Back

Co-clustering documents and words using bipartite spectral graph partitioning

Inderjit S. Dhillon

article2001KDD1,972 citations

Proposes a spectral graph partitioning algorithm that models document collections as bipartite graphs to simultaneously cluster words and documents by finding optimal real relaxations through singular value decomposition.

Listen

Organizing large, unstructured text collections is essential for modern search and information management. Standard approaches typically cluster documents or words independently. However, these two tasks are naturally interconnected: the themes of documents depend on the words they contain, and the meanings of words depend on the documents in which they appear. Failing to group them simultaneously often results in missed contextual relationships and higher computational costs.

The article develops and demonstrates an algorithm that simultaneously clusters words and documents—a process called co-clustering—by framing the task as a graph partitioning problem.

The approach models the relationship between words and documents as a bipartite network, where links exist only between words and the documents containing them. Because finding the globally optimal partition is computationally intractable, the algorithm mathematically approximates the optimal solution by calculating the singular value decomposition of a scaled word-by-document matrix. This maps words and documents into a shared, low-dimensional mathematical space, where standard clustering techniques are applied simultaneously. The framework was evaluated across multiple benchmark text collections ranging from 30 to nearly 3,900 documents, spanning scientific abstracts and real-world news articles.

The evaluation produced several key findings. First, the co-clustering algorithm achieved high accuracy in recovering ground-truth topics; in two-way document splits, misclassifications were kept to fewer than 1% to 2.5% of total documents. Second, the algorithm automatically extracts descriptive word clusters alongside document groupings, providing human-interpretable labels that directly explain the underlying topic of each group. Third, the spectral method proved robust against noise; when tested on raw text containing full sets of uninformative words (such as common stopwords), clustering accuracy degraded only slightly. Finally, the approach performed reliably across small sample sizes (such as 30-document datasets) where traditional clustering techniques frequently fail.

These findings indicate that simultaneous co-clustering improves organizational performance while lowering operational complexity. By transforming a large network problem into a smaller matrix decomposition, the method avoids computationally expensive comparisons between all pairs of documents. Organizations can leverage this capability to automate thematic taxonomy generation, improve document categorization, and provide interpretable search topic labeling.

Decision-makers should consider applying this spectral co-clustering method in workflows that require automated text categorization and interpretability. For initial deployment, practitioners can start with standard two-way or multi-way groupings on moderate collections. When addressing datasets with highly skewed topic sizes, engineering teams should evaluate multi-stage or recursive partitioning strategies, as severe class imbalances in complex real-world news feeds can lead to partial topic overlap.

While the algorithm is mathematically grounded as an optimal continuous relaxation of the partition problem, the final clustering step relies on local search heuristics that require specifying the number of target clusters in advance. Nevertheless, experimental results provide strong confidence in the method's ability to discover coherent dual groupings across diverse, sparse document collections.

University of Texas at Austinbipartite graphsclustering methodsdocument clusteringWordsDocumentsGraph
Dhillon (2001).pdf
  • Paper: Indexing By Latent Semantic Analysis, Scott Deerwester et al. (1990). Provides the foundational algebraic framework for singular value decomposition over term-document co-occurrence matrices that the bipartite spectral method reformulates and solves as a graph cut.
  • Paper: The information bottleneck method, Naftali Tishby et al. (2000). Introduces the theoretical motivation for simultaneously preserving distributional mutual information between words and documents during clustering.
  • Paper: Probabilistic Latent Semantic Analysis, Thomas Hofmann (1999). Establishes statistical co-occurrence modeling between words and documents as an alternative to separate, one-sided clustering approaches.
Cover for Co-clustering documents and words using bipartite spectral graph partitioning

Abstract

Both document clustering and word clustering are well studied problems. Most existing algorithms cluster documents and words separately but not simultaneously. In this paper we present the novel idea of modeling the document collection as a bipartite graph between documents and words, using which the simultaneous clustering problem can be posed as a bipartite graph partitioning problem. To solve the partitioning problem, we use a new spectral co-clustering algorithm that uses the second left and right singular vectors of an appropriately scaled word-document matrix to yield good bipartitionings. The spectral algorithm enjoys some optimality properties; it can be shown that the singular vectors solve a real relaxation to the NP-complete graph bipartitioning problem. We present experimental results to verify that the resulting co-clustering algorithm works well in practice.

Topics

bipartite graphsclustering methodsdocument clusteringelectronic documenttext documentweighted graphhierarchical clustering algorithmsdirected graphspart of speech

Keywords

WordsDocumentsGraphPartitioningCo-clusteringUsingSpectralGraph PartitioningClusteringWords Using Bipartite

Table of Contents

  • 2. BIPARTITE GRAPH MODEL
  • 1. INTRODUCTION
  • 2.1 Simultaneous Clustering
  • 3. GRAPH PARTITIONING
  • 3.1 Spectral Graph Bipartitioning
  • 3.2 Eigenvectors as optimal partition vectors
  • 3.3 Ratio-cut and Normalized-cut objectives
  • 4. THE SVD CONNECTION
  • 4.1 The Bipartitioning Algorithm
  • 4.2 The Multipartitioning Algorithm
  • 5. EXPERIMENTAL RESULTS
  • 5.1 Bipartitioning Results
  • 5.2 Multipartitioning Results
  • 6. CONCLUSIONS
  • 7. REFERENCES

Knowls

  1. Knowl 1 — Bipartite Graph Formulation of Document-Word Co-Clustering

    model/method

    A document collection is modeled as an undirected bipartite graph G=(D,W,E)G = (D, W, E), where D={d1,d2,…,dn}D = \{d_1, d_2, \dots, d_n\} is the set of document vertices and W={w1,w2,…,wm}W = \{w_1, w_2, \dots, w_m\} is the set of word vertices. Edges exist exclusively between document vertices and word vertices with non-negative edge weights Eij=AijE_{ij} = A_{ij} representing the association (such as term frequency or tf-idf weight) between word wiw_i and document djd_j. There are no edges connecting documents to documents or words to words. The adjacency matrix of the bipartite graph is defined as:

    M=[0AAT0]M = \begin{bmatrix} 0 & A \\ A^T & 0 \end{bmatrix}

    where A∈Rm×nA \in \mathbb{R}^{m \times n} is the word-by-document matrix.

    Document clustering and word clustering exhibit a mutual duality: disjoint document clusters D1,…,DkD_1, \dots, D_k determine word clusters W1,…,WkW_1, \dots, W_k by assigning each word to the document cluster with which it shares the greatest total association:

    Wl={wi:∑j∈DlAij≥∑j∈DpAij,∀p=1,…,k}W_l = \left\{ w_i : \sum_{j \in D_l} A_{ij} \ge \sum_{j \in D_p} A_{ij}, \quad \forall p = 1, \dots, k \right\}

    Conversely, word clusters W1,…,WkW_1, \dots, W_k determine document clusters D1,…,DkD_1, \dots, D_k via:

    Dl={dj:∑i∈WlAij≥∑i∈WpAij,∀p=1,…,k}D_l = \left\{ d_j : \sum_{i \in W_l} A_{ij} \ge \sum_{i \in W_p} A_{ij}, \quad \forall p = 1, \dots, k \right\}

    Optimal simultaneous clustering corresponds to partitioning the vertex set V=W∪DV = W \cup D into kk disjoint subsets Vl=Wl∪DlV_l = W_l \cup D_l such that the total weight of cut edges between distinct partitions, cut(V1,…,Vk)=∑i<j∑u∈Vi,v∈VjMuv\text{cut}(V_1, \dots, V_k) = \sum_{i < j} \sum_{u \in V_i, v \in V_j} M_{uv}, is minimized.

  2. Knowl 2 — Equivalence of Bipartite Graph Normalized Laplacian Eigenproblem to SVD of Scaled Word-Document Matrix

    theoretical result

    Let G=(D,W,E)G = (D, W, E) be a connected bipartite graph with word-by-document matrix A∈Rm×nA \in \mathbb{R}^{m \times n}. Define diagonal word-degree matrix D1∈Rm×mD_1 \in \mathbb{R}^{m \times m} and document-degree matrix D2∈Rn×nD_2 \in \mathbb{R}^{n \times n} with entries:

    (D1)ii=∑j=1nAij,(D2)jj=∑i=1mAij(D_1)_{ii} = \sum_{j=1}^n A_{ij}, \qquad (D_2)_{jj} = \sum_{i=1}^m A_{ij}

    Assuming D1D_1 and D2D_2 are non-singular, the graph Laplacian is L=[D1−A−ATD2]L = \begin{bmatrix} D_1 & -A \\ -A^T & D_2 \end{bmatrix} and the total degree matrix is D=[D100D2]D = \begin{bmatrix} D_1 & 0 \\ 0 & D_2 \end{bmatrix}. The generalized eigenvalue problem:

    [D1−A−ATD2][xy]=λ[D100D2][xy]\begin{bmatrix} D_1 & -A \\ -A^T & D_2 \end{bmatrix} \begin{bmatrix} x \\ y \end{bmatrix} = \lambda \begin{bmatrix} D_1 & 0 \\ 0 & D_2 \end{bmatrix} \begin{bmatrix} x \\ y \end{bmatrix}

    with change of variables u=D11/2xu = D_1^{1/2} x and v=D21/2yv = D_2^{1/2} y, transforms directly into the singular value decomposition (SVD) of the normalized matrix An=D1−1/2AD2−1/2A_n = D_1^{-1/2} A D_2^{-1/2}:

    Anv=(1−λ)u,AnTu=(1−λ)vA_n v = (1 - \lambda) u, \qquad A_n^T u = (1 - \lambda) v

    Thus, the generalized eigenvalue λ\lambda and eigenvector z=[xT,yT]Tz = [x^T, y^T]^T correspond to a singular value σ=1−λ\sigma = 1 - \lambda, left singular vector uu, and right singular vector vv of AnA_n. The second smallest generalized eigenvalue λ2\lambda_2 corresponds to the second largest singular value σ2=1−λ2\sigma_2 = 1 - \lambda_2, and the corresponding generalized eigenvector is given by:

    z2=[D1−1/2u2D2−1/2v2]z_2 = \begin{bmatrix} D_1^{-1/2} u_2 \\ D_2^{-1/2} v_2 \end{bmatrix}

    This equivalence avoids solving an eigensystem of size (m+n)×(m+n)(m+n) \times (m+n), replacing it with the computation of singular vectors of the smaller m×nm \times n matrix AnA_n.

  3. Knowl 3 — Bipartite Spectral Graph Multipartitioning Algorithm for Simultaneous Document-Word Clustering

    algorithm

    The multipartitioning algorithm co-clusters a set of mm words and nn documents into kk simultaneous word-document clusters using the top ℓ=⌈log⁡2k⌉\ell = \lceil \log_2 k \rceil non-trivial left and right singular vectors of the normalized word-document matrix.

    Algorithm Multipartition:
    Input: Word-document matrix A∈Rm×nA \in \mathbb{R}^{m \times n}, number of clusters kk
    Output: Document partitions D1,…,DkD_1, \dots, D_k, word partitions W1,…,WkW_1, \dots, W_k
    1. Compute diagonal degree matrices D1=diag(∑jAij)D_1 = \text{diag}(\sum_j A_{ij}) and D2=diag(∑iAij)D_2 = \text{diag}(\sum_i A_{ij})
    2. Form the normalized word-document matrix An=D1−1/2AD2−1/2A_n = D_1^{-1/2} A D_2^{-1/2}
    3. Set ℓ=⌈log⁡2k⌉\ell = \lceil \log_2 k \rceil
    4. Compute the ℓ\ell singular vectors of AnA_n corresponding to the 2nd through (ℓ+1)(\ell+1)-th largest singular values:
       left singular vectors u2,…,uℓ+1u_2, \dots, u_{\ell+1} and right singular vectors v2,…,vℓ+1v_2, \dots, v_{\ell+1}
    5. Form matrices U=[u2,…,uℓ+1]∈Rm×ℓU = [u_2, \dots, u_{\ell+1}] \in \mathbb{R}^{m \times \ell} and V=[v2,…,vℓ+1]∈Rn×ℓV = [v_2, \dots, v_{\ell+1}] \in \mathbb{R}^{n \times \ell}
    6. Construct the (m+n)×ℓ(m+n) \times \ell reduced data matrix:
       Z=[D1−1/2UD2−1/2V]Z = \begin{bmatrix} D_1^{-1/2} U \\ D_2^{-1/2} V \end{bmatrix}
    7. Run the kk-means clustering algorithm on the (m+n)(m+n) rows of ZZ to minimize the sum of squared Euclidean distances to kk cluster centers
    8. Output word clusters W1,…,WkW_1, \dots, W_k from the first mm rows and document clusters D1,…,DkD_1, \dots, D_k from the remaining nn rows
    return (D1,…,Dk),(W1,…,Wk)(D_1, \dots, D_k), (W_1, \dots, W_k)

    The algorithm operates simultaneously on the low-dimensional embeddings of both words and documents, producing paired clusters where document cluster DjD_j is directly associated with word cluster WjW_j.

  4. Knowl 4 — Real Relaxation of Normalized Cut Minimization via Generalized Eigenvectors

    theoretical result

    For a graph G=(V,E)G = (V, E) with Laplacian matrix LL and diagonal degree matrix DD, consider a bipartition of vertices into V1V_1 and V2V_2 (V1∪V2=VV_1 \cup V_2 = V, V1∩V2=∅V_1 \cap V_2 = \emptyset). Let η1=weight(V1)=∑i∈V1Dii\eta_1 = \text{weight}(V_1) = \sum_{i \in V_1} D_{ii} and η2=weight(V2)=∑i∈V2Dii\eta_2 = \text{weight}(V_2) = \sum_{i \in V_2} D_{ii}. Define the generalized partition vector q∈R∣V∣q \in \mathbb{R}^{|V|} by:

    qi={+η2/η1,i∈V1−η1/η2,i∈V2q_i = \begin{cases} +\sqrt{\eta_2 / \eta_1}, & i \in V_1 \\ -\sqrt{\eta_1 / \eta_2}, & i \in V_2 \end{cases}

    This vector satisfies qTDe=0q^T D e = 0 (where e=[1,…,1]Te = [1, \dots, 1]^T) and qTDq=weight(V)q^T D q = \text{weight}(V). The generalized Rayleigh quotient evaluates exactly to the normalized cut objective:

    qTLqqTDq=cut(V1,V2)weight(V1)+cut(V1,V2)weight(V2)=N(V1,V2)\frac{q^T L q}{q^T D q} = \frac{\text{cut}(V_1, V_2)}{\text{weight}(V_1)} + \frac{\text{cut}(V_1, V_2)}{\text{weight}(V_2)} = N(V_1, V_2)

    Finding a discrete generalized partition vector that minimizes this objective is NP-complete. Relaxing qq to take arbitrary real values subject to qTDe=0q^T D e = 0 yields the continuous optimization problem:

    min⁡q≠0,qTDe=0qTLqqTDq\min_{q \ne 0, q^T D e = 0} \frac{q^T L q}{q^T D q}

    The global minimum of this relaxed problem is attained when qq is the eigenvector corresponding to the second smallest eigenvalue λ2\lambda_2 of the generalized eigenvalue problem Lz=λDzL z = \lambda D z.

  5. Knowl 5 — Bipartite Spectral Graph Bipartitioning Algorithm

    algorithm

    The bipartitioning algorithm co-clusters words and documents into two clusters by computing the second singular vectors of the degree-normalized word-document matrix and partitioning the resulting 1-dimensional representation via 2-means clustering.

    Algorithm Bipartition:
    Input: Word-document matrix A∈Rm×nA \in \mathbb{R}^{m \times n}
    Output: Document bipartition (D1,D2)(D_1, D_2), word bipartition (W1,W2)(W_1, W_2)
    1. Compute diagonal degree matrices (D1)ii=∑j=1nAij(D_1)_{ii} = \sum_{j=1}^n A_{ij} and (D2)jj=∑i=1mAij(D_2)_{jj} = \sum_{i=1}^m A_{ij}
    2. Form the normalized matrix An=D1−1/2AD2−1/2A_n = D_1^{-1/2} A D_2^{-1/2}
    3. Compute the second left singular vector u2∈Rmu_2 \in \mathbb{R}^m and second right singular vector v2∈Rnv_2 \in \mathbb{R}^n of AnA_n
    4. Form the 1-dimensional embedding vector:
       z2=[D1−1/2u2D2−1/2v2]∈Rm+nz_2 = \begin{bmatrix} D_1^{-1/2} u_2 \\ D_2^{-1/2} v_2 \end{bmatrix} \in \mathbb{R}^{m+n}
    5. Run 2-means clustering on the 1-dimensional values of z2z_2 to find bi-modal centers m1,m2∈Rm_1, m_2 \in \mathbb{R} minimizing:
       ∑j=12∑z2(i)∈mj(z2(i)−mj)2\sum_{j=1}^2 \sum_{z_2(i) \in m_j} (z_2(i) - m_j)^2
    6. Assign words according to indices 1≤i≤m1 \le i \le m and documents according to indices m+1≤i≤m+nm+1 \le i \le m+n
    return (D1,D2),(W1,W2)(D_1, D_2), (W_1, W_2)

  6. Knowl 6 — Normalized Cut Objective for Balanced Graph Bipartitioning

    definition

    For a graph G=(V,E)G = (V, E) with adjacency matrix entries EijE_{ij}, the cut between two disjoint vertex subsets V1,V2⊂VV_1, V_2 \subset V is defined as cut(V1,V2)=∑i∈V1,j∈V2Eij\text{cut}(V_1, V_2) = \sum_{i \in V_1, j \in V_2} E_{ij}. The vertex weight of a subset is defined as weight(Vl)=∑i∈Vl∑kEik\text{weight}(V_l) = \sum_{i \in V_l} \sum_k E_{ik}.

    The normalized-cut objective function N(V1,V2)N(V_1, V_2) balances cluster sizes by penalizing partitions with small aggregate edge degrees:

    N(V1,V2)=cut(V1,V2)∑i∈V1∑kEik+cut(V1,V2)∑i∈V2∑kEik=2−S(V1,V2)N(V_1, V_2) = \frac{\text{cut}(V_1, V_2)}{\sum_{i \in V_1} \sum_k E_{ik}} + \frac{\text{cut}(V_1, V_2)}{\sum_{i \in V_2} \sum_k E_{ik}} = 2 - S(V_1, V_2)

    where the internal association measure S(V1,V2)S(V_1, V_2) is:

    S(V1,V2)=within(V1)weight(V1)+within(V2)weight(V2)S(V_1, V_2) = \frac{\text{within}(V_1)}{\text{weight}(V_1)} + \frac{\text{within}(V_2)}{\text{weight}(V_2)}

    and within(Vl)\text{within}(V_l) denotes the sum of edge weights with both endpoints in VlV_l. Minimizing N(V1,V2)N(V_1, V_2) is mathematically equivalent to maximizing the proportion of edge weights contained internally within each partition.

  7. Knowl 7 — Robustness of Bipartite Spectral Co-Clustering to Stop Words via Degree Normalization

    empirical result

    Bipartite spectral graph co-clustering exhibits high robustness to the presence of common noise words and stop words without requiring explicit stop-word removal or frequency thresholding. This robustness stems from the diagonal scaling in An=D1−1/2AD2−1/2A_n = D_1^{-1/2} A D_2^{-1/2}, which divides each term entry by the square root of its total word degree (D1)ii=∑jAij(D_1)_{ii} = \sum_j A_{ij}, heavily down-weighting ubiquitous non-discriminative terms.

    When evaluated on collections containing all words including unremoved stop words (MedCran_All: 2,433 documents, 17,162 words; MedCisi_All: 2,493 documents, 19,194 words), Algorithm Bipartition achieved near-perfect separation into the true document classes, nearly matching the results obtained after stop-word filtering:

    MedCran_All MedCisi_All
    Cluster Medline Cranfield Cluster Medline Cisi
    D0D_0 1014 0 D0D_0 925 0
    D1D_1 19 1400 D1D_1 108 1460

    For MedCran_All, 1,014 out of 1,033 Medline documents and all 1,400 Cranfield documents were correctly partitioned. For MedCisi_All, 925 of 1,033 Medline documents and all 1,460 Cisi documents were correctly partitioned.

  8. Knowl 8 — Simultaneous Document-Word Multipartitioning Performance and Concept Word Extraction on Classic3

    empirical result

    Algorithm Multipartition was evaluated on the Classic3 dataset, a mixture of 3,893 documents across 3 classes (1,033 Medline, 1,460 Cisi, 1,400 Cranfield) with 4,303 terms. Using ℓ=⌈log⁡23⌉=2\ell = \lceil \log_2 3 \rceil = 2 singular vectors, the algorithm identified three distinct document clusters D0,D1,D2D_0, D_1, D_2 and three associated word clusters W0,W1,W2W_0, W_1, W_2.

    Cluster Medline Cisi Cranfield
    D0D_0 965 0 0
    D1D_1 65 1458 10
    D2D_2 3 2 1390

    The top 7 words in each word cluster (ranked by highest internal edge weights to the corresponding document cluster) accurately captured the underlying concept domain:

    • W0W_0 (Medical): patients, cells, blood, hormone, renal, cancer, rats
    • W1W_1 (Information Retrieval / Library Science): library, libraries, retrieval, scientific, science, book, system
    • W2W_2 (Aeronautics): boundary layer, heat, shock, mach, supersonic, wing

  9. Knowl 9 — Spectral Co-Clustering Accuracy on Small Document Collections

    empirical result

    Algorithm Multipartition maintains high clustering accuracy even on very small document collections where standard text clustering algorithms such as spherical kk-means typically fail.

    Testing on small random subsets of Medline, Cisi, and Cranfield documents demonstrated near-complete recovery of the original classes:

    Classic3_30docs (10/class) Classic3_150docs (50/class)
    Cluster Medline Cisi Cranfield Cluster Medline Cisi Cranfield
    D0D_0 9 0 0 D0D_0 49 0 0
    D1D_1 0 10 0 D1D_1 0 50 0
    D2D_2 1 0 10 D2D_2 1 0 50

    In Classic3_30docs (30 documents, 1,073 words), 29 of 30 documents were correctly assigned to their respective true classes. In Classic3_150docs (150 documents, 3,652 words), 149 of 150 documents were correctly recovered.

Coverage note — Omitted materials include standard graph partitioning heuristics (Kernighan-Lin, Fiduccia-Mattheyses) and the ratio-cut variant, which were review material from prior literature rather than original contributions.

References

  1. 1.L. D. Baker and A. McCallum. Distributional clustering of words for text classification. In ACM SIGIR, pages 96–103, 1998.
  2. 2.D. Boley. Hierarchical taxonomies using divisive partitioning. Technical Report TR-98-012, University of Minnesota, 1998.
  3. 3.D. Boley, M. Gini, R. Gross, E.-H. Han, K. Hastings, G. Karypis, V. Kumar, B. Mobasher, and J. Moore. Document categorization and query generation on the World Wide Web using WebACE. AI Review, 1998.
  4. 4.C. J. Crouch. A cluster-based approach to thesaurus construction. In ACM SIGIR, pages 309–320, 1988.
  5. 5.D. R. Cutting, D. R. Karger, J. O. Pedersen, and J. W. Tukey. Scatter/gather: A cluster-based approach to browsing large document collections. In ACM SIGIR, 1992.
  6. 6.I. S. Dhillon, J. Fan, and Y. Guan. Efficient clustering of very large document collections. In V. K. R. Grossman, C. Kamath and R. Namburu, editors, Data Mining for Scientific and Engineering Applications. Kluwer Academic Publishers, 2001.
  7. 7.I. S. Dhillon and D. S. Modha. Concept decompositions for large sparse text data using clustering. Machine Learning, 42(1):143–175, January 2001. Also appears as IBM Research Report RJ 10147, July 1999.
  8. 8.W. E. Donath and A. J. Hoffman. Lower bounds for the partitioning of graphs. IBM Journal of Research and Development, 17:420–425, 1973.
  9. 9.R. O. Duda, P. E. Hart, and D. G. Stork. Pattern Classification. John Wiley & Sons, 2000. 2nd Edition.
  10. 10.C. M. Fiduccia and R. M. Mattheyses. A linear time heuristic for improving network partitions. Technical Report 82CRD130, GE Corporate Research, 1982.
  11. 11.M. Fiedler. Algebraic connectivity of graphs. Czecheslovak Mathematical Journal, 23:298–305, 1973.
  12. 12.M. R. Garey and D. S. Johnson. Computers and Intractability: A Guide to the Theory of NP-Completeness. W. H. Freeman & Company, 1979.
  13. 13.G. H. Golub and C. F. V. Loan. Matrix computations. Johns Hopkins University Press, 3rd edition, 1996.
  14. 14.L. Hagen and A. B. Kahng. New spectral methods for ratio cut partitioning and clustering. IEEE Transactions on CAD, 11:1074–1085, 1992.
  15. 15.K. M. Hall. An r-dimensional quadratic placement algorithm. Management Science, 11(3):219–229, 1970.
  16. 16.R. V. Katter. Study of document representations: Multidimensional scaling of indexing terms. System Development Corporation, Santa Monica, CA, 1967.
  17. 17.B. Kernighan and S. Lin. An efficient heuristic procedure for partitioning graphs. The Bell System Technical Journal, 29(2):291–307, 1970.
  18. 18.T. Kohonen. Self-organizing Maps. Springer, 1995.
  19. 19.A. Pothen, H. Simon, and K.-P. Liou. Partitioning sparse matrices with eigenvectors of graphs. SIAM Journal on Matrix Analysis and Applications, 11(3):430–452, July 1990.
  20. 20.G. Salton and M. J. McGill. Introduction to Modern Retrieval. McGraw-Hill Book Company, 1983.
  21. 21.H. Schütze and C. Silverstein. Projections for efficient document clustering. In ACM SIGIR, 1997.
  22. 22.J. Shi and J. Malik. Normalized cuts and image segmentation. IEEE Trans. Pattern Analysis and Machine Intelligence, 22(8):888–905, August 2000.
  23. 23.A. Strehl, J. Ghosh, and R. Mooney. Impact of similarity measures on web-page clustering. In AAAI 2000 Workshop on AI for Web Search, 2000.
  24. 24.C. J. van Rijsbergen. Information Retrieval. Butterworths, London, second edition, 1979.
  25. 25.E. M. Voorhees. The effectiveness and efficiency of agglomerative hierarchic clustering in document retrieval. PhD thesis, Cornell University, 1986.

Citation

MLA
Dhillon, I. S. “Co-clustering Documents and Words Using Bipartite Spectral Graph Partitioning”. Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2001, pp. 269–74, https://doi.org/10.1145/502512.502550.
APA
Dhillon, I. S. (2001). Co-clustering documents and words using bipartite spectral graph partitioning. Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 269–274. https://doi.org/10.1145/502512.502550
Chicago
Dhillon, I. S. 2001. “Co-clustering Documents and Words Using Bipartite Spectral Graph Partitioning”. Proceedings of the Seventh ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 269–74. https://doi.org/10.1145/502512.502550.
Harvard
Dhillon, I.S. (2001) “Co-clustering documents and words using bipartite spectral graph partitioning”, Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, pp. 269–274. Available at: https://doi.org/10.1145/502512.502550.
Vancouver
1. Dhillon IS (2001) Co-clustering documents and words using bipartite spectral graph partitioning. In: Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining. ACM, pp 269–274

BibTeX

@inproceedings{Dhillon_2001, series={KDD01}, title={Co-clustering documents and words using bipartite spectral graph partitioning}, url={http://dx.doi.org/10.1145/502512.502550}, DOI={10.1145/502512.502550}, booktitle={Proceedings of the seventh ACM SIGKDD international conference on Knowledge discovery and data mining}, publisher={ACM}, author={Dhillon, Inderjit S.}, year={2001}, month=Aug, pages={269–274}, collection={KDD01} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF