#### DMCA

## INDEPENDENCE ASSUMPTIONS FOR MULTI-RELATIONAL CLASSIFICATION

### Citations

8896 |
Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ... = x2)/P (X2 = x2). Bayes nets factorize the joint probability over an assignment of values to attributes, as a product over the probability of each attribute value given its parents value assignment =-=[39]-=-. CHAPTER 2. MULTI-RELATIONAL DATA CLASSIFICATION 11 The Markov blanket of a node is the set of its parents, children, and parents’ children. Every node in a BN is conditionally independent of others ... |

6599 |
C4.5: Programs for machine learning
- Quinlan
- 1993
(Show Context)
Citation Context ...tree, tree induction is fairly fast. TILDE (Top-down induction of first-order logical decision tree) [3] induces a logical decision tree from examples by a divide and conquer algorithm extending C4.5 =-=[44]-=- to relational data. Runtime efficiency is a challenge in Tilde as it potentially tests many clauses to find the best candidates. TILDE uses the existential quantifier to deal with the multi-valued at... |

1194 | Foundations of Inductive Logic Programming
- Muggleton
- 1995
(Show Context)
Citation Context ...approach is to upgrade traditional classification algorithms to deal with multirelational data directly. Most methods with upgrading approach fall in the category of Inductive Logic Programming (ILP) =-=[36, 35, 34]-=-. ILP systems like Claudien and ICL are first order upgrades of propositional data mining algorithms that induce association rules. The other classification models are relational rule learners [9] and... |

612 | Learning probabilistic relational models
- Friedman, Getoor, et al.
- 1999
(Show Context)
Citation Context ...ciation rules. The other classification models are relational rule learners [9] and relational decision trees [24] which upgrade their single table model respectively. Probabilistic Relational Models =-=[14]-=- and Bayesian Logic Programs [26] are extensions of Bayesian networks to relational domains, and Relational Markov Random Fields [50] and Markov Logic Networks [10] upgrade Markov Random Fields. Naive... |

415 | Discriminative probabilistic models for relational data
- Taskar, Abbeel, et al.
- 2002
(Show Context)
Citation Context ...r single table model respectively. Probabilistic Relational Models [14] and Bayesian Logic Programs [26] are extensions of Bayesian networks to relational domains, and Relational Markov Random Fields =-=[50]-=- and Markov Logic Networks [10] upgrade Markov Random Fields. Naive Bayes classifier is a well known single table classifier which is easy to learn and implement; relational naive Bayes classifier [6]... |

339 | Introduction to Statistical Relational Learning - Getoor, Taskar - 2007 |

334 |
Principles of Database System
- Ullman
- 1982
(Show Context)
Citation Context ...ng path join T on R1 · · · on Rk on Ek. The symbol on shows the natural join of two tables which is the set of tuples from their cross product that agree on the values of fields common to both tables =-=[51]-=-. The number of attributes in a valid join may become quite large, therefore, Han et al. use only the attributes of the last table in the path, and the class attribute from the target table, which is ... |

257 | FOIL: A midterm report
- Quinlan, Cameron-Jones
- 1993
(Show Context)
Citation Context ...ATIONAL DATA CLASSIFICATION 17 ILP classifier will derive a hypothesised logic program which entails all the positive and none of the negative examples. FOIL (a top down first-order inductive learner)=-=[45]-=- is one of the first and best known ILP systems in the public domain. Foil uses information in a collection of relations to construct existentially quantified logical rules in Prolog. CrossMine [55] i... |

191 |
First-order probabilistic inference.
- Poole
- 2003
(Show Context)
Citation Context ...ail. Parameterized Bayesian Network (PBN): Parametrized Bayes nets specifies a relational Bayes net as a template for a database probability distribution. We follow the original presentation of Poole =-=[40]-=-: A population is a set of individuals, corresponding to a domain or type in logic. A parametrized random variable is of the form f(t1, . . . , tk) where f is a functor (either a function symbol or a ... |

182 |
Link-based classification
- Lu, Getoor
- 2003
(Show Context)
Citation Context ...ti-relational data mining is link-based classification (LBC) which takes advantage of attributes of links and linked entities in addition to attributes of the target entity to predict the class label =-=[31]-=-. Algorithms for classifying a single table of data assume that instances are independent and identically distributed (i.i.d.), but relational data violate this assumption; objects in relational data ... |

161 | Tree induction for probability-based ranking,
- Provost, Domingos
- 2003
(Show Context)
Citation Context ...o select relevant features. The decision trees we consider provide probability estimates in their leaves, rather than simply class labels. Such trees are sometimes called probability estimation trees =-=[43]-=-. We learn a forest of decision trees that contains a tree over attributes of the target entity and a tree over attributes of each of the linked entity tables extended by the class label. To combine t... |

132 | Sound and efficient inference with probabilistic and deterministic dependencies
- Poon, Domingos
- 2006
(Show Context)
Citation Context ... procedure of Weka as its code is not available online. We used MLN with learn-and-join (LAJ) structure learning algorithm [27], discriminative parameter learning [22], and MC-SAT inference algorithm =-=[41]-=- implemented in Alchemy package. We used the LAJ algorithm because (1) its predictive accuracy outperforms other MLN structure learning algorithms, and (2) it is the only current MLN learning algorith... |

131 | Learning Bayesian networks from data: An information-theory based approach
- Cheng, Greiner, et al.
- 2002
(Show Context)
Citation Context ...lows us to use a classifier other than singletable NB to model the conditional distribution P (c(t)|Ji,r). Figure 3.4 illustrates the TNB Assumption for the university schema using a Bayesian network =-=[7]-=-. We now discuss the plausibility and limitations of our assumptions. Building decision trees on the extended tables is more natural than learning them on random subsamples of features like FORF [53].... |

122 | Learning probabilistic models of relational structure.
- Getoor, Friedman, et al.
- 2001
(Show Context)
Citation Context ...her flatten multiple relations into a single table, which is then used as the input of propositional learning methods [28, 49, 24, 42] or upgrade propositional algorithms to deal with relational data =-=[20, 6, 23, 11]-=-. Transforming data into a single table requires considerable time, results in the loss of information [6], and produces a large table with many additional attributes. Upgrading algorithms are faster ... |

118 | Linkage and autocorrelation cause feature selection bias in relational learning - Jensen, Neville - 2002 |

93 | Markov logic: A unifying framework for statistical relational learning.
- Domingos, Richardson
- 2007
(Show Context)
Citation Context ...ly. Probabilistic Relational Models [14] and Bayesian Logic Programs [26] are extensions of Bayesian networks to relational domains, and Relational Markov Random Fields [50] and Markov Logic Networks =-=[10]-=- upgrade Markov Random Fields. Naive Bayes classifier is a well known single table classifier which is easy to learn and implement; relational naive Bayes classifier [6] is its relational version with... |

84 | WEKA: The Waikato Environment for Knowledge Analysis,” in
- Garner
- 1995
(Show Context)
Citation Context ...ng Formula 3.1 5. PIBC: Path Independence Bayesian net Classifier. The implementation of the Simple Decision Forest used many of the procedures of Weka which is a data mining software written in Java =-=[15]-=-. We used J48 as a propositional decision tree learner on each join table. J48 implements the C4.5 decision tree algorithm. We used the probability estimation tree setting that turns off pruning and a... |

79 | Classification with hybrid generative/discriminative models.
- Raina, Shen, et al.
- 2003
(Show Context)
Citation Context ...] CHAPTER 3. DECISION FORESTS FOR RELATIONAL CLASSIFICATION 29 showed that the addition of the scaling weight factors (like rowsi) improves learning weights for Markov Logic Networks. McCallum et al. =-=[46]-=- observed that considering different scaled weights for information from different section of a text improves the results in the text classification problem. Algorithm 2 shows the classification algor... |

76 | 2003 Simple Estimators for Relational Bayesian Classifiers
- Neville, Jensen, et al.
(Show Context)
Citation Context ...plexity of correlation types is to consider model classes with explicitly stated independence assumptions. A prominent example of this approach are multi-relational Naive Bayes net classifiers (NBCs) =-=[37, 6]-=-. NBCs incorporate two different CHAPTER 1. INTRODUCTION 4 kinds of independence assumptions: 1. Across-table independencies: information from different tables is independent given the target class la... |

72 | CrossMine: efficient classification across multiple database relations.
- Yin, Han, et al.
- 2004
(Show Context)
Citation Context ...r)[45] is one of the first and best known ILP systems in the public domain. Foil uses information in a collection of relations to construct existentially quantified logical rules in Prolog. CrossMine =-=[55]-=- is an algorithm proposed to enhance the efficiency of rule-based multi relational classifiers. It uses tuple ID propagation to virtually join tables, and randomly select a subset of dominant instance... |

62 | Probabilistic models of text and link structure for hypertext classification
- Getoor, Segal, et al.
- 2001
(Show Context)
Citation Context ...nt tables is orders of magnitude faster to learn, as our experiments confirm. Probabilistic Relational Models (PRMs) [14] upgrade directed graphical models to deal with relational data. Getoor et al. =-=[18]-=- extend PRMs with link indicator nodes to model link structure between entities, and constrain the PRM such that the probability of a link indicator being true is a function of the attributes of the l... |

61 |
Muggleton and Luc de Raedt. Inductive logic programming: Theory and methods
- Stephen
- 1994
(Show Context)
Citation Context ...approach is to upgrade traditional classification algorithms to deal with multirelational data directly. Most methods with upgrading approach fall in the category of Inductive Logic Programming (ILP) =-=[36, 35, 34]-=-. ILP systems like Claudien and ICL are first order upgrades of propositional data mining algorithms that induce association rules. The other classification models are relational rule learners [9] and... |

56 | Discriminative structure and parameter learning for Markov logic networks.
- Huynh, Mooney
- 2008
(Show Context)
Citation Context ...del, and combined it with a Naive Bayes classifier. Unlike directed graphical models which impose an acyclicity constraint, undirected ones do not have a cyclicity problem and are widely used for LBC =-=[50, 10, 22]-=-. Undirected graphical models can represent essentially arbitrary dependencies [50] compared to directed models. The trade-off for the expressive power of undirected models is higher complexity in lea... |

52 | d-separation: From theorems to algorithms.
- Geiger, Verma, et al.
- 1990
(Show Context)
Citation Context ...its parents, children, and parents’ children. Every node in a BN is conditionally independent of others given its Markov blanket. For the general conditional independence in a Bayesian network, Pearl =-=[16]-=- proposed the dseparation concept. Two sets of nodes X and Y are d-separated in Bayesian networks by a third set Z, excluding X and Y, if and only if for every path between X and Y there is an interme... |

51 | An overview of classifier fusion methods,”
- Ruta, Gabrys
- 2000
(Show Context)
Citation Context ...this context Forest does not mean that we apply an CHAPTER 1. INTRODUCTION 5 ensemble method. For clarification, we provide a list of differences of our method with ensemble methods or fusion methods =-=[47]-=- in the following. 1. Ensemble methods are usually used on propositional data and divide the training set into a number of subsets randomly. We do not use random subsampling techniques to build a set ... |

49 | Finding optimal Bayesian networks
- Chickering, Meek
- 2002
(Show Context)
Citation Context ...puted by decision trees by the simple logistic procedure of Weka. To learn the structure under the Path Independence Assumption, for a single table Bayes net learner we apply the GES search algorithm =-=[8]-=-. We use a generative learner because the Naive Bayes classifier is a generative model, so our experiments avoid conflating the impact of different independence assumptions with the impact of generati... |

43 | Information extraction and integration with Florid: The Mondial case study.
- May
- 1999
(Show Context)
Citation Context ...hip graph of this dataset is depicted in figure 5.2. The number of join tables in the extended database was 4. Mondial Database. This dataset contains data from multiple geographical web data sources =-=[33]-=-. We predict the religion of a country as Christian (positive) with 114 instances vs. all other 40 CHAPTER 5. EVALUATION 41 Figure 5.1: Semantic Relationship Graph for the Financial Dataset. Figure 5.... |

37 | MRDTL: a multi-relational decision tree learning algorithm
- Leiva
- 2002
(Show Context)
Citation Context ...her flatten multiple relations into a single table, which is then used as the input of propositional learning methods [28, 49, 24, 42] or upgrade propositional algorithms to deal with relational data =-=[20, 6, 23, 11]-=-. Transforming data into a single table requires considerable time, results in the loss of information [6], and produces a large table with many additional attributes. Upgrading algorithms are faster ... |

33 | Naive bayesian classification of structured data,”
- Flach, Lachiche
- 2004
(Show Context)
Citation Context ...r assumes that different attributes are independent given the class label. There has been extensive research on multi-relational probabilistic versions of the single-table naive Bayes classifier(NBC) =-=[5, 12, 37, 6]-=-. RNBC is based on the independent value assumption(IVA). To explain the intuition behind the independent value assumption, it is helpful to consider multi-relational classification in terms of multis... |

29 | Logical Bayesian networks and their relation to other probabilistic logical models - Fierens, Blockeel, et al. |

27 |
de Raedt, L.: Bayesian logic programs
- Kersting
- 2000
(Show Context)
Citation Context ...cation models are relational rule learners [9] and relational decision trees [24] which upgrade their single table model respectively. Probabilistic Relational Models [14] and Bayesian Logic Programs =-=[26]-=- are extensions of Bayesian networks to relational domains, and Relational Markov Random Fields [50] and Markov Logic Networks [10] upgrade Markov Random Fields. Naive Bayes classifier is a well known... |

22 |
Blockeel and Luc De Raedt. Top-down induction of logical decision trees
- Hendrik
- 1998
(Show Context)
Citation Context ...expands the feature space and makes it time-consuming to find the best feature in each node of the tree, tree induction is fairly fast. TILDE (Top-down induction of first-order logical decision tree) =-=[3]-=- induces a logical decision tree from examples by a divide and conquer algorithm extending C4.5 [44] to relational data. Runtime efficiency is a challenge in Tilde as it potentially tests many clauses... |

19 | Raedt. How to upgrade propositional learners to first order logic: A case study
- Laer, De
- 2001
(Show Context)
Citation Context ...: 1 CHAPTER 1. INTRODUCTION 2 1. Flattening: join of all the tables to form a single table. 2. Feature construction: create new attributes that summarize or aggregate information from different links =-=[28, 49, 24, 42]-=-. The other approach is to upgrade traditional classification algorithms to deal with multirelational data directly. Most methods with upgrading approach fall in the category of Inductive Logic Progra... |

16 |
Structure learning for markov logic networks with many descriptive attributes
- Khosravi, Man, et al.
(Show Context)
Citation Context ...onal dependencies in directed models. The trade-off for the expressive power of undirected models is the higher complexity in learning, especially scalable model learning is a major challenge in RMNs =-=[27]-=-. Markov Logic Network (MLN): Markov logic networks combine first-order logic with probability theory. An MLN is a collection of formulas from first order logic with a real number, weight, assigned to... |

14 |
First order random forests with complex aggregates. In: ILP,
- Vens, AV, et al.
- 2004
(Show Context)
Citation Context .... We compared our proposed methods with the following Multi-relational classifiers: • TILDE : Top-down Induction of Logical Decision Trees [3]. • FORF-NA: First Order Random Forest with No Aggregates =-=[4]-=-. • Graph-NB: multi-relational naive Bayes classifier [30]. • MLN: Markov Logic Network [10] TILDE and FORF-NA are included in the ACE data mining system [2]. We run TILDE with the default setting. Fo... |

12 | A method for multi-relational classification using single and multi-feature aggregation functions
- Frank, Moser, et al.
- 2007
(Show Context)
Citation Context ...balanced. The number of join tables in the extended database was 8. Hepatitis Database. This dataset is a modified version of the PKDD’02 Discovery Challenge database. We followed the modification of =-=[13]-=-. Biopsy is the target table with 206 instances of Hepatitis B, and 484 cases of Hepatitis C. The inhospital table was modified such that we put each unique test in a column and all tests for a patien... |

10 |
Exploring optimization of semantic relationship graph for multi-relational Bayesian classification. Decision Support Systems
- Chen, Liu, et al.
- 2009
(Show Context)
Citation Context ...50] and Markov Logic Networks [10] upgrade Markov Random Fields. Naive Bayes classifier is a well known single table classifier which is easy to learn and implement; relational naive Bayes classifier =-=[6]-=- is its relational version with one more independence assumption. This thesis describes two new upgrading classification algorithms: Simple Decision Forest that extends decision tree and Path Independ... |

7 |
Luc De Raedt. nfoil: Integrating näıve bayes and foil
- Landwehr, Kersting
- 2005
(Show Context)
Citation Context ... for relational feature generation. Feature generation is done by using aggregate functions to summarize the information in links [28, 49, 24, 42] or learning an existentially quantified logical rule =-=[29]-=-. Several recent systems combine both aggregation and logical conditions (e.g., [53, 42]). However, the predictors in our model are the descriptive attributes as defined in the relational database sch... |

7 | An efficient multi-relational näıve Bayesian classifier based on semantic relationship graph
- Liu, Yin, et al.
- 2005
(Show Context)
Citation Context ...ti-relational classifiers: • TILDE : Top-down Induction of Logical Decision Trees [3]. • FORF-NA: First Order Random Forest with No Aggregates [4]. • Graph-NB: multi-relational naive Bayes classifier =-=[30]-=-. • MLN: Markov Logic Network [10] TILDE and FORF-NA are included in the ACE data mining system [2]. We run TILDE with the default setting. For FORF-NA we used the out-of-bag setting, chose 25% of fea... |

7 |
Relational rule induction with cprogol4.4: A tutorial introduction
- Muggleton, Firth
- 2001
(Show Context)
Citation Context ...approach is to upgrade traditional classification algorithms to deal with multirelational data directly. Most methods with upgrading approach fall in the category of Inductive Logic Programming (ILP) =-=[36, 35, 34]-=-. ILP systems like Claudien and ICL are first order upgrades of propositional data mining algorithms that induce association rules. The other classification models are relational rule learners [9] and... |

7 | Pushing feature selection ahead of join
- She, Wang, et al.
- 2005
(Show Context)
Citation Context ...N 41 Figure 5.1: Semantic Relationship Graph for the Financial Dataset. Figure 5.2: Semantic Relationship Graph for the Hepatitis Dataset. religions with 71 instances. We followed the modification of =-=[48]-=-. To make the classification task more challenging for evaluating learners, we use a subset of the tables and features. Border is a relationship between Country and Country. To create a semantic relat... |

6 |
An efficient relational decision tree classification algorithm
- Guo, Li, et al.
(Show Context)
Citation Context ...e database with a set of conditions on the attributes of the table, and each edge shows that there is at least one record that respects the sets of conditions of the corresponding nodes. Guo et al in =-=[21]-=- speed up the MRDT algorithm by using id propagation to implement a virtual join operation that avoids the cost of physical joins. FORF (First Order Random Forest)[53] is an ensemble of different deci... |

6 | Exploring the power of heuristics and links in multirelational data mining
- Yin, Han
- 2008
(Show Context)
Citation Context ...he multi-relational setting [22]. One of the issues that makes LBC difficult compared to single-table learning is the large number of different types of dependencies that a model may have to consider =-=[54]-=-. A principled way to approach the complexity of correlation types is to consider model classes with explicitly stated independence assumptions. A prominent example of this approach are multi-relation... |

5 |
Feature generation and selection in multirelational learning
- Popescul, Ungar
(Show Context)
Citation Context ...: 1 CHAPTER 1. INTRODUCTION 2 1. Flattening: join of all the tables to form a single table. 2. Feature construction: create new attributes that summarize or aggregate information from different links =-=[28, 49, 24, 42]-=-. The other approach is to upgrade traditional classification algorithms to deal with multirelational data directly. Most methods with upgrading approach fall in the category of Inductive Logic Progra... |

3 |
Annalisa Appice, and Donato Malerba. Mr-sbc: A multi-relational naive Bayes classifier
- Ceci
- 2003
(Show Context)
Citation Context ...r assumes that different attributes are independent given the class label. There has been extensive research on multi-relational probabilistic versions of the single-table naive Bayes classifier(NBC) =-=[5, 12, 37, 6]-=-. RNBC is based on the independent value assumption(IVA). To explain the intuition behind the independent value assumption, it is helpful to consider multi-relational classification in terms of multis... |

3 |
de Raedt. Logical and Relational Learning. Cognitive Technologies
- Luc
- 2008
(Show Context)
Citation Context ...35, 34]. ILP systems like Claudien and ICL are first order upgrades of propositional data mining algorithms that induce association rules. The other classification models are relational rule learners =-=[9]-=- and relational decision trees [24] which upgrade their single table model respectively. Probabilistic Relational Models [14] and Bayesian Logic Programs [26] are extensions of Bayesian networks to re... |

1 |
Learning relational probability trees
- Jennifer, David, et al.
- 2003
(Show Context)
Citation Context ...: 1 CHAPTER 1. INTRODUCTION 2 1. Flattening: join of all the tables to form a single table. 2. Feature construction: create new attributes that summarize or aggregate information from different links =-=[28, 49, 24, 42]-=-. The other approach is to upgrade traditional classification algorithms to deal with multirelational data directly. Most methods with upgrading approach fall in the category of Inductive Logic Progra... |

1 |
A heterogeneous naive-bayesian classifier for relational databases
- Manjunath, Murty, et al.
- 2009
(Show Context)
Citation Context ... the class label to find the most probable class label. Data mining models tend to present independence assumptions in terms of database tables and tuples. Heterogeneous Naive Bayes classifier (HNBC) =-=[32]-=- considers different tables independent of each other given the class label. Its main difference with other relational naive Bayes classifiers is that the naive Bayes assumption is only applied betwee... |

1 |
On discriminative versus generative classifiers: An empirical comparison of logistic regression and Naive Bayes
- Ng, Jordan
- 2002
(Show Context)
Citation Context ...ifiers: Logistic regression fits a curve to a set of data by learning weights for different features. It can be viewed as a discriminative version of the generative naive Bayes classifier (NBC) model =-=[38]-=-. For single-table classification, the advantages of logistic regression over simple CHAPTER 2. MULTI-RELATIONAL DATA CLASSIFICATION 21 NBC have been studied in detail [38], and similar results have b... |

1 |
Propositionalization approaches to relational data mining
- Stefan, Nada, et al.
- 2000
(Show Context)
Citation Context |

1 |
Raedt Luc De, and Dzeroski Saso. On multi-class problems and discretization in inductive logic programming
- Van
- 1997
(Show Context)
Citation Context ...tional classifiers. It uses tuple ID propagation to virtually join tables, and randomly select a subset of dominant instances to overcome the skewness of datasets. ICL (Inductive Classification Logic)=-=[52]-=- is an ILP learning system that learns first order logic formulas from examples which belong to two or more classes. There are two practical barriers that ILP systems may face applied to mining of lar... |