Skip to main content

Über dieses Buch

This three-volume set LNAI 6911, LNAI 6912, and LNAI 6913 constitutes the refereed proceedings of the European conference on Machine Learning and Knowledge Discovery in Databases: ECML PKDD 2011, held in Athens, Greece, in September 2011. The 121 revised full papers presented together with 10 invited talks and 11 demos in the three volumes, were carefully reviewed and selected from about 600 paper submissions. The papers address all areas related to machine learning and knowledge discovery in databases as well as other innovative application domains such as supervised and unsupervised learning with some innovative contributions in fundamental issues; dimensionality reduction, distance and similarity learning, model learning and matrix/tensor analysis; graph mining, graphical models, hidden markov models, kernel methods, active and ensemble learning, semi-supervised and transductive learning, mining sparse representations, model learning, inductive logic programming, and statistical learning. a significant part of the papers covers novel and timely applications of data mining and machine learning in industrial domains.



Regular Papers

Common Substructure Learning of Multiple Graphical Gaussian Models

Learning underlying mechanisms of data generation is of great interest in the scientific and engineering fields amongst others. Finding dependency structures among variables in the data is one possible approach for the purpose, and is an important task in data mining. In this paper, we focus on learning dependency substructures shared by multiple datasets. In many scenarios, the nature of data varies due to a change in the surrounding conditions or non-stationary mechanisms over the multiple datasets. However, we can also assume that the change occurs only partially and some relations between variables remain unchanged. Moreover, we can expect that such commonness over the multiple datasets is closely related to the invariance of the underlying mechanism. For example, errors in engineering systems are usually caused by faults in the sub-systems with the other parts remaining healthy. In such situations, though anomalies are observed in sensor values, the underlying invariance of the healthy sub-systems is still captured by some steady dependency structures before and after the onset of the error. We propose a structure learning algorithm to find such invariances in the case of Graphical Gaussian Models (GGM). The proposed method is based on a block coordinate descent optimization, where subproblems can be solved efficiently by existing algorithms for


and the

continuous quadratic knapsack problem

. We confirm the validity of our approach through numerical simulations and also in applications with real world datasets extracted from the analysis of city-cycle fuel consumption and anomaly detection in car sensors.

Satoshi Hara, Takashi Washio

Mining Research Topic-Related Influence between Academia and Industry

Recently the problem of mining social influence has attracted lots of attention. Given a social network, researchers are interested in problems such as how influence, ideas, information propagate in the network. Similar problems have been proposed on co-authorship networks where the goal is to differentiate the social influences on research topic level and quantify the strength of the influence. In this work, we are interested in the problem of mining topic-specific influence between academia and industry. More specifically, given a co-authorship network, we want to identify which academia researcher is most influential to a given company on specific research topics. Given pairwise influences between researchers, we propose three models (simple additive model, weighted additive model and clustering-based additive model) to evaluate how influential a researcher is to a company. Finally, we illustrate the effectiveness of these three models on real large data set as well as on simulated data set.

Dan He

Typology of Mixed-Membership Models: Towards a Design Method

Presents an analysis of the structure of mixed-membership models into elementary blocks and their numerical properties. By associating such model structures with structures known or assumed in the data, we propose how models can be constructed in a controlled way, using the numerical properties of data likelihood and Gibbs full conditionals as predictors of model behavior. To illustrate this “bottom-up” design method, example models are constructed that may be used for expertise finding from labeled data.

Gregor Heinrich

ShiftTree: An Interpretable Model-Based Approach for Time Series Classification

Efficient algorithms of time series data mining have the common denominator of utilizing the special time structure of the attributes of time series. To accommodate the information of time dimension into the process, we propose a novel instance-level cursor based indexing technique, which is combined with a decision tree algorithm. This is beneficial for several reasons: (a) it is insensitive to the time level noise (for example rendering, time shifting), (b) its working method can be interpreted, making the explanation of the classification process more understandable, and (c) it can manage time series of different length. The implemented algorithm named ShiftTree is compared to the well-known instance-based time series classifier 1-NN using different distance metrics, used over all 20 datasets of a public benchmark time series database and two more public time series datasets. On these benchmark datasets, our experiments show that the new model-based algorithm has an average accuracy slightly better than the most efficient instance-based methods, and there are multiple datasets where our model-based classifier exceeds the accuracy of instance-based methods. We also evaluated our algorithm via blind testing on the 20 datasets of the SIGKDD 2007 Time Series Classification Challenge. To improve the model accuracy and to avoid model overfitting, we provide forest methods as well.

Balázs Hidasi, Csaba Gáspár-Papanek

Image Classification for Age-related Macular Degeneration Screening Using Hierarchical Image Decompositions and Graph Mining

Age-related Macular Degeneration (AMD) is the most common cause of adult blindness in the developed world. This paper describes a new image mining technique to perform automated detection of AMD from colour fundus photographs. The technique comprises a novel hierarchical image decomposition mechanism founded on a circular and angular partitioning. The resulting decomposition is then stored in a tree structure to which a weighted frequent sub-tree mining algorithm is applied. The identified sub-graphs are then incorporated into a feature vector representation (one vector per image) to which classification techniques can be applied. The results show that the proposed approach performs both efficiently and accurately.

Mohd Hanafi Ahmad Hijazi, Chuntao Jiang, Frans Coenen, Yalin Zheng

Online Structure Learning for Markov Logic Networks

Most existing learning methods for Markov Logic Networks (MLNs) use batch training, which becomes computationally expensive and eventually infeasible for large datasets with thousands of training examples which may not even all fit in main memory. To address this issue, previous work has used online learning to train MLNs. However, they all assume that the model’s structure (set of logical clauses) is given, and only learn the model’s parameters. However, the input structure is usually incomplete, so it should also be updated. In this work, we present OSL—the first algorithm that performs both online structure and parameter learning for MLNs. Experimental results on two real-world datasets for natural-language field segmentation show that OSL outperforms systems that cannot revise structure.

Tuyen N. Huynh, Raymond J. Mooney

Fourier-Information Duality in the Identity Management Problem

We compare two recently proposed approaches for representing probability distributions over the space of permutations in the context of multi-target tracking. We show that these two representations, the Fourier approximation and the information form approximation can both be viewed as low dimensional projections of a true distribution, but with respect to different metrics. We identify the strengths and weaknesses of each approximation, and propose an algorithm for converting between the two forms, allowing for a


approach that draws on the strengths of both representations. We show experimental evidence that there are situations where hybrid algorithms are favorable.

Xiaoye Jiang, Jonathan Huang, Leonidas Guibas

Eigenvector Sensitive Feature Selection for Spectral Clustering

Spectral clustering is one of the most popular methods for data clustering, and its performance is determined by the quality of the eigenvectors of the related graph Laplacian. Generally, graph Laplacian is constructed using the full features, which will degrade the quality of the related eigenvectors when there are a large number of noisy or irrelevant features in datasets. To solve this problem, we propose a novel unsupervised feature selection method inspired by perturbation analysis theory, which discusses the relationship between the perturbation of the eigenvectors of a matrix and its elements’ perturbation. We evaluate the importance of each feature based on the average


1 norm of the perturbation of the first


eigenvectors of graph Laplacian corresponding to the k smallest positive eigenvalues, with respect to the feature’s perturbation. Extensive experiments on several high-dimensional multi-class datasets demonstrate the good performance of our method compared with some state-of-the-art unsupervised feature selection methods.

Yi Jiang, Jiangtao Ren

Restricted Deep Belief Networks for Multi-view Learning

Deep belief network (DBN) is a probabilistic generative model with multiple layers of hidden nodes and a layer of visible nodes, where parameterizations between layers obey harmonium or restricted Boltzmann machines (RBMs). In this paper we present

restricted deep belief network

(RDBN) for multi-view learning, where each layer of hidden nodes is composed of




hidden nodes, in order to learn individual and shared hidden spaces from multiple views of data. View-specific hidden nodes are connected to corresponding view-specific hidden nodes in the lower-layer or visible nodes involving a specific view, whereas shared hidden nodes follow inter-layer connections without restrictions as in standard DBNs. RDBN is trained using layer-wise contrastive divergence learning. Numerical experiments on synthetic and real-world datasets demonstrate the useful behavior of the RDBN, compared to the multi-wing harmonium (MWH) which is a two-layer undirected model.

Yoonseop Kang, Seungjin Choi

Motion Segmentation by Model-Based Clustering of Incomplete Trajectories

In this paper, we present a framework for visual object tracking based on clustering trajectories of image key points extracted from a video. The main contribution of our method is that the trajectories are automatically extracted from the video sequence and they are provided directly to a model-based clustering approach. In most other methodologies, the latter constitutes a difficult part since the resulting feature trajectories have a short duration, as the key points disappear and reappear due to occlusion, illumination, viewpoint changes and noise. We present here a sparse, translation invariant regression mixture model for clustering trajectories of variable length. The overall scheme is converted into a Maximum A Posteriori approach, where the Expectation-Maximization (EM) algorithm is used for estimating the model parameters. The proposed method detects the different objects in the input image sequence by assigning each trajectory to a cluster, and simultaneously provides the motion of all objects. Numerical results demonstrate the ability of the proposed method to offer more accurate and robust solution in comparison with the mean shift tracker, especially in cases of occlusions.

Vasileios Karavasilis, Konstantinos Blekas, Christophoros Nikou

PTMSearch: A Greedy Tree Traversal Algorithm for Finding Protein Post-Translational Modifications in Tandem Mass Spectra

Peptide identification by tandem mass spectrometry (MS/MS) and database searching is becoming the standard high-throughput technology in many areas of the life sciences. The analysis of post-translational modifications (PTMs) is a major source of complications in this area, which calls for efficient computational approaches. In this paper we describe PTMSearch, a novel algorithm in which the PTM search space is represented by a tree structure, and a greedy traversal algorithm is used to identify a path within the tree that corresponds to the PTMs that best fit the input data. Tests on simulated and real (experimental) PTMs show that the algorithm performs well in terms of speed and accuracy. Estimates are given for the error caused by the greedy heuristics, for the size of the search space and a scheme is presented for the calculation of statistical significance.

Attila Kertész-Farkas, Beáta Reiz, Michael P. Myers, Sándor Pongor

Efficient Mining of Top Correlated Patterns Based on Null-Invariant Measures

Mining strong correlations from transactional databases often leads to more meaningful results than mining association rules. In such mining, null (transaction)-invariance is an important property of the correlation measures. Unfortunately, some useful null-invariant measures such as




, which can discover correlations even for the very unbalanced cases, lack the (anti)-monotonicity property. Thus, they could only be applied to frequent itemsets as the post-evaluation step. For large datasets and for low supports, this approach is computationally prohibitive. This paper presents new properties for all known null-invariant measures. Based on these properties, we develop efficient pruning techniques and design the Apriori-like algorithm


for mining strongly correlated patterns


. We develop both the threshold-bounded and the top-


variations of the algorithm, where top-


is used when the optimal correlation threshold is not known in advance and to give user control over the output size. We test


on real-life datasets from different application domains, using


as an example of the null-invariant correlation measure. We show that


outperforms support-based approach more than an order of magnitude, and that it is very useful for discovering top correlations in itemsets with low support.

Sangkyum Kim, Marina Barsky, Jiawei Han

Smooth Receiver Operating Characteristics (smROC) Curves

Supervised learning algorithms perform common tasks including classification, ranking, scoring, and probability estimation. We investigate how scoring information, often produced by these models, is utilized by an evaluation measure. The ROC curve represents a visualization of the ranking performance of classifiers. However, they ignore the scores which can be quite informative. While this ignored information is less precise than that given by probabilities, it is much more detailed than that conveyed by ranking. This paper presents a novel method to weight the ROC curve by these scores. We call it the Smooth ROC (


) curve, and we demonstrate how it can be used to visualize the performance of learning models. We report experimental results to show that the


is appropriate for measuring performance similarities and differences between learning models, and is more sensitive to performance characteristics than the standard ROC curve.

William Klement, Peter Flach, Nathalie Japkowicz, Stan Matwin

A Boosting Approach to Multiview Classification with Cooperation

In many fields, such as bioinformatics or multimedia, data may be described using different sets of features (or views) which carry either global or local information. Some learning tasks make use of these several views in order to improve overall predictive power of classifiers through fusion-based methods. Usually, these approaches rely on a weighted combination of classifiers (or selected descriptions), where classifiers are learned independently. One drawback of these methods is that the classifier learned on one view does not communicate its failures within the other views. This paper deals with a novel approach to integrate multiview information. The proposed algorithm, named Mumbo, is based on boosting. Within the boosting scheme, Mumbo maintains one distribution of examples on each view, and at each round, it learns one weak classifier on each view. Within a view, the distribution of examples evolves both with the ability of the dedicated classifier to deal with examples of the corresponding features space, and with the ability of classifiers in other views to process the same examples within their own description spaces. Hence, the principle is to slightly remove the hard examples from the learning space of one view, while their weights get higher in the other views. This way, we expect that examples are urged to be processed by the most appropriate views, when possible. At the end of the iterative learning process, a final classifier is computed by a weighted combination of selected weak classifiers.

This paper provides the Mumbo algorithm in a multiclass and multiview setting, based on recent theoretical advances in boosting. The boosting properties of Mumbo are proved, as well as some results on its generalization capabilities. Several experimental results are reported which point out that complementary views may actually cooperate under some assumptions.

Sokol Koço, Cécile Capponi

ARTEMIS: Assessing the Similarity of Event-Interval Sequences

In several application domains, such as sign language, medicine, and sensor networks, events are not necessarily instantaneous but they can have a time duration. Sequences of interval-based events may contain useful domain knowledge; thus, searching, indexing, and mining such sequences is crucial. We introduce two distance measures for comparing sequences of interval-based events which can be used for several data mining tasks such as classification and clustering. The first measure maps each sequence of interval-based events to a set of vectors that hold information about all concurrent events. These sets are then compared using an existing dynamic programming method. The second method, called


, finds correspondence between intervals by mapping the two sequences into a bipartite graph. Similarity is inferred by employing the Hungarian algorithm. In addition, we present a linear-time lower-bound for


. The performance of both measures is tested on data from three domains: sign language, medicine, and sensor networks. Experiments show the superiority of


in terms of robustness to high levels of artificially introduced noise.

Orestis Kostakis, Panagiotis Papapetrou, Jaakko Hollmén

Unifying Guilt-by-Association Approaches: Theorems and Fast Algorithms

If several friends of


have committed petty thefts, what would you say about


? Most people would not be surprised if


is a hardened criminal.


methods combine weak signals to derive stronger ones, and have been extensively used for anomaly detection and classification in numerous settings (e.g., accounting fraud, cyber-security, calling-card fraud).

The focus of this paper is to compare and contrast several very successful,



Random Walk with Restarts

, Semi-Supervised Learning, and

Belief Propagation


Our main contributions are two-fold: (a) theoretically, we prove that all the methods result in a similar matrix inversion problem; (b) for practical applications, we developed


, a fast algorithm that yields 2× speedup, equal or higher accuracy than BP, and is guaranteed to converge. We demonstrate these benefits using synthetic and real datasets, including YahooWeb, one of the largest graphs ever studied with BP.

Danai Koutra, Tai-You Ke, U. Kang, Duen Horng (Polo) Chau, Hsing-Kuo Kenneth Pao, Christos Faloutsos

Online Clustering of High-Dimensional Trajectories under Concept Drift

Historical transaction data are collected in many applications, e.g., patient histories recorded by physicians and customer transactions collected by companies. An important question is the learning of models upon

the primary objects

(patients, customers) rather than the transactions, especially when these models are subjected to drift.

We address this problem by combining advances of online clustering on multivariate data with the trajectory mining paradigm. We model the measurements of each individual primary object (e.g. its transactions), taken at irregular time intervals, as a trajectory in a high-dimensional feature space. Then, we cluster individuals with similar trajectories to identify sub-populations that evolve similarly, e.g. groups of customers that evolve similarly or groups of employees that have similar careers.

We assume that the multivariate trajectories are generated by drifting Gaussian Mixture Models. We study (i) an EM-based approach that clusters these trajectories incrementally as a reference method that has access to all the data for learning, and propose (ii) an online algorithm based on a Kalman filter that efficiently tracks the trajectories of Gaussian clusters. We show that while both methods approximate the reference well, the algorithm based on a Kalman filter is faster by one order of magnitude compared to the EM-based approach.

Georg Krempl, Zaigham Faraz Siddiqui, Myra Spiliopoulou

Gaussian Logic for Predictive Classification

We describe a statistical relational learning framework called Gaussian Logic capable to work efficiently with combinations of relational and numerical data. The framework assumes that, for a fixed relational structure, the numerical data can be modelled by a multivariate normal distribution. We demonstrate how the Gaussian Logic framework can be applied to predictive classification problems. In experiments, we first show an application of the framework for the prediction of DNA-binding propensity of proteins. Next, we show how the Gaussian Logic framework can be used to find motifs describing highly correlated gene groups in gene-expression data which are then used in a set-level-based classification method.

Ondřej Kuželka, Andrea Szabóová, Matěj Holec, Filip Železný

Toward a Fair Review-Management System

Item reviews are a valuable source of information for potential buyers, who are looking for information on a product’s attributes before making a purchase decision. This search of information is often hindered by overwhelming numbers of available reviews, as well as low-quality and noisy content. While a significant amount of research has been devoted to filtering and organizing review corpora toward the benefit of the buyers, a crucial part of the reviewing process has been overlooked:

reviewer satisfaction

. As in every content-based system, the content-generators, in this case the reviewers, serve as the driving force. Therefore, keeping the reviewers satisfied and motivated to continue submitting high-quality content is essential. In this paper, we propose a system that helps potential buyers by focusing on high-quality and informative reviews, while keeping reviewers content and motivated.

Theodoros Lappas, Evimaria Terzi

Focused Multi-task Learning Using Gaussian Processes

Given a learning task for a data set, learning it together with related tasks (data sets) can improve performance. Gaussian process models have been applied to such multi-task learning scenarios, based on joint priors for functions underlying the tasks. In previous Gaussian process approaches, all tasks have been assumed to be of equal importance, whereas in transfer learning the goal is


: to enhance performance on a target task given all other tasks. In both settings, transfer learning and joint modelling,

negative transfer

is a key problem: performance may actually decrease if the tasks are not related closely enough. In this paper, we propose a Gaussian process model for the asymmetric setting, which learns to “explain away” non-related variation in the additional tasks, in order to focus on improving performance on the target task. In experiments, our model improves performance compared to single-task learning, symmetric multi-task learning using hierarchical Dirichlet processes, and transfer learning based on predictive structure learning.

Gayle Leen, Jaakko Peltonen, Samuel Kaski

Reinforcement Learning through Global Stochastic Search in N-MDPs

Reinforcement Learning (RL) in either fully or partially observable domains usually poses a requirement on the knowledge representation in order to be sound: the underlying stochastic process must be Markovian. In many applications, including those involving interactions between multiple agents (e.g., humans and robots), sources of uncertainty affect rewards and transition dynamics in such a way that a Markovian representation would be computationally very expensive. An alternative formulation of the decision problem involves partially specified behaviors with choice points. While this reduces the complexity of the policy space that must be explored - something that is crucial for realistic autonomous agents that must bound search time - it does render the domain Non-Markovian. In this paper, we present a novel algorithm for reinforcement learning in Non-Markovian domains. Our algorithm, Stochastic Search Monte Carlo, performs a global stochastic search in policy space, shaping the distribution from which the next policy is selected by estimating an upper bound on the value of each action. We experimentally show how, in challenging domains for RL, high-level decisions in Non-Markovian processes can lead to a behavior that is at least as good as the one learned by traditional algorithms, and can be achieved with significantly fewer samples.

Matteo Leonetti, Luca Iocchi, Subramanian Ramamoorthy

Analyzing Word Frequencies in Large Text Corpora Using Inter-arrival Times and Bootstrapping

Comparing frequency counts over texts or corpora is an important task in many applications and scientific disciplines. Given a text corpus, we want to test a hypothesis, such as “word X is frequent”, “word X has become more frequent over time”, or “word X is more frequent in male than in female speech”. For this purpose we need a null model of word frequencies. The commonly used bag-of-words model, which corresponds to a Bernoulli process with fixed parameter, does not account for any structure present in natural languages. Using this model for word frequencies results in large numbers of words being reported as unexpectedly frequent. We address how to take into account the inherent occurrence patterns of words in significance testing of word frequencies. Based on studies of words in two large corpora, we propose two methods for modeling word frequencies that both take into account the occurrence patterns of words and go beyond the bag-of-words assumption. The first method models word frequencies based on the spatial distribution of individual words in the language. The second method is based on bootstrapping and takes into account only word frequency at the text level. The proposed methods are compared to the current gold standard in a series of experiments on both corpora. We find that words obey different spatial patterns in the language, ranging from bursty to non-bursty/uniform, independent of their frequency, showing that the traditional approach leads to many false positives.

Jefrey Lijffijt, Panagiotis Papapetrou, Kai Puolamäki, Heikki Mannila

Discovering Temporal Bisociations for Linking Concepts over Time

Bisociations represent interesting relationships between seemingly unconnected concepts from two or more contexts. Most of the existing approaches that permit the discovery of bisociations from data rely on the assumption that contexts are static or considered as unchangeable domains. Actually, several real-world domains are intrinsically dynamic and can change over time. The same domain can change and can become completely different from what/how it was before: a dynamic domain observed at different time-points can present different representations and can be reasonably assimilated to a series of distinct static domains. In this work, we investigate the task of linking concepts from a dynamic domain through the discovery of bisociations which link concepts over time. This provides us with a means to unearth linkages which have not been discovered when observing the domain as static, but which may have developed over time, when considering the dynamic nature. We propose a computational solution which, assuming a time interval-based discretization of the domain, explores the spaces of association rules mined in the intervals and chains the rules on the basis of the concept generalization and information theory criteria. The application to the literature-based discovery shows how the method can re-discover known connections in biomedical terminology. Experiments and comparisons using alternative techniques highlight the additional peculiarities of this work.

Corrado Loglisci, Michelangelo Ceci

Minimum Neighbor Distance Estimators of Intrinsic Dimension

Most of the machine learning techniques suffer the “curse of dimensionality” effect when applied to high dimensional data. To face this limitation, a common preprocessing step consists in employing a dimensionality reduction technique. In literature, a great deal of research work has been devoted to the development of algorithms performing this task. Often, these techniques require as parameter the number of dimensions to be retained; to this aim, they need to estimate the “intrinsic dimensionality” of the given dataset, which refers to the minimum number of degrees of freedom needed to capture all the information carried by the data. Although many estimation techniques have been proposed, most of them fail in case of noisy data or when the intrinsic dimensionality is too high. In this paper we present a family of estimators based on the probability density function of the normalized nearest neighbor distance. We evaluate the proposed techniques on both synthetic and real datasets comparing their performances with those obtained by state of the art algorithms; the achieved results prove that the proposed methods are promising.

Gabriele Lombardi, Alessandro Rozza, Claudio Ceruti, Elena Casiraghi, Paola Campadelli

Graph Evolution via Social Diffusion Processes

We present a new stochastic process, called as Social Diffusion Process (SDP), to address the graph modeling. Based on this model, we derive a graph evolution algorithm and a series of graph-based approaches to solve machine learning problems, including clustering and semi-supervised learning. SDP can be viewed as a special case of

Matthew effect

, which is a general phenomenon in nature and societies. We use social event as a metaphor of the intrinsic stochastic process for broad range of data. We evaluate our approaches in a large number of frequently used datasets and compare our approaches to other state-of-the-art techniques. Results show that our algorithm outperforms the existing methods in most cases. We also applying our algorithm into the functionality analysis of microRNA and discover biologically interesting cliques. Due to the broad availability of graph-based data, our new model and algorithm potentially have applications in wide range.

Dijun Luo, Chris Ding, Heng Huang

Multi-Subspace Representation and Discovery

This paper presents the multi-subspace discovery problem and provides a theoretical solution which is guaranteed to recover the number of subspaces, the dimensions of each subspace, and the members of data points of each subspace simultaneously. We further propose a data representation model to handle noisy real world data. We develop a novel optimization approach to learn the presented model which is guaranteed to converge to global optimizers. As applications of our models, we first apply our solutions as preprocessing in a series of machine learning problems, including clustering, classification, and semi-supervised learning. We found that our method automatically obtains robust data presentation which preserves the affine subspace structures of high dimensional data and generate more accurate results in the learning tasks. We also establish a robust standalone classifier which directly utilizes our sparse and low rank representation model. Experimental results indicate our methods improve the quality of data by preprocessing and the standalone classifier outperforms some state-of-the-art learning approaches.

Dijun Luo, Feiping Nie, Chris Ding, Heng Huang

A Novel Stability Based Feature Selection Framework for k-means Clustering

Stability of a learning algorithm with respect to small input perturbations is an important property, as it implies the derived models to be robust with respect to the presence of noisy features and/or data sample fluctuations. In this paper we explore the effect of stability optimization in the standard feature selection process for the continuous (PCA-based) k-means clustering problem. Interestingly, we derive that stability maximization naturally introduces a tradeoff between cluster separation and variance, leading to the selection of features that have a high cluster separation index that is not artificially inflated by the feature’s variance. The proposed algorithmic setup is based on a Sparse PCA approach, that selects the features that maximize stability in a greedy fashion. In our study, we also analyze several properties of Sparse PCA relevant to stability that promote Sparse PCA as a viable feature selection mechanism for clustering. The practical relevance of the proposed method is demonstrated in the context of cancer research, where we consider the problem of detecting potential tumor biomarkers using microarray gene expression data. The application of our method to a leukemia dataset shows that the tradeoff between cluster separation and variance leads to the selection of features corresponding to important biomarker genes. Some of them have relative low variance and are not detected without the direct optimization of stability in Sparse PCA based k-means.

Dimitrios Mavroeidis, Elena Marchiori

Link Prediction via Matrix Factorization

We propose to solve the link prediction problem in graphs using a supervised matrix factorization approach. The model learns latent features from the topological structure of a (possibly directed) graph, and is shown to make better predictions than popular unsupervised scores. We show how these latent features may be combined with optional explicit features for nodes or edges, which yields better performance than using either type of feature exclusively. Finally, we propose a novel approach to address the class imbalance problem which is common in link prediction by directly optimizing for a ranking loss. Our model is optimized with stochastic gradient descent and scales to large graphs. Results on several datasets show the efficacy of our approach.

Aditya Krishna Menon, Charles Elkan

On Oblique Random Forests

In his original paper on random forests, Breiman proposed two different decision tree ensembles: one generated from “orthogonal” trees with thresholds on individual features in every split, and one from “oblique” trees separating the feature space by randomly oriented hyperplanes. In spite of a rising interest in the random forest framework, however, ensembles built from orthogonal trees (RF) have gained most, if not all, attention so far.

In the present work we propose to employ “oblique” random forests (oRF) built from multivariate trees which explicitly learn optimal split directions at internal nodes using linear discriminative models, rather than using random coefficients as the original oRF. This oRF outperforms RF, as well as other classifiers, on nearly all data sets but those with discrete factorial features. Learned node models perform distinctively better than random splits. An oRF feature importance score shows to be preferable over standard RF feature importance scores such as Gini or permutation importance. The topology of the oRF decision space appears to be smoother and better adapted to the data, resulting in improved generalization performance. Overall, the oRF propose here may be preferred over standard RF on most learning tasks involving numerical and spectral data.

Bjoern H. Menze, B. Michael Kelm, Daniel N. Splitthoff, Ullrich Koethe, Fred A. Hamprecht

An Alternating Direction Method for Dual MAP LP Relaxation

Maximum a-posteriori (MAP) estimation is an important task in many applications of probabilistic graphical models. Although finding an exact solution is generally intractable, approximations based on linear programming (LP) relaxation often provide good approximate solutions. In this paper we present an algorithm for solving the LP relaxation optimization problem. In order to overcome the lack of strict convexity, we apply an augmented Lagrangian method to the dual LP. The algorithm, based on the alternating direction method of multipliers (ADMM), is guaranteed to converge to the global optimum of the LP relaxation objective. Our experimental results show that this algorithm is competitive with other state-of-the-art algorithms for approximate MAP estimation.

Ofer Meshi, Amir Globerson

Aggregating Independent and Dependent Models to Learn Multi-label Classifiers

The aim of multi-label classification is to automatically obtain models able to tag objects with the labels that better describe them. Despite it could seem like any other classification task, it is widely known that exploiting the presence of certain correlations between labels helps to improve the classification performance. In other words, object descriptions are usually not enough to induce good models, also label information must be taken into account. This paper presents an aggregated approach that combines two groups of classifiers, one assuming independence between labels, and the other considering fully conditional dependence among them. The framework proposed here can be applied not only for multi-label classification, but also in multi-label ranking tasks. Experiments carried out over several datasets endorse the superiority of our approach with regard to other methods in terms of some evaluation measures, keeping competitiveness in terms of others.

Elena Montañés, José Ramón Quevedo, Juan José del Coz

Tensor Factorization Using Auxiliary Information

Most of the existing analysis methods for tensors (or multi-way arrays) only assume that tensors to be completed are of low rank. However, for example, when they are applied to tensor completion problems, their prediction accuracy tends to be significantly worse when only limited entries are observed. In this paper, we propose to use relationships among data as auxiliary information in addition to the low-rank assumption to improve the quality of tensor decomposition. We introduce two regularization approaches using graph Laplacians induced from the relationships, and design iterative algorithms for approximate solutions. Numerical experiments on tensor completion using synthetic and benchmark datasets show that the use of auxiliary information improves completion accuracy over the existing methods based only on the low-rank assumption, especially when observations are sparse.

Atsuhiro Narita, Kohei Hayashi, Ryota Tomioka, Hisashi Kashima

Kernels for Link Prediction with Latent Feature Models

Predicting new links in a network is a problem of interest in many application domains. Most of the prediction methods utilize information on the network’s entities such as nodes to build a model of links. Network structures are usually not used except for the networks with similarity or relatedness semantics. In this work, we use network structures for link prediction with a more general network type with latent feature models. The problem is the difficulty to train these models directly for large data. We propose a method to solve this problem using kernels and cast the link prediction problem into a binary classification problem. The key idea is not to infer latent features explicitly, but to represent these features implicitly in the kernels, making the method scalable to large networks. In contrast to the other methods for latent feature models, our method inherits all the advantages of kernel framework: optimality, efficiency and nonlinearity. We apply our method to real data of protein-protein interactions to show the merits of our method.

Canh Hao Nguyen, Hiroshi Mamitsuka

Frequency-Aware Truncated Methods for Sparse Online Learning

Online supervised learning with



-regularization has gained attention recently because it generally requires less computational time and a smaller space of complexity than batch-type learning methods. However, a simple



-regularization method used in an online setting has the side effect that rare features tend to be truncated more than necessary. In fact, feature frequency is highly skewed in many applications. We developed a new family of



-regularization methods based on the previous updates for loss minimization in linear online learning settings. Our methods can identify and retain low-frequency occurrence but informative features at the same computational cost and convergence rate as previous works. Moreover, we combined our methods with a cumulative penalty model to derive more robust models over noisy data. We applied our methods to several datasets and empirically evaluated the performance of our algorithms. Experimental results showed that our frequency-aware truncated models improved the prediction accuracy.

Hidekazu Oiwa, Shin Matsushima, Hiroshi Nakagawa

A Shapley Value Approach for Influence Attribution

Finding who and what is “important” is an ever-occurring question. Many methods that aim at characterizing important items or influential individuals have been developed in areas such as, bibliometrics, social-network analysis, link analysis, and web search. In this paper we study the problem of attributing influence scores to individuals who accomplish tasks in a collaborative manner. We assume that individuals build small teams, in different and diverse ways, in order to accomplish atomic tasks. For each task we are given an assessment of success or importance score, and the goal is to attribute those team-wise scores to the individuals. The challenge we face is that individuals in strong coalitions are favored against individuals in weaker coalitions, so the objective is to find fair attributions that account for such biasing. We propose an iterative algorithm for solving this problem that is based on the concept of Shapley value. The proposed method is applicable to a variety of scenarios, for example, attributing influence scores to scientists who collaborate in published articles, or employees of a company who participate in projects. Our method is evaluated on two real datasets: ISI Web of Science publication data and the Internet Movie Database.

Panagiotis Papapetrou, Aristides Gionis, Heikki Mannila

Fast Approximate Text Document Clustering Using Compressive Sampling

Document clustering involves repetitive scanning of a document set, therefore as the size of the set increases, the time required for the clustering task increases and may even become impossible due to computational constraints. Compressive sampling is a feature sampling technique that allows us to perfectly reconstruct a vector from a small number of samples, provided that the vector is sparse in some known domain. In this article, we apply the theory behind compressive sampling to the document clustering problem using k-means clustering. We provide a method of computing high accuracy clusters in a fraction of the time it would have taken by directly clustering the documents. This is performed by using the Discrete Fourier Transform and the Discrete Cosine Transform. We provide empirical results showing that compressive sampling provides a 14 times increase in speed with little reduction in accuracy on 7,095 documents, and we also provide a very accurate clustering of a 231,219 document set, providing 20 times increase in speed when compared to performing k-means clustering on the document set. This shows that compressive clustering is a very useful tool that can be used to quickly compute approximate clusters.

Laurence A. F. Park

Ancestor Relations in the Presence of Unobserved Variables

Bayesian networks (BNs) are an appealing model for causal and non-causal dependencies among a set of variables. Learning BNs from observational data is challenging due to the nonidentifiability of the network structure and model misspecification in the presence of unobserved (latent) variables. Here, we investigate the prospects of Bayesian learning of ancestor relations, including arcs, in the presence and absence of unobserved variables. An exact dynamic programming algorithm to compute the respective posterior probabilities is developed, under the complete data assumption. Our experimental results show that ancestor relations between observed variables, arcs in particular, can be learned with good power even when a majority of the involved variables are unobserved. For comparison, deduction of ancestor relations from single maximum a posteriori network structures or their Markov equivalence class appears somewhat inferior to Bayesian averaging. We also discuss some shortcomings of applying existing conditional independence test based methods for learning ancestor relations.

Pekka Parviainen, Mikko Koivisto

ShareBoost: Boosting for Multi-view Learning with Performance Guarantees

Algorithms combining multi-view information are known to exponentially quicken classification, and have been applied to many fields. However, they lack the ability to mine most discriminant information sources (or data types) for making predictions. In this paper, we propose an algorithm based on boosting to address these problems. The proposed algorithm builds base classifiers independently from each data type (view) that provides a partial view about an object of interest. Different from AdaBoost, where each view has its own re-sampling weight, our algorithm uses a single re-sampling distribution for all views at each boosting round. This distribution is determined by the view whose training error is minimal. This shared sampling mechanism restricts noise to individual views, thereby reducing sensitivity to noise. Furthermore, in order to establish performance guarantees, we introduce a randomized version of the algorithm, where a winning view is chosen probabilistically. As a result, it can be cast within a multi-armed bandit framework, which allows us to show that with high probability the algorithm seeks out most discriminant views of data for making predictions. We provide experimental results that show its performance against noise and competing techniques.

Jing Peng, Costin Barbu, Guna Seetharaman, Wei Fan, Xian Wu, Kannappan Palaniappan

Analyzing and Escaping Local Optima in Planning as Inference for Partially Observable Domains

Planning as inference recently emerged as a versatile approach to decision-theoretic planning and reinforcement learning for single and multi-agent systems in fully and partially observable domains with discrete and continuous variables. Since planning as inference essentially tackles a non-convex optimization problem when the states are partially observable, there is a need to develop techniques that can robustly escape local optima. We investigate the local optima of finite state controllers in single agent partially observable Markov decision processes (POMDPs) that are optimized by expectation maximization (EM). We show that EM converges to controllers that are optimal with respect to a one-step lookahead. To escape local optima, we propose two algorithms: the first one adds nodes to the controller to ensure optimality with respect to a multi-step lookahead, while the second one splits nodes in a greedy fashion to improve reward likelihood. The approaches are demonstrated empirically on benchmark problems.

Pascal Poupart, Tobias Lang, Marc Toussaint

Abductive Plan Recognition by Extending Bayesian Logic Programs

Plan recognition is the task of predicting an agent’s top-level plans based on its observed actions. It is an abductive reasoning task that involves inferring cause from effect. Most existing approaches to plan recognition use either first-order logic or probabilistic graphical models. While the former cannot handle uncertainty, the latter cannot handle structured representations. In order to overcome these limitations, we develop an approach to plan recognition using Bayesian Logic Programs (BLPs), which combine first-order logic and Bayesian networks. Since BLPs employ logical deduction to construct the networks, they cannot be used effectively for plan recognition. Therefore, we extend BLPs to use logical abduction to construct Bayesian networks and call the resulting model Bayesian Abductive Logic Programs (BALPs). We learn the parameters in BALPs using the Expectation Maximization algorithm adapted for BLPs. Finally, we present an experimental evaluation of BALPs on three benchmark data sets and compare its performance with the state-of-the-art for plan recognition.

Sindhu Raghavan, Raymond J. Mooney

Higher Order Contractive Auto-Encoder

We propose a novel regularizer when training an auto-encoder for unsupervised feature extraction. We explicitly encourage the latent representation to contract the input space by regularizing the norm of the Jacobian (analytically) and the Hessian (stochastically) of the encoder’s output with respect to its input, at the training points. While the penalty on the Jacobian’s norm ensures robustness to tiny corruption of samples in the input space, constraining the norm of the Hessian extends this robustness when moving further away from the sample. From a manifold learning perspective, balancing this regularization with the auto-encoder’s reconstruction objective yields a representation that varies most when moving along the data manifold in input space, and is most insensitive in directions orthogonal to the manifold. The second order regularization, using the Hessian, penalizes curvature, and thus favors smooth manifold. We show that our proposed technique, while remaining computationally efficient, yields representations that are significantly better suited for initializing deep architectures than previously proposed approaches, beating state-of-the-art performance on a number of datasets.

Salah Rifai, Grégoire Mesnil, Pascal Vincent, Xavier Muller, Yoshua Bengio, Yann Dauphin, Xavier Glorot

The VC-Dimension of SQL Queries and Selectivity Estimation through Sampling

We develop a novel method, based on the statistical concept of VC-dimension, for evaluating the selectivity (output cardinality) of SQL queries – a crucial step in optimizing the execution of large scale database and data-mining operations. The major theoretical contribution of this work, which is of independent interest, is an explicit bound on the VC-dimension of a range space defined by all possible outcomes of a collection (class) of queries. We prove that the VC-dimension is a function of the maximum number of Boolean operations in the selection predicate, and of the maximum number of select and join operations in any individual query in the collection, but it is neither a function of the number of queries in the collection nor of the size of the database. We develop a method based on this result: given a class of queries, it constructs a concise random sample of a database, such that with high probability the execution of


query in the class on the sample provides an accurate estimate for the selectivity of the query on the original large database. The error probability holds


for the selectivity estimates of


queries in the collection, thus the same sample can be used to evaluate the selectivity of multiple queries, and the sample needs to be refreshed only following major changes in the database. The sample representation computed by our method is typically sufficiently small to be stored in main memory. We present extensive experimental results, validating our theoretical analysis and demonstrating the advantage of our technique when compared to complex selectivity estimation techniques used in PostgreSQL and the Microsoft SQL Server.

Matteo Riondato, Mert Akdere, Uǧur Çetintemel, Stanley B. Zdonik, Eli Upfal


Weitere Informationen