Machine Learning and Knowledge Discovery in Databases
Knowledge discovery in databases is not only the nontrivial extraction of implicit, previously unknown and potentially useful information from databases. We argue that in contrast to machine learning, knowledge discovery in databases should be applied to real world databases. Since real world databases are known to be very large, they raise problems of the access. Therefore, real world databases only can be accessed by database management systems and the number of accesses has to be reduced to a minimum. Considering this property, we are forced to use, for example, standard set oriented interfaces of relational database management systems in order to apply methods of knowledge discovery in databases. We present a system for discovering data dependencies, which is build upon a set oriented interface. The point of main effort has been put on the discovery of value restrictions, unary inclusion- and functional dependencies in relational databases. The system also embodies an inference rel...
Machine Learning and Knowledge Discovery in Databases
Knowledge discovery in databases is not only the nontrivial extraction of implicit, previously unknown and potentially useful information from databases. We argue that in contrast to machine learning, knowledge discovery in databases should be applied to real world databases. Since real world databases are known to be very large, they raise problems of the access. Therefore, real world databases only can be accessed by database management systems and the number of accesses has to be reduced to a minimum. Considering this property, we are forced to use, for example, standard set oriented interfaces of relational database management systems in order to apply methods of knowledge discovery in databases. We present a system for discovering data dependencies, which is build upon a set oriented interface. The point of main effort has been put on the discovery of value restrictions, unary inclusion- and functional dependencies in relational databases. The system also embodies an inference rel...
Machine Learning and Knowledge Discovery in Databases
Knowledge discovery in databases is not only the nontrivial extraction of implicit, previously unknown and potentially useful information from databases. We argue that in contrast to machine learning, knowledge discovery in databases should be applied to real world databases. Since real world databases are known to be very large, they raise problems of the access. Therefore, real world databases only can be accessed by database management systems and the number of accesses has to be reduced to a minimum. Considering this property, we are forced to use, for example, standard set oriented interfaces of relational database management systems in order to apply methods of knowledge discovery in databases. We present a system for discovering data dependencies, which is build upon a set oriented interface. The point of main effort has been put on the discovery of value restrictions, unary inclusion- and functional dependencies in relational databases. The system also embodies an inference rel...
The community%2Dsearch problem and how to plan a successful cocktail party.

Evaluating Query Result Significance in Databases via Randomizations.

Query similarity by projecting the query%2Dflow graph.
Defining a measure of similarity between queries is an interesting and difficult problem. A reliable query-similarity measure can be used in a variety of applications such as query recommendation, query expansion, and advertising. In this paper, we exploit the information present in query logs in order to develop a measure of semantic similarity between queries. Our approach relies on the concept of the query-flow graph, a graph-based representation of a query log. The query-flow graph aggregates query reformulations from many users: nodes in the graph represent queries, and two queries are connected if they are likely to appear as part of the same search goal. Our query-similarity measure is obtained by projecting the graph (or appropriate subgraphs extracted from it) on a low-dimensional Euclidean space. Our experiments show that the measure we obtain captures a notion of semantic similarity between queries and it is useful for diversifying query recommendations.
An optimization framework for query recommendation.
Query recommendation is an integral part of modern search engines. The goal of query recommendation is to facilitate users while searching for information. Query recommendation also allows users to explore concepts related to their information needs. In this paper, we present a formal treatment of the problem of query recommendation. In our framework we model the querying behavior of users by a probabilistic reformulation graph, or query-flow graph [Boldi et al. CIKM 2008]. A sequence of queries submitted by a user can be seen as a path on this graph. Assigning score values to queries allows us to define suitable utility functions and to consider the expected utility achieved by a reformulation path on the query-flow graph. Providing recommendations can be seen as adding shortcuts in the query-flow graph that“nudge”the reformulation paths of users, in such a way that users are more likely to follow paths with larger expected utility. We discuss in detail the most important questions that arise in the proposed framework. In particular, we provide examples of meaningful utility functions to optimize, we discuss how to estimate the effect of recommendations on the reformulation probabilities, we address the complexity of the optimization problems that we consider, we suggest efficient algorithmic solutions, and we validate our models and algorithms with extensive experimentation. Our techniques can be applied to other scenarios where user behavior can be modeled as a Markov process.
On the high density of leadership nuclei in endorsement social networks.

Guest editors%27 introduction special issue of selected papers from ECML PKDD 2010.

Learning and Predicting the Evolution of Social Networks.

Special issue for ECML PKDD 2010 Guest editors%27 introduction.

k%2DNearest Neighbors in Uncertain Graphs.
Complex networks, such as biological, social, and communication networks, often entail uncertainty, and thus, can be modeled as probabilistic graphs. Similar to the problem of similarity search in standard graphs, a fundamental problem for probabilistic graphs is to efficiently answer k-nearest neighbor queries (k-NN), which is the problem of computing the k closest nodes to some specific node. In this paper we introduce a framework for processing k-NN queries in probabilistic graphs. We propose novel distance functions that extend well-known graph concepts, such as shortest paths. In order to compute them in probabilistic graphs, we design algorithms based on sampling. During k-NN query processing we efficiently prune the search space using novel techniques. Our experiments indicate that our distance functions outperform previously used alternatives in identifying true neighbors in real-world biological data. We also demonstrate that our algorithms scale for graphs with tens of millions of edges. 1.

Efficient algorithms for large%2Dscale local triangle counting.
In this article, we study the problem of approximate local triangle counting in large graphs. Namely, given a large graph G = (V, E) we want to estimate as accurately as possible the number of triangles incident to every node v ∈ V in the graph. We consider the question both for undirected and directed graphs. The problem of computing the global number of triangles in a graph has been considered before, but to our knowledge this is the first contribution that addresses the problem of approximate local triangle counting with a focus on the efficiency issues arising in massive graphs and that also considers the directed case. The distribution of the local number of triangles and the related local clustering coefficient can be used in many interesting applications. For example, we show that the measures we compute can help detect the presence of spamming activity in largescale Web graphs, as well as to provide useful features for content quality assessment in social networks. For computing the local number of triangles (undirected and directed), we propose two approximation algorithms, which are based on the idea of min-wise independent permutations [Broder et al. 1998]. Our algorithms operate in a semi-streaming fashion, using O(|V |) space in main memory and performing O(log |V |) sequential scans over the edges of the graph. The first algorithm

Fast shortest path distance estimation in large networks.

Mining Graph Evolution Rules.

Taxonomy%2DDriven Lumping for Sequence Mining.
Abstract Given a taxonomy of events and a dataset of sequences of these events, we study the problem of finding efficient and effective ways to produce a compact representation of the sequences. We model sequences with Markov models whose states correspond to nodes in the provided taxonomy, and each state represents the events in the subtree under the corresponding node. By lumping observed events to states that correspond to internal nodes in the taxonomy, we allow more compact models that are easier to understand and visualize, at the expense of a decrease in the data likelihood. We formally define and characterize our problem, and we propose a scalable search method for finding a good trade-off between two conflicting goals: maximizing the data likelihood, and minimizing the model complexity. We implement these ideas in Taxomo, a taxonomy-driven modeler, which we apply in two different domains, query-log mining and mining of moving-object trajectories. The empirical evaluation confirms the feasibility and usefulness of our approach. Keywords Data mining · Sequence analysis · Markov models · Query-log analysis · Spatial-data analysis
Query Significance in Databases via Randomizations

Taxonomy%2Ddriven lumping for sequence mining.
Abstract Given a taxonomy of events and a dataset of sequences of these events, we study the problem of finding efficient and effective ways to produce a compact representation of the sequences. We model sequences with Markov models whose states correspond to nodes in the provided taxonomy, and each state represents the events in the subtree under the corresponding node. By lumping observed events to states that correspond to internal nodes in the taxonomy, we allow more compact models that are easier to understand and visualize, at the expense of a decrease in the data likelihood. We formally define and characterize our problem, and we propose a scalable search method for finding a good trade-off between two conflicting goals: maximizing the data likelihood, and minimizing the model complexity. We implement these ideas in Taxomo, a taxonomy-driven modeler, which we apply in two different domains, query-log mining and mining of moving-object trajectories. The empirical evaluation confirms the feasibility and usefulness of our approach. Keywords Data mining · Sequence analysis · Markov models · Query-log analysis · Spatial-data analysis
k%2DAnonymization with Minimal Loss of Information.

A randomized approximation algorithm for computing bucket orders.
www.elsevier.com/locate/ipl A randomized approximation algorithm for computing bucket orders
Query%2Dlog mining for detecting spam.
Every day millions of users search for information on the web via search engines, and provide implicit feedback to the results shown for their queries by clicking or not onto them. This feedback is encoded in the form of a query log that consists of a sequence of search actions, one per user query, each describing the following information: (i) terms composing a query, (ii) documents returned by the search engine, (iii) documents that have been clicked, (iv) the rank of those documents in the list of results, (v) date and time of the search action/click, (vi) an anonymous identifier for each session, and more. In this work, we investigate the idea of characterizing the documents and the queries belonging to a given query log with the goal of improving algorithms for detecting spam, both at the document level and at the query level. 1.
Dr. Searcher and Mr. Browser a unified hyperlink%2Dclick graph.
We introduce a unified graph representation of the Web, which includes both structural and usage information. We model this graph using a simple union of the Web’s hyperlink and click graphs. The hyperlink graph expresses link structure among Web pages, while the click graph is a bipartite graph of queries and documents denoting users ’ searching behavior extracted from a search engine’s query log. Our most important motivation is to model in a unified way the two main activities of users on the Web: searching and browsing, and at the same time to analyze the effects of random walks on this new graph. The intuition behind this task is to measure how the combination of link structure and usage data provide additional information to that contained in these structures independently. Our experimental results show that both hyperlink and click graphs have strengths and weaknesses when it comes to using their stationary distribution scores for ranking Web pages. Furthermore, our evaluation indicates that the unified graph always generates consistent and robust scores that follow closely the best result obtained from either individual graph, even when applied to “noisy ” data. It is our belief that the unified Web graph has several useful properties for improving current Web document ranking, as well as for generating new rankings of its own. In particular stationary distribution scores derived from the random walks on the combined graph can be used as an indicator of whether structural or usage data are more reliable in different situations. 1.
Searching the wikipedia with contextual information.
We propose a framework for searching the Wikipedia with contextual information. Our framework extends the typical keyword search, by considering queries of the type 〈q, p〉, where q is a set of terms (as in classical Web search), and p is a source Wikipedia document. The query terms q represent the information that the user is interested in finding, and the document p provides the context of the query. The task is to rank other documents in Wikipedia with respect to their relevance to the query terms q given the context document p. By associating a context to the query terms, the search results of a search initiated in a particular page can be made more relevant. We suggest a number of features that extend the classical querysearch model so that the context document p is considered. We then use RankSVM (Joachims 2002) to learn weights for the individual features given suitably constructed training data. Documents are ranked at query time using the inner product of the feature and the weight vectors. The experiments indicate that the proposed method considerably improves results obtained by a more traditional approach that does not take the context into account.
The query%2Dflow graph model and applications.
Query logs record the queries and the actions of the users of search engines, and as such they contain valuable information about the interests, the preferences, and the behavior of the users, as well as their implicit feedback to search-engine results. Mining the wealth of information available in the query logs has many important applications including querylog analysis, user profiling and personalization, advertising, query recommendation, and more. In this paper we introduce the query-flow graph, a graph representation of the interesting knowledge about latent querying behavior. Intuitively, in the query-flow graph a directed edge from query qi to query qj means that the two queries are likely to be part of the same “search mission”. Any path over the query-flow graph may be seen as a searching behavior, whose likelihood is given by the strength of the edges along the path. The query-flow graph is an outcome of query-log mining and, at the same time, a useful tool for it. We propose a methodology that builds such a graph by mining time and textual information as well as aggregating queries from different users. Using this approach we build a real-world queryflow graph from a large-scale query log and we demonstrate its utility in concrete applications, namely, finding logical sessions, and query recommendation. We believe, however, that the usefulness of the query-flow graph goes beyond these two applications.
k%2DAnonymization Revisited.
The concept of k-anonymity, used in the recent literature (e.g., [10, 11, 7, 5, 1]) to formally evaluate the privacy preservation of published tables, was introduced
Mining Large Networks with Subgraph Counting.

Efficient semi%2Dstreaming algorithms for local triangle counting in massive graphs.

Topical query decomposition.
We introduce the problem of query decomposition, where we are given a query and a document retrieval system, and we want to produce a small set of queries whose union of resulting documents corresponds approximately to that of the original query. Ideally, these queries should represent coherent, conceptually well-separated topics. We provide an abstract formulation of the query decomposition problem, and we tackle it from two different perspectives. We first show how the problem can be instantiated as a specific variant of a set cover problem, for which we provide an efficient greedy algorithm. Next, we show how the same problem can be seen as a constrained clustering problem, with a very particular kind of constraint, i.e., clustering with predefined clusters. We develop a two-phase algorithm based on hierarchical agglomerative clustering followed by dynamic programming. Our experiments, conducted on a set of actual queries in a Web scale search engine, confirm the effectiveness of the proposed solutions.
Finding high%2Dquality content in social media.
ABSTRACT: The quality of user-generated content varies drastically from excellent to abuse and spam. As the availability of such content increases, the task of identifying high-quality content in sites based on user contributions—social media sites—becomes increasingly important. Social media in general exhibit a rich variety of information sources: in addition to the content itself, there is a wide array of non-content information available, such as links between items and explicit quality ratings from members of the community. In this paper we investigate methods for exploiting such community feedback to automatically identify high quality content. As a test case, we focus on Yahoo! Answers, a large community question answering portal that is particularly rich in the amount and types of content and social interactions available in it. We introduce a general classification framework for combining the evidence from different sources of information that can be tuned automatically for a given social media type and quality definition. In particular, for the community question answering domain, we show that our system is able to separate high-quality items from the rest with an accuracy close to that of humans. 1
The Discrete Basis Problem.
We consider the Discrete Basis Problem, which can be described as follows: given a collection of Boolean vectors find a collection of k Boolean basis vectors such that the original vectors can be represented using disjunctions of these basis vectors. We show that the decision version of this problem is NP-complete and that the optimization version cannot be approximated within any finite ratio. We also study two variations of this problem, where the Boolean basis vectors must be mutually otrhogonal. We show that the other variation is closely related with the well-known Metric k-median Problem in Boolean space. To solve these problems, two algorithms will be presented. One is designed for the variations mentioned above, and it is solely based on solving the k-median problem, while another is a heuristic intended to solve the general Discrete Basis Problem. We will also study the results of extensive experiments made with these two algorithms with both synthetic and real-world data. The results are twofold: with the synthetic data, the algorithms did rather well, but with the real-world data the results were not as good.
Algorithms for unimodal segmentation with applications to unimodality detection.

Algorithms for unimodal segmentation with applications to unimodality detection.

Optimal segmentation using tree models.
Sequence data are abundant in application areas such as computational biology, environmental sciences, and telecommunication. Many real-life sequences have a strong segmental structure, with segments of different complexities. In this paper we study the description of sequence segments using variable length Markov chains (VLMCs), also known as tree models. We discover the segment boundaries of a sequence and at the same time we obtain a VLMC for each segment. Such a context tree contains the probability distribution vectors that capture the essential features of the corresponding segment. We use the Bayesian Information Criterion (BIC) and the Krichevsky-Trofimov Probability (KT) to select the number of segments of a sequence. On DNA data the method selects segments that closely correspond to the annotated regions of the genes. 1.
Design trade%2Doffs for search engine caching.

k %2DAnonymization with Minimal Loss of Information.

Recurrent Predictive Models for Sequence Segmentation.

Segmentations with Rearrangements.
Abstract. Computational genomics involves comparing sequences based on &amp;quot;similarity &amp;quot; for detecting evolutionary and functional relation-ships. Until very recently, available portions of the human genome sequence (and that of other species) were fairly short and sparse. Mostsequencing effort was focused on genes and other short units; similarity between such sequences was measured based on character level differ-ences. However with the advent of whole genome sequencing technology there is emerging consensus that the measure of similarity between long genome sequences must capture the rearrangements of large segmentsfound in abundance in the human genome. In this paper, we abstract the general problem of computing sequence similarity in the presence of segment rearrangements. This problem isclosely related to computing the smallest grammar for a string or the block edit distance between two strings. Our problem, like these otherproblems, is NP hard. Our main result here is a simple O(1) factor approximation algorithm for this problem. In contrast, best known approxi-mations for the related problems are factor \Omega (log n) off from the optimal. Our algorithm works in linear time, and in one pass. In proving our re-sult, we relate sequence similarity measures based on different segment rearrangements, to each other, tight up to constant factors. 1 Introduction Similarity comparison between biomolecular sequences play an important rolein computational genomics due to the premise that sequence similarity usually
The impact of caching on search engines.

Know your neighbors web spam detection using the web topology.
Web spam can significantly deteriorate the quality of search engine results. Thus there is a large incentive for commercial search engines to detect spam pages efficiently and accurately. In this paper we present a spam detection system that uses the topology of the Web graph by exploiting the link dependencies among the Web pages, and the content of the pages themselves. We find that linked hosts tend to belong to the same class: either both are spam or both are non-spam. We demonstrate three methods of incorporating the Web graph topology into the predictions obtained by our base classifier: (i) clustering the host graph, and assigning the label of all hosts in the cluster by majority vote, (ii) propagating the predicted labels to neighboring hosts, and (iii) using the predicted labels of neighboring hosts as new features and retraining the classifier. The result is an accurate system for detecting Web spam that can be applied in practice to large-scale Web data.
Estimating Number of Citations Using Author Reputation.
Abstract. We study the problem of predicting the popularity of items in a dynamic environment in which authors post continuously new items and provide feedback on existing items. This problem can be applied to predict popularity of blog posts, rank photographs in a photo-sharing system, or predict the citations of a scientific article using author information and monitoring the items of interest for a short period of time after their creation. As a case study, we show how to estimate the number of citations for an academic paper using information about past articles written by the same author(s) of the paper. If we use only the citation information over a short period of time, we obtain a predicted value that has a correlation of r = 0.57 with the actual value. This is our baseline prediction. Our best-performing system can improve that prediction by adding features extracted from the past publishing history of its authors, increasing the correlation between the actual and the predicted values to r = 0.81. 1
The price of validity in dynamic networks.

Clustering aggregation.

Assessing data mining results via swap randomization.
The problem of assessing the significance of data mining results on high-dimensional 0–1 data sets has been studied extensively in the literature. For problems such as mining frequent sets and finding correlations, significance testing can be done by, e.g., chi-square tests, or many other methods. However, the results of such tests depend only on the specific attributes and not on the dataset as a whole. Moreover, the tests are more difficult to apply to sets of patterns or other complex results of data mining. In this paper, we consider a simple randomization technique that deals with this shortcoming. The approach consists of producing random datasets that have the same row and column margins with the given dataset, computing the results of interest on the randomized instances, and comparing them against the results on the actual data. This randomization technique can be used to assess the results of many different types of data mining algorithms, such as frequent sets, clustering, and rankings. To generate random datasets with given margins, we use variations of a Markov chain approach, which is based on a simple swap operation. We give theoretical results on the efficiency of different randomization methods, and apply the swap randomization method to several wellknown datasets. Our results indicate that for some datasets the structure discovered by the data mining algorithms is a random artifact, while for other datasets the discovered structure conveys meaningful information.
Optimal Segmentation Using Tree Models.
Sequence data are abundant in application areas such as computational biology, environmental sciences, and telecommunication. Many real-life sequences have a strong segmental structure, with segments of different complexities. In this paper we study the description of sequence segments using variable length Markov chains (VLMCs), also known as tree models. We discover the segment boundaries of a sequence and at the same time we obtain a VLMC for each segment. Such a context tree contains the probability distribution vectors that capture the essential features of the corresponding segment. We use the Bayesian Information Criterion (BIC) and the Krichevsky-Trofimov Probability (KT) to select the number of segments of a sequence. On DNA data the method selects segments that closely correspond to the annotated regions of the genes. 1.
Assessing data mining results via swap randomization.
The problem of assessing the significance of data mining results on high-dimensional 0–1 data sets has been studied extensively in the literature. For problems such as mining frequent sets and finding correlations, significance testing can be done by, e.g., chi-square tests, or many other methods. However, the results of such tests depend only on the specific attributes and not on the dataset as a whole. Moreover, the tests are more difficult to apply to sets of patterns or other complex results of data mining. In this paper, we consider a simple randomization technique that deals with this shortcoming. The approach consists of producing random datasets that have the same row and column margins with the given dataset, computing the results of interest on the randomized instances, and comparing them against the results on the actual data. This randomization technique can be used to assess the results of many different types of data mining algorithms, such as frequent sets, clustering, and rankings. To generate random datasets with given margins, we use variations of a Markov chain approach, which is based on a simple swap operation. We give theoretical results on the efficiency of different randomization methods, and apply the swap randomization method to several wellknown datasets. Our results indicate that for some datasets the structure discovered by the data mining algorithms is a random artifact, while for other datasets the discovered structure conveys meaningful information.
What is the Dimension of Your Binary Data%3F
Many 0/1 datasets have a very large number of variables; however, they are sparse and the dependency structure of the variables is simpler than the number of variables would suggest. Defining the effective dimensionality of such a dataset is a nontrivial problem. We consider the problem of defining a robust measure of dimension for 0/1 datasets, and show that the basic idea of fractal dimension can be adapted for binary data. However, as such the fractal dimension is difficult to interpret. Hence we introduce the concept of normalized fractal dimension. For a dataset D, its normalized fractal dimension counts the number of independent columns needed to achieve the unnormalized fractal dimension of D. The normalized fractal dimension measures the degree of dependency structure of the data. We study the properties of the normalized fractal dimension and discuss its computation. We give empirical results on the normalized fractal dimension, comparing it against PCA. 1.
Assessing data mining results via swap randomization.
The problem of assessing the significance of data mining results on high-dimensional 0–1 data sets has been studied extensively in the literature. For problems such as mining frequent sets and finding correlations, significance testing can be done by, e.g., chi-square tests, or many other methods. However, the results of such tests depend only on the specific attributes and not on the dataset as a whole. Moreover, the tests are more difficult to apply to sets of patterns or other complex results of data mining. In this paper, we consider a simple randomization technique that deals with this shortcoming. The approach consists of producing random datasets that have the same row and column margins with the given dataset, computing the results of interest on the randomized instances, and comparing them against the results on the actual data. This randomization technique can be used to assess the results of many different types of data mining algorithms, such as frequent sets, clustering, and rankings. To generate random datasets with given margins, we use variations of a Markov chain approach, which is based on a simple swap operation. We give theoretical results on the efficiency of different randomization methods, and apply the swap randomization method to several wellknown datasets. Our results indicate that for some datasets the structure discovered by the data mining algorithms is a random artifact, while for other datasets the discovered structure conveys meaningful information.
Algorithms for discovering bucket orders from data.
Ordering and ranking items of different types are important tasks in various applications, such as query processing and scientific data mining. A total order for the items can be misleading, since there are groups of items that have practically equal ranks. We consider bucket orders, i.e., total orders with ties. They can be used to capture the essential order information without overfitting the data: they form a useful concept class between total orders and arbitrary partial orders. We address the question of finding a bucket order for a set of items, given pairwise precedence information between the items. We also discuss methods for computing the pairwise precedence data. We describe simple and efficient algorithms for finding good bucket orders. Several of the algorithms have a provable approximation guarantee, and they scale well to large datasets. We provide experimental results on artificial and a real data that show the usefulness of bucket orders and demonstrate the accuracy and efficiency of the algorithms.
The Discrete Basis Problem.
We consider the Discrete Basis Problem, which can be described as follows: given a collection of Boolean vectors find a collection of k Boolean basis vectors such that the original vectors can be represented using disjunctions of these basis vectors. We show that the decision version of this problem is NP-complete and that the optimization version cannot be approximated within any finite ratio. We also study two variations of this problem, where the Boolean basis vectors must be mutually otrhogonal. We show that the other variation is closely related with the well-known Metric k-median Problem in Boolean space. To solve these problems, two algorithms will be presented. One is designed for the variations mentioned above, and it is solely based on solving the k-median problem, while another is a heuristic intended to solve the general Discrete Basis Problem. We will also study the results of extensive experiments made with these two algorithms with both synthetic and real-world data. The results are twofold: with the synthetic data, the algorithms did rather well, but with the real-world data the results were not as good.
Segmentation and dimensionality reduction.
Sequence segmentation and dimensionality reduction have been used as methods for studying high-dimensional sequences — they both reduce the complexity of the representation of the original data. In this paper we study the interplay of these two techniques. We formulate the problem of segmenting a sequence while modeling it with a basis of small size, thus essentially reducing the dimension of the input sequence. We give three different algorithms for this problem: all combine existing methods for sequence segmentation and dimensionality reduction. For two of the proposed algorithms we prove guarantees for the quality of the solutions obtained. We describe experimental results on synthetic and real datasets, including data on exchange rates and genomic sequences. Our experiments show that the algorithms indeed discover underlying structure in the data, including both segmental structure and interdependencies between the dimensions. Keywords: segmentation, multidimensional data, PCA, time series 1
Clustering Aggregation.

Parameter%2DFree Spatial Data Mining Using MDL.
Consider spatial data consisting of a set of binary features taking values over a collection of spatial extents (grid cells). We propose a method that simultaneously finds spatial correlation and feature co-occurrence patterns, without any parameters. In particular, we employ the Minimum Description Length (MDL) principle coupled with a natural way of compressing regions. This defines what “good” means: a feature co-occurrence pattern is good, if it helps us better compress the set of locations for these features. Conversely, a spatial correlation is good, if it helps us better compress the set of features in the corresponding region. Our approach is scalable for large datasets (both number of locations and of features). We evaluate our method on both real and synthetic datasets. 1
Mining Chains of Relations.
Traditional data mining applications consider the problem of mining a single relation between two attributes. For example, in a scientific bibliography database, authors are related to papers, and we may be interested in discovering association rules between authors. However, in real life, we often have multiple attributes related though chains of relations. For example, authors write papers, and papers concern one or more topics. Mining such relational chains poses additional challenges. In this paper we consider the following problem: given a chain of two relations R1(A, P) and R2(P, T) we want to find selectors for the objects in T such that the projected relation between A and P satisfies a specific property. The motivation for our approach is that a given property might not hold on the whole dataset, but it might hold when projecting the data on a selector set. We discuss various algorithms and we examine the conditions under which the apriori technique can be used. We experimentally demonstrate the effectiveness of our methods. 1
Dimension induced clustering.

Dimension Induced Clustering.

Unimodal Segmentation of Sequences.

Approximating a collection of frequent sets.
One of the most well-studied problems in data mining is computing the collection of frequent item sets in large transactional databases. One obstacle for the applicability of frequent-set mining is that the size of the output collection can be far too large to be carefully examined and understood by the users. Even restricting the output to the border of the frequent item-set collection does not help much in alleviating the problem.
Geometric and Combinatorial Tiles in 0%2D1 Data.
In this paper we introduce a simple probabilistic model, hierarchical tiles, for 0-1 data. A basic tile (X,Y,p) specifies a subset X of the rows and a subset Y of the columns of the data, i.e., a rectangle, and gives a probability p for the occurrence of 1s in the cells of X x Y. A hierarchical tile has additionally a set of exception tiles that specify the probabilities for subrectangles of the original rectangle. If the rows and columns are ordered and X and Y consist of consecutive elements in those orderings, then the tile is geometric; otherwise it is combinatorial. We give a simple randomized algorithm for finding good geometric tiles. Our main result shows that using spectral ordering techniques one can find good orderings that turn combinatorial tiles into geometric tiles. We give empirical results on the performance of the methods.
The Price of Validity in Dynamic Networks.

Automated Ranking of Database Query Results.
We investigate the problem of ranking answers to a database query when many tuples are returned. We adapt and apply principles of probabilistic models from Information Retrieval for structured data. Our proposed solution is domain independent. It leverages data and workload statistics and correlations. Our ranking functions can be further customized for different applications. We present results of preliminary experiments which demonstrate the efficiency as well as the quality of our ranking system. 1.
Fragments of order.
High-dimensional collections of 0-1 data occur in many applications. The attributes in such data sets are typically considered to be unordered. However, in many cases there is a natural total or partial order # underlying the variables of the data set. Examples of variables for which such orders exist include terms in documents, courses in enrollment data, and paleontological sites in fossil data collections. The observations in such applications are flat, unordered sets; however, the data sets respect the underlying ordering of the variables. By this we mean that if A # B # C are three variables respecting the underlying ordering #, and both of variables A and C appear in an observation, then, up to noise levels, variable B also appears in this observation. Similarly, if A1 # A2 # # A l-1 # A l is a longer sequence of variables, we do not expect to see many observations for which there are indices i &lt; j &lt; k such that A i and Ak occur in the observation but A j does not.
Finding recurrent sources in sequences.
time series display tremendous variability. However, often it is reasonable to assume that the sequence is actually generated by or assembled from a small number of sources, each of which might contribute several segments to the sequence. That is, there are h hidden sources such that the sequence can be written as a concatenation of k ? h pieces, each of which stems from one of the h sources. We define this (k; h)- segmentation problem and show that it is NP-hard in the general case. We give approximation algorithms achieving approximation ratios of 3 for the L1 error measure and    5 for the L2 error measure, and generalize the results to higher dimensions. We give empirical results on real (chromosome 22) and artificial data showing that the methods work well in practice.
XTRACT Learning Document Type Descriptors from XML Document Collections.

DTD Inference from XML Documents The XTRACT Approach.
XML is rapidly emerging as the new standard for data representation and exchange on the Web. Document Type Descriptors (DTDs) contain valuable information on the structure of XML documents and thus have a crucial role in the efficient storage and querying of XML data. Despite their importance, however, DTDs are not mandatory, and it is quite possible for documents in XML databases to not have accompanying DTDs. In this paper, we present an overview of 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 &#034;general &#034; 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.
A combinatorial algorithm for MAX CSP.

Maintaining stream statistics over sliding windows %28extended abstract%29.
Mayur Datar  Aristides Gionis  y  Piotr Indyk  z  Rajeev Motwani  x  Abstract  We consider the problem of maintaining aggregates and statistics over data streams, with respect to the last N data elements seen so far. We refer to this model as the sliding window model. We consider the following basic problem: Given a stream of bits, maintain a count of the number of 1&#039;s in the last N elements seen from the stream. We show that using O(  1  ffl log  2  N) bits of memory, we can estimate the number of 1&#039;s to within a factor of 1 + ffl. We also give a matching lower bound of \Omega\Gamma  1  ffl log  2  N) memory bits for any deterministic or randomized algorithms. We extend our scheme to maintain the sum of the last N positive integers. We provide matching upper and lower bounds for this more general problem as well. We apply our techniques to obtain efficient algorithms for the Lp norms (for p 2 [1; 2]) of vectors under the sliding window model. Using the algorithm for the basic counting problem, one can adapt many other techniques to work for the sliding window model, with a multiplicative overhead of O(  1  ffl log N) in memory and a 1 + ffl factor loss in accuracy. These include maintaining approximate histograms, hash tables, and statistics or aggregates such as sum and averages.
Evaluating strategies for similarity search on the web.
Finding pages on the Web that are similar to a query page (Related Pages) is an important component of modern search engines. A variety of strategies have been proposed for answering Related Pages queries, but comparative evaluation by user studies is expensive, especially when large strategy spaces must be searched (e.g., when tuning parameters). We present a technique for automatically evaluating strategies using Web hierarchies, such as Open Directory, in place of user feedback. We apply this evaluation methodology to a mix of document representation strategies, including the use of text, anchor-text, and links. We discuss the relative advantages and disadvantages of the various approaches examined. Finally, we describe how to eciently construct a similarity index out of our chosen strategies, and provide sample results from our index.
Maintaining Stream Statistics over Sliding Windows.
Mayur Datar  Aristides Gionis  y  Piotr Indyk  z  Rajeev Motwani  x  Abstract  We consider the problem of maintaining aggregates and statistics over data streams, with respect to the last N data elements seen so far. We refer to this model as the sliding window model. We consider the following basic problem: Given a stream of bits, maintain a count of the number of 1&#039;s in the last N elements seen from the stream. We show that using O(  1  ffl log  2  N) bits of memory, we can estimate the number of 1&#039;s to within a factor of 1 + ffl. We also give a matching lower bound of \Omega\Gamma  1  ffl log  2  N) memory bits for any deterministic or randomized algorithms. We extend our scheme to maintain the sum of the last N positive integers. We provide matching upper and lower bounds for this more general problem as well. We apply our techniques to obtain efficient algorithms for the Lp norms (for p 2 [1; 2]) of vectors under the sliding window model. Using the algorithm for the basic counting problem, one can adapt many other techniques to work for the sliding window model, with a multiplicative overhead of O(  1  ffl log N) in memory and a 1 + ffl factor loss in accuracy. These include maintaining approximate histograms, hash tables, and statistics or aggregates such as sum and averages.
Efficient and Tunable Similar Set Retrieval.
model complex data sets. Modern Object Relational systems support set value attributes and allow various query capabilities on them. In this paper we initiate a formal study of indexing techniques for set value attributes based on similarity, for suitably defined notions of similarity between sets. Such techniques are necessary in modern applications such as recommendations through collaborative filtering and automated advertising. Our techniques are probabilistic and approximate in nature. As a design principle we create structures that make use of well known and widely used data structuring techniques, as a means to ease integration with existing infrastructure.
Finding Interesting Associations without Support Pruning.
Abstract Association-rule mining has heretofore relied on the condition of high support to do its work efficiently. In particular, the well-known a-priori algorithm is only effective when the only rules of interest are relationships that occur very frequently. However, there are a number of applications, such as data mining, identification of similar web documents, clustering, and collaborative filtering, where the rules of interest have comparatively few instances in the data. In these cases, we must look for highly correlated items, or possibly even causal relationships between infrequent items. We develop a family of algorithms for solving this problem, employing a combination of random sampling and hashing techniques. We provide analysis of the algorithms developed, and conduct experiments on real and synthetic data to obtain a comparative performance analysis.
Finding Interesting Associations without Support Pruning.
Abstract Association-rule mining has heretofore relied on the condition of high support to do its work efficiently. In particular, the well-known a-priori algorithm is only effective when the only rules of interest are relationships that occur very frequently. However, there are a number of applications, such as data mining, identification of similar web documents, clustering, and collaborative filtering, where the rules of interest have comparatively few instances in the data. In these cases, we must look for highly correlated items, or possibly even causal relationships between infrequent items. We develop a family of algorithms for solving this problem, employing a combination of random sampling and hashing techniques. We provide analysis of the algorithms developed, and conduct experiments on real and synthetic data to obtain a comparative performance analysis.
Scalable Techniques for Clustering the Web.
Clustering is one of the most crucial techniques for dealing with the massive amount of information present on the web. Clustering can either be performed once offline, independent of search queries, or performed online on the results of search queries. Our offline approach aims to efficiently cluster similar pages on the web, using the technique of Locality-Sensitive Hashing (LSH), in which web pages are hashed in such a way that similar pages have a much higher probability of collision than dissimilar pages. Our preliminary experiments on the Stanford WebBase have shown that the hash-based scheme can be scaled to millions of urls. 1.
XTRACT A System for Extracting Document Type Descriptors from XML Documents.

Scalable Techniques for Clustering the Web.
Clustering is one of the most crucial techniques for dealing with the massive amount of information present on the web. Clustering can either be performed once offline, independent of search queries, or performed online on the results of search queries. Our offline approach aims to efficiently cluster similar pages on the web, using the technique of Locality-Sensitive Hashing (LSH), in which web pages are hashed in such a way that similar pages have a much higher probability of collision than dissimilar pages. Our preliminary experiments on the Stanford WebBase have shown that the hash-based scheme can be scaled to millions of urls. 1.
Similarity Search in High Dimensions via Hashing.
The nearest- or near-neighbor query problems arise in a large variety of database applications, usually in the context of similarity searching. Of late, there has been increasing interest in building search/index structures for performing similarity search over high-dimensional data, e.g., image databases, document collections, time-series databases, and genome databases. Unfortunately, all known techniques for solving this problem fall prey to the \curse of dimensionality. &amp;quot; That is, the data structures scale poorly with data dimensionality; in fact, if the number of dimensions exceeds 10 to 20, searching in k-d trees and related structures involves the inspection of a large fraction of the database, thereby doing no better than brute-force linear search. It has been suggested that since the selection of features and the choice of a distance metric in typical applications is rather heuristic, determining an approximate nearest neighbor should su ce for most practical purposes. In this paper, we examine a novel scheme for approximate similarity search based on hashing. The basic idea is to hash the points
Surface Reflectance Recovery under Point Light Illumination.
In this paper, a novel algorithm for colour recovery is presented. It  assumes that the 3-D geometry of the scene is known. The spectral  power distribution of a point illumination source, and the response  function of the sensor are calibrated jointly. This algorithm has been  used for the colour recovery part of an integrated system, developed  in our laboratory, for environmental modelling. The geometry of the  scene is recovered using a laser stripe range-finder and this information  is exploited by the colour recovery algorithm. A point light source,  attached to the whole system, has been used for the illumination of the  scene, in order to confine undesirable side-effects of the ambient light.  The joint spectral power distribution of this point light source and the  response function of the camera are obtained with off-line calibration.  1 Introduction  The objective of this paper is to present a new algorithm for obtaining surface colour information in a controled enviroment. ...
