W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....the search space. The declarative bias is expressed using clause models that define the syntax of the clauses that can appear in hypothesis. The refinement operator adopts these models to generate only the clauses that are allowed by the syntax. The formalism is described in details in [ADRB95, VLDDR94] 40 Abductive Reasoning in Learning As discussed in section 1.1, the problem of learning from an incomplete background knowledge is still an open issue in ILP research. In real world problems, the knowledge acquisition process is often imperfect and some relevant pieces of information may be ....

W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994.

....(see Section 5.4 and Section 7.3) While CN2 has a specific handling of unknown values ( and ) ICL just assumes the closed world assumption. 4.8. HISTORICAL NOTE 87 4. 8 Historical note Our initial research focussed on the development and application of CLAUDIEN [De Raedt and Bruynooghe, 1993; Van Laer et al. 1994; De Raedt and Dehaspe, 1997a] CLAUDIEN iS a clausal discovery engine which, in its original setting, learns a set of first order clauses (i.e. a CNF hypothesis) which is true in a given logical interpretation. Each clause can be seen as an integrity constraint valid for the complete data set. ....

W. Van Laer, L. Dehaspe, and L. De Raedt. Applica- tions of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263-274, 1994.

....intersected into 6 parts, the class of node b11 is 6. With this description of the structures, we obtain a training set containing 5 graphs named A E with classified nodes. The table 2 shows the results for the systems FOIL, MFOIL, GOLEM (column Gol. and Claudien (column Clau. taken from [ Dehaspe et al. 1994 ] on this training set. For references to FOIL, MFOIL and Claudien see [ Dehaspe et al. 1994 ] The column labelled I1 shows the results of INDIGO on the deterministic (cf. Dolsak and Muggleton, 1992 ] version of the data. Column I2 shows the results on the indeterministic version. Row A ....

....we obtain a training set containing 5 graphs named A E with classified nodes. The table 2 shows the results for the systems FOIL, MFOIL, GOLEM (column Gol. and Claudien (column Clau. taken from [ Dehaspe et al. 1994 ] on this training set. For references to FOIL, MFOIL and Claudien see [ Dehaspe et al. 1994 ] The column labelled I1 shows the results of INDIGO on the deterministic (cf. Dolsak and Muggleton, 1992 ] version of the data. Column I2 shows the results on the indeterministic version. Row A shows the absolute number of correctly classified nodes of structure A, when using the ....

Dehaspe, L., van Laer, W., and De Raedt, L. (1994). Applications of a logical discovery engine. In Proceedings of the Fourth International Workshop on Inductive Logic Programming (ILP-94), GMD-Studien Nr. 237.

....only initial attribute the accuracy is at most 80 on the 188 data set, i.e. the performance is improved significantly by using contextual attributes. On the mesh design data, TRITOP performs with 37 predictive accuracy better than the systems FOIL (12 ) MFOIL (21 ) GOLEM (19 ) Claudien (28 , (Dehaspe et al. 1994), MILP (32 ) FOSSIL (32 ) FORS (31 ) SRT (24 , Kramer, 1995) CILLG (22 , Kietz, 1996) and INDIGO (34 ) on the deterministic data. When learning illegal chess end game positions (see e.g. Quinlan, 1990) TRITOP reaches an accuracy of 99:3 on a 1000 instance training set using a ....

Dehaspe, L., W. van Laer and L. De Raedt (1994). Applications of a logical discovery engine. In: Proceedings of the Fourth International Workshop on Inductive Logic Programming (ILP-94), GMD-Studien Nr. 237.

....better than PROGOL and FOIL on the same version of the dataset. The performance of TRITOP is increased significantly by using contextual attributes. On the mesh design data, TRITOP performs with 37 predictive accuracy better than the systems FOIL (12 ) MFOIL (21 ) GOLEM (19 ) Claudien (28 , [3]) MILP (32 ) FOSSIL (32 ) FORS (31 ) SRT (24 , 8] CILLG (22 , 7] and INDIGO (34 ) on the deterministic data. When learning illegal chess end game positions (e.g. 11] TRITOP reaches an accuracy of 99:3 on a 1000 instance training set using a 10 fold cross validation. In accuracy ....

L. Dehaspe, W. van Laer, and L. De Raedt. Applications of a logical discovery engine. In Proc. 4th Int. Workshop on ILP, GMD-Studien Nr. 237, 1994.

.... for the mesh data Structure Foil MFoil Golem Claudien I1 I2 A 17 22 17 31 21 21 B 5 12 9 9 9 14 C 7 9 5 5 9 9 D 0 6 11 19 41 18 E 5 10 10 15 27 33 Total 34 59 52 79 107 95 Percentage 12 21 19 28 38 34 The table 2 showsthe results for the systemsFOIL, MFOIL,GOLEM, and Claudien taken from [1] (for references to the systems see [1] The column labelled I1 shows the results for INDIGO on a deterministic (cf. 2] version of the data. Column I2 shows the results on an indeterministic version. Row A shows the absolute number of correctly classified nodes of structure A, when using the ....

.... Golem Claudien I1 I2 A 17 22 17 31 21 21 B 5 12 9 9 9 14 C 7 9 5 5 9 9 D 0 6 11 19 41 18 E 5 10 10 15 27 33 Total 34 59 52 79 107 95 Percentage 12 21 19 28 38 34 The table 2 showsthe results for the systemsFOIL, MFOIL,GOLEM, and Claudien taken from [1] for references to the systems see [1]) The column labelled I1 shows the results for INDIGO on a deterministic (cf. 2] version of the data. Column I2 shows the results on an indeterministic version. Row A shows the absolute number of correctly classified nodes of structure A, when using the structures B, C, D, E for learning. The ....

L. Dehaspe, W. van Laer, and L. De Raedt, `Applications of a logical discovery engine', in Proceedings of the Fourth International Workshop on Inductive Logic Programming(ILP-94), GMD-StudienNr. 237, (1994).

....algorithmic learning. Heuristic based ILP systems like FOIL [21] or LINUS [5, 15] aim to describe large, may be noisy example sets by the rst order logic formula. The resulting formula is not necessary complete nor consistent with the learning set. Algorithmic ILP systems like MIS [22] CLAUDINE [4] or Markus [13, 14] learn from a few, necessary correct examples (as a rule recursive) logic programs. The most developed algorithmic systems can be used as inductive tools in (automatic) programming. In the latter, the teacher is a human being who knows the learned concept and is assumed to ....

....D cannot. Just to complete the summary, we mention two more systems employed in database technology. INDEX , a program for inducing attribute dependencies and an interactive decomposition of database relations, is described in [6] Improvements of that program were presented in [10] CLAUDINE [4] can learn both data dependencies and integrity constraints in relational databases. I.4.6 8 Conclusion We have presented an application of inductive logic programming in database schema design in the area of deductive object oriented databases. Algorithmic ILP programs seem to be powerfull ....

Dehaspe L., Van Laer W., De Raedt L.: Applications of a logical discovery engine. In: Wrobel S.(ed.): Proc. of 4th Workshop on Inductive Logic Programming ILP'94, Bonn Germany, 1994.

....top down search must decide about which predicate to learn before specializing. Delaying this decision helps to avoid some search problems [36] In the Claudien system [35] this operator has been extended further to be able to generate general clauses, also based on a declarative bias language [47, 16] Enumerative generation Whereas top down and bottom up search for clauses have a fixed search direction and are usually controlled by heuristic criteria (see below) the ILP project has also investigated an enumerative full search approach to finding new clauses. In the Tracy system [10] new ....

....is based on a refinement operator that refines head and body literals during its search. The search space is defined in a declarative bias language, and heuristically searched in an iterative deepening fashion. Claudien has successfully discovered integrity constraints in many test applications [47]. Both Index and Claudien are based on the alternative non monotonic semantics of ILP (also a result of the ILP project [21, 35, 15, 18] In this semantics, individual clauses can be learned independently, which has been used to parallelize Claudien [16] Similarly, the ICDT system [19] is ....

Wim Van Laer, Luc Dehaspe, and Luc De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, 1994.

....but uses a somewhat simpler confirmation criterion that does not include the Nap component. De Raedt and Bruynooghe remark that in Claudien, learning from knowledge bases with many predicates is computationally intractable. Therefore, they recently extended Claudien by so called clause models [DLR94]. Clause models are used to define the syntactic structure of the hypotheses, as the IC schemes proposed here. The foremost topic for further research would now be to examine the two systems empirically to see how the different search strategies and confirmation criteria affect performance when ....

Lud Dehaspe, Wim Van Laer, and Luc De Raedt. Applications of a logical discovery engine. In Proc. of AAAI Workshop on Knowlege Discovery in Databases, 1994. No. 231.

....1 (Most General Cover) The set of functional dependencies F is a most general cover if for every dependency X A 2 F , there exists no Y with Y ae X and Y A 2 F . Our presented system can be seen at the first glance as an optimized version of CLAUDIEN regarding functional dependencies, [Dehaspe et al. 1994]. But there are differences: first, in CLAUDIEN the relationship between the dependencies is based on subsumption and the verification of the hypotheses on theorem proving. In our approach, the relationship of the dependencies is based on an axiomatization of FDs and UINDs. The verification is ....

Dehaspe, L., Laer, W. V., and Raedt, L. D. (1994). Applications of a logical discovery engine. In Wrobel, S., editor, Proc. of the Fourth International Workshop on Inductive Logic Programming, GMD-Studien Nr. 237, pages 291--304, St. Augustin, Germany. GMD.

....in clausal form. If the set of data is an (purely extensional) instance of a database, then the clauses that are found can be used to find intensional definitions and integrity constraints for the database. For our application, we have used the ILP system Claudien [ De Raedt and Bruynooghe, 1993; Van Laer et al. 1994 ] This system takes as input a database D (which can contain intensional as well as extensional definitions) a background knowledge B and a language bias L. It then induces clauses that are true in the least Herbrand model of D[B and that belong to L. In our application, D contains only ....

W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994.

....patterns in complex structured domains. These databases usually involve several levels of objects and complex relations among them. At the same time new and more efficient ILP algorithms have been developed that are making feasible the application of ILP methods to real complex datasets [8, 12, 7]. However, the regularities that can be obtained from them with the current applications of KDD are static, i.e. they represent relations holding in the database at a certain moment in time. However, most of the real databases are inherently temporal, i.e. data have a temporal dimension, i.e. ....

....can be obtained from books or from an expert. These discrete values can replace the numerical arguments in the set of observations or can be used as candidate thresholds in a template (see below) A similar process is used by CN2 for the management of numerical descriptors. age gestation 1 1: [5,12,18,20,23,30,36] 6 Conclusions Many ILP applications of KDD in increasingly large and complex databases are probable to appear in the coming years. Following this direction, we have shown in this paper the adequacy of clausal discovery algorithms for the discovery of knowledge in temporal databases and an ....

Dehaspe L., W. Van Laer, and L. De Raedt. Applications of a logical discovery engine. Proceedings of the 4th International Workshop on Inductive Logic Programming, 1994.

....These algorithms represent the efforts to integrate static domain knowledge with induction. The researchers in inductive logical programming have also addressed the problem of learning relationbased patterns. Different from the algorithms described above, these programs (see for example [2, 15, 18]) are focused on inventing new predicates from positive and negative training examples. Some of them (for example [17] also use semantic knowledge as guidance. Our research in this area is focused on discovering relation based patterns from data that are commonly found in databases. These data ....

Wim Van Laer, Luc Dehaspe, and Luc De Raedt. Applications of a logical discovery engine. In Proceedings of 1994 AAAI Workshop on Knowledge Discovery in Databases, 1994.

....noun( E 4 = s( the; girl; wants; the; boy] H 4 = noun(sing3( Gamma [girl] verb(sing3( trans) Gamma [wants] Figure 3: Example of abduction 2.1.2 Nonmonotonic setting: knowledge discovery The less common nonmonotonic 5 setting is used (f.i. in the system Claudien [5, 6]. The aim here is not to discriminate between different classes, but to discover properties that are valid with respect to the knowledge base as a whole. Definition 3 (nonmonotonic explanation) Given knowledge base KB and formal language L, target hypothesis H is a maximal subset 6 of L such ....

L. Dehaspe, W. Van Laer, and L. De Raedt. Applications of a logical discovery engine. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMD-Studien, pages 291--304. Gesellschaft fur Mathematik und Datenverarbeitung MBH, 1994.

....of clauses that can appear in the hypotheses. From these models, one can automatically derive a refinement operator that only generates clauses that are allowed by the syntax. A full discussion of this declarative bias mechanism is outside the scope of this paper, but see [ Ad e et al. 1995; Van Laer et al. 1994 ] for more details. 3 The Mutagenesis Problem A classification problem that has received some attention lately, is that of classifying nitroaromatic molecules into mutagenic and non mutagenic ones. As in a lot of chemical problems, the data from which to learn are structured. One cannot predict ....

W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994.

....used to define the syntax of clauses that can appear in the hypotheses. From these models, one can automatically derive a refinement operator that only generates clauses that are allowed by the syntax. A full discussion of this declarative bias mechanism is outside the scope of this paper, but see [1, 18] for more details. 3 The Mutagenesis Problem A classification problem that has received some attention lately, is that of classifying nitroaromatic molecules into mutagenic and non mutagenic ones. As in a lot of chemical problems, the data from which to learn are structured. One cannot predict ....

W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994.

....cards on the main line satifies the secret rule. These can be used as positive examples. Each card on a side line is an illegal successor of the corresponding card on the main line. So these two cards can be seen as a negative example. Each example (a set of two cards) is translated (as in [ Van Laer et al. 1994 ] into one fact of the form: canfollow(R 2 ; S 2 ; R 1 ; S 1 ) which states that a card of rank R 2 and suit S 2 can follow a sequence ending with a card of rank R 1 and suit S 1 . In this way, each interpretation contains one fact. For example in the first sequence, one positive model might ....

....best clause is shown in figure 5. As in CN2, we use beam search as search strategy. A classical refinement operator under subsumption (Plokin 70, Shapiro 83) is used together with CLAUDIEN s mechanism to specify the declarative bias (i.e. the syntax of well formed clauses in hypotheses) cf. Van Laer et al. 1994; Ad e et al. 1995 ] and below. Notice that whenever a clause c is true in an interpretation I, all refinements of c are also true in I. The search starts with the most general clause of the refinement graph (namely false true, for wich each example model is false) During the search, ICL ....

[Article contains additional citation context not shown here]

W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994. This article was processed using the L A T E X macro package with LLNCS style

....database designer, who should be given the opportunity to reject the definitions proposed by the system. The paper is organised as follows: in Section 2, we review some concepts of deductive databases and logic programming, in Section 3, we review the inductive logic programming system Claudien [1, 8], which will be adapted for use in our inductive database design tool, in Section 4, we address the problem of finding intensional definitions for predicates, in Section 5, we present an experiment, and finally, in Section 6, we conclude and touch upon related work. 2 Logic Programming Concepts ....

....theory T , one can run the query B 1 ; Bn ; A 1 ; Am on a database containing T . If the query finitely fails, the clause is valid, otherwise it is invalid. 3 Inductive Logic Programming In this section, we give a brief overview of the inductive logic programming system Claudien [1, 8, 2, 3]. The Claudien system starts from a definite clause theory T and a language L (which is a set of well formed clauses) and finds a set of maximally general clauses that are valid in the given theory. Definition7. Clause c 1 is more general than clause c 2 iff there exists a substitution such ....

W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994. This article was processed using the L A T E X macro package with LLNCS style

.... two problem specifications for ILP, and end up with theoretical constraints on Parallel Inductive Logic Programming (PILP) Given these constraints, the second part of the paper then focuses on the comparative evaluation (Section 3) of a parallel version of the clausal discovery system Claudien [1, 4]. Finally, we round up with some conclusions in Section 4. 2 Partitioning the ILP task Roughly speaking, ILP starts from an initial background theory B, some evidence (or examples) E, and a language bias L, which defines the set of well formed clauses. The aim is then to induce a hypothesis H ae ....

L. Dehaspe, W. Van Laer, and L. De Raedt. Applications of a logical discovery engine. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMD-Studien, pages 291--304. Gesellschaft fur Mathematik und Datenverarbeitung MBH, 1994.

.... Dehaspe, 1995] ffl declarative bias formalism Dlab: Dehaspe and De Raedt, 1995a; Dehaspe and De Raedt, 1996] ffl problem setting: De Raedt and Bruynooghe, 1993; De Raedt and Lavrac, 1993; Muggleton and De Raedt, 1994] ffl PAC learning results: De Raedt and Dzeroski, 1994] ffl applications: [Dehaspe et al. 1994] ffl parallel version: Dehaspe and De Raedt, 1995b] The rest of this user s guide is organized as follows. In Section 2 the inputs and outputs of Claudien ard briefly introduced as well as basic instructions for intstalling, starting and stopping the system. Section 3 then contains a detailed ....

L. Dehaspe, W. Van Laer, and L. De Raedt. Applications of a logical discovery engine. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMD-Studien, pages 291--304, Sankt Augustin, Germany, 1994. Gesellschaft fur Mathematik und Datenverarbeitung MBH.

....the paper then focuses on the algorithmic description (Section 3.1) prototypical implementation (Section 3.2) and comparative evaluation 1 Notice that the aim is not to extend the solvable problem space of a problem. Section 4) of a parallel version of the clausal discovery system Claudien [2, 6]. Finally, we round up with some conclusions in Section 5. 2 Partitioning the ILP task Roughly speaking, ILP starts from an initial background theory B, some evidence (or examples) E, and a language bias L, which defines the set of well formed clauses. The aim is then to induce a hypothesis H ae ....

....is put back, and unpromising items are pruned away. Through the instantiation of the Delete and Prune parameters, the user can tailor searching and pruning strategies to the application at hand. For more information on theoretical background, PAC learning results and applications of Claudien, see [2, 3, 6]. The notion of specialization imposes a graph structure on the hypothesis space, such that the overly general clauses c in QC each represent the subgraph of clauses that can be construed via repeated applications of refinement operator ae. This property of the hypothesis space offers a simple ....

L. Dehaspe, W. Van Laer, and L. De Raedt. Applications of a logical discovery engine. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMD-Studien, pages 291--304. Gesellschaft fur Mathematik und Datenverarbeitung MBH, 1994.

....be seen in Figure 3, inducing facts in a natural language processing context corresponds to extending the lexicon on the basis of new language material and a reliable grammar. Nonmonotonic Setting: Knowledge Discovery The less common nonmonotonic 5 setting is used (f.i. in the system Claudien [6, 5, 8]. The aim here is 5 The name nonmonotonic for this setting was introduced by Helft [9] and relates to the fact that the closed world assumption is used. Thus, the addition of new facts to the knowledge base might falsify previously inferred rules. B2 = det Gamma [the] noun(sing3( ....

L. Dehaspe, W. Van Laer, and L. De Raedt. Applications of a logical discovery engine. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMD-Studien, pages 291--304, Sankt Augustin, Germany, 1994. Gesellschaft fur Mathematik und Datenverarbeitung MBH.

No context found.

Dehaspe L., Van Laer W. and De Raedt L. Applications of a logical discovery engine. In S. Wrobel, editor, Proceedings of the 4th International Workshop on Inductive Logic Programming, volume 237 of GMD-Studien, pp 291-304, Sankt Augustin, Germany, 1994. Gesellschaft fur Mathematik und Datenverarbeitung MBH.

No context found.

W. Van Laer, L. Dehaspe, and L. De Raedt. Applications of a logical discovery engine. In Proceedings of the AAAI Workshop on Knowledge Discovery in Databases, pages 263--274, 1994.

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