Efficient search for statistically significant dependency rules in binary data

Analyzing statistical dependencies is a fundamental problem in all empirical science. Dependencies help us understand causes and effects, create new scientific theories, and invent cures to problems. Nowadays, large amounts of data is available, but efficient computational tools for analyzing the data are missing. In this research, we develop efficient algorithms for a commonly occurring search problem - searching for the statistically most significant dependency rules in binary data. We consider dependency rules of the form X->A or X->not A, where X is a set of positive-valued attributes and A is a single attribute. Such rules describe which factors either increase or decrease the probability of the consequent A. A classical example are genetic and environmental factors, which can either cause or prevent a disease. The emphasis in this research is that the discovered dependencies should be genuine - i.e. they should also hold in future data. This is an important distinction from the traditional association rules, which - in spite of their name and a similar appearance to dependency rules - do not necessarily represent statistical dependencies at all or represent only spurious connections, which occur by chance. Therefore, the principal objective is to search for the rules with statistical significance measures. Another important objective is to search for only non-redundant rules, which express the real causes of dependence, without any occasional extra factors. The extra factors do not add any new information on the dependence, but can only blur it and make it less accurate in future data. The problem is computationally very demanding, because the number of all possible rules increases exponentially with the number of attributes. In addition, neither the statistical dependency nor the statistical significance are monotonic properties, which means that the traditional pruning techniques do not work. As a solution, we first derive the mathematical basis for pruning the search space with any well-behaving statistical significance measures. The mathematical theory is complemented by a new algorithmic invention, which enables an efficient search without any heuristic restrictions. The resulting algorithm can be used to search for both positive and negative dependencies with any commonly used statistical measures, like Fisher's exact test, the chi-squared measure, mutual information, and z scores. According to our experiments, the algorithm is well-scalable, especially with Fisher's exact test. It can easily handle even the densest data sets with 10000-20000 attributes. Still, the results are globally optimal, which is a remarkable improvement over the existing solutions. In practice, this means that the user does not have to worry whether the dependencies hold in future data or if the data still contains better, but undiscovered dependencies.

Mining Cell Transition Data

Cell transition data is obtained from a cellular phone that switches its current serving cell tower. The data consists of a sequence of transition events, which are pairs of cell identifiers and transition times. The focus of this thesis is applying data mining methods to such data, developing new algorithms, and extracting knowledge that will be a solid foundation on which to build location-aware applications. In addition to a thorough exploration of the features of the data, the tools and methods developed in this thesis provide solutions to three distinct research problems. First, we develop clustering algorithms that produce a reliable mapping between cell transitions and physical locations observed by users of mobile devices. The main clustering algorithm operates in online fashion, and we consider also a number of offline clustering methods for comparison. Second, we define the concept of significant locations, known as bases, and give an online algorithm for determining them. Finally, we consider the task of predicting the movement of the user, based on historical data. We develop a prediction algorithm that considers paths of movement in their entirety, instead of just the most recent movement history. All of the presented methods are evaluated with a significant body of real cell transition data, collected from about one hundred different individuals. The algorithms developed in this thesis are designed to be implemented on a mobile device, and require no extra hardware sensors or network infrastructure. By not relying on external services and keeping the user information as much as possible on the user s own personal device, we avoid privacy issues and let the users control the disclosure of their location information.

Learning Nonlinear Visual Processing from Natural Images

The paradigm of computational vision hypothesizes that any visual function -- such as the recognition of your grandparent -- can be replicated by computational processing of the visual input. What are these computations that the brain performs? What should or could they be? Working on the latter question, this dissertation takes the statistical approach, where the suitable computations are attempted to be learned from the natural visual data itself. In particular, we empirically study the computational processing that emerges from the statistical properties of the visual world and the constraints and objectives specified for the learning process. This thesis consists of an introduction and 7 peer-reviewed publications, where the purpose of the introduction is to illustrate the area of study to a reader who is not familiar with computational vision research. In the scope of the introduction, we will briefly overview the primary challenges to visual processing, as well as recall some of the current opinions on visual processing in the early visual systems of animals. Next, we describe the methodology we have used in our research, and discuss the presented results. We have included some additional remarks, speculations and conclusions to this discussion that were not featured in the original publications. We present the following results in the publications of this thesis. First, we empirically demonstrate that luminance and contrast are strongly dependent in natural images, contradicting previous theories suggesting that luminance and contrast were processed separately in natural systems due to their independence in the visual data. Second, we show that simple cell -like receptive fields of the primary visual cortex can be learned in the nonlinear contrast domain by maximization of independence. Further, we provide first-time reports of the emergence of conjunctive (corner-detecting) and subtractive (opponent orientation) processing due to nonlinear projection pursuit with simple objective functions related to sparseness and response energy optimization. Then, we show that attempting to extract independent components of nonlinear histogram statistics of a biologically plausible representation leads to projection directions that appear to differentiate between visual contexts. Such processing might be applicable for priming, \ie the selection and tuning of later visual processing. We continue by showing that a different kind of thresholded low-frequency priming can be learned and used to make object detection faster with little loss in accuracy. Finally, we show that in a computational object detection setting, nonlinearly gain-controlled visual features of medium complexity can be acquired sequentially as images are encountered and discarded. We present two online algorithms to perform this feature selection, and propose the idea that for artificial systems, some processing mechanisms could be selectable from the environment without optimizing the mechanisms themselves. In summary, this thesis explores learning visual processing on several levels. The learning can be understood as interplay of input data, model structures, learning objectives, and estimation algorithms. The presented work adds to the growing body of evidence showing that statistical methods can be used to acquire intuitively meaningful visual processing mechanisms. The work also presents some predictions and ideas regarding biological visual processing.

Offloading content routing cost from routers

The publish/subscribe paradigm has lately received much attention. In publish/subscribe systems, a specialized event-based middleware delivers notifications of events created by producers (publishers) to consumers (subscribers) interested in that particular event. It is considered a good approach for implementing Internet-wide distributed systems as it provides full decoupling of the communicating parties in time, space and synchronization. One flavor of the paradigm is content-based publish/subscribe which allows the subscribers to express their interests very accurately. In order to implement a content-based publish/subscribe middleware in way suitable for Internet scale, its underlying architecture must be organized as a peer-to-peer network of content-based routers that take care of forwarding the event notifications to all interested subscribers. A communication infrastructure that provides such service is called a content-based network. A content-based network is an application-level overlay network. Unfortunately, the expressiveness of the content-based interaction scheme comes with a price - compiling and maintaining the content-based forwarding and routing tables is very expensive when the amount of nodes in the network is large. The routing tables are usually partially-ordered set (poset) -based data structures. In this work, we present an algorithm that aims to improve scalability in content-based networks by reducing the workload of content-based routers by offloading some of their content routing cost to clients. We also provide experimental results of the performance of the algorithm. Additionally, we give an introduction to the publish/subscribe paradigm and content-based networking and discuss alternative ways of improving scalability in content-based networks. ACM Computing Classification System (CCS): C.2.4 [Computer-Communication Networks]: Distributed Systems - Distributed applications

Storage and retrieval of individual genomes

A repetitive sequence collection is one where portions of a base sequence of length n are repeated many times with small variations, forming a collection of total length N. Examples of such collections are version control data and genome sequences of individuals, where the differences can be expressed by lists of basic edit operations. Flexible and efficient data analysis on a such typically huge collection is plausible using suffix trees. However, suffix tree occupies O(N log N) bits, which very soon inhibits in-memory analyses. Recent advances in full-text self-indexing reduce the space of suffix tree to O(N log σ) bits, where σ is the alphabet size. In practice, the space reduction is more than 10-fold, for example on suffix tree of Human Genome. However, this reduction factor remains constant when more sequences are added to the collection. We develop a new family of self-indexes suited for the repetitive sequence collection setting. Their expected space requirement depends only on the length n of the base sequence and the number s of variations in its repeated copies. That is, the space reduction factor is no longer constant, but depends on N / n. We believe the structures developed in this work will provide a fundamental basis for storage and retrieval of individual genomes as they become available due to rapid progress in the sequencing technologies.

Collecting data from distributed FOSS projects

A key trait of Free and Open Source Software (FOSS) development is its distributed nature. Nevertheless, two project-level operations, the fork and the merge of program code, are among the least well understood events in the lifespan of a FOSS project. Some projects have explicitly adopted these operations as the primary means of concurrent development. In this study, we examine the effect of highly distributed software development, is found in the Linux kernel project, on collection and modelling of software development data. We find that distributed development calls for sophisticated temporal modelling techniques where several versions of the source code tree can exist at once. Attention must be turned towards the methods of quality assurance and peer review that projects employ to manage these parallel source trees. Our analysis indicates that two new metrics, fork rate and merge rate, could be useful for determining the role of distributed version control systems in FOSS projects. The study presents a preliminary data set consisting of version control and mailing list data.

Collecting data from distributed FOSS projects

A key trait of Free and Open Source Software (FOSS) development is its distributed nature. Nevertheless, two project-level operations, the fork and the merge of program code, are among the least well understood events in the lifespan of a FOSS project. Some projects have explicitly adopted these operations as the primary means of concurrent development. In this study, we examine the effect of highly distributed software development, is found in the Linux kernel project, on collection and modelling of software development data. We find that distributed development calls for sophisticated temporal modelling techniques where several versions of the source code tree can exist at once. Attention must be turned towards the methods of quality assurance and peer review that projects employ to manage these parallel source trees. Our analysis indicates that two new metrics, fork rate and merge rate, could be useful for determining the role of distributed version control systems in FOSS projects. The study presents a preliminary data set consisting of version control and mailing list data.

Data fusion and matching by maximizing statistical dependencies

The core aim of machine learning is to make a computer program learn from the experience. Learning from data is usually defined as a task of learning regularities or patterns in data in order to extract useful information, or to learn the underlying concept. An important sub-field of machine learning is called multi-view learning where the task is to learn from multiple data sets or views describing the same underlying concept. A typical example of such scenario would be to study a biological concept using several biological measurements like gene expression, protein expression and metabolic profiles, or to classify web pages based on their content and the contents of their hyperlinks. In this thesis, novel problem formulations and methods for multi-view learning are presented. The contributions include a linear data fusion approach during exploratory data analysis, a new measure to evaluate different kinds of representations for textual data, and an extension of multi-view learning for novel scenarios where the correspondence of samples in the different views or data sets is not known in advance. In order to infer the one-to-one correspondence of samples between two views, a novel concept of multi-view matching is proposed. The matching algorithm is completely data-driven and is demonstrated in several applications such as matching of metabolites between humans and mice, and matching of sentences between documents in two languages.

Patterns in permuted binary matrices

Reorganizing a dataset so that its hidden structure can be observed is useful in any data analysis task. For example, detecting a regularity in a dataset helps us to interpret the data, compress the data, and explain the processes behind the data. We study datasets that come in the form of binary matrices (tables with 0s and 1s). Our goal is to develop automatic methods that bring out certain patterns by permuting the rows and columns. We concentrate on the following patterns in binary matrices: consecutive-ones (C1P), simultaneous consecutive-ones (SC1P), nestedness, k-nestedness, and bandedness. These patterns reflect specific types of interplay and variation between the rows and columns, such as continuity and hierarchies. Furthermore, their combinatorial properties are interlinked, which helps us to develop the theory of binary matrices and efficient algorithms. Indeed, we can detect all these patterns in a binary matrix efficiently, that is, in polynomial time in the size of the matrix. Since real-world datasets often contain noise and errors, we rarely witness perfect patterns. Therefore we also need to assess how far an input matrix is from a pattern: we count the number of flips (from 0s to 1s or vice versa) needed to bring out the perfect pattern in the matrix. Unfortunately, for most patterns it is an NP-complete problem to find the minimum distance to a matrix that has the perfect pattern, which means that the existence of a polynomial-time algorithm is unlikely. To find patterns in datasets with noise, we need methods that are noise-tolerant and work in practical time with large datasets. The theory of binary matrices gives rise to robust heuristics that have good performance with synthetic data and discover easily interpretable structures in real-world datasets: dialectical variation in the spoken Finnish language, division of European locations by the hierarchies found in mammal occurrences, and co-occuring groups in network data. In addition to determining the distance from a dataset to a pattern, we need to determine whether the pattern is significant or a mere occurrence of a random chance. To this end, we use significance testing: we deem a dataset significant if it appears exceptional when compared to datasets generated from a certain null hypothesis. After detecting a significant pattern in a dataset, it is up to domain experts to interpret the results in the terms of the application.

Planar subgraphs without low-degree nodes

We study the following problem: given a geometric graph G and an integer k, determine if G has a planar spanning subgraph (with the original embedding and straight-line edges) such that all nodes have degree at least k. If G is a unit disk graph, the problem is trivial to solve for k = 1. We show that even the slightest deviation from the trivial case (e.g., quasi unit disk graphs or k = 1) leads to NP-hard problems.

Algorithms for Exact Structure Discovery in Bayesian Networks

Bayesian networks are compact, flexible, and interpretable representations of a joint distribution. When the network structure is unknown but there are observational data at hand, one can try to learn the network structure. This is called structure discovery. This thesis contributes to two areas of structure discovery in Bayesian networks: space--time tradeoffs and learning ancestor relations. The fastest exact algorithms for structure discovery in Bayesian networks are based on dynamic programming and use excessive amounts of space. Motivated by the space usage, several schemes for trading space against time are presented. These schemes are presented in a general setting for a class of computational problems called permutation problems; structure discovery in Bayesian networks is seen as a challenging variant of the permutation problems. The main contribution in the area of the space--time tradeoffs is the partial order approach, in which the standard dynamic programming algorithm is extended to run over partial orders. In particular, a certain family of partial orders called parallel bucket orders is considered. A partial order scheme that provably yields an optimal space--time tradeoff within parallel bucket orders is presented. Also practical issues concerning parallel bucket orders are discussed. Learning ancestor relations, that is, directed paths between nodes, is motivated by the need for robust summaries of the network structures when there are unobserved nodes at work. Ancestor relations are nonmodular features and hence learning them is more difficult than modular features. A dynamic programming algorithm is presented for computing posterior probabilities of ancestor relations exactly. Empirical tests suggest that ancestor relations can be learned from observational data almost as accurately as arcs even in the presence of unobserved nodes.