Adopting an Object-Oriented Data Model in Inductive Logic Programming

The increasing amount of infm’mation to be manage.d in knowledge-based systems has promoted, on one hand, the exploitation of machine lem’ning for the automated acquisition of knowledge mid. on the other hand. the adoption of object-oriented representation models for easing the maintenmlce. In this comext, adopting techniques for structuring hmwledge represematiou in machine learning seems particularly appealing. Inductive Logic Programming {ILP) is a promising approach tbr the automated discovery of rules it, kr, owledge based systems. We Dropose an object-orietd ed extension of ILP employing multi-theory logic prog,’ants as the representation language. We define st new Icarning problem ~md propose the corresponding learlting algorithm. Our approadl enables ILP ,o benefit o[ object-oriented domain modelling in ,he lemning process, such as allowing st.ructured clomains to be directly mapped onto progrmn constructs, or easing the management of large knowledge bases. Introduction As the application of knowledge-based systems in real world situations is becoming more azKI Inore common, the amount of knowledge that. must be ~:quired and maintained is growing larger and larger, h, order to manage knowledge bases of consic.lerable size representing real world domains, ol).i(’ct-based representations such as frames (Mii,sky 1975), description logic (Woods & Schrnolze 1991). semantic m.tworks {Br~u:hman 1979), inheritance networks (Touretzky. liorty, Thomason 1988) have been proven lo be etft.ctive t.echniques. In these formalisms, each ol)je,’t in t.he domain is directly represented by means of apl)ropriate s.vntt~’tic structures, and objects sharing similar properties are grouped into classes. On their turn, classes are organised in a hierar,:hy and inheritance is exploited in ordtq’ to represent Olfly om:e propertit~s that at’,’ shared by all objects of a ,’lass. Inductive Logic Programming (ILP) (Muggleton 1991) is a promising approach for th,. automated discovery of rules in knowledge ba.sod systems (Morik ,t C.opyright @1999, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. al. 19.(}3). Therefore. it seems appealing to adopt techt,iq,ws for structuring knowledge representation in ILP. ILl’ is com’crl,ed with the problcu, of learning logic programs from a background knowlcdge and a set of positiw’ anti ltegal.ive examples. However. logic programming does not provide all)’ SUl)i),)rl for structuring domain ,’epresent ation. lu this paper, we aim at investigating the imp~v.’t of a,i object-oriented data model on II,1’. By Vml)loying mtdli-l.btory logic lal~yua.qc (Bugliesi. Lamina. &" Mello 199.:1: M,:(’abe 1992). the knowleclge is strm’turt,d as colh,ction (,f SelJarate logic theorie.~ orga,,is,.d i,, a hieral’clJy. Ea,’h mJde. ill the Itierarchy r(.l)r(’.-.vl|lS a (:la.ss ol)jvcts whil,’ each h.af rcl)r¢.senls a sil,gl,. ,,I,.i,’ct. ’l’lw coltl ril)ution of this pap~’r is twofold: (i) tim clefinition of a h,;.)rning problem in a multi-theory f, nx.’ironnwat: (it) the description of the corresponding learning algorit hm. The advantages of ~Klopl.ing a nnllli-theory logic language are twofold: from the knowledge represenl.ation viewpoint and from the learning process viowpoiut. As regar(ls the representation of the domain knowledge. stru(:tured domains can be direc:tly mapped ont.o program constructs. The standard ILP approach does not allow the notions of class and attrtbuh to I,e distinguish¢’d, since they are uniformly r,-’l)res,’t,l,.’d I)3" means of I)redi,’ate symbols. An ol)je,’t-o,’ientcd data model. ins! fad. provides intrinsic Sttl)l)ort for I.he trot ion of class and ,’lass hierar,:hy, where~Ls obj,.(-t’s attl’ibu!,,s are rei)resented as i)redicates. Moreover, inheritat,t’t’ provides an effici,.ut and compact way of r,’l)reseuiing properties shared by a se! of objc,’ts. As regards th," b’arning process, ca,:it ,’la.ss constilut,’s a uatural IJoun,lary for the learnittg pro,’~’ss, aud providf.s it with a tinc’r granularity lev,’l. Inheril.an,’e ],lechanisms p,’ovich, for new generalisation and specialisation Ol)el’;~l.ors alld (’al, 1)(-" t~xl)loited bol hiH Ih(’ learni,,g I)roc~’ss. aml when using lhe h.ar,.’(l program.