H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....data has been extensively studied by the inductive logic programming (ILP) community [Lavrac and Dzeroski, 1994] In general, ILP learns models from a richer class than our work (for example, learning recursive concepts) but is also generally believed to be very inefficient for large databases. Blockeel et al. 1999] developed a scalable ILP system named TILDE that effectively flattens relational data into what they call interpretations and then uses a version of FOIL [Quinlan, 1990] modified to make efficient access to data from disk, on these interpretations. TILDE was evaluated on a synthetic data set ....

H. Blockeel, L. D. Raedt, and N. Jacobs. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.

....linear time with the number of examples and the number of features. Experiments on the traditional musk1 dataset provided good results using a 10 fold crossvalidation for NAIVE RIPPERMI (with an accuracy of 0.88) compared to ITERATED DISCRIMAPR [4] 0.92) DIVERSE DENSITY [6] 0. 89) TILDE [1] (0.87) ALL POS APR [4] 0.80) and MULTIINST (0.77) For a more detailed description of these experiments, see [2] 4 Analysis of RIPPERMI algorithms The purpose of this section is to analyze and to understand the behavior of the algorithm presented earlier as NAIVE RIPPERMI. This analysis will ....

....The average classification error is ploted on figure 1. For example, on datasets containing 15 instances per bag, NAIVERIPPERMI obtains an average classification error rate of 26.5 , and the induction phase lasts less than a three seconds on a Sun SparcStation 4 computer. The ILP learners TILDE [1] and FOIL were also run on these datasets in order to evaluate the ability of ILP tools on multiple instance data. The top curves of figure 1 show their accuracy with various numbers of instances per bag. On this particular task, they are outperformed by NAIVERIPPERMI in terms of accuracy. 4.2 ....

[Article contains additional citation context not shown here]

Hendrik Blockeel, Luc De Raedt, Nico Jacobs, and Bart Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.

....An important issue of ILP systems in general is its efficiency and scalability (mainly w.r.t. time complexity) For ILP systems based on our upgrading method, these can be addressed by applying (optimization) ideas from propos itional learning, data mining and databases. For example, in [Blockeel et at. 1999] one applies some ideas of [Mehta et al. 1996] to TILDE. And in WARMR some ideas of [Agrawal et at. 1996; Agrawal et at. 1993] have been applied (like turning the algorithm inside out) In Chapter 7 we come back to some of these efficiency issues. 3.11 Related Work Our upgrading method is ....

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

....for patterns. In particular, a refinement step consists of adding to the pattern to be refined one or more Datalog atoms in the language of patterns. Recently, the ILP setting of learning from interpretations has been proposed as a promising way of scaling up ILP algorithms in KDD applications [3]. The notion of spatial observation in SPADA, though only virtually supported in the current version of the system, adapts the notion of interpretation to the case of spatial databases. Example 5 The system SPADA has been run on the mining task in Example 1 with thresholds minsup[1] 0.3 and ....

....in the current version of the system, adapts the notion of interpretation to the case of spatial databases. Example 5 The system SPADA has been run on the mining task in Example 1 with thresholds minsup[1] 0.3 and minconf[1] 0.8 at level 1, minsup[2] 0.25 and minconf [2] 0. 7 at level 2, and minsup[3]=0.2 and minconf[3] 0.6 at level 3. The specification of the language bias and other settings is given in input to the system together with D(large town) The specification defines the alphabet A by listing all the valid atoms as facts like lb atom(close to(old ro, diff tro) It also supplies ....

[Article contains additional citation context not shown here]

Blockeel, H., De Raedt, L., Jacobs, N., Demoen, B.: Scaling Up Inductive Logic Programming by Learning from Interpretations. Data Mining and Knowledge Discovery 3(1) (1999) 59-93

....the sense that the key values k are retrieved one by one, the subsequent subdatabases r k are activated once in (2.a) and all queries are evaluated locally with respect to r k in (2. b) An experimental evaluation of this localisation of information in a related data mining task can be found in [7]. Consider as an example our database D with customer information. Each subset r k of this database would contain a fact customer(cid k ) and zero or more facts parent(wid k ,atype) This subset of the database indeed suffices for solving queries Q k built with predicates customer and ....

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.

....by Jun et al. 16] The algorithms mentioned here do not generate interpretable hypotheses such as rule sets, which is our purpose. In the following, a method for inducing multipleinstance rules with a modified traditional rule learner will be presented. Note that Blockeel and De Raedt [4] already presented a method for extending propositional learners to handle relational data. The extension of a propositional learner to the multiple instance case is less complex, and yields specific multiple instance issues, as will be shown in the following. 3 Extending a propositional learner ....

....datasets, in order to design improvements. Learner Musk1 Musk2 Model ITERATED DISCRIM APR [8] 0.92 0.89 Axis parallel Rectangle CITATION KNN [16] 0.92 0.86 k nearest neighbour DIVERSE DENSITY [12] 0.89 0.82 Points in # RIPPERMI 0.88 0.77 rule set NAIVE RIPPERMI 0.88 0. 77 rule set TILDE [4] 0.87 0.79 horn clauses ALL POS APR [8] 0.80 0.73 APR MULTIINST [3] 0.77 0.84 APR Table 1. Compared accuracy of MI learners on both musk datasets. 4 Analysis of RIPPERMI algorithms The purpose of this section is to analyze and to understand the behavior of the algorithm presented earlier as ....

[Article contains additional citation context not shown here]

Hendrik Blockeel, Luc De Raedt, Nico Jacobs, and Bart Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.

....express the importance of those items for that user. Various types of preference functions may exist. The type of a preference function characterises the structure 17 of profile (e.g. see [23] 37] Another profile learning approach, based on Inductive Logic Programming, can be found in [5] [6], 11] Several collaborative based recommendation systems have been introduced in which the preferences of users are modelled automatically. Examples of online recommendation systems that employ a collaborative approach are MovieFinder [39] and FireFly [13] The preferences of a user are ....

BLOCKEEL, H., De RAEDT, L. , JACOBS, N. , DEMOEN, B. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

....relationally, the results of the mining also may be represented relationally. Because learning relational descriptions is harder than learning propositional ones, relational algorithms are considerably slower than propositional ones, and the scaling problem is correspondingly harder. However, Blockeel, De Raedt, Jacobs, and Demoen (1999) observe that the full power of standard ILP is not used for most practical applications. Therefore, a general approach to speeding up learning with relational data is to avoid expensive but little used constructs. Aronis et al. 1996) investigate induction from featurevector based data items ....

....up learning with relational data is to avoid expensive but little used constructs. Aronis et al. 1996) investigate induction from featurevector based data items linked to relational background knowledge. For the sake of efficiency, they purposely avoid n ary and recursive relational terms. Blockeel et al. 1999) study an efficient subset of ILP known as learning from interpretations. In particular, they study scaling up first order logical decision trees (FOLDTs) which are more expressive than propositional decision trees, but also avoid the most expensive ILP constructs. Of particular note, FOLDTs ....

Blockeel, H., L. De Raedt, N. Jacobs, and B. Demoen (1999). Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery . To appear.

....Specifically, not only are the data represented relationally, the results of the mining also may be represented relationally. Because learning relational descriptions is harder than learning propositional ones, the scaling problem is correspondingly harder. Of particular note is the recent work of Blockeel, De Raedt, Jacobs, and Demoen (1999), who note that the full power of standard ILP is not used for most practical applications, and who therefore study an efficient subset of ILP known as learning from interpretations. In particular, they study scaling up first order logical decision trees (FOLDTs) which are more expressive than ....

Blockeel, H., L. De Raedt, N. Jacobs, and B. Demoen (1999). Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery . To appear.

No context found.

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.

No context found.

Blockeel, H., L. De Raedt, N. Jacobs, and B. Demoen (1999). Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery 3 (1), 59--93.

No context found.

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.

No context found.

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

.... Scalability to large knowledge bases is obtained in practical ILP systems either (1) by connecting the ILP system to an external database management system (DBMS) 1, 15, 13] or (2) by using independent examples that can be loaded one by one, i.e. the Learning from Interpretations setting [3] in which each example is de ned as a small relational database (typically implemented as a set of facts, i.e. an interpretation) The bene t of (2) over (1) is that (1) must rely for eciency on the caching mechanism provided by the DBMS which lacks information about the speci c way in which ....

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

....the output hypothesis. We can use ILP techniques to automatically build a model of the behavior of the user. The input of the system consists of actions performed by the user together with a description of the state of the environment if relevant. In the learning from interpretations setting [1] we represent input as facts in rst order logic. The output is a general description of the behavior of the user represented as a rst order logic theory. Because of this, it is easy to express relational data, which is important in analyzing shell usage. In the rest of the paper we will focus ....

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

....of classes C, a set of classified examples E, a background theory B Find a hypothesis H such that: for all e # E, H # e # B # c, and H # e # B # c where c is the class of the example e and c # C c . To make the discussion more concrete we focus on one ILP system: Tilde (see [4] [5]) This system performs predictive induction in the learning from interpretation setting by inducing logical decision trees from classified examples and background theory. Consider for example the background knowledge that is mentioned above. Suppose also a set of observations describing ....

....given examples. The Tilde system has the benefits (like most ILP systems) of being able to build complex hypotheses (using first order logic) and using background knowledge in finding these hypotheses. Moreover experiments have shown that the Tilde system scales nicely on large datasets (see [5]) More details of the Tilde system can be found in [4] and [5] 7.1. User Preference Modelling with Tilde Following Section 6, we may define a user preferences model for a certain user u as a function j u mapping an observation o from the set of possible observations O onto a preference ....

[Article contains additional citation context not shown here]

BLOCKEEL, H. , De RAEDT, L. , JACOBS, N. , DEMOEN, B. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

....98 team. This latter team was based on a reactive agent architecture. This team competed in the 98 RoboCup world cup but was also used in experiments to learn strategies using a genetic algorithm [4] experiments on verification and validation of multi agent systems [3] and data mining research [1]. The major problem we experienced with KULRoT 98 was however a slow reaction time. The cause of this problem was related to insufficient synchronization between the changing environment and the soccer players. To solve this problem, we created an improved architecture that gives a player a much ....

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 1998. To appear.

....user. The input of the system consists of actions performed by the user together Copyright c 2000, American Association for Arti cial Intelligence (www.aaai.org) All rights reserved. with a description of the state of the environment if relevant. In the learning from interpretations setting (Blockeel et al. 1999) we represent input as facts in rst order logic. The output is a general description of the behavior of the user represented as a rst order logic theory. In the rest of the paper we will focus on modeling a command shell user. Although this is a very old and basic user interface, it is still ....

Blockeel, H.; De Raedt, L.; Jacobs, N.; and Demoen, B. 1999. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery 3(1):59-93.

....the learning process. By this we can for example use a lexicon (such as WordNet [9] in the learning process. In the experiments the ILP system Tilde [3] was used because it is very ecient in constructing classi ers in the ILP setting. Moreover, experiments show that it can handle large datasets [4]. Tilde builds a rst order equivalent of a decision tree. Figure 2 shows an example of a decision tree built by Tilde. If the page contains ftp links the duration of the visit will be medium. Else it depends on 2 stop words like the , a , are left out 3 general information which holds ....

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

....single example before moving on to the next example: this could be called the examples in outer loop strategy, as opposed to the queries in outer loop strategy used by most ILP systems. The examples in outer loop strategy has important advantages when processing large data sets; see, e.g. [12, 5]. 3.4 Computational Complexity Lower and upper bounds on the speedup factor that can be achieved by executing a pack instead of separate queries can be obtained as follows. For a pack containing n queries q i = a; b i ) let T i be the time needed to compute the first answer substitution of q i ....

....implementation as described in [4] Since queries are represented as terms, each evaluation of a query involves a meta call in Prolog. 2. An Examples in outer loop implementation of Tilde, where examples are considered one by one and for each example all queries are run one after another (see [5]) Each call of a query, as in the previous setting, involves a meta call. 3. Disjoint execution of packs: a query pack is executed in which all queries in the pack are put beside one another; i.e. common parts are not shared by the queries. The computational redundancy in executing such a pack ....

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59--93, 1999.

No context found.

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling Up Inductive Logic Programming by Learning from Interpretations. Data Mining and Knowledge Discovery, 3:59--93, 1999.

No context found.

H. Blockeel, L. De Raedt, N. Jacobs, and B. Demoen. Scaling up inductive logic programming by learning from interpretations. Data Mining and Knowledge Discovery, 3(1):59-93, 1999.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC