Kyuseok Shim

CURE: an efficient clustering algorithm for large databases

Clustering, in data mining, is useful for discovering groups and identifying interesting distributions in the underlying data. Traditional clustering algorithms either favor clusters with spherical shapes and similar sizes, or are very fragile in the presence of outliers. We propose a new clustering algorithm called CURE that is more robust to outliers, and identifies clusters having non-spherical shapes and wide variances in size. CURE achieves this by representing each cluster by a certain fixed number of points that are generated by selecting well scattered points from the cluster and then shrinking them toward the center of the cluster by a specified fraction. Having more than one representative point per cluster allows CURE to adjust well to the geometry of non-spherical shapes and the shrinking helps to dampen the effects of outliers. To handle large databases, CURE employs a combination of random sampling and partitioning. A random sample drawn from the data set is first partitioned and each partition is partially clustered. The partial clusters are then clustered in a second pass to yield the desired clusters. Our experimental results confirm that the quality of clusters produced by CURE is much better than those found by existing algorithms. Furthermore, they demonstrate that random sampling and partitioning enable CURE to not only outperform existing algorithms but also to scale well for large databases without sacrificing clustering quality.

Efficient algorithms for mining outliers from large data sets

In this paper, we propose a novel formulation for distance-based outliers that is based on the distance of a point from its kth nearest neighbor. We rank each point on the basis of its distance to its kth nearest neighbor and declare the top n points in this ranking to be outliers. In addition to developing relatively straightforward solutions to finding such outliers based on the classical nested-loop join and index join algorithms, we develop a highly efficient partition-based algorithm for mining outliers. This algorithm first partitions the input data set into disjoint subsets, and then prunes entire partitions as soon as it is determined that they cannot contain outliers. This results in substantial savings in computation. We present the results of an extensive experimental study on real-life and synthetic data sets. The results from a real-life NBA database highlight and reveal several expected and unexpected aspects of the database. The results from a study on synthetic data sets demonstrate that the partition-based algorithm scales well with respect to both data set size and data set dimensionality.

ROCK: a robust clustering algorithm for categorical attributes

Abstract Clustering, in data mining, is useful to discover distribution patterns in the underlying data. Clustering algorithms usually employ a distance metric based (e.g., euclidean) similarity measure in order to partition the database such that data points in the same partition are more similar than points in different partitions. In this paper, we study clustering algorithms for data with boolean and categorical attributes. We show that traditional clustering algorithms that use distances between points for clustering are not appropriate for boolean and categorical attributes. Instead, we propose a novel concept of links to measure the similarity/proximity between a pair of data points. We develop a robust hierarchical clustering algorithm ROCK that employs links and not distances when merging clusters. Our methods naturally extend to non-metric similarity measures that are relevant in situations where a domain expert/similarity table is the only source of knowledge. In addition to presenting detailed complexity results for ROCK, we also conduct an experimental study with real-life as well as synthetic data sets to demonstrate the effectiveness of our techniques. For data with categorical attributes, our findings indicate that ROCK not only generates better quality clusters than traditional algorithms, but it also exhibits good scalability properties.

Approximate query processing using wavelets

Abstract. Approximate query processing has emerged as a cost-effective approach for dealing with the huge data volumes and stringent response-time requirements of todays decision support systems (DSS). Most work in this area, however, has so far been limited in its query processing scope, typically focusing on specific forms of aggregate queries. Furthermore, conventional approaches based on sampling or histograms appear to be inherently limited when it comes to approximating the results of complex queries over high-dimensional DSS data sets. In this paper, we propose the use of multi-dimensional wavelets as an effective tool for general-purpose approximate query processing in modern, high-dimensional applications. Our approach is based on building wavelet-coefficient synopses of the data and using these synopses to provide approximate answers to queries. We develop novel query processing algorithms that operate directly on the wavelet-coefficient synopses of relational tables, allowing us to process arbitrarily complex queries entirely in the wavelet-coefficient domain. This guarantees extremely fast response times since our approximate query execution engine can do the bulk of its processing over compact sets of wavelet coefficients, essentially postponing the expansion into relational tuples until the end-result of the query. We also propose a novel wavelet decomposition algorithm that can build these synopses in an I/O-efficient manner. Finally, we conduct an extensive experimental study with synthetic as well as real-life data sets to determine the effectiveness of our wavelet-based approach compared to sampling and histograms. Our results demonstrate that our techniques: (1) provide approximate answers of better quality than either sampling or histograms; (2) offer query execution-time speedups of more than two orders of magnitude; and (3) guarantee extremely fast synopsis construction times that scale linearly with the size of the data.

Cure: an efficient clustering algorithm for large databases

Abstract Clustering, in data mining, is useful for discovering groups and identifying interesting distributions in the underlying data. Traditional clustering algorithms either favor clusters with spherical shapes and similar sizes, or are very fragile in the presence of outliers. We propose a new clustering algorithm called CURE that is more robust to outliers, and identifies clusters having non-spherical shapes and wide variances in size. CURE achieves this by representing each cluster by a certain fixed number of points that are generated by selecting well scattered points from the cluster and then shrinking them toward the center of the cluster by a specified fraction. Having more than one representative point per cluster allows CURE to adjust well to the geometry of non-spherical shapes and the shrinking helps to dampen the effects of outliers. To handle large databases, CURE employs a combination of random sampling and partitioning . A random sample drawn from the data set is first partitioned and each partition is partially clustered. The partial clusters are then clustered in a second pass to yield the desired clusters. Our experimental results confirm that the quality of clusters produced by CURE is much better than those found by existing algorithms. Furthermore, they demonstrate that random sampling and partitioning enable CURE to not only outperform existing algorithms but also to scale well for large databases without sacrificing clustering quality.

Data-streams and histograms

Histograms have been used widely to capture data distribution, to represent the data by a small number of step functions. Dynamic programming algorithms which provide optimal construction of these histograms exist, albeit running in quadratic time and linear space. In this paper we provide linear time construction of 1 + ε approximation of optimal histograms, running in polylogarithmic space. Our results extend to the context of data-streams, and in fact generalize to give 1 + ε approximation of several problems in data-streams which require partitioning the index set into intervals. The only assumptions required are that the cost of an interval is monotonic under inclusion (larger interval has larger cost) and that the cost can be computed or approximated in small space. This exhibits a nice class of problems for which we can have near optimal data-stream algorithms.

XTRACT: a system for extracting document type descriptors from XML documents

XML is rapidly emerging as the new standard for data representation and exchange on the Web. An XML document can be accompanied by a Document Type Descriptor (DTD) which plays the role of a schema for an XML data collection. DTDs contain valuable information on the structure of documents and thus have a crucial role in the efficient storage of XML data, as well as the effective formulation and optimization of XML queries. In this paper, we propose XTRACT, a novel system for inferring a DTD schema for a database of XML documents. Since the DTD syntax incorporates the full expressive power of regular expressions, naive approaches typically fail to produce concise and intuitive DTDs. Instead, the XTRACT inference algorithms employ a sequence of sophisticated steps that involve: (1) finding patterns in the input sequences and replacing them with regular expressions to generate “general” candidate DTDs, (2) factoring candidate DTDs using adaptations of algorithms from the logic optimization literature, and (3) applying the Minimum Description Length (MDL) principle to find the best DTD among the candidates. The results of our experiments with real-life and synthetic DTDs demonstrate the effectiveness of XTRACTs approach in inferring concise and semantically meaningful DTD schemas for XML databases.

Optimization of queries with user-defined predicates

Relational databases provide the ability to store user-defined functions and predicates which can be invoked in SQL queries. When evaluation of a user-defined predicate is relatively expensive, the traditional method of evaluating predicates as early as possible is no longer a sound heuristic. There are two previous approaches for optimizing such queries. However, neither is able to guarantee the optimal plan over the desired execution space. We present efficient techniques that are able to guarantee the choice of an optimal plan over the desired execution space. The optimization algorithm with complete rank-ordering improves upon the naive optimization algorithm by exploiting the nature of the cost formulas for join methods and is polynomial in the number of user-defined predicates (for a given number of relations.) We also propose pruning rules that significantly reduce the cost of searching the execution space for both the naive algorithm as well as for the optimization algorithm with complete rank-ordering, without compromising optimality. We also propose a conservative local heuristic that is simpler and has low optimization overhead. Although it is not always guaranteed to find the optimal plans, it produces close to optimal plans in most cases. We discuss how, depending on application requirements, to determine the algorithm of choice. It should be emphasized that our optimization algorithms handle user-defined selections as well as user-defined join predicates uniformly. We present complexity analysis and experimental comparison of the algorithms.

PUBLIC: A Decision Tree Classifier that Integrates Building and Pruning

Classification is an important problem in data mining. Given a database of records, each with a class label, a classifier generates a concise and meaningful description for each class that can be used to classify subsequent records. A number of popular classifiers construct decision trees to generate class models. These classifiers first build a decision tree and then prune subtrees from the decision tree in a subsequent pruning phase to improve accuracy and prevent “overfitting”.Generating the decision tree in two distinct phases could result in a substantial amount of wasted effort since an entire subtree constructed in the first phase may later be pruned in the next phase. In this paper, we propose PUBLIC, an improved decision tree classifier that integrates the second “pruning” phase with the initial “building” phase. In PUBLIC, a node is not expanded during the building phase, if it is determined that it will be pruned during the subsequent pruning phase. In order to make this determination for a node, before it is expanded, PUBLIC computes a lower bound on the minimum cost subtree rooted at the node. This estimate is then used by PUBLIC to identify the nodes that are certain to be pruned, and for such nodes, not expend effort on splitting them. Experimental results with real-life as well as synthetic data sets demonstrate the effectiveness of PUBLICs integrated approach which has the ability to deliver substantial performance improvements.

Parametric Query Optimization

Abstract. In most database systems, the values of many important run-time parameters of the system, the data, or the query are unknown at query optimization time. Parametric query optimization attempts to identify at compile time several execution plans, each one of which is optimal for a subset of all possible values of the run-time parameters. The goal is that at run time, when the actual parameter values are known, the appropriate plan should be identifiable with essentially no overhead. We present a general formulation of this problem and study it primarily for the buffer size parameter. We adopt randomized algorithms as the main approach to this style of optimization and enhance them with a sideways information passing feature that increases their effectiveness in the new task. Experimental results of these enhanced algorithms show that they optimize queries for large numbers of buffer sizes in the same time needed by their conventional versions for a single buffer size, without much sacrifice in the output quality and with essentially zero run-time overhead.