Authors
Data Clustering: Algorithms and Applications
2
Contents
1 Grid-based Clustering Wei Cheng, Wei Wang, and Sandra Batista 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 The Classical Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 Earliest Approaches: GRIDCLUS and BANG . . . . . . . . . . . . 1.2.2 STING and STING+: The Statistical Information Grid Approach 1.2.3 WaveCluster: Wavelets in Grid-based Clustering . . . . . . . . . . 1.3 Adaptive Grid-based Algorithms . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 AMR: Adaptive Mesh Refinement Clustering . . . . . . . . . . . . 1.4 Axis-shifting Grid-based Algorithms . . . . . . . . . . . . . . . . . . . . . 1.4.1 NSGC: New Shifting Grid Clustering Algorithm . . . . . . . . . . 1.4.2 ADCC: Adaptable Deflect and Conquer Clustering . . . . . . . . . 1.4.3 ASGC: Axis-Shifted Grid-Clustering . . . . . . . . . . . . . . . . . 1.4.4 GDILC: Grid-based Density-IsoLine Clustering Algorithm . . . . . 1.5 High Dimensional Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 1.5.1 CLIQUE: The Classical High-Dimensional Algorithm . . . . . . . . 1.5.2 Variants of CLIQUE . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.2.1 ENCLUS: Entropy-based Approach . . . . . . . . . . . . 1.5.2.2 MAFIA: Adaptive Grids in High Dimensions . . . . . . . 1.5.3 OptiGrid: Density-based Optimal Grid Partitioning . . . . . . . . 1.5.4 Variants of the OptiGrid Approach . . . . . . . . . . . . . . . . . . 1.5.4.1 O-Cluster: A Scalable Approach . . . . . . . . . . . . . . 1.5.4.2 CBF: Cell-based Filtering . . . . . . . . . . . . . . . . . . 1.6 Conclusions and Summary . . . . . . . . . . . . . . . . . . . . . . . . . . Bibliography
1 . . . . . . . . . . . . . . . . . . . . . .
1 5 5 7 8 9 9 10 10 11 12 13 14 14 15 15 15 16 18 18 19 20 23
i
ii
Chapter 1 Grid-based Clustering Wei Cheng Computer Science Department University of North Carolina at Chapel Hill Chapel Hill, NC 27599 Wei Wang Computer Science Department University of California, Los Angeles Los Angeles, CA 90095 Sandra Batista Statistical Science Department Duke University Durham, NC 27710 1.1 1.2
1.3 1.4
1.5
1.6
1.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . The Classical Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.1 Earliest Approaches: GRIDCLUS and BANG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2.2 STING and STING+: The Statistical Information Grid Approach . . . . . . . . . . . . . . 1.2.3 WaveCluster: Wavelets in Grid-based Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Adaptive Grid-based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 AMR: Adaptive Mesh Refinement Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Axis-shifting Grid-based Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.1 NSGC: New Shifting Grid Clustering Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.2 ADCC: Adaptable Deflect and Conquer Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.3 ASGC: Axis-Shifted Grid-Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.4 GDILC: Grid-based Density-IsoLine Clustering Algorithm . . . . . . . . . . . . . . . . . . . . . . High Dimensional Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.1 CLIQUE: The Classical High-Dimensional Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.2 Variants of CLIQUE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.2.1 ENCLUS: Entropy-based Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.2.2 MAFIA: Adaptive Grids in High Dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.3 OptiGrid: Density-based Optimal Grid Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.4 Variants of the OptiGrid Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.4.1 O-Cluster: A Scalable Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.4.2 CBF: Cell-based Filtering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Conclusions and Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1 5 5 7 8 9 9 10 10 11 12 12 13 14 15 15 15 16 17 18 19 19
Introduction
Grid-based clustering algorithms are efficient in mining large multidimensional data sets. These algorithms partition the data space into a finite number of cells to form a grid structure and then form clusters from the cells in the grid structure. Clusters correspond to regions that are more dense in data points than their surroundings. Grids were initially 1
2
Data Clustering: Algorithms and Applications
TABLE 1.1: Grid-based algorithms that use hierarchical clustering or subspace clustering hierarchical clustering GRIDCLUS, BANG-clustering, AMR, STING, STING+ subspace clustering MAFIA, CLIQUE, ENCLUS
proposed by Warnekar and Krishna [29] to organize the feature space, e.g., in GRIDCLUS [25], and increased in popularity after STING [28], CLIQUE [1], and WaveCluster [27] were introduced. The great advantage of grid-based clustering is a significant reduction in time complexity, especially for very large data sets. Rather than clustering the data points directly, grid-based approaches cluster the neighborhood surrounding the data points represented by cells. In most applications since the number of cells is significantly smaller than the number of data points, the performance of grid-based approaches is significantly improved. Grid-based clustering algorithms typically involve the following five steps [9, 10]: 1. Creating the grid structure, i.e., partitioning the data space into a finite number of cells. 2. Calculating the cell density for each cell. 3. Sorting of the cells according to their densities. 4. Identifying cluster centers. 5. Traversal of neighbor cells. Since cell density often needs to be calculated in order to sort cells and select cluster centers, most grid-based clustering algorithms may also be considered density-based. Some grid-based clustering algorithms also combine hierarchical clustering or subspace clustering in order to organize cells based on their density. Table 1.1 lists several representative gridbased algorithms which also use hierarchical clustering or subspace clustering. Grid-based clustering is susceptible to the following data challenges: 1. Non-Uniformity: Using a single inflexible, uniform grid may not be sufficient to achieve desired clustering quality or efficiency for highly irregular data distributions. 2. Locality: If there are local variations in the shape and density of the distribution of data points, the effectiveness of grid-based clustering is limited by predefined cell sizes, cell borders, and the density threshold for significant cells. 3. Dimensionality: Since performance depends on the size of the grid structures and the size of grid structures may increase significantly with more dimensions, grid-based approaches may not be scalable for clustering very high-dimensional data. In addition, there are aspects of the “curse of dimensionality” including filtering noise and selecting the most relevant attributes that are increasingly difficult with more dimensions in a grid-based clustering approach. To overcome the challenge of non-uniformity, adaptive grid-based clustering algorithms that divide the feature space at multiple resolutions, e.g., AMR [14] and MAFIA [21], were proposed. The varying grid sizes can cluster well data with non-uniform distributions. For example, as illustrated in Figure 1.1(a), the data is dispersed throughout the spatial domain with several more dense nested regions in the shape of a circle, square, and rectangle. A single resolution uniform grid would have difficulty identifying those more dense, nested regions as clusters as shown in Figure 1.1(b). In contrast, an adaptive algorithm, such as
Grid-based Clustering
(a) Original data
3
(b) Uniform grid
FIGURE 1.1: Non-uniformity example with nested clusters
TABLE 1.2: Grid-based Algorithms Addressing Non-uniformity (Adaptive), Locality (Axis-Shifting), and Dimensionality Adaptive MAFIA, AMR Axis-shifting NSGC, ADCC, ASGC, GDILC High-dimension CLIQUE, MAFIA, ENCLUS, OptiGrid, O-cluster, CBF
AMR, that permits higher resolution throughout the space can recognize those nested, more dense clusters with centers at the most clear, dense shapes. (Figure 1.1 is adapted from Figure 1 from Liao et al.[14] and is only illustrative, not based on real data.) To address locality, axis-shifting algorithms were introduced. These methods adopt axisshifted partitioning strategies to identify areas of high density in the feature space. For instance, in Figure 1.2(a), traditional grid-based algorithms will have difficulty adhering to the border and continuity of the most dense regions because of the predefined grids and the threshold of significant cells. The clustering from using a single uniform grid, shown in Figure 1.2(b), demonstrates that some clusters are divided into several smaller clusters because the continuity of the border of the dense regions is disturbed by cells with low density. To remedy this, axis-shifting algorithms, such as ASGC [16], shift the coordinate axis by half a cell width in each dimension creating a new grid structure. This shifting yields a clustering that recognizes more dense regions adjacent to lower density cells as shown in Figure 1.2(c). By combining the clustering from both axes, algorithms can recognize dense regions as clusters as shown in Figure 1.2(d). (Figures 1.2(a), 1.2(b), 1.2(c), and 1.2(d) are adapted from Figures 11, 12, 14, and 18 respectively from Lin et al. [16] and are only illustrative, not based on real data or real clustering algorithm results.) For handling high dimensional data, there are several grid-based approaches. For example, the CLIQUE algorithm selects appropriate subspaces rather than the whole feature space for finding the dense regions. In contrast, the OptiGrid algorithm uses density estimations. A summary of grid-based algorithms that address these three challenges is presented in Table 1.2.
4
Data Clustering: Algorithms and Applications
v
v
Cluster 1
Cluster 2 Cluster 3
(a) original data
(b) 3 clusters found by 31 × 38 cells
Cluster 1
v
Cluster 3
Cluster 2
(c) 3 clusters found by 31 × 38 cells after axisshifting
v
Cluster 1
Cluster 2
(d) final clusters by combining b and c
FIGURE 1.2: Locality Example: Axis-shifting grid-based clustering
Grid-based Clustering
5
In the remainder of this chapter we survey classical grid-based clustering algorithms as well as those algorithms that directly address the challenges of non-uniformity, locality, and high dimensionality. First, we discuss some classical grid-based clustering algorithms in Section 1.2. These classical grid-based clustering algorithms include the earliest approaches: GRIDCLUS, STING, WaveCluster, and variants of them. We present an adaptive grid-based algorithm, AMR, in Section 1.3. Several axis-shifting algorithms are evaluated in Section 1.4. In Section 1.5, we discuss high dimensional grid-based algorithms, including CLIQUE, OptiGrid, and their variants. We offer our conclusions and summary in Section 1.6.
1.2
The Classical Algorithms
In this section, we introduce three classical grid-based clustering algorithms together with their variants: GRIDCLUS, STING, and WaveCluster.
1.2.1
Earliest Approaches: GRIDCLUS and BANG
Schikuta et al. [25] introduced the first GRID-based hierarchical CLUStering algorithm called GRIDCLUS. The algorithm partitions the data space into a grid structure comprised of disjoint d-dimensional hyper rectangles or blocks. Data points are considered points in d-dimensional space and are designated to blocks in the grid structure such that their topological distributions are maintained. Once the data is assigned to blocks, clustering is done by a neighbor search algorithm. In some respects, GRIDCLUS is the canonical grid-based clustering algorithm and its basic steps coincide with those given for grid-based algorithms in Section 1.1. Namely, GRIDCLUS inserts points into blocks in its grid structure, calculates the resultant density of the blocks, sorts the blocks according to their density, recognizes the most dense blocks as cluster centers, and constructs the rest of clusters using a neighbor search on the blocks. The grid structure has a scale for each dimension, a grid directory, and the set of data blocks. Each scale is used to partition the entire d-dimensional space and this partitioning is stored in the grid directory. The data blocks contain the data points and there is an upper bound on the number of points per block. The blocks must be non-empty, cover all the data points, and not have any data points in common. Hinrichs offers a more thorough discussion of the grid file structure used[15]. The density index of a block, B, is defined as the number of points in the block divided by the spatial volume of the block, i.e., DB =
pB , VB
(1.1)
where pB is the number of data points in the block B and VB is the spatial volume of the block B, i.e., d ∏ VB = e Bi . (1.2) i=1
where d is the number of dimensions and eBi is the extent of the block in the i dimension. GRIDCLUS sorts the blocks according to their density and those with the highest density are chosen as the cluster centers. The blocks are clustered in order of descending density iteratively to create a nested sequence of nonempty, disjoint clusters. Starting from cluster centers only neighboring blocks are merged into clusters. The neighbor search is
6
Data Clustering: Algorithms and Applications
Algorithm 1 GRIDCLUS Algorithm 1: Set u := 0, W [] := {},C[] := {}{initialization}; 2: Create the grid structure and calculate the block density indices; 3: Generate a sorted block sequence B1′ , B2′ , ..., Bb′ and mark all blocks “not active” and “not clustered”; 4: while a “not active” block exists do 5: u ⇐ u + 1; 6: mark first B1′ , B2′ , ..., Bj ′ with equal density index “active”; 7: for each “not clustered” block Bl′ := B1′ , B2′ , ..., Bj ′ do 8: Create a new cluster set C[u]; 9: W [u] ⇐ W [u] + 1,C[u, W [u]] ← Bl′ ; 10: Mark block Bl′ clustered; 11: N EIGHBOR SEARCH(Bl′ ,C[u, W [u]]); 12: end for 13: for each “not active” block B do 14: W [u] ⇐ W [u] + 1,C[u, W [u]] ← B; 15: end for 16: Mark all blocks “not clustered”; 17: end while Algorithm 2 Procedure NEIGHBOR SEARCH(B,C) 1: for each “active” and “not clustered” neighbor B ′ of B do 2: C ← B′; 3: Mark block B ′ “clustered”; 4: N EIGHBOR SEARCH(B ′ , C); 5: end for
done recursively starting at the cluster center, checking for adjacent blocks that should be added to the cluster, and for only those neighboring blocks added to the cluster, continuing the search. The GRIDCLUS algorithm is described in Algorithm 1 and the function N EIGHBOR SEARCH is the recursive procedure described in Algorithm 2 [10, 25]. While no explicit time complexity analysis is given for GRIDCLUS in the original paper, the algorithm may not have time complexity much better than other hierarchical clustering algorithms in the worst case. The number of blocks in the worst case is O(n) where n is the number of data points and sorting the blocks by density is O(n log n). However, this complexity would still be better than hierarchical clustering. The problem is that step 4 can also require O(n) if all the blocks have different densities and step 7 can also require O(n) if all the blocks have the same density. In addition, while the number of neighbors of any block is a function of the number of dimensions, the depth of the recursive calls to the neighbor search function can also be O(n). This can occur if the blocks are adjacent in a single place analogous to a spanning tree that is a straight line. Without any discriminatory density thresholds, the pathological case of step 7 could also apply and the complexity would be O(n2 ). (Granted average case complexity for several distributions may be significantly better (i.e., O(n)) and that may be a better analysis to consider.) The BANG algorithm introduced by Schikuta and Erhart [26] is an extension of the GRIDCLUS algorithm. It addresses some of the inefficiencies of the GRIDCLUS algorithm in terms of grid structure size, searching for neighbors, and managing blocks by their density. BANG also places data points in blocks and uses a variant of the grid directory called a BANG structure to maintain blocks. Neighbor search and processing the blocks in decreasing
Grid-based Clustering
7
Algorithm 3 BANG-clustering Algorithm 1: Partition the feature space into rectangular blocks which contains up to a maximum of pmax data points. 2: Build a binary tree to maintain the populated blocks, in which the partition level corresponds to the node depth in the tree. 3: Calculate the dendrogram in which the density indices of all blocks are calculated and sorted in decreasing order. 4: Starting with the highest density index, all neighbor blocks are determined and classified in decreasing order. BANG-clustering places the found regions in the dendrogram to the right of the original blocks. 5: Repeat step 4 for the remaining blocks of the dendrogram.
TABLE 1.3: Statistical Information in STING n number of objects (points) in the cell mean mean of each dimension in this cell std standard deviation of each dimension in this cell min the minimum value of each dimension in this cell max the maximum value of each dimension in this cell dist the distribution of points in this cell
order of density are also used for clustering blocks. Nearness of neighbors is determined by the maximum dimensions shared by a common face between blocks. A binary tree is used to store the grid structure, so that neighbor searching can be done more efficiently. From this tree in the grid directory and the sorted block densities, the dendrogram is calculated. Centers of clusters are still the most highly dense blocks in the clustering phase. The BANG algorithm is summarized in Algorithm 3 [10]. While both GRIDCLUS and BANG can discern nested clusters efficiently, BANG has been shown to be more efficient than GRIDCLUS on large data sets because of its significantly reduced growth of grid structure size [26].
1.2.2
STING and STING+: The Statistical Information Grid Approach
Wang et al. [28] proposed a STatistical INformation Grid-based clusterinG method (STING) to cluster spatial databases and to facilitate region oriented queries. STING divides the spatial area into rectangular cells and stores the cells in a hierarchical grid structure tree. Each cell (except leaves in the tree) is partitioned into 4 child cells at the next level with each child corresponding to a quadrant of the parent cell. A parent cell is the union of its children; the root cell at level 1 corresponds to the whole spatial area. The leaf level cells are of uniform size, determined globally from the average density of objects. For each cell, both attribute-dependent and attribute-independent parameters of the statistical information are maintained. These parameters are defined in Table 1.3. STING maintains summary statistics for each cell in its hierarchical tree. As a result, statistical parameters of parent cells can easily be computed from the parameters of child cells. Note that the distribution types may be normal, uniform, exponential and none. Value of dist may either be assigned by the user or obtained by hypothesis tests such as the χ2 test. Even though measures of these statistical parameters are calculated in a bottom-up fashion from any leaf node, the STING algorithm adopts a top-down approach for clustering
8
Data Clustering: Algorithms and Applications
Algorithm 4 STING Algorithm 1: Determine a level to begin with. 2: For each cell of this level, we calculate the confidence interval (or estimated range) of probability that this cell is relevant to the query. 3: From the interval calculated above, we label the cell as relevant or not relevant. 4: If this level is the leaf level, go to Step 6; otherwise, go to Step 5. 5: We go down the hierarchy structure by one level. Go to Step 2 for those cells that form the relevant cells of the higher level. 6: If the specification of the query is met, go to Step 8; otherwise, go to Step 7. 7: Retrieve those data fall into the relevant cells and do further processing. Return the result that meet the requirement of the query. Go to Step 9. 8: Find the regions of relevant cells. Return those regions that meet the requirement of the query. Go to Step 9. 9: Stop.
and query by starting from the root of its hierarchical grid structure tree. The algorithm is summarized in Algorithm 4 [10, 28]. The tree can be constructed in O(N ) time, where N is the total number of data points. Dense cells are identified and clustered by examining the density of these cells in a similar vein to the density-based DBSCAN algorithm[7]. If the cell tree has K leaves, then the complexity of spatial querying and clustering for STING is O(K), which is O(N ) in the worst case since cells that would be empty never need to be materialized and stored in the tree. A common misconception is that K would be O(2d ) where d is the number of dimensions and that this would be problematic in high dimensions. STING may have problems with higher dimensional data common to all grid-based algorithms (e.g., handling noise and selecting most relevant attributes) [11], but scalability of the grid structure is not one of them. There are several advantages of STING. First, it is a query-independent approach since the statistical information exists independent of queries. The computational complexity of STING for clustering is O(K) and this is quite efficient in clustering large data sets especially when K ≪ N . The algorithm is readily parallelizable and allows for multiple resolutions for examining the data in its hierarchical grid structure. In addition, incremental data updating is supported, so there is lower overhead for incorporating new data points. Wang et al. extended STING to STING+ so that it is able to process dynamically evolving spatial databases. In addition, STING+ enables active data mining by supporting user-defined trigger conditions.
1.2.3
WaveCluster: Wavelets in Grid-based Clustering
Sheikholeslami et al. [27] proposed a grid-based and density-based clustering approach that uses wavelet transforms: WaveCluster. This algorithm applies wavelet transforms to the data points and then uses the transformed data to find clusters. A wavelet transform is a signal processing technique that decomposes a signal into different frequency subbands. The insight to using the wavelet transforms is that data points are considered d-dimensional signals where d is the number of dimensions. The high-frequency parts of a signal correspond to the more sparse data regions such as boundaries of the clusters whereas the low-frequency high-amplitude parts of a signal correspond to the more dense data regions such as cluster interiors [3]. By examining in different frequency subbands, clustering results may be achieved at different resolutions and scales from fine to coarse. Data are transformed to preserve relative distance between objects at different levels of resolution. A hat-shaped filter
Grid-based Clustering
9
Algorithm 5 WaveCluster Algorithm INPUT: Multidimensional data objects’ feature vectors OUTPUT: cluster objects 1: 2: 3: 4: 5: 6:
First bin the feature space, then assign objects to the units, and compute unit summaries. Apply wavelet transform on the feature space. Find connected components (clusters) in the subbands of transformed feature space, at multiple levels. Assign labels to the units in the connected components. Make the lookup table. Map the objects to the clusters.
is used to emphasize regions where points cluster and to suppress weaker information in their boundaries. This makes natural clusters more distinguishable and eliminates outliers simultaneously. As input the algorithm requires as parameters the number of grid cells for each dimension, the wavelet, and the number of applications of the wavelet transform. The algorithm is summarized in Algorithm 5 [27]. WaveCluster offers several advantages. The time complexity is O(N ) where N is the number of data points; this is very efficient for large spatial databases. The clustering results are insensitive to outliers and the data input order. The algorithm can accurately discern arbitrarily shaped clusters such as those with concavity and nesting. The wavelet transformation permits multiple levels of resolution, so that clusters may be detected more accurately. This algorithm is primarily only suited for low dimensional data. However, in the case of very high dimensional data, PCA may be applied to the data to reduce the number of dimensions, so that N > mf where m is the number of intervals in each dimension and f is the number of dimensions selected after PCA. After this, WaveCluster may be applied to the data to cluster it and still achieve linear time efficiency [27].
1.3
Adaptive Grid-based Algorithms
When a single inflexible, uniform grid is used, it may be difficult to achieve the desired clustering quality or efficiency for highly irregular data distributions. In such instances, adaptive algorithms that modify the uniform grid may be able to overcome this weakness in uniform grids. In this section, we introduce an adaptive grid-based clustering algorithm: AMR. Another adaptive algorithm (MAFIA) will be discussed in Section 1.5.2.2.
1.3.1
AMR: Adaptive Mesh Refinement Clustering
Liao et al. [14] proposed a grid-based clustering algorithm AMR using a Adaptive Mesh Refinement technique that applies higher resolution grids to the localized denser regions. Different from traditional grid-based clustering algorithms, such as CLIQUE and GRIDCLUS, which use a single resolution mesh grid, AMR divides the feature space at multiple resolutions. While STING also offers multiple resolutions, it does so over the entire space, not localized regions. AMR creates a hierarchical tree constructed from the grids at multiple resolutions. Using this tree, this algorithm can discern clusters, especially nested ones, that
10
Data Clustering: Algorithms and Applications
may be difficult to discover without clustering several levels of resolutions at once. AMR is very fit for data mining problems with highly irregular data distributions. AMR clustering algorithm mainly contains two steps summarized here from [13]: 1. Grid Construction: First grids are created at multiple resolutions based on regional density. The grid hierarchy tree contains nested grids of increasing resolution since the grid construction is done recursively. The construction of the AMR tree starts with a uniform grid covering the entire space and for those cells that exceed a density threshold, the grid is refined into higher resolution grids at each recursive step. The new child grids created as part of the refinement step are connected in the tree to parent grid cells whose density exceeds the threshold. 2. Clustering: To create clusters, each leaf node is considered to be the center of an individual cluster. The algorithm recursively assigns objects in the parent nodes to clusters until the root node is reached. Cells are assigned to clusters based on the minimum distance to the clusters under the tree branch. h
h
d 1−q The overall complexity for constructing the AMR tree is O(dtN 1−p 1−p + (dtk + 6 )r 1−q ), where N is the number of data points, d is the dimensionality, t is the number of attributes in each dimension, h is the AMR tree height, and p represents the average percentage of data points to be refined at each level, r is the mesh size at the root, and q is the average ratio of mesh sizes between two grid levels [14]. Like most grid-based methods, AMR exhibits insensitivity to the order of input data. The AMR clustering algorithm may be applied to to any collection of attributes with numerical values even those with very irregular or very concentrated data distributions. However, like GDILC, it cannot be scaled to high-dimensional databases because of its overall complexity.
1.4
Axis-shifting Grid-based Algorithms
The effectiveness of a grid-based clustering algorithm is seriously limited by the size of the predefined grids, the borders of the cells, and the density threshold of the significant cells in the face of local variations in shape and density in a data space. These challenges motivate another kind of grid-based algorithms: axis-shifting algorithms. In this section, we introduce four axis-shifting algorithms: NSGC, ADCC, ASGC, and GDILC.
1.4.1
NSGC: New Shifting Grid Clustering Algorithm
Fixed grids may suffer from the boundary effect. To alleviate this, Ma and Chow [18] proposed a New Shifting Grid Clustering algorithm (NSGC). NSGC is both density-based and grid-based. To form its grid structure, the algorithm divides each dimension of the space into an equal number of intervals. NSGC shifts the whole grid structure and uses the shifted grid along with the original grid to determine the density of cells. This reduces the influence of the size and borders of the cells. It then clusters the cells rather than the points. Specifically, NSGC consists of four main steps summarized in Algorithm 6 [18]. NSGC repeats the steps above until the result of the previous iteration and that of the current iteration are smaller than a specified accepted error threshold. The complexity of NSGC is O((2w)d ), where d is the dimensionality and w is the number of iterations of the algorithm. While it is claimed that this algorithm is non-parametric, its performance is dependent upon the choice of the number of iterations, w, and the accepted error threshold.
Grid-based Clustering
11
Algorithm 6 NSGC Algorithm 1: Cell construction: It divides each dimension of the space into 2w intervals, where w is the number of iterations. 2: Cell assignment: It first finds the data points belonging to a cell, then shifts by half cell-size of the corresponding dimension, and finds the data points belonging to shifted cells. 3: Cell density computation: It uses both the density of the cell itself and its nearest neighborhood to obtain a descriptive density profile. 4: Group assignment(clustering): It starts when the considered cell or one of its neighbor cells has no group assigned. Otherwise, the next cell is considered until all non-empty cells are assigned. Algorithm 7 ADCC Algorithm 1: Generate the first grid structure. 2: Identify the significant cells. 3: Generate the first set of clusters. 4: Transform the grid structure. 5: Generate the second set of clusters. 6: Revise the original clusters. In this case, the first set of clusters and second set of clusters are combined recursively. 7: Generate the final clustering result.
If w is set too low (or high) or the error threshold too high, then clustering results may not be accurate; there is no a priori way to know the best values of these parameters for specific data. NSGC is susceptible to errors caused by cell sizes that are too small also. As the size of cells decreases (and the number of iterations increases), the total number of cells and the number of clusters reported both increase. The reported clusters may be too small and not correspond to clusters in the original data. The strongest advantage of NSGC is that its grid shifting strategy permits it to recognize clusters of very arbitrary boundary shapes with great accuracy.
1.4.2
ADCC: Adaptable Deflect and Conquer Clustering
The clustering quality of grid-based clustering algorithms often depends on the size of the predefined grids and the density threshold. To reduce their influence, Lin et al. adopted “deflect and conquer techniques” to propose a new grid-based clustering algorithm ADCC (Adaptable Deflect and Conquer Clustering) [17]. Very similar to NSGC, the idea of ADCC is to utilize the predefined grids and predefined threshold to identify significant cells. Nearby cells that are also significant can be merged to develop a cluster. Next, the grids are deflected half a cell size in all directions and the significant cells are identified again. Finally, the newly generated significant cells and the initial set of significant cells are merged to improve the clustering of both phases. Specifically, ADCC is summarized in Algorithm 7. The overall complexity of ADCC is O(md + N ), where m is the number of intervals in each dimension, d is the dimensionality of data, and N is the number of data points. While ADCC is very similar to NSGC in its axis-shifting strategy, it is quite different in how it constructs clusters from the sets of grids. Rather than examining a neighborhood of the two grids at once as NSGC does, ADCC examines the two grids recursively looking for consensus in the significance of cells in both clusterings especially those that overlap a previous clustering to make a determination about the final clustering. This step can
12
Data Clustering: Algorithms and Applications
Algorithm 8 ASGC Algorithm 1: Generate the first grid structure: the entire feature space is divided into non overlapping cells thus forming the first grid structure. 2: Identify the significant cells: These are cells whose density is more than a predefined threshold. 3: Generate the first set of clusters: all neighboring significant cells are grouped together to form clusters. 4: Transform the grid structure: the original coordinate origin is shifted by distance ξ in each dimension of the feature space to obtain a new grid structure. 5: Generate the second set of clusters: new clusters are generated using steps 2 and 3. 6: Revise the original clusters: the clusters generated from the shifted grid structures can be used to revise the clusters generated from the original grid structure. 7: Generate the final clustering result.
actually help to separate clusters more effectively especially if there is only a small distance with very little data between them. Both methods are suspectible to errors caused by small cell sizes, but can for the most part handle arbitrary borders and shapes in clusters very well. ADCC is not dependent on many parameters to determine its termination. It is only dependent on the choice of the number of intervals per dimension, m.
1.4.3
ASGC: Axis-Shifted Grid-Clustering
Another attempt by Chang et al. [16] to minimize the impact of the size and borders of the cells is ASGC (Axis-Shifted Grid-Clustering) (also referred to as ACICA+ ). After creating an original grid structure and initial clustering from that grid structure, the original grid structure is shifted in each dimension and another clustering is done. The shifted grid structure can be translated an arbitrary distance to be specified. The effect of this is to implicitly change the size of the original cells. It also offers greater flexibility to adjust to boundaries of clusters in the original data and minimize the effect of the boundary cells. The clusters generated from this shifted grid structure can be used to revise the original clusters. Specifically, the ASGC algorithm involves 7 steps, and is summarized in Algorithm 8 from [13]. The complexity of ASGC is the same as that of ADCC, which is O(md + N ), where N is the number of data points, d is the dimensionality of the data, and m is the number of intervals in each dimension. The main difference between ADCC and ASGC is that the consensus method to revise clusters is bi-directional in ASGC: using the overlapping cells the clusters from the first phase can be used to modify the clusters of the second phase and vice versa. When a cluster of the first clustering overlaps a cluster of the second clustering, the combined cluster of union of both can then be modified in order to generate the final clustering. This permits great flexibility in handling arbitrary shapes of clusters in the original data and minimizes the extent to which either grid structure will separate clusters. By essentially translating the original grid structure an arbitrary distance to create the second grid and overlapping it with the original grid structure, a different resolution (and implicitly different cell size) is also achieved by this translation. While this method is less susceptible to the effects of cell sizes and cell density thresholds than other axis-shifting grid clustering methods, it still requires careful initial choice of cell size and cell density threshold.
Grid-based Clustering
13
Algorithm 9 GDILC Algorithm 1: Cells are initialized by dividing each dimension into m intervals. 2: The distances between sample points and those in neighboring cells are calculated. The distance threshold T is computed. 3: The density vector and density threshold τ are computed. 4: At first, GDILC takes each data point whose density is more than the density threshold τ as a cluster. Then, for each data point x, check, for every data point whose density is more than the density threshold τ in the neighbor cells of x, whether its distance to x is less than the distance threshold T . If so, GDILC then combines the two clusters containing those two data points. The algorithm continues until all point pairs have been checked. 5: Outliers are removed.
1.4.4
GDILC: Grid-based Density-IsoLine Clustering Algorithm
Zhao and Song [31] proposed a Grid-based Density-IsoLine Clustering algorithm (GDILC) to perform clustering by making use of the density-isoline figure. It assumes that all data samples have been normalized. All attributes are numerals and are in the range of [0, 1]. This is for the convenience of distance and density calculation. GDILC first implicitly calculates a density-isoline figure, the contour figure of the density of data points. Then clusters are discovered from the density-isoline figure. GDILC computes the density of a data point by counting the number of points in its neighbor region. Specifically, the density of a data point x is defined as follows: Density(x) = |{y : Dist(x, y) ≤ T }|,
(1.3)
where T is a given distance threshold and Dist(x, y) is a distance function (e.g., Euclidean distance) used to measure the dissimilarity between data points x and y. The densityisoline figure is never drawn, but is obtained from the density vectors. The density vectors are computed by counting the elements of each row of the distance matrix that are less than the radius of the neighbor region, T . To avoid enumerating all data points for calculating the density vector, GDILC employs a grid-based method. The grid-based method first partitions each dimension into several intervals creating hyper-rectangular cells. Then, to calculate the density of data point x, GDILC only considers data points in the same cell with x and those data points in its neighbor cells; this is identical to axis shifting. The GDILC algorithm is shown in Algorithm 9 [10]. For many data sets, this grid-based method significantly reduces the search space of calculating the point pair distances; the complexity may appear nearly linear. In the worst case the time complexity of GDILC remains O(N 2 ) (i.e., for the pathological case when all the points cluster in the neighborhood of a constant number of cells). However, it cannot be scaled to high-dimensional data because the space is divided into md cells, where m is the number of intervals in each dimension, and d is the dimensionality. When the dimensionality d is very large, md is significantly large, and the data points in each cell is very sparse, the GDILC algorithm will no longer work. (There will be difficulty computing any distances or thresholds.) There are two significant advantages to this algorithm. First, it can handle outliers explicitly and this can be refined as desired. Second, it computes necessary thresholds such as those for density and distance directly from the data. These can be fined tuned as needed (i.e., they don’t need to be guessed at any point). In essence this algorithm dynamically learns the data distribution of samples while learning the parameters for thresholds in addition to discerning the clustering in the data.
14
1.5
Data Clustering: Algorithms and Applications
High Dimensional Algorithms
The scalability of grid-based approaches is a significant problem in higher dimensional data because of the increase in the size of the grid structure and the resultant time complexity increase. Moreover, inherent issues in clustering high dimensional data such as filtering noise and identifying the most relevant attributes or dimensions that represent the most dense regions must be addressed inherently in the grid structure creation as well as the actual clustering algorithm. In this section, we examine carefully a subspace clustering approach presented in CLIQUE and a density estimation approach presented by OptiGrid. This section is greatly influenced by the survey of Berkhin [3] with additional insights on the complexity, strengths, and weaknesses of each algorithm presented.
1.5.1
CLIQUE: The Classical High-Dimensional Algorithm
Agrawal et al. [1] proposed a hybrid density-based, grid-based clustering algorithm, CLIQUE (Clustering In QUEst), to find automatically subspace clustering of high dimensional numerical data. It locates clusters embedded in subspaces of high dimensional data without much user intervention to discern significant subclusters. In order to present the clustering results in an easily interpretable format, each cluster is given a minimal description as a disjunctive normal form (DNF) expression. CLIQUE first partitions its numerical space into units for its grid structure. More specifically, let A = {A1 , A2 , ..., Ad } be a set of bounded, totally ordered domains (attributes) and S = A1 × A2 × ..., ×Ad be a d -dimensional numerical space. By partitioning every dimension Ai (1 ≤ i ≤ d) into m intervals of equal length, CLIQUE divides the d -dimensional data space into md non-overlapping rectangular units. A d -dimensional data point, v, is considered in a unit, u, if the value of v in each attribute, is greater than or equal to the left boundary of that attribute in u and less than the right boundary of that attribute in u. The selectivity of a unit is defined to be the fraction of total data points in the unit. Only units whose selectivity is greater than a parameter τ are viewed as dense and retained. The definition of dense units applies to all subspaces of the original d-dimensional space. To identify dense units to retain and subspaces that contain clusters, CLIQUE considers projections of the subspaces from the bottom up (i.e., the least dimensional subspaces to those of increasing dimension). Given a projection subspace, At1 ×At2 ×...×Atp , where p < d and ti < tj if i < j, a unit is the intersection of an interval in each dimension. By leveraging the Apriori algorithm, CLIQUE employs a bottom-up scheme because monotonicity holds: if a collection of points is a cluster in a p-dimensional space, then this collection of points is also part of a cluster in any (p − 1)-dimensional projections of this space. In CLIQUE, the recursive step from (p − 1)-dimensional units to p-dimensional units involves a self-join of the p−1 units sharing first common (p−2)-dimensions [3]. To reduce the time complexity of the Apriori process, CLIQUE prunes the pool of candidates, only keeping the set of dense units to form the candidate units in the next level of the dense unit generation algorithm. To prune the candidates, all the subspaces are sorted by their coverage, i.e., the fraction of the database that is covered by the dense units in it. The less covered subspaces are pruned. The cut point between retained and pruned subspaces is selected based on the MDL [24] principle in information theory. CLIQUE then forms clusters from the remaining candidate units. Two p-dimensional units u1 , u2 are connected if they have a common face or if there exists another p-dimensional unit us such that u1 is connected to us and u2 is connected to us . A cluster is a maximal set of connected dense units in p-dimensions. Finding clusters is equivalent to finding connected
Grid-based Clustering
15
components in the graph defined to represent the dense units as the vertices and edges between vertices existing if and only if the units share a common face. In the worst case, this can be done in quadratic time in the number of dense units. After finding all the clusters, CLIQUE uses a DNF expression to specify a finite set of maximal segments (regions) whose union is the cluster. Finding the minimal descriptions for the clusters is equivalent to finding an optimal cover of the clusters; this is NP-hard. In light of this, instead, CLIQUE adopts a greedy approach to cover the cluster in regions and then discards redundant regions. By integrating density-based, grid-based, and subspace clustering, CLIQUE discovers clusters embedded in subspaces of high dimensional data without requiring users to select subspaces of interest. The DNF expressions for the clusters give a clear representation of clustering results. The time complexity of CLIQUE is O(cp + pN ), where p is the highest subspace dimension selected, N is the number of input points, and c is a constant; this grows exponentially with respect to p. The algorithm offers an effective, efficient method of pruning the space of dense units in order to counter the inherent exponential nature of the problem. However, there is a trade-off for the pruning of dense units in the subspaces with low coverage. While the algorithm is faster, there is an increased likelihood of missing clusters. In addition, while CLIQUE does not require users to select subspaces of interest, its susceptibility to noise and ability to identify relevant attributes is highly dependent on the user’s choice of unit intervals, m, and sensitivity threshold, τ .
1.5.2
Variants of CLIQUE
There are two aspects of the CLIQUE algorithm that can be improved. The first one is the criterion for the subspace selection. The second is the size and resolution of the grid structure. The former is addressed by the ENCLUS algorithm by using entropy as subspace selection criterion. The latter is addressed by the MAFIA algorithm by using adaptive grids for fast subspace clustering. 1.5.2.1
ENCLUS: Entropy-based Approach
The algorithm ENCLUS (ENtropy-based CLUStering) [6] is an adaptation of the CLIQUE that uses a different, entropy-based criterion for subspace selection. Rather than using the fraction of total points in a subspace as a criterion to select subspaces, ENCLUS uses an entropy criteria and only those subspaces spanned by attributes A1 , ..., Ap with entropy H(A1 , ..., Ap ) < ϖ(a threshold) are selected for clustering. A low-entropy subspace corresponds to a more dense region of units. An analogous monotonicity condition or Apriori property also exists in terms of entropy. If a p-dimensional subspace has low entropy, then so does any (p − 1)-dimensional projections of this subspace: H(A1 , ..., Ap−1 ) = H(A1 , ..., Ap ) − H(Ap |A1 , ..., Ap−1 ) < ϖ.
(1.4)
A significant limitation of ENCLUS is its extremely high computational cost, especially in terms of computing the entropy of subspaces. However, this cost also yields the benefit that this approach has increased sensitivity to detect clusters especially extremely dense small ones. 1.5.2.2
MAFIA: Adaptive Grids in High Dimensions
MAFIA (Merging of Adaptive Finite IntervAls) proposed by Goil et al. [21] is a descendant of CLIQUE. Instead of using a fixed size cell grid structure with an equal number of bins in each dimension, MAFIA constructs adaptive grids to improve subspace clustering and also uses parallelism on a shared-nothing architecture to handle massive data sets.
16
Data Clustering: Algorithms and Applications
Algorithm 10 MAFIA Algorithm 1: Do one scan of the data to construct adaptive grids in each dimension. 2: Compute the histograms by reading blocks of data into memory using bins. 3: Using the histograms to merge bins into a smaller number of adaptive variable-size bins, where adjacent bins with similar histogram values are combined to form larger bins. The bins that have low density of data are pruned. 4: Select bins that are α-times (α is a parameter called the cluster dominance factor) more densely populated than average as p (p = 1 now) candidate dense units (CDUs). 5: Iteratively scan data for higher dimensions, and construct new p-CDU from two (p − 1)CDUs if they share any (p − 2)-face, and merge adjacent CDUs into clusters. 6: Generate minimal DNF expressions for each cluster
MAFIA proposes an adaptive grid of bins in each dimension. Then using an Apriori algorithm, dense intervals are merged to create clusters in the higher dimensional space. The adaptive grid is created by partitioning each dimension independently based on the distribution (i.e., the histogram) observed in that dimension, merging intervals that have the same observed distribution, and pruning those intervals with low density. This pruning during the construction of the adaptive grid reduces the overall computation of the clustering step. The steps of MAFIA are summarized from [3] in Algorithm 10. If p is the highest dimensionality of a candidate dense unit (CDU), N is the number of data points, and m is a constant, the algorithm’s complexity is O (mp +pN ), still exponential in the dimension as CLIQUE also is. However, the performance results on real data sets show that MAFIA is 40 to 50 times faster than CLIQUE because of the use of adaptive grids and their ability to select a smaller set of interesting CDUs [6]. Parallel MAFIA further offers the ability to obtain a highly scalable clustering for large data sets. Since the adaptive grid permits not only variable resolution because of the variable bin size, but also variable, adaptive grid boundaries, MAFIA yields with greater accuracy cluster boundaries that are very close to grid boundaries and are readily expressed as minimal DNF expressions.
1.5.3
OptiGrid: Density-based Optimal Grid Partitioning
Hinneburg and Keim proposed OptiGrid (OPTimal GRID-Clustering) [12] to address several aspects of the “curse of dimensionality”: noise, scalability of the grid construction, and selecting relevant attributes by optimizing the density function over the data space. OptiGrid uses density estimations to determine the centers of clusters as the clustering was done for the DENCLUE algorithm [11]. A cluster is a region of concentrated density centered around a strong density attractor or local maximum of the density function with density above the noise threshold. Clusters may also have multiple centers if the centers are strong density attractors and there exists a path between them above the noise threshold. By recursively partitioning the feature space into multidimensional grids, OptiGrid creates an optimal grid-partition by constructing the best cutting hyperplanes of the space. These cutting planes cut the space in areas of low density (i.e. local minima of the density function) and preserve areas of high density or clusters, specifically the cluster centers (i.e. local maxima of the density function). The cutting hyperplanes are found using a set of contracting linear projections of the feature space. The contracting projections create upper bounds for the density of the planes orthogonal to them. Namely, for any point, x, in a contracting projection, P , then for any point y such that P (y) = x, the density of y is at most the density of x. To define the grid more precisely, we present the definitions offered in [12] as summarized
Grid-based Clustering
17
in [10]. A cutting plane is a (d − 1)-dimensional hyperplane consisting of all points y that ∑d satisfy the equation i=1 wi yj = 1. The cutting plane partitions Rd into two half spaces. The decision function H(x) determines the half space, where a point x ∈ R is located: { H(x) =
1 0
∑d if i=1 wi xj ≥ 1, otherwise.
(1.5)
Then, a multi-dimensional grid G for the feature space S is defined to be a set H = {H1 , H2 , ..., Hk } of (d − 1)-dimensional cutting planes. The coding function cG : S → N is defined as: k ∑ x ∈ S, c(x) = 2i · Hi (x). (1.6) i=1
OptiGrid uses a density function to determine the best cutting places and to locate clusters. The density function fˆD is defined as 1 ∑ x − xi ), fˆD = KD( nh i=1 h n
(1.7)
where D is a set of N d-dimensional points, h is the smoothness level, and KD is the kernel density estimator. Clusters are defined as the maxima of the density function, which are above a certain noise level ξ. A center-based cluster for a maximum x∗ of the density function fˆD is the subset C ⊆ D, with x ∈ C being density-attracted by x∗ and fˆD (x∗ ) ≥ ξ. Points x ∈ D are called outliers if they are density-attracted by a local maximum x∗0 with fˆD (x∗0 ) < ξ. OptiGrid selects a set of contracting projections. These projections are then used to find the optimal cutting planes. The projections are useful because they concentrate the density of points and the cutting planes, in contrast, will have low density. Each cutting plane is selected to have minimal point density and to separate two dense half spaces. After each step of constructing a multi-dimensional grid defined by the best cutting planes, OptiGrid finds the clusters using the density function. The algorithm is then applied recursively to the clusters. In each round of recursion, OptiGrid only maintains data objects in the dense grids from the previous round of recursion. The algorithm OptiGrid is described in Algorithm 11 [10, 12]. The time complexity of OptiGrid is O(d · N · log N ), where N is the number of data points and d is the dimensionality of the data. This is very efficient for clustering large highdimensional databases. However, it may perform poorly in locating clusters embedded in a low-dimensional subspace of a very high-dimensional database, because its recursive method only reduces the dimensions by one at every step [10]. In addition, it suffers sensitivity to parameter choice and does not efficiently handle grid sizes that exceed available memory [20]. Moreover, OptiGrid requires very careful selection of the projections, density estimate, and determination of what constitutes a best or optimal cutting plane from users. The difficulty of this is only determined on a case by case basis on the data being studied. However, a special case of applying this algorithm can be considered a more efficient variant of CLIQUE and MAFIA. Namely, if the projections used are the projection maps in each dimension, the density estimate is uniform, and there are sufficient cutting planes to separate each density attractor on each dimension, then a more efficient and accurate clustering can be achieved that circumvents the difficulties of CLIQUE, i.e, the time complexity is no longer exponential in the dimensions and clusters are not missed.
18
Data Clustering: Algorithms and Applications
Algorithm 11 OptiGrid Algorithm INPUT: data set D, q, min cut score 1: Determine a set of contracting projections P = {P0 , P1 , ..., Pk } and calculate all the projections of the data set D : Pi (D), i = 1, 2, ..., k; 2: Initialize a list of cutting planes BEST CU T ⇐ Φ, CU T ⇐ Φ; 3: for i=0 to k do 4: CU T ⇐best local cuts Pi (D); 5: CU T SCORE ⇐Score best local cuts Pi (D); 6: Insert all the cutting planes with a score ≥ min cuts core into BEST CU T ; 7: if BEST CU T = Φ then 8: return D as a cluster; 9: else 10: Select the q cutting planes of the highest score from BEST CU T and construct a multidimensional grid G using the q cutting planes; 11: Insert all data points in D into G and determine the highly populated grid cells in G; add these cells to the set of clusters C; 12: Refine C; 13: for all clusters Ci in C do 14: Do the same process with data set Ci ; 15: end for 16: end if 17: end for
1.5.4
Variants of the OptiGrid Approach
In this section, we consider variants of OptiGrid that were introduced to address the issues of the scalabity of the grid structure, especially with respect to available memory, and a clear criterion for the selection of cutting planes. 1.5.4.1
O-Cluster: A Scalable Approach
Milenova et al. proposed a O-cluster (Orthogonal partitioning CLUSTERing) [20] to address three limitations of OptiGrid: scalability in terms of data relative to memory size, lack of clear criterion to determine if a cutting plane is optimal or not, and sensitivity to threshold parameters for noise and cut plane density. O-Clusters addresses the first limitation by using a random sampling technique on the original data and a small buffer size. Only partitions that are not resolved (i.e., ambiguous) have data points maintained in the buffer. As a variant of OptiGrid, O-Cluster uses an axis-parallel partitioning strategy to locate high density areas in the data. To do so, O-Cluster uses contracting projections, but also proposes the use of a statistical test to validate the quality of a cutting plane. The statistical test checks for statistical significance between the difference in the density of the peaks and a valley when the valley separates the two peaks using a standard χ2 test. If statistical significance is found, the cutting plane would then be through such a valley. O-Cluster is also a recursive method. After testing the splitting points for all projections in a partition, the optimal one is chosen to partition the data. The algorithm then searches for cutting planes in the new partitions. A hierarchical tree structure is used to divide the data into rectangular regions. The main steps are summarized in Algorithm 12. The time complexity of O-Cluster is approximated to be O(N d), where N is the number of data points and d is the number of dimensions. However, Hinneburg and Keim claim a superlinear lower bound for the time complexity of clustering high dimensional data with noise [12]. Their proof sketch addresses the issue that given an O(N ) amount of noise in data
Grid-based Clustering
19
Algorithm 12 O-cluster Algorithm 1: Load data buffer. 2: Compute histograms for active partitions. 3: Find “best” splits for active partitions. 4: Flag ambiguous and “frozen” partitions. 5: Split active partitions. 6: Reload buffer.
that has been read and must be searched, there is not a constant time way to do that even for random noise. Moreover, the time complexity of the OptiGrid algorithm is dominated by the insertion time into the grid structure. They assert that the insertion time for axis parallel planes is O(N qI) where q is the number of planes and I is the insertion time and that the insertion time is the minimum of q and log N . Since the O-Cluster algorithm uses a binary clustering tree to load the buffer, it is possible that its running time is dominated by this and is O(N s) where s is the depth of the binary clustering tree, but clear worst case analysis of that depth was not offered. However, empirically O-Cluster has shown good scalability results even using a small buffer size. While O-Cluster handles uniform noise well, its performance degrades with increasing noise. O-Cluster, unlike OptiGrid, also has a tendency to split clusters because it tends to oversplit the data space and uses histograms that do not sufficiently smooth the distribution density. However, like OptiGrid, it may have difficulty with clusters embedded in a lowdimensional subspace of a very high-dimensional feature space because of its recursion step [10]. 1.5.4.2
CBF: Cell-based Filtering
CBF (Cell-Based Filtering) [4] proposed by Chang and Jin focuses on the scalability of the grid structure, handling large data sets in memory, and the efficiency of insertion and retrieval of clusters from the grid structure. It also offers a clear criteria for a cutting plane. CBF creates its grid structure by splitting each dimension into a set of partitions using a split index. Once a dimension is split into classes, the split index is equal to the sum of the relative density squared of each class. CBF finds the optimal split section in each dimension by repeatedly examining the the split value index for partitions along the potential split points in the dimension until the maximum value of a split index for a partitioning is found. Cells are created from the overlapping regions of the partitions in each dimension. This cell creation algorithm results in many fewer cells in the grid structure than other high dimensional grid based algorithms we have discussed. Cells with higher density than a given threshold are inserted into the clusters in the index structure. There is a second layer of the index structure built upon the cluster information file that provides approximation information of the clusters by grouping them into sections. For sections of clusters, if the density of the section is greater than a threshold, then this secondary index is set to true for that section and false otherwise. By using this approximation technique, the index structure is able to filter queries by first examining the section information and then accessing the cell and cluster information as necessary. This filtering-based index structure achieves good retrieval performance on clusters by minimizing the I/O access to acluster information file. While CBF is very efficient in its cell grid creation and query retrieval time, the tradeoff for such efficiency is lower precision for clustering.
20
1.6
Data Clustering: Algorithms and Applications
Conclusions and Summary
The efficiency of grid-based clustering algorithms comes from how data points are grouped into cells and clustered collectively rather than individually. This results in drastic time complexity improvements because often data is grouped into far few cells than there are data points. The general approach of these algorithms is to divide the data space into grid data structures, summarize each cell in the grids by a statistic such as density, and then cluster the grid cells. The greatest challenge for using these algorithms is then determining the best strategy of constructing the grid structure. The size of the grid structure and time for its construction largely determines the time complexity of grid-based algorithms. The size of the grid structure is directly influenced by the size of the cell and the size of the cell determines the resolution at which data may be clustered. The resolution of the grid determines the clustering accuracy and the ability to recognize diverse data cluster boundaries. We have presented classical grid-based algorithms that use uniform grids and several variant classes that deal with specific challenges of using uniform grid structures. Adaptive algorithms permit grid structures that offer finer resolutions over some regions of the data space in order to deal with data that has highly irregular or concentrated data distributions. Axis shifting algorithms deal with local variations in data shape clusters and density of the data by translating the original grid structure across the data space. Grid-based algorithms for high dimensional data deal with a problem that is inherently exponential in nature; constructing a grid structure that partitions the data in each dimension will create one that is exponential in the number of dimensions. These algorithms choose ways to filter the grid structure to only investigate relevant subspaces or to filter the original subspace in order to only select the most relevant attributes and filter noise. The time complexities of several grid-based algorithms discussed in this chapter are summarized in Table 1.4. Figure 1.6 summarizes the grid-based algorithms we have discussed in this chapter and their relationships. CGA is a low dimensional grid-based clustering algorithm that can be applied to high dimensional data after PCA is applied to the data [32]. GCHL is a variant of GRIDCLUS that is suitable for high dimensional data [23]. Some current research on grid-based clustering focuses on parallel implementation of grid-based clustering such as PGMCLU[30] and GridDBSCAN [19], grid-based clustering of fuzzy query results[2], and domain-specific applications such as grid-based clustering on data streams such as SGCS, D-Stream , and DUCstream [22, 5, 8].
Grid-based Clustering
21
FIGURE 1.3: Relationships among grid-based algorithms
TABLE 1.4: Time complexity of grid based algorithms Algorithm Name Time Complexity STING O(K), K is the number of cells at bottom layer WaveCluster O(N ), N is the number of data objects d NSGC O((2w) ), d is the dimensionality, w is number of iterations ADCC,ASGC O(md ) + O(N ), m is the number of intervals in each dimension GDILC O(N ) p CLIQUE,MAFIA O(c + pN ), p is the highest subspace dimension selected, c is a constant OptiGrid O(d · N · log N ) O-Cluster Ω(N · d)
22
Data Clustering: Algorithms and Applications
Bibliography
[1] Rakesh Agrawal, Johannes Gehrke, Dimitrios Gunopulos, and Prabhakar Raghavan. Automatic subspace clustering of high dimensional data for data mining applications. In Proceedings of the 1998 ACM SIGMOD international conference on Management of data, SIGMOD ’98, pages 94–105, New York, NY, USA, 1998. ACM. [2] Mounir Bechchi, Amenel Voglozin, Guillaume Raschia, and Noureddine Mouaddib. Multi-Dimensional Grid-Based Clustering of Fuzzy Query Results. Rapport de recherche, INRIA, 2008. [3] Pavel Berkhin. Survey of clustering data mining techniques. Technical report, Accrue Software, Inc., San Jose, CA, 2002. [4] Jae-Woo Chang and Du-Seok Jin. A new cell-based clustering method for large, highdimensional data in data mining applications. In Proceedings of the 2002 ACM symposium on Applied computing, SAC ’02, pages 503–507. ACM, 2002. [5] Yixin Chen and Li Tu. Density-based clustering for real-time stream data. In Proceedings of the 13th ACM SIGKDD international conference on Knowledge discovery and data mining, KDD ’07, pages 133–142, 2007. [6] Chun-Hung Cheng. ENCLUS: Entropy-based Subspace Clustering for Mining Numerical Data. 1999. [7] Martin Ester, Hans-Peter Kriegel, Jorg Sander, and Xiaowei Xu. A density-based algorithm for discovering clusters in large spatial databases with noise. In KDD’96, pages 226–231, 1996. [8] Jing Gao, Jianzhong Li, Zhaogong Zhang, and Pang-Ning Tan. An incremental data stream clustering algorithm based on dense units detection. In Proceedings of the 9th Pacific-Asia conference on Advances in Knowledge Discovery and Data Mining, PAKDD’05, pages 420–425, Berlin, Heidelberg, 2005. Springer-Verlag. [9] Peter Grabusts and Arkady Borisov. Using grid-clustering methods in data classification, 2002. [10] Chaoqun Ma Guojun Gan and Jianhong Wu. Data Clustering: Theory, Algorithms, and Applications. SIAM, 2007. [11] Alexander Hinneburg, Er Hinneburg, and Daniel A. Keim. An efficient approach to clustering in large multimedia databases with noise. In Proc. 4rd Int. Conf. on Knowledge Discovery and Data Mining, pages 58–65. AAAI Press, 1998. [12] Alexander Hinneburg and Daniel A. Keim. Optimal grid-clustering: Towards breaking the curse of dimensionality in high-dimensional clustering. In VLDB’99, pages 506–517. Morgan Kaufmann, 1999. 23
24
Data Clustering: Algorithms and Applications
[13] MR ILANGO and Dr V MOHAN. A survey of grid based clustering algorithms. In International Journal of Engineering Science and Technology, pages 3441–3446, 2010. [14] Wei keng Liao, Ying Liu, and Alok Choudhary. A grid-based clustering algorithm using adaptive mesh refinement, 2004. [15] Hinrichs Klaus Helmer. The grid file system: implementation and case studies of applications. Technical report, ETH Zurich, 1985. [16] Nancy P. Lin, Chung-I Chang, Nien-yi Jan, Hao-En Chueh, Hung-Jen Chen, and WeiHua Hao. An axis-shifted crossover-imaged clustering algorithm. WTOS, 7(3):175–184, March 2008. [17] Nancy P. Lin, Chung-I Chang, and Chao-Lung Pan. An adaptable deflect and conquer clustering algorithm. In Proceedings of the 6th Conference on WSEAS International Conference on Applied Computer Science - Volume 6, ACOS’07, pages 155–159, 2007. [18] Eden W.M. Ma and Tommy W.S. Chow. A new shifting grid clustering algorithm. Pattern Recognition, 37(3):503–514, 2004. [19] S. Mahran and K. Mahar. Using grid for accelerating density-based clustering. In 8th IEEE International Conference on Computer and Information Technology, pages 35–40, 2008. [20] Boriana L. Milenova and Marcos M. Campos. O-cluster: scalable clustering of large high dimensional data sets. In Proceedings from the IEEE International Conference on Data Mining, pages 290–297, 2002. [21] H. Nagesh, S. Goil, and A. Choudhary. MAFIA: Efficient and scalable subspace clustering for very large data sets, 1999. [22] Nam Hun Park and Won Suk Lee. Statistical grid-based clustering over data streams. SIGMOD Rec., 33(1):32–37, March 2004. [23] Abdol Hamid Pilevar and M. Sukumar. Gchl: A grid-clustering algorit hm for highdimensional very large spatial data bases. Pattern Recognition Letters, pages 999–1010, 2005. [24] Jorma Rissanen. Stochastic Complexity in Statistical Inquiry Theory. World Scientific Publishing Co., Inc., River Edge, NJ, USA, 1989. [25] E. Schikuta. Grid-clustering: an efficient hierarchical clustering method for very large data sets. Proceedings of the 13th International Conference on Pattern Recognition, 2:101–105 vol.2, 1996. [26] Erich Schikuta and Martin Erhart. The bang-clustering system: Grid-based data analysis. In Proc. Sec. Int. Symp. IDA-97, pages 513–524. Springer-Verlag, 1997. [27] Gholamhosein Sheikholeslami, Surojit Chatterjee, and Aidong Zhang. Wavecluster: A multi-resolution clustering approach for very large spatial databases. In VLDB’98, pages 428–439, 1998. [28] Wei Wang, Jiong Yang, and Richard R. Muntz. Sting: A statistical information grid approach to spatial data mining. In VLDB’97, Proceedings of 23rd International Conference on Very Large Data Bases, August 25-29, 1997, Athens, Greece, pages 186–195. Morgan Kaufmann, 1997.
Grid-based Clustering
25
[29] C. S. Warnekar and G. Krishna. A heuristic clustering algorithm using union of overlapping pattern-cells. Pattern Recognition, 11(2):85–93, 1979. [30] Chen Xiaoyun, Chen Yi, Qi Xiaoli, Yue Min, and He Yanshan. Pgmclu: A novel parallel grid-based clustering algorithm for multi-density datasets. In SWS ’09: 1st IEEE Symposium on Web Society, pages 166–171, 2009. [31] Zhao Yanchang and Song Junde. Gdilc: a grid-based density-isoline clustering algorithm. In International Conferences on Info-tech and Info-net, volume 3, pages 140 –145 vol.3, 2001. [32] Zhiwen Yu and Hau-San Wong. Gca: A real-time grid-based clustering algorithm for large data set. In ICPR (2), pages 740–743. IEEE Computer Society, 2006.