Sunita Sarawagi

Interactive deduplication using active learning

Deduplication is a key operation in integrating data from multiple sources. The main challenge in this task is designing a function that can resolve when a pair of records refer to the same entity in spite of various data inconsistencies. Most existing systems use hand-coded functions. One way to overcome the tedium of hand-coding is to train a classifier to distinguish between duplicates and non-duplicates. The success of this method critically hinges on being able to provide a covering and challenging set of training pairs that bring out the subtlety of deduplication function. This is non-trivial because it requires manually searching for various data inconsistencies between any two records spread apart in large lists.We present our design of a learning-based deduplication system that uses a novel method of interactively discovering challenging training pairs using active learning. Our experiments on real-life datasets show that active learning significantly reduces the number of instances needed to achieve high accuracy. We investigate various design issues that arise in building a system to provide interactive response, fast convergence, and interpretable output.

Modeling multidimensional databases

The authors propose a data model and a few algebraic operations that provide semantic foundation to multidimensional databases. The distinguishing feature of the proposed model is the symmetric treatment not only of all dimensions but also measures. The model provides support for multiple hierarchies along each dimension and support for ad hoc aggregates. The proposed operators are composable, reorderable, and closed in application. These operators are also minimal in the sense that none can be expressed in terms of others nor can any one be dropped without sacrificing functionality. They make possible the declarative specification and optimization of multidimensional database queries that are currently specified operationally. The operators have been designed to be translated to SQL and can be implemented either on top of a relational database system or within a special purpose multidimensional database engine. In effect, they provide an algebraic application programming interface (API) that allows the separation of the front end from the back end. Finally, the proposed model provides a framework in which to study multidimensional databases and opens several new research problems.

Discovery-Driven Exploration of OLAP Data Cubes

A method for locating data anomalies in a k dimensional data cube that includes the steps of associating a surprise value with each cell of a data cube, and indicating a data anomaly when the surprise value associated with a cell exceeds a predetermined exception threshold. According to one aspect of the invention, the surprise value associated with each cell is a composite value that is based on at least one of a Self-Exp value for the cell, an In-Exp value for the cell and a Path-Exp value for the cell. Preferably, the step of associating the surprise value with each cell includes the steps of determining a Self-Exp value for the cell, determining an In-Exp value for the cell, determining a Path-Exp value for the cell, and then generating the surprise value for the cell based on the Self-Exp value, the In-Exp value and the Path-value.

Discriminative Methods for Multi-labeled Classification

In this paper we present methods of enhancing existing discriminative classifiers for multi-labeled predictions. Discriminative methods like support vector machines perform very well for uni-labeled text classification tasks. Multi-labeled classification is a harder task subject to relatively less attention. In the multi-labeled setting, classes are often related to each other or part of a is-a hierarchy. We present a new technique for combining text features and features indicating relationships between classes, which can be used with any discriminative algorithm. We also present two enhancements to the margin of SVMs for building better models in the presence of overlapping classes. We present results of experiments on real world text benchmark datasets. Our new methods beat accuracy of existing methods with statistically significant improvements.

Integrating association rule mining with relational database systems: alternatives and implications

Data mining on large data warehouses is becoming increasingly important. In support of this trend, we consider a spectrum of architectural alternatives for coupling mining with database systems. These alternatives include: loose-coupling through a SQL cursor interface; encapsulation of a mining algorithm in a stored procedure; caching the data to a file system on-the-fly and mining; tight-coupling using primarily user-defined functions; and SQL implementations for processing in the DBMS. We comprehensively study the option of expressing the mining algorithm in the form of SQL queries using Association rule mining as a case in point. We consider four options in SQL-92 and six options in SQL enhanced with object-relational extensions (SQL-OR). Our evaluation of the different architectural alternatives shows that from a performance perspective, the Cache-Mine option is superior, although the performance of the SQL-OR option is within a factor of two. Both the Cache-Mine and the SQL-OR approaches incur a higher storage penalty than the loose-coupling approach which performance-wise is a factor of 3 to 4 worse than Cache-Mine. The SQL-92 implementations were too slow to qualify as a competitive option. We also compare these alternatives on the basis of qualitative factors like automatic parallelization, development ease, portability and inter-operability.

Efficient set joins on similarity predicates

In this paper we present an efficient, scalable and general algorithm for performing set joins on predicates involving various similarity measures like intersect size, Jaccard-coefficient, cosine similarity, and edit-distance. This expands the existing suite of algorithms for set joins on simpler predicates such as, set containment, equality and non-zero overlap. We start with a basic inverted index based probing method and add a sequence of optimizations that result in one to two orders of magnitude improvement in running time. The algorithm folds in a data partitioning strategy that can work efficiently with an index compressed to fit in any available amount of main memory. The optimizations used in our algorithm generalize to several weighted and unweighted measures of partial word overlap between sets.

Record linkage: similarity measures and algorithms

This tutorial provides a comprehensive and cohesive overview of the key research results in the area of record linkage methodologies and algorithms for identifying approximate duplicate records, and available tools for this purpose. It encompasses techniques introduced in several communities including databases, information retrieval, statistics and machine learning. It aims to identify similarities and differences across the techniques as well as their merits and limitations.

Exploiting dictionaries in named entity extraction: combining semi-Markov extraction processes and data integration methods

We consider the problem of improving named entity recognition (NER) systems by using external dictionaries---more specifically, the problem of extending state-of-the-art NER systems by incorporating information about the similarity of extracted entities to entities in an external dictionary. This is difficult because most high-performance named entity recognition systems operate by sequentially classifying words as to whether or not they participate in an entity name; however, the most useful similarity measures score entire candidate names. To correct this mismatch we formalize a semi-Markov extraction process, which is based on sequentially classifying segments of several adjacent words, rather than single words. In addition to allowing a natural way of coupling high-performance NER methods and high-performance similarity functions, this formalism also allows the direct use of other useful entity-level features, and provides a more natural formulation of the NER problem than sequential word classification. Experiments in multiple domains show that the new model can substantially improve extraction performance over previous methods for using external dictionaries in NER.

Automatic segmentation of text into structured records

In this paper we present a method for automatically segmenting unformatted text records into structured elements. Several useful data sources today are human-generated as continuous text whereas convenient usage requires the data to be organized as structured records. A prime motivation is the warehouse address cleaning problem of transforming dirty addresses stored in large corporate databases as a single text field into subfields like “City” and “Street”. Existing tools rely on hand-tuned, domain-specific rule-based systems. We describe a tool DATAMOLD that learns to automatically extract structure when seeded with a small number of training examples. The tool enhances on Hidden Markov Models (HMM) to build a powerful probabilistic model that corroborates multiple sources of information including, the sequence of elements, their length distribution, distinguishing words from the vocabulary and an optional external data dictionary. Experiments on real-life datasets yielded accuracy of 90% on Asian addresses and 99% on US addresses. In contrast, existing information extraction methods based on rule-learning techniques yielded considerably lower accuracy.

Functional Sites in Protein Families Uncovered via an Objective and Automated Graph Theoretic Approach

We report a method for detection of recurring side-chain patterns (DRESPAT) using an unbiased and automated graph theoretic approach. We first list all structural patterns as sub-graphs where the protein is represented as a graph. The patterns from proteins are compared pair-wise to detect patterns common to a protein pair based on content and geometry criteria. The recurring pattern is then detected using an automated search algorithm from the all-against-all pair-wise comparison data of proteins. Intra-protein pattern comparison data are used to enable detection of patterns recurring within a protein. A method has been proposed for empirical calculation of statistical significance of recurring pattern. The method was tested on 17 protein sets of varying size, composed of non-redundant representatives from SCOP superfamilies. Recurring patterns in serine proteases, cysteine proteases, lipases, cupredoxin, ferredoxin, ferritin, cytochrome c, aspartoyl proteases, peroxidases, phospholipase A2, endonuclease, SH3 domain, EF-hand and lectins show additional residues conserved in the vicinity of the known functional sites. On the basis of the recurring patterns in ferritin, EF-hand and lectins, we could separate proteins or domains that are structurally similar yet different in metal ion-binding characteristics. In addition, novel recurring patterns were observed in glutathione-S-transferase, phospholipase A2 and ferredoxin with potential structural/functional roles. The results are discussed in relation to the known functional sites in each family. Between 2000 and 50,000 patterns were enumerated from each protein with between ten and 500 patterns detected as common to an evolutionarily related protein pair. Our results show that unbiased extraction of functional site pattern is not feasible from an evolutionarily related protein pair but is feasible from protein sets comprising five or more proteins. The DRESPAT method does not require a user-defined pattern, size or location of the pattern and therefore, has the potential to uncover new functional sites in protein families.