A Logic-Based Approach for Adaptive Information Filtering ... - CiteSeerX
A Logic-Based Approach for Adaptive Information Filtering Agents Raymond Lau1, Arthur H.M. ter Hofstede1 , and Peter D. Bruza2 1 Cooperative Information Systems Research Centre
Queensland University of Technology, Brisbane, Qld 4001, Australia fraymond,
[email protected]
2 Distributed Systems Technology Centre, The University of Queensland
Brisbane, Qld 4072, Australia
[email protected]
Abstract. Adaptive information ltering agents have been developed to alleviate the problem of information overload on the Internet by learning a user's changing information needs and autonomously scanning through the incoming stream of information according to the user's latest information needs. However, the explanatory power and the learning autonomy of these agents should be improved. Applying a logic-based framework for representation, learning, and matching in adaptive information ltering agents is promising since users' changing information needs can automatically be deduced by the agents. In addition, the inferred changes can be explained and justi ed based on formal deduction. This paper proposes a logic-based model for adaptive information ltering agents. Particularly, it illustrates how the AGM paradigm of belief revision can be applied to the learning processes of these agents. The impact of a logic-based framework on agents' learning autonomy and explanatory power is discussed through examples.
1 Introduction Information ltering (IF ) and information retrieval (IR) are "two sides of the same coin" [4]. However, IF is more concerned with the removal of irrelevant information from a stream of incoming information. With the explosive growth of the Internet and the World Wide Web (Web ), it is becoming increasingly dicult for users to retrieve relevant information. This is the so-called problem of information overload on the Internet. Augmenting existing Internet search engines with personalised information ltering tools is one possible method to alleviate this problem. Adaptive information ltering agents are computer systems situated on the Web. They autonomously lter the incoming stream of information on behalf of the users. Users' information needs will change over time, and information ltering agents must be able to revise their beliefs about the users' information needs so that the accuracy of the ltering process can be maintained. The AGM belief revision paradigm [1] provides a rich and rigorous foundation for modelling such revision processes. It enables an agent to modify its beliefs in a rational and minimal way. In addition, since semantic relationships among information items can be taken into account during the belief revision process, the agent can automatically deduce a user's changing information needs. Therefore, less users' direct relevance feedback [19] is required to train the ltering agents, and hence a higher level of learning autonomy can be achieved. This paper focuses on the application of the AGM paradigm of belief revision to the development of learning mechanisms in adaptive information ltering agents.
2 Related Work News Dude[5] is a multi-agent system which can learn a user's daily news interests incrementally. A hybrid approach of weighted keyword vector and Bayesian feature is employed to develop the user model. Basically, each feature indicates the presence or absence of a particular keyword. News ltering is conducted in two stages. Firstly, incoming news is compared with the vector representation of the user's information needs by the nearest neighbor algorithm, and the cosine similarity measure developed in IR.[20] Secondly, the new item will be classi ed again by the Bayesian classi er if it is not considered close enough to the vector user model. With the similarity score from stage one and the probability score from stage two, the recommender agent can rank the news item and present the highly rated items to a user. Agent adaptation is conducted through the reinforcement learning process where the users' relevance feedback is used to update the weights of the keyword vector and the probability distribution of the Bayesian classi er. Amalthaea[15] is a multi-agent system for information discovery and ltering on the Web. It employs keyword vectors to represent users' information needs and Web pages. Learning and adaptation is by relevance feedback and genetic algorithms. Each keyword is mapped to a ltering agent (i.e. a gene). A user's relevance feedback is converted to a numeric score and assigned to the ltering agents that carry the keywords appearing in the judged Web document. Then, a genetic process takes place (i.e. mutation and cross-over) among the highly rated ltering agents to produce the next generation of tter agents. Thereby, the Web pages presented by Amalthaea can converge to the user's taste. Fab[3] is a multi-agent system that retrieves and lters Web pages on behalf of the users. Fab's ltering agent employs both content based ltering as well as collaborative ltering.[17] For content based ltering, Fab is based on the vector space model[20] for knowledge representation and ltering. Learning is based on the users' relevance feedback and the Rocchio method[18] that updates the weights in a keyword vector, which represent the user's information needs. The adaptive information agents described so far are based on the users' direct relevance feedback to re ne their user models. This method produces high learning accuracy, but the disadvantage is that it demonstrates low learning autonomy (i.e. involving many human interventions). There are information agents[14, 2] that employ users' implicit feedback for learning and adaptation. Based on pre-de ned heuristic, these agents can infer users' relevance judgements by observing their online browsing behavior. For instance, if a user add a Web page to the hot list of the Web Browser, the agents will infer that the user considers this Web page as relevant. Basically, the agent extracts some tokens (e.g. keywords) from the Web pages stored in the hot list, and updates these tokens to the keyword vector, which represents a user's information needs. Thereby, the agent can infer what Web documents should be recommended to its user in future. The advantage of this type of learning approach is that less human interventions are required, and a higher level of learning autonomy is achieved. However, since the implicit heuristic and the keyword vector are two distinct knowledge representation schemes, it is still dicult to combine these two representations in order to explain an agent's behavior (i.e. why selecting a document and why not the others). Moreover, it is dicult to maintain the knowledge base when more heuristic is required by the agent.
3 System Architecture Figure 1 depicts the system architecture of AIFS [13] in which the adaptive information ltering agent is examined. A user communicates with AIFS via the interface agent. The retrieval agents
(RA) retrieve Web pages from external sources. The data management agent is responsible for house-keeping and characterisation of Web documents. The focus of this paper is on the adaptive information ltering agent. Particularly, the learning component of the ltering agent. Learning in the ltering agent consists of two steps. Firstly, based on a user's relevance feedback and the statistical data stored in AIFS's output queue, the learning component induces beliefs [8] about a user's information needs. Secondly, these beliefs are revised into the ltering agent's memory in a consistent and minimal way. The AGM belief revision framework [1] and the corresponding computational algorithms[21, 22] provides a sound and rigorous mechanism to carry out such an operation. Moreover, a user's changing information needs can be inferred by the agent through the belief revision process. The matching component carries out the ltering function based on logical deduction. Web documents deemed relevant by the ltering agent are transferred to the output queue via the Data Management Agent. Adaptive Filtering Agent
Retrieval Agents
User: Information needs
RA Interface Agent
Relevance Feedback
Learning Component
Memory
RA
External Agents
RA
External IR Sys.
Matching Component
Filtered Web Documents
Internet Search Engines
Data Management Agent
Filtered Web Documents
Web Documents
Output Queue
Input Queue
Fig. 1. The System Architecture of AIFS
4 The AGM Belief Revision Paradigm The AGM framework [1] formalises consistent and minimal belief changes by sets of postulates and belief functions e.g. expansion (K +), contraction (K ? ), and revision (K ). One of the constructions of these functions is by epistemic entrenchment (6) [9]. Beliefs with the lowest degree of epistemic entrenchment are given up when inconsistency arises because of applying changes. Nevertheless, for computer-based implementation, nite partial entrenchment ranking (B) which ranks a nite subset of beliefs with the minimum possible degree of entrenchment (6B ), and maxiadjustment [21, 22] which repeatedly transmutes (B) using an absolute measure of minimal change
under maximal information inertia, have been proposed. Williams [21] has formally de ned nite partial entrenchment ranking as a function B that maps a nite subset of sentences in L (e.g. classical propositional language) into the interval [0; O], where O is a suciently large ordinal such that the following conditions hold for all 2 dom(B): (PER1) f 2 dom(B) : B( ) < B( )g 6` ; (PER2) If ` : then B( ) = 0; (PER3) B( ) = O if and only if ` . The set of all partial entrenchment rankings is denoted B. B( ) is referred as the degree of acceptance of . The explicit beliefs of B 2 B is f 2 dom(B) : B( ) > 0g, and is denoted exp(B). Similarly, the implicit beliefs represented by B 2 B is Cn(exp(B)), and is denoted content(B). Cn is the classical consequence operation. The degree of acceptance of implicit belief is de ned as [21]: degree(B; ) = largest j such that f 2 exp(B) : B( ) j g ` if 2 content(B); otherwise degree(B; ) = 0. Let B 2 B be nite. The range of B is enumerated in ascending order as j0 ; j1 ; j2 ; : : : ; jO . Let be a contingent sentence, jm = degree(B; ) and 0 i < O. Then the ( ; i) maxi-adjustment of B is B? ( ; i) de ned by [21]: ? ? B ( ; i) = (B ( ; i))
if i jm (B? (: ; 0))+ ( ; i) otherwise
where for all 2 dom(B), B? ( ; i) is de ned as follows: 1. For with B( ) > jm ; B? ( ; i)( ) = B( ). 2. For with i < B( ) jm , suppose B? ( ; i)( ) for is de ned with B( ) jm?k for k = ?1; 0; 1; 2; : : :; n ? 1, then for with B( ) = jm?n , 8 i if ` or > > > > 6` and 2 ? > > < where ? is a minimal subset of ? B ( ; i)( ) = > f
: B( ) = jm?n g such that > > > f
: B? ( ; i)( ) > jm?n g [ ? ` > > : B( ) otherwise
3. For with B( ) i ; B? ( ; i)( ) = B( ). For all 2 dom(B) [ f g; B+ ( ; i) is de ned as follows:
8 B( ) if B( ) > i > > B( )if i< degree (B; ! ) > :
degree(B; ! ) otherwise
5 Knowledge Representation In IR, a retrieval situation was proposed to be the basis for evaluating document relevance [16]. A retrieval situation usually refers to situational factors such as the searcher's state of knowledge, background, intentions and so on. For the discussion in this paper, a retrieval situation also implies the searcher's topical information needs. The ltering agent's memory stores the abstraction of a retrieval situation based on the notion of beliefs. Each belief is represented as a formula of a classical rst-order language L. This paper focuses on two main elements of a retrieval situation, namely a user's information needs and their background knowledge.
5.1 Web documents A Web page is pre-processed based on traditional IR techniques [20]. For instance, a Web document is rst converted to plain text. The text is indexed resulting in a set of tokens. Stop words (e.g. "the", "a", etc.) are removed, the remaining tokens are stemmed. The stemmed tokens are then weighted according to the tf-idf measure [20], and those above a xed threshold are deemed to be suciently descriptive to represent the text. At the symbolic level, each selected token k is mapped to the ground term of the positive keyword predicate pkw e.g. pkw(k). Each pkw(k) is in fact a proposition since its interpretation is either true or false. Technically it is an atomic formula of a classical rst-order language L. For example, if fbusiness; commerce; trade; : : :g are the top n extracted tokens (e.g. keywords) from a document, the corresponding symbolic representation will be: fpkw(business); pkw(commerce); pkw(trade); : : :g.
5.2 Users' topical information needs Kindo's keyword classi er [12] is modi ed to induce a user's information need. In a relevance feedback environment, a user's information needs can be induced from a set of relevant documents D+ and a set of non-relevant documents D? judged by the user [6, 19]. The basic idea is that a keyword (i.e. token) appearing more frequently in D+ is probably an indicator of the user's information preference. On the other hand, a keyword appears frequently in D? may be an indicator of what the user does not want to retrieve. Therefore, the notions of positive, neutral, and negative keywords are proposed [12]. Intuitively, positive keywords represent the information items that a user requires, whereas negative keywords represent the information items that the user does not want. Neutral keywords are not useful for determining a user's information need. Accordingly, a positive keyword k is represented as a formula pkw(k), whereas a negative keyword is represented as :pkw(k). Eq.(1) is used to classify the keywords from D+ and D? , and to induce the preference value pre(k) for each keyword k. Whenever a Web page is judged by the user, the preference values of the set of keywords representing that page can be computed. The corresponding beliefs will be updated in the agent's memory via the belief revision mechanism.
pre(k) = tanh df(k) p(krel ) tanh p(pkrelrel ) ? (1 ? p(krel )) tanh (1(1??p(pkrelrel)))
(1)
where is used to restrict the range of pre(k) in the interval ?1 < pre(k) < 1. The examples illustrated in this paper assume that = 0:9. df (k) is the sum of the number of relevant documents df (krel ) and the number of non-relevant documents df (knrel ) that contains the keyword k, and tanh is the hyperbolic tangent. The rarity parameter is used to control rare or new keywords and is expressed as int(logN + 1), where N is the total number of Web documents judged by a user, and int is an integer function that truncates the decimal values. p(krel ) is the estimated probability (krel ) that a document containing keyword k is relevant and is expressed as the fraction df (kreldf)+ df (knrel ) . prel is the estimated probability that a document is relevant. In our system, it is assumed that the probability that a Web document presented by the ltering agent and judged as relevant by a user is prel = 0:5. A positive value of pre(k) implies that the associated keyword is positive, whereas a negative value of pre(k) indicates a negative keyword. If pre(k) is below a threshold value , the associated keyword is considered neutral. It is assumed that = 0:5 for the examples demonstrated
in this paper. Basically a positive keyword k is mapped to pkw(k), and a negative keyword k is mapped to :pkw(k). There is no need to create the symbolic representations for neutral keywords. For pkw(k) or :pkw(k), the degree of acceptance B( k ) of the corresponding formula k is de ned as:
jpre(k)j B( k ) = 0jpre(k)j ifotherwise
(2)
Table 1 shows the results of applying Eq.(1) and Eq.(2) to induce the beliefs and the associated degree of acceptance based on D+ and D? . It is assumed that there are a set of ve documents having been judged as relevant (i.e. jD+ j = 5) and another set of ve documents having been judged as non-relevant by a user (i.e. jD? j = 5). Each document is characterised by a set of keywords e.g. d1 = fbusiness; commerce; systemg; d2 = fart; sculptureg.
Table 1. Inducing a user's information preferences Keywords D+ D? pre(k) Formula: k B( k ) business commerce system art sculpture insurance
5 4 2 0 0 1
0 0 2 5 3 0
0.856 pkw(business) 0.836 pkw(commerce) 0 -0.856 :pkw(art) -0.785 :pkw(sculpture) 0.401 -
0.856 0.836 0.856 0.785 -
5.3 Background knowledge A user's state of knowledge includes their understanding about a particular application domain. For example, in the domain of databases, classi cation knowledge such as (O-O databases ! databases) can be used by a searcher. Linguistic knowledge such as (VDU monitor) may also be used in a search. In our current framework, background knowledge is assumed to be elicited from a particular user or to be transferred from a thesaurus of a particular domain. A manual procedure is involved to encode these knowledge in the form of rst-order formulae. The following is an example of background knowledge used for the discussion in this paper. In the context of IR, these rules can be interpreted as: an information item about "business" implies that it is about "commerce" and vice versa; "sculpture" is a kind of Arts.
pkw(business) pkw(commerce); 1:000 pkw(sculpture) ! pkw(art); 1:000
6 Learning and Adaptation Whenever a user provides relevance feedback for a presented Web document, the belief revision process can be invoked to learn the user's information preferences. Conceptually, the ltering
agent's learning and adaptation mechanism is characterised by the belief revision and contraction processes. For example, if ? = fa1; a2 ; : : : ; an g is a set of formulae representing a Web document d 2 D+ , the belief revision process (((Ka1 )a2 ) : : :)an is invoked for each ai 2 ? , where K is the belief set stored in the ltering agent's memory. On the other hand, the belief contraction process (((Ka?1 )?a2 ) : : :)?an is applied for each ai 2 ? and d 2 D? . At the computational level, belief revision is actually taken as the adjustment of entrenchment ranking B in the theory base exp(B). Particularly, maxi-adjustment [21, 23] is employed by our system to modify the ranking in an absolute minimal way under maximal information inertia. As the input to the maxi-adjustment algorithm includes an ordinal i representing the new degree of acceptance of a sentence , the procedure described in Section 5.2 is used to obtain the new degree for each 2 ? . Moreover, for our implementation of the maxi-adjustment algorithm, the maximal ordinal O is chosen as 1. Therefore, any formula assigned the entrenchment value 1 will not be removed from the theory base exp(B).
Example 1:
The rst example shows how adding one belief to the agent's memory will automatically raise the entrenchment rank of another related belief. It is assumed that the belief B(:pkw(sculpture)) = 0:785 and the belief B(pkw(business)) = 0:856 have been learnt by the ltering agent. If several Web documents characterised by the keyword art are also judged as non-relevant by the user later on, the preference value of the keyword art can be induced according to Eq.(1). Assuming that pre(art) = ?0:856, the corresponding entrenchment rank can be computed as B(:pkw(art)) = 0:856 according to Eq.(2). By applying B? (:pkw(art); 0:856) to the theory base exp(B), the before and after images of the agent's explicit beliefs (i.e. exp(B)) can be tabulated in Table 2. Based on the maxi-adjustment algorithm, B+ ( ; i)( ) = i if B( ) i < degree(B; ! ). * B(:pkw(sculpture)) 0:856 < degree(B; :pkw(art) ! :pkw(sculpture)) ) B+ (:pkw(art); 0:856)(:pkw(sculpture)) = 0:856
The implicit belief :pkw(art) ! :pkw(sculpture) in content(B) is derived from the explicit belief pkw(sculpture) ! pkw(art) in the theory base exp(B), and its degree of acceptance is 1 according to the de nition of degree. As the belief :pkw(art) implies the belief :pkw(sculpture) and the agent believes in :pkw(art), the belief :pkw(sculpture) should be at least as entrenched as the belief :pkw(art) according to (PER1). In other words, whenever the agent believes that the user is not interested in art (i.e. :pkw(art)), it must be prepared to accept that the user is also not interested in sculpture at least to the degree of the former. The proposed learning and adaptation framework is more e ective than other learning approaches that can not take into account the semantic relationships among information items. This example demonstrates the automatic revision of the agent's beliefs about related keywords given the relevance feedback for a particular keyword. Therefore, less users' relevance feedback is required during reinforcement learning. Consequently, learning autonomy of the ltering agent can be enhanced.
Example 2:
The second example illustrates the belief contraction process. In particular, how the contraction of one belief will automatically remove another related belief from the agent's memory if there is a semantic relationship between the underlying keywords. Assuming that more Web documents characterised by the keyword sculpture are judged as relevant by the user at a later stage, the
Table 2. Raising related beliefs Formula: B( ) Before B( ) After pkw(business) pkw(commerce) 1.000 1.000 pkw(sculpture) ! pkw(art) 1.000 1.000 pkw(business) 0.856 0.856 :pkw(sculpture) 0.785 0.856 :pkw(art) 0 0.856 belief B(pkw(sculpture)) = 0:785 could be induced. As B? ( ; i) = (B? (: ; 0))+ ( ; i) if i > jm , where i = 0:785 and jm = 0 in this example, B? ( pkw(sculpture); 0:785) leads to the contraction of the belief :pkw(sculpture) from the theory base exp(B). Moreover, B? ( : pkw(sculpture); 0) (: pkw(art)) = 0 is computed because B( : pkw( art)) = 0:856 = jm?n and the set f : B? ( : pkw(sculpture); 0) ( ) > 0:856 g [ f : pkw(art) g ` : pkw(sculpture) is obtained. The before and after images of the ltering agent's explicit beliefs are tabulated in Table 3.
Table 3. Contracting related beliefs Formula: B( ) Before B( ) After pkw(business) pkw(commerce) 1.000 1.000 pkw(sculpture) ! pkw(art) 1.000 1.000 pkw(business) 0.856 0.856 pkw(sculpture) 0 0.785 :pkw(sculpture) 0.856 0 :pkw(art) 0.856 0 The explanation of the above learning processing is that given a user's new information preference about sculpture, the previous requirement of not sculpture should be removed from the agent's memory since it is impossible for the user to require Web documents about sculpture, and not require Web documents about sculpture at the same time. Moreover, as sculpture implies art, a user who rejects Web documents about sculpture may also reject documents about art. The belief revision mechanism exactly carries out this kind of reasoning to infer the possible changes of a user's information needs. As can be seen, the ltering agent's behavior can be explained based on logical deduction and the explicit semantic relationships among information items. The process is more transparent than examining the weight changes in a keyword vector. In summary, the agent's memory consists of the following beliefs before the learning process:
pkw(business) pkw(commerce); 1:000 pkw(sculpture) ! pkw(art); 1:000 pkw(business); 0:856 :pkw(sculpture); 0:856 :pkw(art); 0:856 However, after the belief revision process, the ltering agent's memory becomes:
pkw(business) pkw(commerce); 1:000 pkw(sculpture) ! pkw(art); 1:000 pkw(business); 0:856 pkw(sculpture); 0:785
7 Filtering Web Documents In our current framework, the matching function of the ltering agent is modelled as logical deduction. Moreover, a Web document is taken as the conjunction of a set of formulae [7, 11]. The following example illustrates the agent's deduction process with reference to previous examples. The resulting belief sets K = Cn(exp(B)) from examples 1 and 2 in the previous section are used to determine the relevance of the following three Web documents:
= fpkw(business) ^ pkw(art)g ' = fpkw(sculpture) ^ pkw(art)g = fpkw(business) ^ pkw(commerce)g The ltering agent's conclusions about the relevance of the Web documents are summarised as follows: Time: (t1)
Time: (t2)
K = Cn(fpkw(business) pkw(commerce); pkw(sculpture) ! pkw(art); pkw(business); :pkw(sculpture); :pkw(art)g) ) K 6` ; K 6` '; K ` K = Cn(fpkw(business) pkw(commerce); pkw(sculpture) ! pkw(art); pkw(business); pkw(sculpture)g) ) K ` ; K ` '; K `
At time (t1), the ltering agent believes that the user does not need Web documents about art. Therefore, documents and ' are rejected. Moreover, since business is equivalent to commerce for this particular information searcher, and she is interested in business, a Web document about business and commerce should also be relevant to her. At time (t2), after the belief revision process, the agent learns that the user is interested in sculpture. As sculpture is a kind of art, the agent can deduce that she may want Web documents about art as well. Therefore, documents and ' are considered as relevant at this time. So, a logic-based framework for representation and reasoning facilitates the explanation of an agent's decisions for information retrieval since the decisions can be explained by logical deduction.
8 Conclusions and Future Work The expressive power of logic enables domain knowledge properly captured and reasoned about in information agents. Accordingly, a symbolic framework for modelling the agents' learning and adaptation processes is desirable. The AGM belief revision paradigm o ers a sound and robust formalism to develop the learning components of adaptive information ltering agents. Since semantic relationships among information items are taken into account during the agent's learning and adaptation process, changes about a user's information requirements can automatically be deduced. Therefore, the proposed symbolic learning and adaptation framework as presented in this account seems to provide better learning autonomy than other reinforcement learning approaches that cannot, or do not, take semantic relationships into consideration. Moreover, the explicit representation of these semantic relationships also facilitate the explanation of the agents' learning and ltering behavior. However, quantitative evaluation of the e ectiveness and the eciency of these ltering agents are necessary. Future work involves simplifying the maxi-adjustment algorithm to gain higher computational eciency. Moreover, further extension of the proposed system may apply data mining agents to discover the association rules among the information items stored in AIFS. Thereby, the ltering agent's learning mechanism is able to acquire the background domain knowledge automatically.
Acknowledgments The work reported in this paper has been funded in part by the Cooperative Research Centres Program through the Department of the Prime Minister and Cabinet of Australia.
References 1. C.E. Alchourron, P. Gardenfors, and D. Makinson. On the logic of theory change: partial meet contraction and revision functions. Journal of Symbolic Logic, 50:510{530, 1985. 2. R. Armstrong, D. Freitag, T. Joachims, and T. Mitchell. Webwatcher: A learning apprentice for the world wide web. In AAAI Spring Symposium on Information Gathering, pages 6{12, 1995. 3. M. Balabanovic. An adaptive web page recommendation service. In W. Lewis Johnson and Barbara Hayes-Roth, editors, Proceedings of the First International Conference on Autonomous Agents (Agents'97), pages 378{385, New York, February 5{8, 1997. ACM Press. 4. N. Belkin and W. Croft. Information Filtering and Information Retrieval: Two sides of the same coin? Communications of the ACM, 35(12):29{38, 1992. 5. D. Billsus and M.J. Pazzani. A personal news agent that talks, learns and explains. In Proceedings of the Third International Conference on Autonomous Agents (Agents'99), pages 268{275, Seattle, WA, 1999. ACM Press. 6. C. Buckley, G. Salton, and J. Allan. The e ect of adding relevance information in a relevance feedback environment. In Proceedings of the Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Routing, pages 292{300, 1994. 7. Y. Chiaramella and J. P. Chevallet. About retrieval models and logic. The Computer Journal, 35(3):233{242, June 1992. 8. P. Gardenfors. Knowledge in ux: modeling the dynamics of epistemic states. The MIT Press, Cambridge, Massachusetts, 1988. 9. P. Gardenfors and D. Makinson. Revisions of knowledge systems using epistemic entrenchment. In Moshe Y. Vardi, editor, Proceedings of the Second Conference on Theoretical Aspects of Reasoning About Knowledge, pages 83{95, San Francisco, CA, 1988. Morgan Kaufmann Inc. 10. P. Gardenfors and D. Makinson. Nonmonotonic inference based on expectations. Arti cial Intelligence, 65(2):197{245, 1994.
11. A. Hunter. Using default logic in information retrieval. In C Froidevaux and J Kohlas, editors, Symbolic and Quantitative Approaches to Uncertainty, volume 946 of Lecture Notes in Computer Science, pages 235{242, 1995. 12. T. Kindo, H. Yoshida, T. Morimoto, and T. Watanabe. Adaptive personal information ltering system that organizes personal pro les automatically. In Martha E. Pollack, editor, Proceedings of the Fifteenth International Joint Conference on Arti cial Intelligence, pages 716{721, San Francisco, CA, August 23 { 29 1997. Morgan Kaufmann publishers Inc. 13. R. Lau, A. H. M. Hofstede, and P. D. Bruza. A Study of Belief Revision in the Context of Adaptive Information Filtering. In Proceedings of the Fifth International Computer Science Conference (ICSC'99), volume 1749 of Lecture Notes in Computer Science, pages 1{10, Berlin, Germany, 1999. Springer. 14. H. Lieberman. Letizia: An agent that assists web browsing. In Chris S. Mellish, editor, Proceedings of the Fourteenth International Joint Conference on Arti cial Intelligence, pages 924{929. Morgan Kaufmann publishers Inc.: San Mateo, CA, USA, August 20{ 25 1995. 15. A. Moukas and P. Maes. Amalthaea: An evolving information ltering and discovery system for the WWW. Journal of Autonomous Agents and Multi-Agent Systems, 1(1):59{88, 1998. 16. J.Y. Nie, M. Brisebois, and F. Lepage. Information retrieval as counterfactual. The Computer Journal, 38(8):643{657, 1995. 17. D.W. Oard. The state of the art in text ltering. User Modeling and User-Adapted Interaction, 7(3):141{178, 1997. 18. J. Rocchio. Relevance Feedback in Information Retrieval. In G. Salton, editor, The SMART retrieval system : experiments in automatic document processing, pages 313{323. Prentice-Hall, Englewood Cli s, N.J., 1971. 19. G. Salton and C. Buckley. Improving retrieval performance by relevance feedback. Journal of American Society for Information Science, 41(4):288{297, 1990. 20. G. Salton and M.J. McGill. Introduction to Modern Information Retrieval. McGraw-Hill, New York, 1983. 21. M.-A. Williams. Towards a practical approach to belief revision: Reason-based change. In Luigia Carlucci Aiello, Jon Doyle, and Stuart Shapiro, editors, KR'96: Principles of Knowledge Representation and Reasoning, pages 412{420, San Francisco, CA, 1996. Morgan Kaufmann Publishers Inc. 22. M.-A. Williams. Anytime belief revision. In Martha E. Pollack, editor, Proceedings of the Fifteenth International Joint Conference on Arti cial Intelligence, pages 74{79, San Francisco, CA, August 23 { 29 1997. Morgan Kaufmann Publishers Inc. 23. M.-A. Williams. Applications of belief revision. Lecture Notes in Computer Science, 1472:287{316, 1998.
Queensland University of Technology, Brisbane, Qld 4001, Australia fraymond,
[email protected]
2 Distributed Systems Technology Centre, The University of Queensland
Brisbane, Qld 4072, Australia
[email protected]
Abstract. Adaptive information ltering agents have been developed to alleviate the problem of information overload on the Internet by learning a user's changing information needs and autonomously scanning through the incoming stream of information according to the user's latest information needs. However, the explanatory power and the learning autonomy of these agents should be improved. Applying a logic-based framework for representation, learning, and matching in adaptive information ltering agents is promising since users' changing information needs can automatically be deduced by the agents. In addition, the inferred changes can be explained and justi ed based on formal deduction. This paper proposes a logic-based model for adaptive information ltering agents. Particularly, it illustrates how the AGM paradigm of belief revision can be applied to the learning processes of these agents. The impact of a logic-based framework on agents' learning autonomy and explanatory power is discussed through examples.
1 Introduction Information ltering (IF ) and information retrieval (IR) are "two sides of the same coin" [4]. However, IF is more concerned with the removal of irrelevant information from a stream of incoming information. With the explosive growth of the Internet and the World Wide Web (Web ), it is becoming increasingly dicult for users to retrieve relevant information. This is the so-called problem of information overload on the Internet. Augmenting existing Internet search engines with personalised information ltering tools is one possible method to alleviate this problem. Adaptive information ltering agents are computer systems situated on the Web. They autonomously lter the incoming stream of information on behalf of the users. Users' information needs will change over time, and information ltering agents must be able to revise their beliefs about the users' information needs so that the accuracy of the ltering process can be maintained. The AGM belief revision paradigm [1] provides a rich and rigorous foundation for modelling such revision processes. It enables an agent to modify its beliefs in a rational and minimal way. In addition, since semantic relationships among information items can be taken into account during the belief revision process, the agent can automatically deduce a user's changing information needs. Therefore, less users' direct relevance feedback [19] is required to train the ltering agents, and hence a higher level of learning autonomy can be achieved. This paper focuses on the application of the AGM paradigm of belief revision to the development of learning mechanisms in adaptive information ltering agents.
2 Related Work News Dude[5] is a multi-agent system which can learn a user's daily news interests incrementally. A hybrid approach of weighted keyword vector and Bayesian feature is employed to develop the user model. Basically, each feature indicates the presence or absence of a particular keyword. News ltering is conducted in two stages. Firstly, incoming news is compared with the vector representation of the user's information needs by the nearest neighbor algorithm, and the cosine similarity measure developed in IR.[20] Secondly, the new item will be classi ed again by the Bayesian classi er if it is not considered close enough to the vector user model. With the similarity score from stage one and the probability score from stage two, the recommender agent can rank the news item and present the highly rated items to a user. Agent adaptation is conducted through the reinforcement learning process where the users' relevance feedback is used to update the weights of the keyword vector and the probability distribution of the Bayesian classi er. Amalthaea[15] is a multi-agent system for information discovery and ltering on the Web. It employs keyword vectors to represent users' information needs and Web pages. Learning and adaptation is by relevance feedback and genetic algorithms. Each keyword is mapped to a ltering agent (i.e. a gene). A user's relevance feedback is converted to a numeric score and assigned to the ltering agents that carry the keywords appearing in the judged Web document. Then, a genetic process takes place (i.e. mutation and cross-over) among the highly rated ltering agents to produce the next generation of tter agents. Thereby, the Web pages presented by Amalthaea can converge to the user's taste. Fab[3] is a multi-agent system that retrieves and lters Web pages on behalf of the users. Fab's ltering agent employs both content based ltering as well as collaborative ltering.[17] For content based ltering, Fab is based on the vector space model[20] for knowledge representation and ltering. Learning is based on the users' relevance feedback and the Rocchio method[18] that updates the weights in a keyword vector, which represent the user's information needs. The adaptive information agents described so far are based on the users' direct relevance feedback to re ne their user models. This method produces high learning accuracy, but the disadvantage is that it demonstrates low learning autonomy (i.e. involving many human interventions). There are information agents[14, 2] that employ users' implicit feedback for learning and adaptation. Based on pre-de ned heuristic, these agents can infer users' relevance judgements by observing their online browsing behavior. For instance, if a user add a Web page to the hot list of the Web Browser, the agents will infer that the user considers this Web page as relevant. Basically, the agent extracts some tokens (e.g. keywords) from the Web pages stored in the hot list, and updates these tokens to the keyword vector, which represents a user's information needs. Thereby, the agent can infer what Web documents should be recommended to its user in future. The advantage of this type of learning approach is that less human interventions are required, and a higher level of learning autonomy is achieved. However, since the implicit heuristic and the keyword vector are two distinct knowledge representation schemes, it is still dicult to combine these two representations in order to explain an agent's behavior (i.e. why selecting a document and why not the others). Moreover, it is dicult to maintain the knowledge base when more heuristic is required by the agent.
3 System Architecture Figure 1 depicts the system architecture of AIFS [13] in which the adaptive information ltering agent is examined. A user communicates with AIFS via the interface agent. The retrieval agents
(RA) retrieve Web pages from external sources. The data management agent is responsible for house-keeping and characterisation of Web documents. The focus of this paper is on the adaptive information ltering agent. Particularly, the learning component of the ltering agent. Learning in the ltering agent consists of two steps. Firstly, based on a user's relevance feedback and the statistical data stored in AIFS's output queue, the learning component induces beliefs [8] about a user's information needs. Secondly, these beliefs are revised into the ltering agent's memory in a consistent and minimal way. The AGM belief revision framework [1] and the corresponding computational algorithms[21, 22] provides a sound and rigorous mechanism to carry out such an operation. Moreover, a user's changing information needs can be inferred by the agent through the belief revision process. The matching component carries out the ltering function based on logical deduction. Web documents deemed relevant by the ltering agent are transferred to the output queue via the Data Management Agent. Adaptive Filtering Agent
Retrieval Agents
User: Information needs
RA Interface Agent
Relevance Feedback
Learning Component
Memory
RA
External Agents
RA
External IR Sys.
Matching Component
Filtered Web Documents
Internet Search Engines
Data Management Agent
Filtered Web Documents
Web Documents
Output Queue
Input Queue
Fig. 1. The System Architecture of AIFS
4 The AGM Belief Revision Paradigm The AGM framework [1] formalises consistent and minimal belief changes by sets of postulates and belief functions e.g. expansion (K +), contraction (K ? ), and revision (K ). One of the constructions of these functions is by epistemic entrenchment (6) [9]. Beliefs with the lowest degree of epistemic entrenchment are given up when inconsistency arises because of applying changes. Nevertheless, for computer-based implementation, nite partial entrenchment ranking (B) which ranks a nite subset of beliefs with the minimum possible degree of entrenchment (6B ), and maxiadjustment [21, 22] which repeatedly transmutes (B) using an absolute measure of minimal change
under maximal information inertia, have been proposed. Williams [21] has formally de ned nite partial entrenchment ranking as a function B that maps a nite subset of sentences in L (e.g. classical propositional language) into the interval [0; O], where O is a suciently large ordinal such that the following conditions hold for all 2 dom(B): (PER1) f 2 dom(B) : B( ) < B( )g 6` ; (PER2) If ` : then B( ) = 0; (PER3) B( ) = O if and only if ` . The set of all partial entrenchment rankings is denoted B. B( ) is referred as the degree of acceptance of . The explicit beliefs of B 2 B is f 2 dom(B) : B( ) > 0g, and is denoted exp(B). Similarly, the implicit beliefs represented by B 2 B is Cn(exp(B)), and is denoted content(B). Cn is the classical consequence operation. The degree of acceptance of implicit belief is de ned as [21]: degree(B; ) = largest j such that f 2 exp(B) : B( ) j g ` if 2 content(B); otherwise degree(B; ) = 0. Let B 2 B be nite. The range of B is enumerated in ascending order as j0 ; j1 ; j2 ; : : : ; jO . Let be a contingent sentence, jm = degree(B; ) and 0 i < O. Then the ( ; i) maxi-adjustment of B is B? ( ; i) de ned by [21]: ? ? B ( ; i) = (B ( ; i))
if i jm (B? (: ; 0))+ ( ; i) otherwise
where for all 2 dom(B), B? ( ; i) is de ned as follows: 1. For with B( ) > jm ; B? ( ; i)( ) = B( ). 2. For with i < B( ) jm , suppose B? ( ; i)( ) for is de ned with B( ) jm?k for k = ?1; 0; 1; 2; : : :; n ? 1, then for with B( ) = jm?n , 8 i if ` or > > > > 6` and 2 ? > > < where ? is a minimal subset of ? B ( ; i)( ) = > f
: B( ) = jm?n g such that > > > f
: B? ( ; i)( ) > jm?n g [ ? ` > > : B( ) otherwise
3. For with B( ) i ; B? ( ; i)( ) = B( ). For all 2 dom(B) [ f g; B+ ( ; i) is de ned as follows:
8 B( ) if B( ) > i > > B( )if i< degree (B; ! ) > :
degree(B; ! ) otherwise
5 Knowledge Representation In IR, a retrieval situation was proposed to be the basis for evaluating document relevance [16]. A retrieval situation usually refers to situational factors such as the searcher's state of knowledge, background, intentions and so on. For the discussion in this paper, a retrieval situation also implies the searcher's topical information needs. The ltering agent's memory stores the abstraction of a retrieval situation based on the notion of beliefs. Each belief is represented as a formula of a classical rst-order language L. This paper focuses on two main elements of a retrieval situation, namely a user's information needs and their background knowledge.
5.1 Web documents A Web page is pre-processed based on traditional IR techniques [20]. For instance, a Web document is rst converted to plain text. The text is indexed resulting in a set of tokens. Stop words (e.g. "the", "a", etc.) are removed, the remaining tokens are stemmed. The stemmed tokens are then weighted according to the tf-idf measure [20], and those above a xed threshold are deemed to be suciently descriptive to represent the text. At the symbolic level, each selected token k is mapped to the ground term of the positive keyword predicate pkw e.g. pkw(k). Each pkw(k) is in fact a proposition since its interpretation is either true or false. Technically it is an atomic formula of a classical rst-order language L. For example, if fbusiness; commerce; trade; : : :g are the top n extracted tokens (e.g. keywords) from a document, the corresponding symbolic representation will be: fpkw(business); pkw(commerce); pkw(trade); : : :g.
5.2 Users' topical information needs Kindo's keyword classi er [12] is modi ed to induce a user's information need. In a relevance feedback environment, a user's information needs can be induced from a set of relevant documents D+ and a set of non-relevant documents D? judged by the user [6, 19]. The basic idea is that a keyword (i.e. token) appearing more frequently in D+ is probably an indicator of the user's information preference. On the other hand, a keyword appears frequently in D? may be an indicator of what the user does not want to retrieve. Therefore, the notions of positive, neutral, and negative keywords are proposed [12]. Intuitively, positive keywords represent the information items that a user requires, whereas negative keywords represent the information items that the user does not want. Neutral keywords are not useful for determining a user's information need. Accordingly, a positive keyword k is represented as a formula pkw(k), whereas a negative keyword is represented as :pkw(k). Eq.(1) is used to classify the keywords from D+ and D? , and to induce the preference value pre(k) for each keyword k. Whenever a Web page is judged by the user, the preference values of the set of keywords representing that page can be computed. The corresponding beliefs will be updated in the agent's memory via the belief revision mechanism.
pre(k) = tanh df(k) p(krel ) tanh p(pkrelrel ) ? (1 ? p(krel )) tanh (1(1??p(pkrelrel)))
(1)
where is used to restrict the range of pre(k) in the interval ?1 < pre(k) < 1. The examples illustrated in this paper assume that = 0:9. df (k) is the sum of the number of relevant documents df (krel ) and the number of non-relevant documents df (knrel ) that contains the keyword k, and tanh is the hyperbolic tangent. The rarity parameter is used to control rare or new keywords and is expressed as int(logN + 1), where N is the total number of Web documents judged by a user, and int is an integer function that truncates the decimal values. p(krel ) is the estimated probability (krel ) that a document containing keyword k is relevant and is expressed as the fraction df (kreldf)+ df (knrel ) . prel is the estimated probability that a document is relevant. In our system, it is assumed that the probability that a Web document presented by the ltering agent and judged as relevant by a user is prel = 0:5. A positive value of pre(k) implies that the associated keyword is positive, whereas a negative value of pre(k) indicates a negative keyword. If pre(k) is below a threshold value , the associated keyword is considered neutral. It is assumed that = 0:5 for the examples demonstrated
in this paper. Basically a positive keyword k is mapped to pkw(k), and a negative keyword k is mapped to :pkw(k). There is no need to create the symbolic representations for neutral keywords. For pkw(k) or :pkw(k), the degree of acceptance B( k ) of the corresponding formula k is de ned as:
jpre(k)j B( k ) = 0jpre(k)j ifotherwise
(2)
Table 1 shows the results of applying Eq.(1) and Eq.(2) to induce the beliefs and the associated degree of acceptance based on D+ and D? . It is assumed that there are a set of ve documents having been judged as relevant (i.e. jD+ j = 5) and another set of ve documents having been judged as non-relevant by a user (i.e. jD? j = 5). Each document is characterised by a set of keywords e.g. d1 = fbusiness; commerce; systemg; d2 = fart; sculptureg.
Table 1. Inducing a user's information preferences Keywords D+ D? pre(k) Formula: k B( k ) business commerce system art sculpture insurance
5 4 2 0 0 1
0 0 2 5 3 0
0.856 pkw(business) 0.836 pkw(commerce) 0 -0.856 :pkw(art) -0.785 :pkw(sculpture) 0.401 -
0.856 0.836 0.856 0.785 -
5.3 Background knowledge A user's state of knowledge includes their understanding about a particular application domain. For example, in the domain of databases, classi cation knowledge such as (O-O databases ! databases) can be used by a searcher. Linguistic knowledge such as (VDU monitor) may also be used in a search. In our current framework, background knowledge is assumed to be elicited from a particular user or to be transferred from a thesaurus of a particular domain. A manual procedure is involved to encode these knowledge in the form of rst-order formulae. The following is an example of background knowledge used for the discussion in this paper. In the context of IR, these rules can be interpreted as: an information item about "business" implies that it is about "commerce" and vice versa; "sculpture" is a kind of Arts.
pkw(business) pkw(commerce); 1:000 pkw(sculpture) ! pkw(art); 1:000
6 Learning and Adaptation Whenever a user provides relevance feedback for a presented Web document, the belief revision process can be invoked to learn the user's information preferences. Conceptually, the ltering
agent's learning and adaptation mechanism is characterised by the belief revision and contraction processes. For example, if ? = fa1; a2 ; : : : ; an g is a set of formulae representing a Web document d 2 D+ , the belief revision process (((Ka1 )a2 ) : : :)an is invoked for each ai 2 ? , where K is the belief set stored in the ltering agent's memory. On the other hand, the belief contraction process (((Ka?1 )?a2 ) : : :)?an is applied for each ai 2 ? and d 2 D? . At the computational level, belief revision is actually taken as the adjustment of entrenchment ranking B in the theory base exp(B). Particularly, maxi-adjustment [21, 23] is employed by our system to modify the ranking in an absolute minimal way under maximal information inertia. As the input to the maxi-adjustment algorithm includes an ordinal i representing the new degree of acceptance of a sentence , the procedure described in Section 5.2 is used to obtain the new degree for each 2 ? . Moreover, for our implementation of the maxi-adjustment algorithm, the maximal ordinal O is chosen as 1. Therefore, any formula assigned the entrenchment value 1 will not be removed from the theory base exp(B).
Example 1:
The rst example shows how adding one belief to the agent's memory will automatically raise the entrenchment rank of another related belief. It is assumed that the belief B(:pkw(sculpture)) = 0:785 and the belief B(pkw(business)) = 0:856 have been learnt by the ltering agent. If several Web documents characterised by the keyword art are also judged as non-relevant by the user later on, the preference value of the keyword art can be induced according to Eq.(1). Assuming that pre(art) = ?0:856, the corresponding entrenchment rank can be computed as B(:pkw(art)) = 0:856 according to Eq.(2). By applying B? (:pkw(art); 0:856) to the theory base exp(B), the before and after images of the agent's explicit beliefs (i.e. exp(B)) can be tabulated in Table 2. Based on the maxi-adjustment algorithm, B+ ( ; i)( ) = i if B( ) i < degree(B; ! ). * B(:pkw(sculpture)) 0:856 < degree(B; :pkw(art) ! :pkw(sculpture)) ) B+ (:pkw(art); 0:856)(:pkw(sculpture)) = 0:856
The implicit belief :pkw(art) ! :pkw(sculpture) in content(B) is derived from the explicit belief pkw(sculpture) ! pkw(art) in the theory base exp(B), and its degree of acceptance is 1 according to the de nition of degree. As the belief :pkw(art) implies the belief :pkw(sculpture) and the agent believes in :pkw(art), the belief :pkw(sculpture) should be at least as entrenched as the belief :pkw(art) according to (PER1). In other words, whenever the agent believes that the user is not interested in art (i.e. :pkw(art)), it must be prepared to accept that the user is also not interested in sculpture at least to the degree of the former. The proposed learning and adaptation framework is more e ective than other learning approaches that can not take into account the semantic relationships among information items. This example demonstrates the automatic revision of the agent's beliefs about related keywords given the relevance feedback for a particular keyword. Therefore, less users' relevance feedback is required during reinforcement learning. Consequently, learning autonomy of the ltering agent can be enhanced.
Example 2:
The second example illustrates the belief contraction process. In particular, how the contraction of one belief will automatically remove another related belief from the agent's memory if there is a semantic relationship between the underlying keywords. Assuming that more Web documents characterised by the keyword sculpture are judged as relevant by the user at a later stage, the
Table 2. Raising related beliefs Formula: B( ) Before B( ) After pkw(business) pkw(commerce) 1.000 1.000 pkw(sculpture) ! pkw(art) 1.000 1.000 pkw(business) 0.856 0.856 :pkw(sculpture) 0.785 0.856 :pkw(art) 0 0.856 belief B(pkw(sculpture)) = 0:785 could be induced. As B? ( ; i) = (B? (: ; 0))+ ( ; i) if i > jm , where i = 0:785 and jm = 0 in this example, B? ( pkw(sculpture); 0:785) leads to the contraction of the belief :pkw(sculpture) from the theory base exp(B). Moreover, B? ( : pkw(sculpture); 0) (: pkw(art)) = 0 is computed because B( : pkw( art)) = 0:856 = jm?n and the set f : B? ( : pkw(sculpture); 0) ( ) > 0:856 g [ f : pkw(art) g ` : pkw(sculpture) is obtained. The before and after images of the ltering agent's explicit beliefs are tabulated in Table 3.
Table 3. Contracting related beliefs Formula: B( ) Before B( ) After pkw(business) pkw(commerce) 1.000 1.000 pkw(sculpture) ! pkw(art) 1.000 1.000 pkw(business) 0.856 0.856 pkw(sculpture) 0 0.785 :pkw(sculpture) 0.856 0 :pkw(art) 0.856 0 The explanation of the above learning processing is that given a user's new information preference about sculpture, the previous requirement of not sculpture should be removed from the agent's memory since it is impossible for the user to require Web documents about sculpture, and not require Web documents about sculpture at the same time. Moreover, as sculpture implies art, a user who rejects Web documents about sculpture may also reject documents about art. The belief revision mechanism exactly carries out this kind of reasoning to infer the possible changes of a user's information needs. As can be seen, the ltering agent's behavior can be explained based on logical deduction and the explicit semantic relationships among information items. The process is more transparent than examining the weight changes in a keyword vector. In summary, the agent's memory consists of the following beliefs before the learning process:
pkw(business) pkw(commerce); 1:000 pkw(sculpture) ! pkw(art); 1:000 pkw(business); 0:856 :pkw(sculpture); 0:856 :pkw(art); 0:856 However, after the belief revision process, the ltering agent's memory becomes:
pkw(business) pkw(commerce); 1:000 pkw(sculpture) ! pkw(art); 1:000 pkw(business); 0:856 pkw(sculpture); 0:785
7 Filtering Web Documents In our current framework, the matching function of the ltering agent is modelled as logical deduction. Moreover, a Web document is taken as the conjunction of a set of formulae [7, 11]. The following example illustrates the agent's deduction process with reference to previous examples. The resulting belief sets K = Cn(exp(B)) from examples 1 and 2 in the previous section are used to determine the relevance of the following three Web documents:
= fpkw(business) ^ pkw(art)g ' = fpkw(sculpture) ^ pkw(art)g = fpkw(business) ^ pkw(commerce)g The ltering agent's conclusions about the relevance of the Web documents are summarised as follows: Time: (t1)
Time: (t2)
K = Cn(fpkw(business) pkw(commerce); pkw(sculpture) ! pkw(art); pkw(business); :pkw(sculpture); :pkw(art)g) ) K 6` ; K 6` '; K ` K = Cn(fpkw(business) pkw(commerce); pkw(sculpture) ! pkw(art); pkw(business); pkw(sculpture)g) ) K ` ; K ` '; K `
At time (t1), the ltering agent believes that the user does not need Web documents about art. Therefore, documents and ' are rejected. Moreover, since business is equivalent to commerce for this particular information searcher, and she is interested in business, a Web document about business and commerce should also be relevant to her. At time (t2), after the belief revision process, the agent learns that the user is interested in sculpture. As sculpture is a kind of art, the agent can deduce that she may want Web documents about art as well. Therefore, documents and ' are considered as relevant at this time. So, a logic-based framework for representation and reasoning facilitates the explanation of an agent's decisions for information retrieval since the decisions can be explained by logical deduction.
8 Conclusions and Future Work The expressive power of logic enables domain knowledge properly captured and reasoned about in information agents. Accordingly, a symbolic framework for modelling the agents' learning and adaptation processes is desirable. The AGM belief revision paradigm o ers a sound and robust formalism to develop the learning components of adaptive information ltering agents. Since semantic relationships among information items are taken into account during the agent's learning and adaptation process, changes about a user's information requirements can automatically be deduced. Therefore, the proposed symbolic learning and adaptation framework as presented in this account seems to provide better learning autonomy than other reinforcement learning approaches that cannot, or do not, take semantic relationships into consideration. Moreover, the explicit representation of these semantic relationships also facilitate the explanation of the agents' learning and ltering behavior. However, quantitative evaluation of the e ectiveness and the eciency of these ltering agents are necessary. Future work involves simplifying the maxi-adjustment algorithm to gain higher computational eciency. Moreover, further extension of the proposed system may apply data mining agents to discover the association rules among the information items stored in AIFS. Thereby, the ltering agent's learning mechanism is able to acquire the background domain knowledge automatically.
Acknowledgments The work reported in this paper has been funded in part by the Cooperative Research Centres Program through the Department of the Prime Minister and Cabinet of Australia.
References 1. C.E. Alchourron, P. Gardenfors, and D. Makinson. On the logic of theory change: partial meet contraction and revision functions. Journal of Symbolic Logic, 50:510{530, 1985. 2. R. Armstrong, D. Freitag, T. Joachims, and T. Mitchell. Webwatcher: A learning apprentice for the world wide web. In AAAI Spring Symposium on Information Gathering, pages 6{12, 1995. 3. M. Balabanovic. An adaptive web page recommendation service. In W. Lewis Johnson and Barbara Hayes-Roth, editors, Proceedings of the First International Conference on Autonomous Agents (Agents'97), pages 378{385, New York, February 5{8, 1997. ACM Press. 4. N. Belkin and W. Croft. Information Filtering and Information Retrieval: Two sides of the same coin? Communications of the ACM, 35(12):29{38, 1992. 5. D. Billsus and M.J. Pazzani. A personal news agent that talks, learns and explains. In Proceedings of the Third International Conference on Autonomous Agents (Agents'99), pages 268{275, Seattle, WA, 1999. ACM Press. 6. C. Buckley, G. Salton, and J. Allan. The e ect of adding relevance information in a relevance feedback environment. In Proceedings of the Seventeenth Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Routing, pages 292{300, 1994. 7. Y. Chiaramella and J. P. Chevallet. About retrieval models and logic. The Computer Journal, 35(3):233{242, June 1992. 8. P. Gardenfors. Knowledge in ux: modeling the dynamics of epistemic states. The MIT Press, Cambridge, Massachusetts, 1988. 9. P. Gardenfors and D. Makinson. Revisions of knowledge systems using epistemic entrenchment. In Moshe Y. Vardi, editor, Proceedings of the Second Conference on Theoretical Aspects of Reasoning About Knowledge, pages 83{95, San Francisco, CA, 1988. Morgan Kaufmann Inc. 10. P. Gardenfors and D. Makinson. Nonmonotonic inference based on expectations. Arti cial Intelligence, 65(2):197{245, 1994.
11. A. Hunter. Using default logic in information retrieval. In C Froidevaux and J Kohlas, editors, Symbolic and Quantitative Approaches to Uncertainty, volume 946 of Lecture Notes in Computer Science, pages 235{242, 1995. 12. T. Kindo, H. Yoshida, T. Morimoto, and T. Watanabe. Adaptive personal information ltering system that organizes personal pro les automatically. In Martha E. Pollack, editor, Proceedings of the Fifteenth International Joint Conference on Arti cial Intelligence, pages 716{721, San Francisco, CA, August 23 { 29 1997. Morgan Kaufmann publishers Inc. 13. R. Lau, A. H. M. Hofstede, and P. D. Bruza. A Study of Belief Revision in the Context of Adaptive Information Filtering. In Proceedings of the Fifth International Computer Science Conference (ICSC'99), volume 1749 of Lecture Notes in Computer Science, pages 1{10, Berlin, Germany, 1999. Springer. 14. H. Lieberman. Letizia: An agent that assists web browsing. In Chris S. Mellish, editor, Proceedings of the Fourteenth International Joint Conference on Arti cial Intelligence, pages 924{929. Morgan Kaufmann publishers Inc.: San Mateo, CA, USA, August 20{ 25 1995. 15. A. Moukas and P. Maes. Amalthaea: An evolving information ltering and discovery system for the WWW. Journal of Autonomous Agents and Multi-Agent Systems, 1(1):59{88, 1998. 16. J.Y. Nie, M. Brisebois, and F. Lepage. Information retrieval as counterfactual. The Computer Journal, 38(8):643{657, 1995. 17. D.W. Oard. The state of the art in text ltering. User Modeling and User-Adapted Interaction, 7(3):141{178, 1997. 18. J. Rocchio. Relevance Feedback in Information Retrieval. In G. Salton, editor, The SMART retrieval system : experiments in automatic document processing, pages 313{323. Prentice-Hall, Englewood Cli s, N.J., 1971. 19. G. Salton and C. Buckley. Improving retrieval performance by relevance feedback. Journal of American Society for Information Science, 41(4):288{297, 1990. 20. G. Salton and M.J. McGill. Introduction to Modern Information Retrieval. McGraw-Hill, New York, 1983. 21. M.-A. Williams. Towards a practical approach to belief revision: Reason-based change. In Luigia Carlucci Aiello, Jon Doyle, and Stuart Shapiro, editors, KR'96: Principles of Knowledge Representation and Reasoning, pages 412{420, San Francisco, CA, 1996. Morgan Kaufmann Publishers Inc. 22. M.-A. Williams. Anytime belief revision. In Martha E. Pollack, editor, Proceedings of the Fifteenth International Joint Conference on Arti cial Intelligence, pages 74{79, San Francisco, CA, August 23 { 29 1997. Morgan Kaufmann Publishers Inc. 23. M.-A. Williams. Applications of belief revision. Lecture Notes in Computer Science, 1472:287{316, 1998.
