaining millions of constraints. For the exact solutions shown in Section 4.1, CPLEX 7.1 software was utilized which takes advantage of branch-and-bound techniques and Gomory’s cutting-plane method =-=[42]-=-. Other possible approaches to exemplar-based clustering borrow from techniques employed for minimizing the sum of cut weights while partitioning graphs (graph cuts) or its dual formulation, maximizin... |

279 |
An algorithmic approach to network location problems, part ii: the p-medians,
- Kariv, Hakimi
- 1979
(Show Context)
Citation Context ...APTER 2. BACKGROUND 18 K. 2.4.1 Linear Programming Relaxation Maximizing the net similarity objective function in equation (2.9)—or even Sdata for that matter—has been shown to be N P-hard in general =-=[56]-=-. Linear programming relaxations can, however, be employed to find optimal solutions in small problems where N < 1000; this is outlined in the 0–1 integer program of equations (2.10–2.11). 0–1 INTEGER... |

253 |
A best possible heuristic for the k-Center problem,”
- Hochbaum, Shmoys
- 1985
(Show Context)
Citation Context ...oming this. Furthest-first traversal Parametric clustering algorithms are sensitive to the initial set of cluster centers, µ (0) , so a common initialization that often lies near a good solution (see =-=[49]-=- for theory) is to construct an initial set of centers with a furthest-first traversal. Specifically, the center µ (0) 1 is a random data point xi1, and subsequent centers, µ (0) k , are set to the “f... |

249 | A constant-factor approximation algorithm for the k-median problem
- Charikar, Guha, et al.
- 1999
(Show Context)
Citation Context ...unction if possible, i.e. maxℓ [D(L)−D(L∪n\ℓ)]. This process is repeated until convergence, when no cost-reducing substitutions are possible. Some algorithms have provable worst-case guarantees (e.g. =-=[14]-=-), whereby their solution’s cost D(L) is related to the optimal cost D(L∗ ∣ ∣ ) by a constant factor as follows: ∣≤ε. ∣ D(L)−D(L∗) D(L) Values of ε are rarely small and often much larger than the typi... |

243 | Real-time object detection for “smart” vehicles. Paper presented at the Computer Vision,
- Gavrila, Philomin
- 1999
(Show Context)
Citation Context ...e been used with success in a variety of vision tasks, including image synthesis [27, 101], super-resolution [33, 92], image and video completion [52, 105], and combined tracking and object detection =-=[40, 97]-=-. The use of exemplars is attractive for several reasons. A relatively small number of representative exemplars can capture high-order statistics, since each exemplar can simultaneously express depend... |

217 |
On the statistical analysis of dirty pictures (with discussions):
- Besag
- 1986
(Show Context)
Citation Context ...search space is only O(N 2 ). There are many approximate techniques for finding general K-way graph cuts, such as simulated annealing [13, 62, 75], Gibbs sampling [41], and iterated conditional modes =-=[6]-=-, but more recent techniques such as using expansion moves and swap moves [9] have shown greatly improved performance. For example, α-expansion moves involve iteratively solving binary subproblems con... |

202 |
Optimum Locations of Switching Centers and the Absolute Centers and Medians of a Graph.
- Hakimi
- 1964
(Show Context)
Citation Context ...sely related to the p-median problem. The p-median problem was formally defined and investigated in literature from the early 1960s with notable contributions from Cooper ([15],[17],[16]) and Hakimi (=-=[45]-=-,[46]); for a more recent survey of approaches to the problem see [78]. It is defined as follows: given a set, N , of possible facility locations and a set, M, of customers to be serviced, select a su... |

197 |
Markov random fields and their applications,
- Kindermann, Snell
- 1980
(Show Context)
Citation Context ...nctly expressing and visualizing the structure and dependencies present in networks of variables. 2.6.1 Factor Graphs Standard graphical models such as Bayesian networks [83] and Markov random fields =-=[60]-=- have long been used for modeling hierarchical dependencies and energy-based models, respectively. A more recent innovation is the factor graph [65], a graphical model that provides a natural way of r... |

196 | Analytic and algorithmic solution of random satisfiability problems,
- Mezard, Parisi, et al.
- 2002
(Show Context)
Citation Context ...iques (see Section 2.6) that have previously been used to approach Shannon’s limit in error-correcting decoding [5] and solve random satisfiability problems with an order-ofmagnitude increase in size =-=[76]-=-. The objective function it maximizes is the net similarity, S, which is the sum of the similarities of non-exemplar data points to their exemplars plus the sum of exemplar preferences (negative costs... |

149 |
Location of Bank Accounts to Optimize Float: An Analytic Study of Exact and Approximate Algorithms,"
- Cornuejols, Fisher, et al.
- 1977
(Show Context)
Citation Context ...has p been shown to be N P-hard in general [56]. An exact solution is possible for many problems with hundreds of facilities based on linear programming relaxations of the integer programming problem =-=[11,86]-=-. MN binary variables {bmn} are introduced to indicate which facilities serve each customer (i.e., bmn =1 if customer m is served by facility n, and bmn =0 otherwise), and N binary variables {an} indi... |

142 | Space-time video completion
- Wexler, Irani
- 2004
(Show Context)
Citation Context ...processing step for subsequent analysis. Exemplars have been used with success in a variety of vision tasks, including image synthesis [27, 101], super-resolution [33, 92], image and video completion =-=[52, 105]-=-, and combined tracking and object detection [40, 97]. The use of exemplars is attractive for several reasons. A relatively small number of representative exemplars can capture high-order statistics, ... |

141 | Learning segmentation by random walks.
- Meila, Shi
- 2001
(Show Context)
Citation Context ... results.CHAPTER 4. BENCHMARKING AFFINITY PROPAGATION 72 tive similarity) between points to partition the dataset into clusters. For these simulations, a spectral clustering algorithm [80] (based on =-=[74, 104]-=-) is used. Briefly, the K largest eigenvectors are computed (using MATLAB’s svds command) for an N ×N normalized distance matrix with diagonal elements set to zero and stacked into an N ×K matrix whos... |

124 |
A heuristic program for locating warehouses.
- Kuehn, Hamburger
- 1963
(Show Context)
Citation Context ...lutions via linear programming relaxations are usually unavailable with current computing technology so the task is left to heuristics. Standard facility-location heuristics include: Greedy Heuristic =-=[66]-=-: Initialize the set of open facilities, L (0), to be the empty set. Perform p rounds during which an unopened facility nt ∈ M\L is opened during the tth round (L(t) = L (t−1) ∪ nt) so that the cost d... |

118 | A.: Epitomic analysis of appearance and shape. In:
- Jojic, Frey, et al.
- 2003
(Show Context)
Citation Context ...processing step for subsequent analysis. Exemplars have been used with success in a variety of vision tasks, including image synthesis [27, 101], super-resolution [33, 92], image and video completion =-=[52, 105]-=-, and combined tracking and object detection [40, 97]. The use of exemplars is attractive for several reasons. A relatively small number of representative exemplars can capture high-order statistics, ... |

112 | Tightening LP relaxations for MAP using message passing.
- Sontag, Meltzer, et al.
- 2008
(Show Context)
Citation Context ...uestions for further research: The relationship between max-product belief propagation and linear programming relaxations is not well-understood but is beginning to be more widely investigated (e.g., =-=[94, 109]-=-). In [88], a linear programming relaxation for the weighted matching problem is compared to max-product belief propagation with a proof that “if the [linear programming] relaxation isCHAPTER 6. CONC... |

111 |
Step-wise clustering procedures.
- King
- 1967
(Show Context)
Citation Context ...n in Fig. 4.185 . Hierarchical Agglomerative Clustering Hierarchical agglomerative clustering is shown alongside CPLEX and affinity propagation in Figure 4.19. Hierarchical agglomerative clustering ( =-=[61, 93, 103]-=-) involves creating a linkage structure (easily visualized as a dendrogram) for a dataset containing a series of nested clusters, beginning with N clusters and ending with one large cluster. At each o... |

105 |
Heuristic methods for estimating the generalized vertex median of a weighted graph.
- Teitz, Bart
- 1968
(Show Context)
Citation Context ...ng users to the closestCHAPTER 2. BACKGROUND 23 opened facility and open facilities are replaced by new facilities nearest to median of their customers’ location. Vertex Substitution Heuristic (VSH) =-=[96]-=-: Randomly initialize L to contain p facilities. For each unopened facility n∈M\L, find the open facility, ℓ∈L to substitute with it so as to most-improve the cost function if possible, i.e. maxℓ [D(L... |

95 |
Bayesian Networks: A Model of Self-Activated Memory for Evidential Reasoning.
- Pearl
- 1985
(Show Context)
Citation Context ... are a useful device for succinctly expressing and visualizing the structure and dependencies present in networks of variables. 2.6.1 Factor Graphs Standard graphical models such as Bayesian networks =-=[83]-=- and Markov random fields [60] have long been used for modeling hierarchical dependencies and energy-based models, respectively. A more recent innovation is the factor graph [65], a graphical model th... |

93 |
Optimum distribution of switching centers in a communication network and some related graph theoretic problems.
- Hakimi
- 1965
(Show Context)
Citation Context ...related to the p-median problem. The p-median problem was formally defined and investigated in literature from the early 1960s with notable contributions from Cooper ([15],[17],[16]) and Hakimi ([45],=-=[46]-=-); for a more recent survey of approaches to the problem see [78]. It is defined as follows: given a set, N , of possible facility locations and a set, M, of customers to be serviced, select a subset... |

84 |
Concept acquisition through representational adjustment. Unpublished Doctoral dissertation,
- Schlimmer
- 1987
(Show Context)
Citation Context ...ts, to uniform-loss embeddable in 121-dimensional space for mushrooms, to extremely-high dimension as is the case for the sparse movie ratings data. 4.2.1 Mushroom data (N =8124) The mushroom dataset =-=[90]-=- is available at the UCI Machine Learning Respository8 . It contains 22 discrete, non-ordinal attributes for each of 8124 mushrooms as illustrated in Figure 4.24. These attributes include color, shape... |

80 |
Location-allocation problems.
- Cooper
- 1963
(Show Context)
Citation Context ...escribed in Section 2 is closely related to the p-median problem. The p-median problem was formally defined and investigated in literature from the early 1960s with notable contributions from Cooper (=-=[15]-=-,[17],[16]) and Hakimi ([45],[46]); for a more recent survey of approaches to the problem see [78]. It is defined as follows: given a set, N , of possible facility locations and a set, M, of customers... |

80 | Approximation algorithms for geometric median problems
- Lin, Vitter
- 1992
(Show Context)
Citation Context ...] where ∀i, j : ˆ bij ∈ R and 0 ≤ ˆ bij ≤ 1 or implemented in optimization software packages such as CPLEX [19]. If the resulting solution is non-integer, stochastic rounding techniques or heuristics =-=[69]-=- have been shown to produce satisfactory results. With current computing technology, such approaches are feasible for problems up to about 1000 data points containing millions of constraints. For the ... |

78 |
Statistical theory of superlattices
- Bethe
- 1935
(Show Context)
Citation Context ...l justification for loopy belief propagation was later shown [111], where update equations (2.17–2.19) were related to minimizing a KL-divergence [18] computed using Bethe’s free energy approximation =-=[7]-=- from 1935. Given a factor graph describing a probability density of P(x) ∝ ∏ m fm(xN(m)), one can search for a simpler approximating distribution, Q(x), such that the KL-divergence between them, D(Q(... |

66 | Probabilistic tracking with exemplars in a metric space,”
- Toyama, Blake
- 2002
(Show Context)
Citation Context ...e been used with success in a variety of vision tasks, including image synthesis [27, 101], super-resolution [33, 92], image and video completion [52, 105], and combined tracking and object detection =-=[40, 97]-=-. The use of exemplars is attractive for several reasons. A relatively small number of representative exemplars can capture high-order statistics, since each exemplar can simultaneously express depend... |

65 | Space-time super-resolution.
- Shechtman, Caspi, et al.
- 2005
(Show Context)
Citation Context ...ures as output or require it as a preprocessing step for subsequent analysis. Exemplars have been used with success in a variety of vision tasks, including image synthesis [27, 101], super-resolution =-=[33, 92]-=-, image and video completion [52, 105], and combined tracking and object detection [40, 97]. The use of exemplars is attractive for several reasons. A relatively small number of representative exempla... |

64 |
A two-round variant of em for gaussian mixtures.
- Dasgupta, Schulman
- 2000
(Show Context)
Citation Context ...ROUND 15 or splits clusters (e.g., the center describing data with the lowest probability). The specifics of split-and-merge criteria are described in [98]. k·log(k) heuristic Dasgupta et al. show in =-=[21, 22]-=- for high-dimensional Gaussians (where dimension D ≫ ln K) that the EM algorithm for a mixture of Gaussians can avoid many poor local minima by initializing the algorithm with O(K ln K) Gaussians and ... |

58 |
Variable neighborhood search for the p-median.
- Hansen, Mladenovic
- 1997
(Show Context)
Citation Context ...to selecting an algorithm [78]. The vertex substitution heuristic [96] has been the standard for comparison for four decades and provides the basis for the variable neighborhood search meta-heuristic =-=[47]-=- that was compared with affinity propagation in [10, 35]. Variable-neighborhood search utilizes speedups to the original vertex substitution heuristic by storing all nearest and second-nearest open fa... |

52 | Non-metric affinity propagation for unsupervised image categorization.
- Dueck, Frey
- 2007
(Show Context)
Citation Context ...i max [0, r(i′ , k)], for k=i ; ( min − max i ′:i′/∈{i,k} min [0, r(i′ , k)], r(k, k) + ∑ max [0, r(i ′ ) , k)] , for k ̸=i . i ′ :i ′ /∈{i,k} (3.20) This algorithm, while still O(N 2 ), was shown in =-=[25]-=- to have quite inferior performance due to the extra constraint unnecessarily preventing the algorithm from moving through regions of the search space on the way to better solutions. To compare the tw... |

49 |
On the location of supply points to minimize transport costs.
- Maranzana
- 1964
(Show Context)
Citation Context ...es, L (0) to be N . Perform M−p rounds during which one open facility nt ∈ L (t) is closed so that the cost increase between rounds, ∣ ∣ D(L (t) )−D(L (t−1) ) ∣ ∣, is minimized. Alternating Heuristic =-=[72]-=-: The alternating heuristic is identical to k-medoids clustering in equation (2.7), whereby there are alternating phases of assigning users to the closestCHAPTER 2. BACKGROUND 23 opened facility and ... |

45 |
central facilities location’,
- Revelle, SWAin
- 1970
(Show Context)
Citation Context ...has p been shown to be N P-hard in general [56]. An exact solution is possible for many problems with hundreds of facilities based on linear programming relaxations of the integer programming problem =-=[11,86]-=-. MN binary variables {bmn} are introduced to indicate which facilities serve each customer (i.e., bmn =1 if customer m is served by facility n, and bmn =0 otherwise), and N binary variables {an} indi... |

44 |
Heuristic methods for location–allocation problems,”
- Cooper
- 1964
(Show Context)
Citation Context ...n Section 2 is closely related to the p-median problem. The p-median problem was formally defined and investigated in literature from the early 1960s with notable contributions from Cooper ([15],[17],=-=[16]-=-) and Hakimi ([45],[46]); for a more recent survey of approaches to the problem see [78]. It is defined as follows: given a set, N , of possible facility locations and a set, M, of customers to be ser... |

40 |
DR: NCBI Reference Sequence project: update and current status. Nucleic Acids Res
- KD, Tatusova, et al.
- 2003
(Show Context)
Citation Context ...ies, DNA segments assigned to exemplars other than the non-transcribed exemplar were considered to be parts of genes. All DNA segments were separately mapped to the RefSeq database of annotated genes =-=[85]-=- to produce labels used for reporting true positive and false positive rates. These results are compared in Figure 5.5, where the true-positive (TP) rate is plotted against the false-positive (FP) rat... |

38 | Learning to estimate scenes from images
- Freeman, Pasztor
- 1999
(Show Context)
Citation Context ...ures as output or require it as a preprocessing step for subsequent analysis. Exemplars have been used with success in a variety of vision tasks, including image synthesis [27, 101], super-resolution =-=[33, 92]-=-, image and video completion [52, 105], and combined tracking and object detection [40, 97]. The use of exemplars is attractive for several reasons. A relatively small number of representative exempla... |

37 | Convex clustering with exemplar-based models,”
- Lashkari, Golland
- 2008
(Show Context)
Citation Context ...gure 4.18.CHAPTER 4. BENCHMARKING AFFINITY PROPAGATION 71 equivalent to the definition of similarity. Convex Clustering (Lashkari-Golland, 2007) The convex clustering method (Lashkari-Golland, 2007) =-=[67]-=- is shown alongside CPLEX and affinity propagation in Figure 4.20. This algorithm is based on the idea that instead of max∑N i=1 log [ ∑K k=1 πkf(xi; ] µk) (where f is imizing a typical mixture-model ... |

33 | A probabilistic analysis of EM for mixtures of separated, spherical gaussians.
- Dasgupta, Schulman
- 2007
(Show Context)
Citation Context ...ROUND 15 or splits clusters (e.g., the center describing data with the lowest probability). The specifics of split-and-merge criteria are described in [98]. k·log(k) heuristic Dasgupta et al. show in =-=[21, 22]-=- for high-dimensional Gaussians (where dimension D ≫ ln K) that the EM algorithm for a mixture of Gaussians can avoid many poor local minima by initializing the algorithm with O(K ln K) Gaussians and ... |

32 | Clustering by soft-constraint affinity propagation: applications to gene-expression data,
- Leone, Sumedha, et al.
- 2007
(Show Context)
Citation Context ...gation has spawned a growing volume of research, such as: • Vector quantization codebook design (Jiang et al. in [51]) • Soft-constraint affinity propagation for gene expression data (Leone et al. in =-=[68]-=-) • Multiple view image segmentation (Xiao et al. in [108]) • Finding light sources using images (An et al. in [2]) • Image categorization and normalized mutual information analysis (Grira et al. in [... |

31 |
A fast algorithm for the greedy interchange of large-scale clustering and median location problems.
- Whitaker
- 1983
(Show Context)
Citation Context ... speedups to the original vertex substitution heuristic by storing all nearest and second-nearest open facilities for each customer and only recomputing certain elements in these lists when necessary =-=[106]-=- (i.e., a pertinent substitution is made). It also restructures the p(N −p) possible interchanges to involve fewer comparisons with early exit conditions, and randomly chooses higher-order π-ary inter... |

30 | Linear programming analysis of loopy belief propagation for weighted matching.
- Sanghavi, Malioutov, et al.
- 2007
(Show Context)
Citation Context ...rther research: The relationship between max-product belief propagation and linear programming relaxations is not well-understood but is beginning to be more widely investigated (e.g., [94, 109]). In =-=[88]-=-, a linear programming relaxation for the weighted matching problem is compared to max-product belief propagation with a proof that “if the [linear programming] relaxation isCHAPTER 6. CONCLUSIONS AN... |

29 |
Graphs and the Sum-Product Algorithm.
- Factor
- 2001
(Show Context)
Citation Context ... as Bayesian networks [83] and Markov random fields [60] have long been used for modeling hierarchical dependencies and energy-based models, respectively. A more recent innovation is the factor graph =-=[65]-=-, a graphical model that provides a natural way of representing global functions or probability distributions that can be factored into simpler local functions. A factor graph is a bi-partite graph co... |

29 | The p-median problem: A survey of metaheuristic approaches,
- Mladenovic, Brimberg, et al.
- 2007
(Show Context)
Citation Context ...y defined and investigated in literature from the early 1960s with notable contributions from Cooper ([15],[17],[16]) and Hakimi ([45],[46]); for a more recent survey of approaches to the problem see =-=[78]-=-. It is defined as follows: given a set, N , of possible facility locations and a set, M, of customers to be serviced, select a subsetCHAPTER 2. BACKGROUND 21 L⊆M of those facilities to open (where p... |

28 | Sparse nonnegative matrix factorization for clustering.
- Kim, Park
- 2008
(Show Context)
Citation Context ... et al. in [112]) • Analysis of hydrophobic-polar protein model (Santana et al. in [89]) • Face recognition with linear discriminant analysis (Du et al. in [24]) • Clustering text data (Kim et al. in =-=[59]-=-) • Adaptive extensions of affinity propagation (Wang et al. in [102]) • Knowledge discovery in medical data sources (Senf et al. in [91]) • Analysis of land-use and land-cover data (Cardille et al. i... |

27 | Face Identification and Feature Extraction using Hidden Markov Models.
- Samaria, Fallside
- 1993
(Show Context)
Citation Context ...rks affinity propagation alongside 15 other clustering methods for a range of small (N < 1000) and large (N > 5000) datasets. 4.1 Olivetti faces: Clustering a small dataset The Olivetti faces dataset =-=[87]-=- is a collection of 400 64×64 greyscale images of human faces (10 from each of 40 people) with varying facial expressions and lighting conditions. The complete dataset is available at http://www.cs.to... |

26 | Mixture modeling by affinity propagation
- Frey, Dueck
- 2006
(Show Context)
Citation Context ...her computational complexity (3.4.2-3.4.4) than affinity propagation and are thus not developed further. 3.4.1 Affinity propagation with added non-empty cluster constraint In its earliest formulation =-=[34]-=-, affinity propagation clustering employed a factor graph with identical topology but more stringent constraint functions, {f ′ k (c)}Nk=1 . Not only did this function disallow exemplar-less clusters ... |

24 | Genome-wide analysis of mouse transcripts using exon microarrays and factor graphs
- Frey, Mohammad, et al.
- 2005
(Show Context)
Citation Context ...indicating the putative exon’s function. Also, when nearby segments of DNA undergo coordinated transcription across multiple tissues, they are likely to come from transcribed regions of the same gene =-=[36]-=-. By grouping together feature vectors for nearby probes, we can detect genes and variations of genes. Figure 5.3(A) shows a normalized subset of the data and gives three examples of groups of nearby ... |

24 |
Image synthesis from a single example image
- Vetter, Poggio
- 1996
(Show Context)
Citation Context ...e a clustering of input features as output or require it as a preprocessing step for subsequent analysis. Exemplars have been used with success in a variety of vision tasks, including image synthesis =-=[27, 101]-=-, super-resolution [33, 92], image and video completion [52, 105], and combined tracking and object detection [40, 97]. The use of exemplars is attractive for several reasons. A relatively small numbe... |

20 | Split and merge em algorithm for improving gaussian mixture density estimates,”
- Ueda, Nakano, et al.
- 2000
(Show Context)
Citation Context ...tically decrease the likelihood)CHAPTER 2. BACKGROUND 15 or splits clusters (e.g., the center describing data with the lowest probability). The specifics of split-and-merge criteria are described in =-=[98]-=-. k·log(k) heuristic Dasgupta et al. show in [21, 22] for high-dimensional Gaussians (where dimension D ≫ ln K) that the EM algorithm for a mixture of Gaussians can avoid many poor local minima by ini... |

18 |
Warehouse location under continuous economies of scale.
- Feldman, Lehrer, et al.
- 1966
(Show Context)
Citation Context ...s during which an unopened facility nt ∈ M\L is opened during the tth round (L(t) = L (t−1) ∪ nt) so that the cost decrease between rounds, ∣ ∣ (t) (t−1) D(L )−D(L ) ∣, is maximized. Stingy Heuristic =-=[29]-=-: Initialize the set of open facilities, L (0) to be N . Perform M−p rounds during which one open facility nt ∈ L (t) is closed so that the cost increase between rounds, ∣ ∣ D(L (t) )−D(L (t−1) ) ∣ ∣,... |

14 | Using “epitomes” to model genetic diversity: Rational design of HIV vaccine cocktails. In
- Jojic, Jojic, et al.
- 2006
(Show Context)
Citation Context ...mers are used). See Figure 5.7 for more details. The utility u(T, R) of a strain T for a fragment R would ideally be set to its potential for immunological protection, but following the approaches in =-=[30,53,54,81]-=-, it is set to the frequency of the fragment in the database of HIV sequences if fragment R is present in strain T , and zero otherwise, as in equation (5.3). ⎧ ⎪⎨ frequency of R in HIV sequence datab... |

13 |
Solutions of generalized locational equilibrium models,”
- Cooper
- 1967
(Show Context)
Citation Context ...bed in Section 2 is closely related to the p-median problem. The p-median problem was formally defined and investigated in literature from the early 1960s with notable contributions from Cooper ([15],=-=[17]-=-,[16]) and Hakimi ([45],[46]); for a more recent survey of approaches to the problem see [78]. It is defined as follows: given a set, N , of possible facility locations and a set, M, of customers to b... |

11 | Response to comment on ”clustering by passing messages between data points
- Frey, Dueck
- 2008
(Show Context)
Citation Context ...ion heuristic [96] has been the standard for comparison for four decades and provides the basis for the variable neighborhood search meta-heuristic [47] that was compared with affinity propagation in =-=[10, 35]-=-. Variable-neighborhood search utilizes speedups to the original vertex substitution heuristic by storing all nearest and second-nearest open facilities for each customer and only recomputing certain ... |

11 | Unwrapping phase images by propagating probabilities across graphs
- Koetter, Frey, et al.
(Show Context)
Citation Context ...coding theory [5] were shown to be a special case of belief propagation in this loopy case [73]. Additional factor graphs (involving applications such as medical diagnostics [79] and phase-unwrapping =-=[63]-=-) were investigated at the time; positive results gave additional credence to the idea that pseudo-marginals provided useful enough approximations that argmax xn q(xn) = argmax xn P(xn) for many nodes... |

10 | Flexible priors for exemplar-based clustering
- Tarlow, Zemel, et al.
- 2008
(Show Context)
Citation Context ...nalysis of land-use and land-cover data (Cardille et al. in [12]) • Customer micro-targeting (Jiang et al. in [50]) An interesting and recent research thrust is Dirichlet process affinity propagation =-=[95]-=- which involves adapting the graphical model in Figure 3.5 to incorporate a Dirichlet prior over the size of clusters into the factor graph. This representation can then be viewed as maximum a posteri... |

9 | Comment on ”clustering by passing messages between data points
- Brusco, Kohn
- 2008
(Show Context)
Citation Context ...ion heuristic [96] has been the standard for comparison for four decades and provides the basis for the variable neighborhood search meta-heuristic [47] that was compared with affinity propagation in =-=[10, 35]-=-. Variable-neighborhood search utilizes speedups to the original vertex substitution heuristic by storing all nearest and second-nearest open facilities for each customer and only recomputing certain ... |

9 | Constructing treatment portfolios using affinity propagation. International conference on research in computational molecular biology
- Dueck, Frey, et al.
- 2008
(Show Context)
Citation Context ...cussed and selected manually. On the other hand, if the number of possibilities is large, a computational approach may be needed to select the appropriate options. Affinity propagation has been shown =-=[26]-=- to be an effective approach to this task. 5.3.1 Treatment Portfolio Design For concreteness, the possible set of options is referred to as ‘treatments’ and the assays used to measure the suitability ... |

7 | Using pairs of data-points to define splits for decision trees
- Hinton, Revow
- 1996
(Show Context)
Citation Context ...are represented efficiently as pointers into the training data (e.g., a subset of image features), so the number of bits of information needing to be specified during exemplar learning is quite small =-=[48]-=-. 5.1.1 Augmenting the Olivetti dataset The Olivetti dataset (§4.1) was modified for computer vision experiments as follows: to examine the effect of a wider range in image variation for each individu... |

7 | A decoupled approach to exemplar-based unsupervised learning.
- Nowozin, Bakir
- 2008
(Show Context)
Citation Context ...omponents, which turns out to be a multiplicative adjustment of the negative similarity for this example; it controls the number of exemplars found by the algorithm. Consistent with other experiments =-=[82]-=-, the convex clustering algorithm seems to have poor performance in practice 6 . j=1 Markov Clustering Algorithm (van Dongen, 2000) The Markov clustering algorithm [99] is a graph-based clustering alg... |

7 |
T.: Adaptive affinity propagation clustering
- Wang, Zhang, et al.
- 2008
(Show Context)
Citation Context ...tana et al. in [89]) • Face recognition with linear discriminant analysis (Du et al. in [24]) • Clustering text data (Kim et al. in [59]) • Adaptive extensions of affinity propagation (Wang et al. in =-=[102]-=-) • Knowledge discovery in medical data sources (Senf et al. in [91]) • Analysis of land-use and land-cover data (Cardille et al. in [12]) • Customer micro-targeting (Jiang et al. in [50]) An interest... |

6 | Finding novel transcripts in high-resolution genome-wide microarray data using the GenRate model
- Frey, Morris, et al.
- 2005
(Show Context)
Citation Context ...parsity: Exon Detection An important problem in current genomics research is the discovery of genes and gene variants that are expressed as messenger RNAs (mRNAs) in normal tissues. In a recent study =-=[37]-=-, DNA-based techniques were used to identify more than 800,000 possible exons (‘putative exons’) in the mouse genome. For each putative exon, an Agilent microarray probe matching a 60-base long DNA se... |

6 |
The Western Australian HIV Cohort Study,
- Mallal
- 1998
(Show Context)
Citation Context ...pitopes recognizable by a single patient are shown in a single color; mutations marked by red arrows escape MHC I binding. molecules of five different patients taken from the Western Australia cohort =-=[71]-=-. Epitopes recognizable by a single patient are shown in a single color, and each patient is assigned a different color. Some mutations (marked by red arrows) ‘escape’ MHC I binding. For example, the ... |

4 |
Exact maximum aposteriori estimation for binary images
- Greig, Porteous, et al.
- 1989
(Show Context)
Citation Context ...imizing the sum of cut weights while partitioning graphs (graph cuts) or its dual formulation, maximizing network flow [31, 32]. The optimal two-way (binary) graph-cut can be found in polynomial time =-=[43]-=-, which corresponds to finding a K = 2 clustering solution whose search space is only O(N 2 ). There are many approximate techniques for finding general K-way graph cuts, such as simulated annealing [... |

3 | ISOMAP based metrics for Clustering
- Baya, Granitto
- 2007
(Show Context)
Citation Context ...sis (Zhang et al. in [113]) • Gene3D: Protein analysis (Yeats et al. in [110]) • Protein sequence clustering (Wittkop et al. in [107]) • Affinity propagation with isomap-based metrics (Baya et al. in =-=[4]-=-) • Data streaming and analysis of grid computing jobs (Zhang et al. in [114]) • Analysis of cuticular hydrocarbons (Kent et al. in [58]) • Analysis of brain tissue MRI data (Verma et al. in [100]) • ... |

3 |
Integrating affinity propagation clustering method with linear discriminant analysis for face recognition
- Du, Yang, et al.
- 2007
(Show Context)
Citation Context ... clustering for text detection in images (Yi et al. in [112]) • Analysis of hydrophobic-polar protein model (Santana et al. in [89]) • Face recognition with linear discriminant analysis (Du et al. in =-=[24]-=-) • Clustering text data (Kim et al. in [59]) • Adaptive extensions of affinity propagation (Wang et al. in [102]) • Knowledge discovery in medical data sources (Senf et al. in [91]) • Analysis of lan... |

3 |
of both: a hybridized centroid-medoid clustering heuristic
- Best
- 2007
(Show Context)
Citation Context ...]) • Multiple view image segmentation (Xiao et al. in [108]) • Finding light sources using images (An et al. in [2]) • Image categorization and normalized mutual information analysis (Grira et al. in =-=[44]-=-) • Semi-supervised object classification (Fu et al. in [39]) • Image-audio dataset analysis (Zhang et al. in [113]) • Gene3D: Protein analysis (Yeats et al. in [110]) • Protein sequence clustering (W... |

3 |
Where are the exemplars
- Mézard
- 2007
(Show Context)
Citation Context ...(The Calling of St. Matthew, [20]) is an artistic depiction of identifying exemplars based on the direction of gestures, gazes, and even lighting in the painting. This interpretation was suggested in =-=[77]-=-. xChapter 1 Introduction Clustering or discovering meaningful partitions of data based on a measure of similarity is a critical step in scientific data analysis and a fundamental problem in computer... |

2 |
Single and double vertex substitution in heuristic procedures for the p-median problem
- Eilon, Galvão
- 1978
(Show Context)
Citation Context ...ange search space for each iteration has size π π , which grows impractically large for interesting problems where N > 1000 and p is non-trivial (5 <p<N −5). Experiments for π = 2 have been conducted =-=[28]-=- but only for N ≤30.CHAPTER 2. BACKGROUND 24 X 3 f 1 f 2 X 1 X 2 Figure 2.3: A sample factor graph showing a relationship between three variables, X1, X2, and X3, and two connecting function nodes, f... |

2 | Dynamic micro-targeting: fitness-based approach to predicting individual preferences
- Jiang, Tuzhilin
- 2008
(Show Context)
Citation Context ...g et al. in [102]) • Knowledge discovery in medical data sources (Senf et al. in [91]) • Analysis of land-use and land-cover data (Cardille et al. in [12]) • Customer micro-targeting (Jiang et al. in =-=[50]-=-) An interesting and recent research thrust is Dirichlet process affinity propagation [95] which involves adapting the graphical model in Figure 3.5 to incorporate a Dirichlet prior over the size of c... |

2 | An Affinity Propagation Based Method for Vector Quantization,” Eprint arXiv 0710.2037, http:// rxiv.org/abs/0710.2037v2
- Wu, Ding, et al.
- 2007
(Show Context)
Citation Context ...echniques in a matter of hours. In the two years since its introduction, affinity propagation has spawned a growing volume of research, such as: • Vector quantization codebook design (Jiang et al. in =-=[51]-=-) • Soft-constraint affinity propagation for gene expression data (Leone et al. in [68]) • Multiple view image segmentation (Xiao et al. in [108]) • Finding light sources using images (An et al. in [2... |

2 | Algorithms for rational vaccine design
- Jojic
- 2007
(Show Context)
Citation Context ...mers are used). See Figure 5.7 for more details. The utility u(T, R) of a strain T for a fragment R would ideally be set to its potential for immunological protection, but following the approaches in =-=[30,53,54,81]-=-, it is set to the frequency of the fragment in the database of HIV sequences if fragment R is present in strain T , and zero otherwise, as in equation (5.3). ⎧ ⎪⎨ frequency of R in HIV sequence datab... |

2 |
A polynomial algorithm in linear programming
- Kachiyan
- 1979
(Show Context)
Citation Context ... BACKGROUND 19 can be included ( ∑ N k=1 bkk = K) if the net similarity, S, is being minimized and not just the data-point similarity, Sdata. A common approach is to solve a linear program relaxation =-=[55, 57]-=- where ∀i, j : ˆ bij ∈ R and 0 ≤ ˆ bij ≤ 1 or implemented in optimization software packages such as CPLEX [19]. If the resulting solution is non-integer, stochastic rounding techniques or heuristics [... |

2 |
Learning factorizations in estimation of distribution algorithms using affinity propagation
- Santana, Larrañaga, et al.
- 2010
(Show Context)
Citation Context ...Clustering speakers from audio data (Zhang et al. in [115]) • Color-based clustering for text detection in images (Yi et al. in [112]) • Analysis of hydrophobic-polar protein model (Santana et al. in =-=[89]-=-) • Face recognition with linear discriminant analysis (Du et al. in [24]) • Clustering text data (Kim et al. in [59]) • Adaptive extensions of affinity propagation (Wang et al. in [102]) • Knowledge ... |

1 |
Acquiring Critical Light Points for Illumination Subspaces of Face Images by Affinity Propagation Clustering
- An, Liu, et al.
- 2007
(Show Context)
Citation Context ...1]) • Soft-constraint affinity propagation for gene expression data (Leone et al. in [68]) • Multiple view image segmentation (Xiao et al. in [108]) • Finding light sources using images (An et al. in =-=[2]-=-) • Image categorization and normalized mutual information analysis (Grira et al. in [44]) • Semi-supervised object classification (Fu et al. in [39]) • Image-audio dataset analysis (Zhang et al. in [... |

1 |
Widespread human signature in representative U.S. landscapes
- Cardille, Lambois
- 2008
(Show Context)
Citation Context ...• Adaptive extensions of affinity propagation (Wang et al. in [102]) • Knowledge discovery in medical data sources (Senf et al. in [91]) • Analysis of land-use and land-cover data (Cardille et al. in =-=[12]-=-) • Customer micro-targeting (Jiang et al. in [50]) An interesting and recent research thrust is Dirichlet process affinity propagation [95] which involves adapting the graphical model in Figure 3.5 t... |

1 |
Laplacian Affinity Propagation for SemiSupervised Object Classification
- Fu, Li, et al.
- 2007
(Show Context)
Citation Context ... • Finding light sources using images (An et al. in [2]) • Image categorization and normalized mutual information analysis (Grira et al. in [44]) • Semi-supervised object classification (Fu et al. in =-=[39]-=-) • Image-audio dataset analysis (Zhang et al. in [113]) • Gene3D: Protein analysis (Yeats et al. in [110]) • Protein sequence clustering (Wittkop et al. in [107]) • Affinity propagation with isomap-b... |

1 |
A Model-Based Analysis of Chemical and Temporal
- Kent, Azanchi, et al.
- 2007
(Show Context)
Citation Context ...• Affinity propagation with isomap-based metrics (Baya et al. in [4]) • Data streaming and analysis of grid computing jobs (Zhang et al. in [114]) • Analysis of cuticular hydrocarbons (Kent et al. in =-=[58]-=-) • Analysis of brain tissue MRI data (Verma et al. in [100]) • Clustering speakers from audio data (Zhang et al. in [115]) • Color-based clustering for text detection in images (Yi et al. in [112]) •... |

1 |
A Statistical Algorithm to Discover Knowledge in Medical Data Sources
- Senf, Leonard, et al.
- 2007
(Show Context)
Citation Context ...ysis (Du et al. in [24]) • Clustering text data (Kim et al. in [59]) • Adaptive extensions of affinity propagation (Wang et al. in [102]) • Knowledge discovery in medical data sources (Senf et al. in =-=[91]-=-) • Analysis of land-use and land-cover data (Cardille et al. in [12]) • Customer micro-targeting (Jiang et al. in [50]) An interesting and recent research thrust is Dirichlet process affinity propaga... |

1 |
On Detecting Subtle Pathology via Tissue Clustering of Multiparametric Data using Affinity Propagation
- Verma, Wang
- 2007
(Show Context)
Citation Context ...l. in [4]) • Data streaming and analysis of grid computing jobs (Zhang et al. in [114]) • Analysis of cuticular hydrocarbons (Kent et al. in [58]) • Analysis of brain tissue MRI data (Verma et al. in =-=[100]-=-) • Clustering speakers from audio data (Zhang et al. in [115]) • Color-based clustering for text detection in images (Yi et al. in [112]) • Analysis of hydrophobic-polar protein model (Santana et al.... |