380:
clustering of the masses. For example, assuming that all the springs and the masses are identical in the 2-dimensional spring system pictured, one would intuitively expect that the loosest connected masses on the right-hand side of the system would move with the largest amplitude and in the opposite direction to the rest of the masses when the system is shaken β and this expectation will be confirmed by analyzing components of the eigenvectors of the graph
Laplacian corresponding to the smallest eigenvalues, i.e., the smallest
218:. There are many different ways to define a Laplacian which have different mathematical interpretations, and so the clustering will also have different interpretations. The eigenvectors that are relevant are the ones that correspond to several smallest eigenvalues of the Laplacian except for the smallest eigenvalue which will have a value of 0. For computational efficiency, these eigenvectors are often computed as the eigenvectors corresponding to the largest several eigenvalues of a function of the Laplacian.
229:
28:
20:
1845:, which estimate or sample a neighborhood of a given data point for nearest neighbors, and compute non-zero entries of the adjacency matrix by comparing only pairs of the neighbors. The number of the selected nearest neighbors thus determines the number of non-zero entries, and is often fixed so that the memory footprint of the
2546:
Ideas and network measures related to spectral clustering also play an important role in a number of applications apparently different from clustering problems. For instance, networks with stronger spectral partitions take longer to converge in opinion-updating models used in sociology and economics.
379:
The masses that are tightly connected by the springs in the mass-spring system evidently move together from the equilibrium position in low-frequency vibration modes, so that the components of the eigenvectors corresponding to the smallest eigenvalues of the graph
Laplacian can be used for meaningful
394:
The goal of normalization is making the diagonal entries of the
Laplacian matrix to be all unit, also scaling off-diagonal entries correspondingly. In a weighted graph, a vertex may have a large degree because of a small number of connected edges but with large weights just as well as due to a large
1744:
Moreover, a normalized
Laplacian has exactly the same eigenvectors as the normalized adjacency matrix, but with the order of the eigenvalues reversed. Thus, instead of computing the eigenvectors corresponding to the smallest eigenvalues of the normalized Laplacian, one can equivalently compute the
1700:
The graph
Laplacian can be and commonly is constructed from the adjacency matrix. The construction can be performed matrix-free, i.e., without explicitly forming the matrix of the graph Laplacian and no AO. It can also be performed in-place of the adjacency matrix without increasing the memory
1248:
represents, one cluster data points identified with mutually strongly connected masses would move together in one direction, while in the complement cluster data points identified with remaining masses would move together in the opposite direction. The algorithm can be used for
2485:
The ideas behind spectral clustering may not be immediately obvious. It may be useful to highlight relationships with other methods. In particular, it can be described in the context of kernel clustering methods, which reveals several similarities with other approaches.
236:
Spectral clustering is well known to relate to partitioning of a mass-spring system, where each mass is associated with a data point and each spring stiffness corresponds to a weight of an edge describing a similarity of the two related data points, as in the
1225:, thus bi-partitioning the graph and labeling the data points with two labels. This sign-based approach follows the intuitive explanation of spectral clustering via the mass-spring model β in the low frequency vibration mode that the
2530:
of each cluster (in the clustering) was at least Ξ± and the weight of the inter-cluster edges was at most Ξ΅ fraction of the total weight of all the edges in the graph. They also look at two approximation algorithms in the same paper.
1833:
entries of the matrix. Nystrom method can be used to approximate the similarity matrix, but the approximate matrix is not elementwise positive, i.e. cannot be interpreted as a distance-based similarity.
241:. Specifically, the classical reference explains that the eigenvalue problem describing transversal vibration modes of a mass-spring system is exactly the same as the eigenvalue problem for the graph
578:
757:
367:
2526:
Ravi Kannan, Santosh
Vempala and Adrian Vetta proposed a bicriteria measure to define the quality of a given clustering. They said that a clustering was an (Ξ±, Ξ΅)-clustering if the
1692:
The need to construct the graph
Laplacian is common for all distance- or correlation-based clustering methods. Computing the eigenvectors is specific to spectral clustering only.
677:
63:
in fewer dimensions. The similarity matrix is provided as an input and consists of a quantitative assessment of the relative similarity of each pair of points in the dataset.
1476:
has not already been explicitly constructed, the efficiency of spectral clustering may be improved if the solution to the corresponding eigenvalue problem is performed in a
460:
923:
136:
830:
2156:
2120:
2807:
Acer, Seher; Boman, Erik G.; Glusa, Christian A.; Rajamanickam, Sivasankaran (2021). "Sphynx: A parallel multi-GPU graph partitioner for distributed-memory systems".
2350:
2044:
1280:
278:
1831:
1804:
1777:
1223:
1196:
2408:
2379:
2284:
1970:
1941:
1912:
2208:
2182:
1102:
1647:
1149:
2428:
2324:
2304:
2255:
2235:
2084:
2064:
2018:
1998:
1883:
1863:
1739:
1719:
1687:
1667:
1624:
1604:
1584:
1564:
1540:
1474:
1444:
1424:
1404:
1381:
1361:
1341:
1318:
1246:
1169:
1122:
1073:
1053:
1033:
1013:
993:
973:
953:
873:
853:
783:
601:
483:
302:
216:
176:
156:
100:
2872:. Clustering Large Data Sets; Third IEEE International Conference on Data Mining (ICDM 2003) Melbourne, Florida: IEEE Computer Society. pp. 59β62.
2086:
graph
Laplacian matrix by a vector, which varies greatly whether the graph Laplacian matrix is dense or sparse. For the dense case the cost thus is
400:
3144:
Donath, William; Hoffman, Alan (1972). "Algorithms for partitioning of graphs and computer logic based on eigenvectors of connections matrices".
1566:. No matter the algorithm of the spectral clustering, the two main costly items are the construction of the graph Laplacian and determining its
2502:-means problem shares the objective function with the spectral clustering problem, which can be optimized directly by multi-level methods.
1745:
eigenvectors corresponding to the largest eigenvalues of the normalized adjacency matrix, without even talking about the
Laplacian matrix.
2438:
with distributed memory, resulting not only in high quality clusters, which spectral clustering is famous for, but also top performance.
2886:
Multiscale
Spectral Image Segmentation Multiscale preconditioning for computing eigenvalues of graph Laplacians in image segmentation
2778:
Wang, S.; Gittens, A.; Mahoney, M.W. (2019). "Scalable Kernel K-Means Clustering with Nystrom Approximation: Relative-Error Bounds".
3021:
Dhillon, Inderjit; Guan, Yuqiang; Kulis, Brian (November 2007). "Weighted Graph Cuts without Eigenvectors: A Multilevel Approach".
67:
1701:
footprint. Either way, the costs of constructing the graph Laplacian is essentially determined by the costs of constructing the
3286:
Golub, Benjamin; Jackson, Matthew O. (2012-07-26). "How Homophily Affects the Speed of Learning and Best-Response Dynamics".
2925:
3326:
1514:, and dimension reduction techniques such as locally-linear embedding can be used to reduce errors from noise or outliers.
499:
2561:
381:
2996:
2511:
2046:) matrix of selected eigenvectors of the graph Laplacian is normally proportional to the cost of multiplication of the
1511:
1496:
3207:
Daniel A. Spielman and Shang-Hua Teng (1996). "Spectral Partitioning Works: Planar graphs and finite element meshes".
691:
388:
490:
3331:
413:
2852:
313:
2687:
Arias-Castro, E.; Chen, G.; Lerman, G. (2011), "Spectral clustering based on local linear approximations.",
2954:
2591:
2466:
3192:
Guattery, Stephen; Miller, Gary L. (1995). "On the performance of spectral graph partitioning methods".
2514:β the optimal clusters with no edges cut β spectral clustering is also related to a spectral version of
617:
3094:
Kannan, Ravi; Vempala, Santosh; Vetta, Adrian (2004). "On Clusterings : Good, Bad and Spectral".
1503:
method. Spectral clustering has been successfully applied on large graphs by first identifying their
3035:
3008:
Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining
2474:
419:
56:
878:
105:
1842:
1250:
36:
3030:
2446:
Free software implementing spectral clustering is available in large open source projects like
796:
762:
and can also be used for spectral clustering. A mathematically equivalent algorithm takes the
2125:
2089:
2914:. Workshop on Algorithms for Modern Massive Datasets Stanford University and Yahoo! Research.
2571:
2329:
2023:
1259:
251:
2969:
2556:
1809:
1782:
1755:
1201:
1174:
44:
3226:
2384:
2355:
2260:
1946:
1917:
1888:
8:
2623:
2527:
2187:
2161:
1586:
eigenvectors for the spectral embedding. The last step β determining the labels from the
1504:
1477:
1081:
2973:
1629:
1131:
3111:
3056:
2834:
2816:
2787:
2760:
2714:
2696:
2664:
2638:"Data reduction for spectral clustering to analyze high throughput flow cytometry data"
2637:
2610:
2413:
2309:
2289:
2240:
2220:
2069:
2049:
2003:
1983:
1868:
1848:
1724:
1704:
1672:
1652:
1609:
1589:
1569:
1549:
1525:
1459:
1449:
1429:
1409:
1389:
1366:
1346:
1326:
1303:
1231:
1154:
1107:
1058:
1038:
1018:
998:
978:
958:
938:
858:
838:
768:
586:
468:
287:
201:
161:
141:
85:
2869:
Modern preconditioned eigensolvers for spectral image segmentation and graph bisection
1546:
and compute time, or number of arithmetic operations (AO) performed, as a function of
3303:
3268:
3115:
3048:
2909:
2884:
2867:
2838:
2752:
2669:
1481:
1480:(without explicitly manipulating or even computing the similarity matrix), as in the
79:
52:
3129:
Cheeger, Jeff (1969). "A lower bound for the smallest eigenvalue of the Laplacian".
2718:
3295:
3260:
3172:
3103:
3060:
3040:
2977:
2890:
2826:
2764:
2744:
2706:
2659:
2649:
2566:
2540:
1749:
1543:
1488:
928:
790:
612:
373:
242:
222:
195:
179:
60:
2939:
2286:
non-zero entries, the cost of the matrix-vector product and thus of computing the
2981:
2470:
2458:
1492:
305:
2894:
2830:
995:
of selected eigenvectors, mapping β called spectral embedding β of the original
2748:
2211:
1226:
1125:
409:
3264:
2184:
and is clearly misleading, since, e.g., in a hierarchical spectral clustering
3320:
3307:
3272:
3177:
3160:
2654:
1838:
238:
3044:
2732:
3052:
2756:
2673:
2626:", Neural Information Processing Systems 13 (NIPS 2000), 2001, pp. 873β879.
2447:
1499:
is a key technology accelerating the convergence, e.g., in the matrix-free
3107:
3077:
The Relationship of DBSCAN to Matrix Factorization and Spectral Clustering
3075:
3299:
1972:
non-zero entries, and the calculations can be trivially run in parallel.
1626:
matrix of eigenvectors β is typically the least expensive requiring only
1487:
For large-sized graphs, the second eigenvalue of the (normalized) graph
1128:, corresponds to the second smallest eigenvalue. Using the components of
763:
604:
463:
191:
48:
3131:
Proceedings of the Princeton Conference in Honor of Professor S. Bochner
2592:"CS267: Notes for Lecture 23, April 9, 1999, Graph Partitioning, Part 2"
228:
138:
represents a measure of the similarity between data points with indices
2543:
method was popularized by Shi & Malik and Ng, Jordan, & Weiss.
786:
608:
486:
2953:
Filippone, M.; Camastra, F.; Masulli, F.; Rovetta, S. (January 2008).
2710:
66:
In application to image segmentation, spectral clustering is known as
3206:
2455:
3240:
2821:
2792:
178:. The general approach to spectral clustering is to use a standard
2701:
2539:
Spectral clustering has a long history. Spectral clustering as a
3249:"Persuasion Bias, Social Influence, and Unidimensional Opinions"
3248:
2866:
Knyazev, Andrew V. (2003). Boley; Dhillon; Ghosh; Kogan (eds.).
2952:
2515:
2462:
2451:
2431:
1500:
1495:, leading to slow convergence of iterative eigenvalue solvers.
1283:
3023:
IEEE Transactions on Pattern Analysis and Machine Intelligence
2737:
IEEE Transactions on Pattern Analysis and Machine Intelligence
2430:
data points. Moreover, matrix-free eigenvalue solvers such as
2410:β both are the optimal low bounds of complexity of clustering
855:
of the symmetrically normalized Laplacian and the eigenvector
2911:
Multiscale Spectral Graph Partitioning and Image Segmentation
2636:
Zare, Habil; Shooshtari, P.; Gupta, A.; Brinkman, R. (2010).
2480:
1448:
Cluster the graph nodes based on these features (e.g., using
875:
of the left normalized Laplacian are related by the identity
1752:, e.g., using the RBF kernel, make it dense, thus requiring
1253:
by repeatedly partitioning the subsets in the same fashion.
3279:
2635:
2435:
1943:
sequential arithmetic operations are needed to compute the
27:
2806:
927:
398:
A popular normalized spectral clustering technique is the
19:
3074:
Schubert, Erich; Hess, Sibylle; Morik, Katharina (2018).
1055:. Now the analysis is reduced to clustering vectors with
3209:
Annual IEEE Symposium on Foundations of Computer Science
2955:"A survey of kernel and spectral methods for clustering"
2926:"Clustering - RDD-based API - Spark 3.2.0 Documentation"
1837:
Algorithms to construct the graph adjacency matrix as a
3247:
DeMarzo, P. M.; Vayanos, D.; Zwiebel, J. (2003-08-01).
3225:
Ng, Andrew Y.; Jordan, Michael I.; Weiss, Yair (2002).
2889:. Fast Manifold Learning Workshop, WM Williamburg, VA.
2686:
2613:, IEEE Transactions on PAMI, Vol. 22, No. 8, Aug 2000.
3246:
2876:
2416:
2387:
2358:
2332:
2312:
2292:
2263:
2243:
2223:
2190:
2164:
2128:
2092:
2072:
2052:
2026:
2006:
1986:
1949:
1920:
1891:
1871:
1851:
1812:
1785:
1758:
1727:
1707:
1675:
1655:
1632:
1612:
1592:
1572:
1552:
1528:
1462:
1432:
1412:
1392:
1369:
1349:
1329:
1306:
1282:, any vector clustering technique can be used, e.g.,
1262:
1234:
1204:
1177:
1157:
1134:
1110:
1084:
1061:
1041:
1021:
1001:
981:
961:
941:
881:
861:
841:
799:
771:
694:
620:
589:
573:{\displaystyle L^{\text{norm}}:=I-D^{-1/2}AD^{-1/2}.}
502:
471:
422:
316:
290:
254:
204:
164:
144:
108:
88:
2518:
clustering that finds density-connected components.
1343:
eigenvectors (the eigenvectors corresponding to the
3227:"On spectral clustering: analysis and an algorithm"
2946:
2434:can efficiently run in parallel, e.g., on multiple
3093:
2777:
2534:
2521:
2422:
2402:
2373:
2344:
2318:
2298:
2278:
2249:
2229:
2202:
2176:
2150:
2114:
2078:
2058:
2038:
2012:
1992:
1964:
1935:
1906:
1877:
1857:
1825:
1798:
1771:
1733:
1713:
1681:
1661:
1641:
1618:
1598:
1578:
1558:
1534:
1468:
1438:
1418:
1398:
1375:
1355:
1335:
1312:
1274:
1240:
1217:
1190:
1163:
1143:
1116:
1096:
1067:
1047:
1027:
1007:
987:
967:
947:
917:
867:
847:
824:
777:
751:
671:
595:
572:
477:
454:
361:
296:
272:
210:
170:
150:
130:
94:
3234:Advances in Neural Information Processing Systems
3073:
2988:
2122:. The very commonly cited in the literature cost
387:
3318:
3194:Annual ACM-SIAM Symposium on Discrete Algorithms
3020:
3001:-means: spectral clustering and normalized cuts"
2994:
1075:components, which may be done in various ways.
752:{\displaystyle L^{\text{rw}}:=D^{-1}L=I-D^{-1}A}
3294:(3). Oxford University Press (OUP): 1287β1338.
1695:
3191:
187:
3224:
3143:
395:number of connected edges with unit weights.
3285:
2940:"Kernlab: Kernel-Based Machine Learning Lab"
2733:"Spectral grouping using the Nystrom method"
2489:
2469:for pseudo-eigenvector clustering using the
1151:one can place all points whose component in
1035:-dimensional vector space using the rows of
78:Given an enumerated set of data points, the
23:An example connected graph, with 6 vertices.
2995:Dhillon, I.S.; Guan, Y.; Kulis, B. (2004).
1426:-th row defines the features of graph node
2481:Relationship with other clustering methods
1522:Denoting the number of the data points by
1510:Spectral clustering is closely related to
683:random walk (or left) normalized Laplacian
3176:
3034:
2820:
2791:
2700:
2663:
2653:
2505:
1975:
1386:Consider the matrix formed by the first
2611:"Normalized Cuts and Image Segmentation"
227:
68:segmentation-based object categorization
26:
18:
3259:(3). Oxford University Press: 909β968.
3158:
3128:
2907:
2882:
2865:
2730:
362:{\displaystyle D_{ii}=\sum _{j}A_{ij},}
3319:
2800:
2381:, with the memory footprint also only
31:Partitioning into two connected graphs
3220:
3218:
2624:Learning Segmentation by Random Walks
485:corresponding to the second-smallest
416:. It partitions points into two sets
182:method (there are many such methods,
82:may be defined as a symmetric matrix
2780:Journal of Machine Learning Research
2605:
2603:
2601:
607:corresponding to the second-largest
2562:Kernel principal component analysis
2510:In the trivial case of determining
2352:matrix of selected eigenvectors is
1507:, and then clustering communities.
221:
13:
3288:The Quarterly Journal of Economics
3253:The Quarterly Journal of Economics
3215:
3161:"Algebraic connectivity of graphs"
2589:
1542:, it is important to estimate the
1512:nonlinear dimensionality reduction
1104:, the selected single eigenvector
672:{\displaystyle D^{-1/2}AD^{-1/2}.}
14:
3343:
3165:Czechoslovak Mathematical Journal
3146:IBM Technical Disclosure Bulletin
2598:
1748:Naive constructions of the graph
611:of the symmetrically normalized
2689:Electronic Journal of Statistics
2622:Marina MeilΔ & Jianbo Shi, "
1689:vector of the labels in memory.
791:random walk normalized adjacency
3200:
3185:
3152:
3137:
3122:
3087:
3067:
3014:
2932:
2918:
2901:
2859:
2609:Jianbo Shi and Jitendra Malik,
2535:History and related literatures
2522:Measures to compare clusterings
1885:graph adjacency matrix is only
2845:
2771:
2724:
2680:
2629:
2616:
2583:
2397:
2391:
2368:
2362:
2273:
2267:
2145:
2132:
2109:
2096:
1959:
1953:
1930:
1924:
1901:
1895:
1015:data points is performed to a
491:symmetric normalized Laplacian
449:
423:
389:Laplacian matrix normalization
232:A 2-dimensional spring system.
73:
1:
2577:
1320:(or the normalized Laplacian)
1289:
785:corresponding to the largest
455:{\displaystyle (B_{1},B_{2})}
408:introduced by Jianbo Shi and
2982:10.1016/j.patcog.2007.05.018
2257:graph Laplacian matrix with
1806:AO to determine each of the
1696:Constructing graph Laplacian
918:{\displaystyle D^{-1/2}v=u.}
131:{\displaystyle A_{ij}\geq 0}
7:
3327:Cluster analysis algorithms
2908:Knyazev, Andrew V. (2006).
2895:10.13140/RG.2.2.35280.02565
2883:Knyazev, Andrew V. (2006).
2831:10.1016/j.parco.2021.102769
2550:
2441:
43:techniques make use of the
10:
3348:
3159:Fiedler, Miroslav (1973).
2749:10.1109/TPAMI.2004.1262185
2512:connected graph components
2217:In the sparse case of the
1980:The cost of computing the
3265:10.1162/00335530360698469
3083:. LWDA. pp. 330β334.
1841:are typically based on a
1456:If the similarity matrix
1300:Calculate the Laplacian
825:{\displaystyle P=D^{-1}A}
401:normalized cuts algorithm
3178:10.21136/CMJ.1973.101168
2655:10.1186/1471-2105-11-403
2151:{\displaystyle O(n^{3})}
2115:{\displaystyle O(n^{2})}
1741:graph adjacency matrix.
1517:
1363:smallest eigenvalues of
57:dimensionality reduction
3045:10.1109/tpami.2007.1115
1843:nearest neighbor search
1649:AO and creating just a
1251:hierarchical clustering
1171:is positive in the set
55:of the data to perform
37:multivariate statistics
3332:Algebraic graph theory
2506:Relationship to DBSCAN
2424:
2404:
2375:
2346:
2345:{\displaystyle k\ll n}
2320:
2300:
2280:
2251:
2231:
2204:
2178:
2152:
2116:
2080:
2060:
2040:
2039:{\displaystyle k\ll n}
2014:
1994:
1976:Computing eigenvectors
1966:
1937:
1908:
1879:
1859:
1827:
1800:
1773:
1735:
1715:
1683:
1663:
1643:
1620:
1600:
1580:
1560:
1536:
1470:
1440:
1420:
1400:
1377:
1357:
1337:
1314:
1276:
1275:{\displaystyle k>1}
1242:
1219:
1192:
1165:
1145:
1118:
1098:
1069:
1049:
1029:
1009:
989:
969:
949:
931:via Spectral Embedding
919:
869:
849:
826:
779:
753:
673:
597:
574:
479:
456:
363:
298:
274:
273:{\displaystyle L:=D-A}
233:
212:
172:
152:
132:
96:
32:
24:
3108:10.1145/990308.990313
2572:Spectral graph theory
2425:
2405:
2376:
2347:
2321:
2301:
2281:
2252:
2232:
2210:as determined by the
2205:
2179:
2153:
2117:
2081:
2061:
2041:
2015:
1995:
1967:
1938:
1909:
1880:
1860:
1828:
1826:{\displaystyle n^{2}}
1801:
1799:{\displaystyle n^{2}}
1774:
1772:{\displaystyle n^{2}}
1736:
1716:
1684:
1664:
1644:
1621:
1601:
1581:
1561:
1537:
1471:
1441:
1421:
1401:
1378:
1358:
1338:
1315:
1277:
1243:
1220:
1218:{\displaystyle B_{-}}
1193:
1191:{\displaystyle B_{+}}
1166:
1146:
1119:
1099:
1078:In the simplest case
1070:
1050:
1030:
1010:
990:
970:
950:
920:
870:
850:
827:
780:
754:
674:
598:
575:
480:
457:
382:vibration frequencies
364:
299:
275:
231:
213:
173:
153:
133:
97:
30:
22:
2557:Affinity propagation
2498:The weighted kernel
2414:
2403:{\displaystyle O(n)}
2385:
2374:{\displaystyle O(n)}
2356:
2330:
2310:
2290:
2279:{\displaystyle O(n)}
2261:
2241:
2221:
2188:
2162:
2158:comes from choosing
2126:
2090:
2070:
2050:
2024:
2004:
1984:
1965:{\displaystyle O(n)}
1947:
1936:{\displaystyle O(n)}
1918:
1907:{\displaystyle O(n)}
1889:
1869:
1849:
1810:
1783:
1756:
1725:
1705:
1673:
1653:
1630:
1610:
1590:
1570:
1550:
1526:
1460:
1430:
1410:
1390:
1367:
1347:
1327:
1323:Calculate the first
1304:
1260:
1256:In the general case
1232:
1202:
1175:
1155:
1132:
1108:
1082:
1059:
1039:
1019:
999:
979:
959:
939:
879:
859:
839:
797:
769:
692:
618:
587:
500:
469:
420:
412:, commonly used for
314:
288:
252:
202:
186:-means is discussed
162:
142:
106:
86:
2974:2008PatRe..41..176F
2962:Pattern Recognition
2942:. 12 November 2019.
2731:Fowlkes, C (2004).
2203:{\displaystyle k=1}
2177:{\displaystyle k=n}
1505:community structure
1478:matrix-free fashion
1406:eigenvectors; the
1097:{\displaystyle k=1}
406:ShiβMalik algorithm
41:spectral clustering
3300:10.1093/qje/qjs021
3096:Journal of the ACM
2809:Parallel Computing
2642:BMC Bioinformatics
2490:Relationship with
2420:
2400:
2371:
2342:
2316:
2296:
2276:
2247:
2227:
2200:
2174:
2148:
2112:
2076:
2056:
2036:
2010:
1990:
1962:
1933:
1904:
1875:
1855:
1823:
1796:
1769:
1731:
1711:
1679:
1659:
1642:{\displaystyle kn}
1639:
1616:
1596:
1576:
1556:
1532:
1466:
1450:k-means clustering
1436:
1416:
1396:
1373:
1353:
1333:
1310:
1272:
1238:
1215:
1188:
1161:
1144:{\displaystyle v,}
1141:
1114:
1094:
1065:
1045:
1025:
1005:
985:
965:
945:
915:
865:
845:
822:
775:
749:
669:
593:
570:
475:
452:
414:image segmentation
359:
342:
294:
270:
234:
208:
168:
148:
128:
92:
33:
25:
16:Clustering methods
3029:(11): 1944β1957.
3010:. pp. 551β6.
2853:"2.3. Clustering"
2711:10.1214/11-ejs651
2423:{\displaystyle n}
2319:{\displaystyle k}
2299:{\displaystyle n}
2250:{\displaystyle n}
2230:{\displaystyle n}
2079:{\displaystyle n}
2059:{\displaystyle n}
2013:{\displaystyle k}
1993:{\displaystyle n}
1878:{\displaystyle n}
1858:{\displaystyle n}
1734:{\displaystyle n}
1714:{\displaystyle n}
1682:{\displaystyle 1}
1662:{\displaystyle n}
1619:{\displaystyle k}
1599:{\displaystyle n}
1579:{\displaystyle k}
1559:{\displaystyle n}
1535:{\displaystyle n}
1482:Lanczos algorithm
1469:{\displaystyle A}
1439:{\displaystyle l}
1419:{\displaystyle l}
1399:{\displaystyle k}
1376:{\displaystyle L}
1356:{\displaystyle k}
1336:{\displaystyle k}
1313:{\displaystyle L}
1241:{\displaystyle v}
1164:{\displaystyle v}
1117:{\displaystyle v}
1068:{\displaystyle k}
1048:{\displaystyle V}
1028:{\displaystyle k}
1008:{\displaystyle n}
988:{\displaystyle V}
968:{\displaystyle k}
948:{\displaystyle n}
868:{\displaystyle u}
848:{\displaystyle v}
778:{\displaystyle u}
702:
596:{\displaystyle v}
510:
478:{\displaystyle v}
333:
297:{\displaystyle D}
211:{\displaystyle A}
171:{\displaystyle j}
151:{\displaystyle i}
95:{\displaystyle A}
80:similarity matrix
53:similarity matrix
3339:
3312:
3311:
3283:
3277:
3276:
3244:
3238:
3237:
3231:
3222:
3213:
3212:
3204:
3198:
3197:
3189:
3183:
3182:
3180:
3156:
3150:
3149:
3141:
3135:
3134:
3126:
3120:
3119:
3091:
3085:
3084:
3082:
3071:
3065:
3064:
3038:
3018:
3012:
3011:
3005:
2992:
2986:
2985:
2959:
2950:
2944:
2943:
2936:
2930:
2929:
2922:
2916:
2915:
2905:
2899:
2898:
2880:
2874:
2873:
2863:
2857:
2856:
2849:
2843:
2842:
2824:
2804:
2798:
2797:
2795:
2775:
2769:
2768:
2728:
2722:
2721:
2704:
2684:
2678:
2677:
2667:
2657:
2633:
2627:
2620:
2614:
2607:
2596:
2595:
2587:
2567:Cluster analysis
2541:machine learning
2429:
2427:
2426:
2421:
2409:
2407:
2406:
2401:
2380:
2378:
2377:
2372:
2351:
2349:
2348:
2343:
2325:
2323:
2322:
2317:
2305:
2303:
2302:
2297:
2285:
2283:
2282:
2277:
2256:
2254:
2253:
2248:
2236:
2234:
2233:
2228:
2209:
2207:
2206:
2201:
2183:
2181:
2180:
2175:
2157:
2155:
2154:
2149:
2144:
2143:
2121:
2119:
2118:
2113:
2108:
2107:
2085:
2083:
2082:
2077:
2065:
2063:
2062:
2057:
2045:
2043:
2042:
2037:
2019:
2017:
2016:
2011:
1999:
1997:
1996:
1991:
1971:
1969:
1968:
1963:
1942:
1940:
1939:
1934:
1913:
1911:
1910:
1905:
1884:
1882:
1881:
1876:
1864:
1862:
1861:
1856:
1832:
1830:
1829:
1824:
1822:
1821:
1805:
1803:
1802:
1797:
1795:
1794:
1778:
1776:
1775:
1770:
1768:
1767:
1750:adjacency matrix
1740:
1738:
1737:
1732:
1720:
1718:
1717:
1712:
1688:
1686:
1685:
1680:
1668:
1666:
1665:
1660:
1648:
1646:
1645:
1640:
1625:
1623:
1622:
1617:
1605:
1603:
1602:
1597:
1585:
1583:
1582:
1577:
1565:
1563:
1562:
1557:
1544:memory footprint
1541:
1539:
1538:
1533:
1489:Laplacian matrix
1475:
1473:
1472:
1467:
1445:
1443:
1442:
1437:
1425:
1423:
1422:
1417:
1405:
1403:
1402:
1397:
1382:
1380:
1379:
1374:
1362:
1360:
1359:
1354:
1342:
1340:
1339:
1334:
1319:
1317:
1316:
1311:
1281:
1279:
1278:
1273:
1247:
1245:
1244:
1239:
1224:
1222:
1221:
1216:
1214:
1213:
1198:and the rest in
1197:
1195:
1194:
1189:
1187:
1186:
1170:
1168:
1167:
1162:
1150:
1148:
1147:
1142:
1123:
1121:
1120:
1115:
1103:
1101:
1100:
1095:
1074:
1072:
1071:
1066:
1054:
1052:
1051:
1046:
1034:
1032:
1031:
1026:
1014:
1012:
1011:
1006:
994:
992:
991:
986:
974:
972:
971:
966:
954:
952:
951:
946:
929:Cluster analysis
924:
922:
921:
916:
902:
901:
897:
874:
872:
871:
866:
854:
852:
851:
846:
835:The eigenvector
831:
829:
828:
823:
818:
817:
784:
782:
781:
776:
758:
756:
755:
750:
745:
744:
720:
719:
704:
703:
700:
678:
676:
675:
670:
665:
664:
660:
641:
640:
636:
613:adjacency matrix
602:
600:
599:
594:
579:
577:
576:
571:
566:
565:
561:
542:
541:
537:
512:
511:
508:
484:
482:
481:
476:
461:
459:
458:
453:
448:
447:
435:
434:
374:adjacency matrix
368:
366:
365:
360:
355:
354:
341:
329:
328:
303:
301:
300:
295:
279:
277:
276:
271:
243:Laplacian matrix
223:Laplacian matrix
217:
215:
214:
209:
196:Laplacian matrix
177:
175:
174:
169:
157:
155:
154:
149:
137:
135:
134:
129:
121:
120:
101:
99:
98:
93:
3347:
3346:
3342:
3341:
3340:
3338:
3337:
3336:
3317:
3316:
3315:
3284:
3280:
3245:
3241:
3229:
3223:
3216:
3205:
3201:
3190:
3186:
3157:
3153:
3142:
3138:
3127:
3123:
3092:
3088:
3080:
3072:
3068:
3036:10.1.1.131.2635
3019:
3015:
3003:
2993:
2989:
2957:
2951:
2947:
2938:
2937:
2933:
2924:
2923:
2919:
2906:
2902:
2881:
2877:
2864:
2860:
2851:
2850:
2846:
2805:
2801:
2776:
2772:
2729:
2725:
2685:
2681:
2634:
2630:
2621:
2617:
2608:
2599:
2588:
2584:
2580:
2553:
2537:
2524:
2508:
2496:
2483:
2471:power iteration
2459:preconditioning
2444:
2415:
2412:
2411:
2386:
2383:
2382:
2357:
2354:
2353:
2331:
2328:
2327:
2311:
2308:
2307:
2291:
2288:
2287:
2262:
2259:
2258:
2242:
2239:
2238:
2222:
2219:
2218:
2189:
2186:
2185:
2163:
2160:
2159:
2139:
2135:
2127:
2124:
2123:
2103:
2099:
2091:
2088:
2087:
2071:
2068:
2067:
2051:
2048:
2047:
2025:
2022:
2021:
2005:
2002:
2001:
1985:
1982:
1981:
1978:
1948:
1945:
1944:
1919:
1916:
1915:
1890:
1887:
1886:
1870:
1867:
1866:
1850:
1847:
1846:
1817:
1813:
1811:
1808:
1807:
1790:
1786:
1784:
1781:
1780:
1763:
1759:
1757:
1754:
1753:
1726:
1723:
1722:
1706:
1703:
1702:
1698:
1674:
1671:
1670:
1654:
1651:
1650:
1631:
1628:
1627:
1611:
1608:
1607:
1591:
1588:
1587:
1571:
1568:
1567:
1551:
1548:
1547:
1527:
1524:
1523:
1520:
1497:Preconditioning
1493:ill-conditioned
1461:
1458:
1457:
1431:
1428:
1427:
1411:
1408:
1407:
1391:
1388:
1387:
1368:
1365:
1364:
1348:
1345:
1344:
1328:
1325:
1324:
1305:
1302:
1301:
1295:Basic Algorithm
1292:
1261:
1258:
1257:
1233:
1230:
1229:
1209:
1205:
1203:
1200:
1199:
1182:
1178:
1176:
1173:
1172:
1156:
1153:
1152:
1133:
1130:
1129:
1109:
1106:
1105:
1083:
1080:
1079:
1060:
1057:
1056:
1040:
1037:
1036:
1020:
1017:
1016:
1000:
997:
996:
980:
977:
976:
960:
957:
956:
940:
937:
936:
933:
893:
886:
882:
880:
877:
876:
860:
857:
856:
840:
837:
836:
810:
806:
798:
795:
794:
770:
767:
766:
737:
733:
712:
708:
699:
695:
693:
690:
689:
656:
649:
645:
632:
625:
621:
619:
616:
615:
588:
585:
584:
557:
550:
546:
533:
526:
522:
507:
503:
501:
498:
497:
470:
467:
466:
443:
439:
430:
426:
421:
418:
417:
392:
347:
343:
337:
321:
317:
315:
312:
311:
306:diagonal matrix
289:
286:
285:
253:
250:
249:
226:
203:
200:
199:
163:
160:
159:
143:
140:
139:
113:
109:
107:
104:
103:
87:
84:
83:
76:
17:
12:
11:
5:
3345:
3335:
3334:
3329:
3314:
3313:
3278:
3239:
3214:
3199:
3184:
3171:(2): 298β305.
3151:
3136:
3121:
3102:(3): 497β515.
3086:
3066:
3013:
2987:
2968:(1): 176β190.
2945:
2931:
2917:
2900:
2875:
2858:
2844:
2799:
2770:
2723:
2679:
2628:
2615:
2597:
2581:
2579:
2576:
2575:
2574:
2569:
2564:
2559:
2552:
2549:
2536:
2533:
2523:
2520:
2507:
2504:
2495:
2488:
2482:
2479:
2443:
2440:
2419:
2399:
2396:
2393:
2390:
2370:
2367:
2364:
2361:
2341:
2338:
2335:
2315:
2295:
2275:
2272:
2269:
2266:
2246:
2226:
2212:Fiedler vector
2199:
2196:
2193:
2173:
2170:
2167:
2147:
2142:
2138:
2134:
2131:
2111:
2106:
2102:
2098:
2095:
2075:
2055:
2035:
2032:
2029:
2009:
1989:
1977:
1974:
1961:
1958:
1955:
1952:
1932:
1929:
1926:
1923:
1903:
1900:
1897:
1894:
1874:
1854:
1820:
1816:
1793:
1789:
1766:
1762:
1730:
1710:
1697:
1694:
1678:
1658:
1638:
1635:
1615:
1595:
1575:
1555:
1531:
1519:
1516:
1465:
1454:
1453:
1446:
1435:
1415:
1395:
1384:
1372:
1352:
1332:
1321:
1309:
1297:
1296:
1291:
1288:
1271:
1268:
1265:
1237:
1227:Fiedler vector
1212:
1208:
1185:
1181:
1160:
1140:
1137:
1126:Fiedler vector
1113:
1093:
1090:
1087:
1064:
1044:
1024:
1004:
984:
964:
944:
932:
926:
914:
911:
908:
905:
900:
896:
892:
889:
885:
864:
844:
821:
816:
813:
809:
805:
802:
774:
760:
759:
748:
743:
740:
736:
732:
729:
726:
723:
718:
715:
711:
707:
698:
685:is defined as
668:
663:
659:
655:
652:
648:
644:
639:
635:
631:
628:
624:
592:
581:
580:
569:
564:
560:
556:
553:
549:
545:
540:
536:
532:
529:
525:
521:
518:
515:
506:
474:
451:
446:
442:
438:
433:
429:
425:
410:Jitendra Malik
391:
386:
370:
369:
358:
353:
350:
346:
340:
336:
332:
327:
324:
320:
293:
282:
281:
269:
266:
263:
260:
257:
225:
220:
207:
190:) on relevant
167:
147:
127:
124:
119:
116:
112:
91:
75:
72:
15:
9:
6:
4:
3:
2:
3344:
3333:
3330:
3328:
3325:
3324:
3322:
3309:
3305:
3301:
3297:
3293:
3289:
3282:
3274:
3270:
3266:
3262:
3258:
3254:
3250:
3243:
3235:
3228:
3221:
3219:
3210:
3203:
3195:
3188:
3179:
3174:
3170:
3166:
3162:
3155:
3147:
3140:
3132:
3125:
3117:
3113:
3109:
3105:
3101:
3097:
3090:
3079:
3078:
3070:
3062:
3058:
3054:
3050:
3046:
3042:
3037:
3032:
3028:
3024:
3017:
3009:
3002:
3000:
2991:
2983:
2979:
2975:
2971:
2967:
2963:
2956:
2949:
2941:
2935:
2927:
2921:
2913:
2912:
2904:
2896:
2892:
2888:
2887:
2879:
2871:
2870:
2862:
2854:
2848:
2840:
2836:
2832:
2828:
2823:
2818:
2814:
2810:
2803:
2794:
2789:
2785:
2781:
2774:
2766:
2762:
2758:
2754:
2750:
2746:
2743:(2): 214β25.
2742:
2738:
2734:
2727:
2720:
2716:
2712:
2708:
2703:
2698:
2694:
2690:
2683:
2675:
2671:
2666:
2661:
2656:
2651:
2647:
2643:
2639:
2632:
2625:
2619:
2612:
2606:
2604:
2602:
2593:
2586:
2582:
2573:
2570:
2568:
2565:
2563:
2560:
2558:
2555:
2554:
2548:
2544:
2542:
2532:
2529:
2519:
2517:
2513:
2503:
2501:
2493:
2487:
2478:
2476:
2472:
2468:
2464:
2460:
2457:
2453:
2449:
2439:
2437:
2433:
2417:
2394:
2388:
2365:
2359:
2339:
2336:
2333:
2313:
2293:
2270:
2264:
2244:
2224:
2215:
2213:
2197:
2194:
2191:
2171:
2168:
2165:
2140:
2136:
2129:
2104:
2100:
2093:
2073:
2053:
2033:
2030:
2027:
2007:
1987:
1973:
1956:
1950:
1927:
1921:
1898:
1892:
1872:
1852:
1844:
1840:
1839:sparse matrix
1835:
1818:
1814:
1791:
1787:
1764:
1760:
1751:
1746:
1742:
1728:
1708:
1693:
1690:
1676:
1656:
1636:
1633:
1613:
1593:
1573:
1553:
1545:
1529:
1515:
1513:
1508:
1506:
1502:
1498:
1494:
1490:
1485:
1483:
1479:
1463:
1451:
1447:
1433:
1413:
1393:
1385:
1370:
1350:
1330:
1322:
1307:
1299:
1298:
1294:
1293:
1287:
1285:
1269:
1266:
1263:
1254:
1252:
1235:
1228:
1210:
1206:
1183:
1179:
1158:
1138:
1135:
1127:
1124:, called the
1111:
1091:
1088:
1085:
1076:
1062:
1042:
1022:
1002:
982:
962:
942:
930:
925:
912:
909:
906:
903:
898:
894:
890:
887:
883:
862:
842:
833:
819:
814:
811:
807:
803:
800:
792:
788:
772:
765:
746:
741:
738:
734:
730:
727:
724:
721:
716:
713:
709:
705:
696:
688:
687:
686:
684:
679:
666:
661:
657:
653:
650:
646:
642:
637:
633:
629:
626:
622:
614:
610:
606:
590:
567:
562:
558:
554:
551:
547:
543:
538:
534:
530:
527:
523:
519:
516:
513:
504:
496:
495:
494:
492:
488:
472:
465:
462:based on the
444:
440:
436:
431:
427:
415:
411:
407:
403:
402:
396:
390:
385:
383:
377:
375:
372:and A is the
356:
351:
348:
344:
338:
334:
330:
325:
322:
318:
310:
309:
308:
307:
291:
267:
264:
261:
258:
255:
248:
247:
246:
244:
240:
239:spring system
230:
224:
219:
205:
197:
193:
189:
185:
181:
165:
145:
125:
122:
117:
114:
110:
89:
81:
71:
69:
64:
62:
58:
54:
50:
46:
42:
38:
29:
21:
3291:
3287:
3281:
3256:
3252:
3242:
3233:
3208:
3202:
3193:
3187:
3168:
3164:
3154:
3145:
3139:
3130:
3124:
3099:
3095:
3089:
3076:
3069:
3026:
3022:
3016:
3007:
2998:
2990:
2965:
2961:
2948:
2934:
2920:
2910:
2903:
2885:
2878:
2868:
2861:
2847:
2812:
2808:
2802:
2783:
2779:
2773:
2740:
2736:
2726:
2692:
2688:
2682:
2645:
2641:
2631:
2618:
2585:
2545:
2538:
2525:
2509:
2499:
2497:
2491:
2484:
2473:method, and
2448:scikit-learn
2445:
2216:
1979:
1836:
1747:
1743:
1699:
1691:
1521:
1509:
1486:
1455:
1255:
1077:
935:Knowing the
934:
834:
761:
682:
680:
603:is also the
582:
405:
399:
397:
393:
378:
371:
283:
245:defined as
235:
192:eigenvectors
183:
77:
65:
40:
34:
2695:: 1537β87,
2590:Demmel, J.
2528:conductance
1779:memory and
764:eigenvector
605:eigenvector
583:The vector
493:defined as
464:eigenvector
74:Definitions
49:eigenvalues
3321:Categories
2822:2105.00578
2815:: 102769.
2793:1706.02803
2578:References
1290:Algorithms
787:eigenvalue
609:eigenvalue
487:eigenvalue
180:clustering
61:clustering
3308:0033-5533
3273:0033-5533
3116:207558562
3031:CiteSeerX
2839:233481603
2702:1001.1323
2456:multigrid
2337:≪
2031:≪
1491:is often
1211:−
888:−
812:−
739:−
731:−
714:−
651:−
627:−
552:−
528:−
520:−
335:∑
265:−
123:≥
51:) of the
3053:17848776
2997:"Kernel
2786:: 1β49.
2757:15376896
2719:88518155
2674:20667133
2551:See also
2442:Software
102:, where
45:spectrum
3061:9402790
2970:Bibcode
2765:2384316
2665:2923634
2648:: 403.
1914:, only
975:matrix
793:matrix
789:of the
489:of the
304:is the
59:before
3306:
3271:
3114:
3059:
3051:
3033:
2837:
2763:
2755:
2717:
2672:
2662:
2516:DBSCAN
2494:-means
2463:ARPACK
2452:LOBPCG
2450:using
2432:LOBPCG
2020:(with
1501:LOBPCG
1284:DBSCAN
284:where
3230:(PDF)
3112:S2CID
3081:(PDF)
3057:S2CID
3004:(PDF)
2958:(PDF)
2835:S2CID
2817:arXiv
2788:arXiv
2761:S2CID
2715:S2CID
2697:arXiv
2467:MLlib
2454:with
2326:with
1518:Costs
194:of a
188:below
3304:ISSN
3269:ISSN
3049:PMID
2753:PMID
2670:PMID
2436:GPUs
2306:-by-
2237:-by-
2066:-by-
2000:-by-
1865:-by-
1721:-by-
1669:-by-
1606:-by-
1267:>
955:-by-
681:The
509:norm
158:and
3296:doi
3292:127
3261:doi
3257:118
3173:doi
3104:doi
3041:doi
2978:doi
2891:doi
2827:doi
2813:106
2745:doi
2707:doi
2660:PMC
2650:doi
2461:or
404:or
198:of
35:In
3323::
3302:.
3290:.
3267:.
3255:.
3251:.
3232:.
3217:^
3169:23
3167:.
3163:.
3110:.
3100:51
3098:.
3055:.
3047:.
3039:.
3027:29
3025:.
3006:.
2976:.
2966:41
2964:.
2960:.
2833:.
2825:.
2811:.
2784:20
2782:.
2759:.
2751:.
2741:26
2739:.
2735:.
2713:,
2705:,
2691:,
2668:.
2658:.
2646:11
2644:.
2640:.
2600:^
2477:.
2465:,
2214:.
1484:.
1286:.
832:.
706::=
701:rw
514::=
384:.
376:.
259::=
70:.
39:,
3310:.
3298::
3275:.
3263::
3236:.
3211:.
3196:.
3181:.
3175::
3148:.
3133:.
3118:.
3106::
3063:.
3043::
2999:k
2984:.
2980::
2972::
2928:.
2897:.
2893::
2855:.
2841:.
2829::
2819::
2796:.
2790::
2767:.
2747::
2709::
2699::
2693:5
2676:.
2652::
2594:.
2500:k
2492:k
2475:R
2418:n
2398:)
2395:n
2392:(
2389:O
2369:)
2366:n
2363:(
2360:O
2340:n
2334:k
2314:k
2294:n
2274:)
2271:n
2268:(
2265:O
2245:n
2225:n
2198:1
2195:=
2192:k
2172:n
2169:=
2166:k
2146:)
2141:3
2137:n
2133:(
2130:O
2110:)
2105:2
2101:n
2097:(
2094:O
2074:n
2054:n
2034:n
2028:k
2008:k
1988:n
1960:)
1957:n
1954:(
1951:O
1931:)
1928:n
1925:(
1922:O
1902:)
1899:n
1896:(
1893:O
1873:n
1853:n
1819:2
1815:n
1792:2
1788:n
1765:2
1761:n
1729:n
1709:n
1677:1
1657:n
1637:n
1634:k
1614:k
1594:n
1574:k
1554:n
1530:n
1464:A
1452:)
1434:l
1414:l
1394:k
1383:)
1371:L
1351:k
1331:k
1308:L
1270:1
1264:k
1236:v
1207:B
1184:+
1180:B
1159:v
1139:,
1136:v
1112:v
1092:1
1089:=
1086:k
1063:k
1043:V
1023:k
1003:n
983:V
963:k
943:n
913:.
910:u
907:=
904:v
899:2
895:/
891:1
884:D
863:u
843:v
820:A
815:1
808:D
804:=
801:P
773:u
747:A
742:1
735:D
728:I
725:=
722:L
717:1
710:D
697:L
667:.
662:2
658:/
654:1
647:D
643:A
638:2
634:/
630:1
623:D
591:v
568:.
563:2
559:/
555:1
548:D
544:A
539:2
535:/
531:1
524:D
517:I
505:L
473:v
450:)
445:2
441:B
437:,
432:1
428:B
424:(
357:,
352:j
349:i
345:A
339:j
331:=
326:i
323:i
319:D
292:D
280:,
268:A
262:D
256:L
206:A
184:k
166:j
146:i
126:0
118:j
115:i
111:A
90:A
47:(
Text is available under the Creative Commons Attribution-ShareAlike License. Additional terms may apply.