#### DMCA

## Knowledge Discovery in Databases (1996)

### Cached

### Download Links

- [piggy.cs.nthu.edu.tw]
- [ftp.cs.ust.hk]
- [ftp.cs.ust.hk]
- DBLP

### Other Repositories/Bibliography

Citations: | 9 - 4 self |

### Citations

6592 |
C4.5: Programs for Machine Learning
- Quinlan
- 1993
(Show Context)
Citation Context ...tion), and FQL . The basic work on symbolic Machine learning has been carried out using the knowledge representation tool of a decision tree which is equivalent to propositional deduction rules (e.g. =-=[26, 33]-=-). More recent approaches use Datalog as knowledge representation language [27, 10]. Datalog is more expressive than propositional rules, however, unlike the later, Datalog is not able to handle uncer... |

6205 |
Fuzzy sets
- Zadeh
- 1965
(Show Context)
Citation Context ...sing how to extend Datalog and FQL to handle probabilistic information. 3 UNCERTAINTY 37 3.2 Other Approaches to Uncertainty We briefly give an overview on fuzzy sets whose roots go back to the paper =-=[55]-=-. A similar, a bit longer such overview can be found in [17]. Given is an universe of discourse U , a set of elements. Each subset S of U is characterized by a membership functionsS : U ! [0; 1]. S is... |

5951 |
Classification and Regression Trees
- Breiman, Friedman, et al.
- 1984
(Show Context)
Citation Context ...ty for data mining than previously developed systems and techniques. The features of DISCOVER are compared with the features of some other techniques. These techniques include ffl decision trees (DT) =-=[16, 77]-=- ffl regression analysis (RA) [48] CHAPTER 9. LEARNING PROBABILISTIC FQL* 119 ffl Inductive Logic Programming (ILP) [62, 27] ffl neural nets (NN) [35, 47] Figure 9.1: Starting DISCOVER. 9.1 Mining Tec... |

4361 | Induction of decision trees
- Quinlan
- 1986
(Show Context)
Citation Context ...question is of course when and where to prune. There seems to be no generally accepted and applicable pruning strategy, hence when and where to prune has more or less to be heuristically decided (see =-=[31]-=- for more on simplifying propositional rules the equivalent of pruning decision trees). 5 LEARNING PROPOSITIONAL RULES AND DECISION TREES 67 We already anticipate here that the mean squared error, pro... |

3998 |
Introduction to Modern Information Retrieval
- SALTON, MCGILL
- 1983
(Show Context)
Citation Context ...put. These days, neural net techniques seem to be the most popular approaches to deal with weighted input and weighted training examples. Finally, we mention that in the area of information retrieval =-=[36]-=- weighting has always been used for retrieving information and predicting for instance what document fits the library user's need best. 5.4 Further Issues An often used notion in symbolic machine lear... |

3600 | Fast algorithms for mining association rules
- Agrawal, Srikant
(Show Context)
Citation Context ... discussed in chapter 6. In contrast to those ideas the novelties are as follows. ffl The knowledge representation language FQL is much more expressive and powerful than Datalog and association rules =-=[5]-=- used by other learning systems. This is mainly due to the fact that we deal with probabilistic information whereas other systems cannot handle probabilistic information. ffl The user can fine-tune th... |

2932 | Maintaining knowledge about temporal intervals
- Allen
- 1983
(Show Context)
Citation Context ...ll predict the competitiveness from the premises. The expert classified each premise of a company into one of seven columns (very bad to very good, see figure 2). We simply divide the closed interval =-=[0; 1]-=- equally by the seven numbers 0; 0:17; 0:33; 0:5; 0:67; 0:83; 1 and attach these numbers as weights to the premise facts instead. The McKinsey matrix has 3 columns, so that a firm can be seen as nonco... |

2206 |
Introduction to the Theory of Neural Computation
- Hertz, Krogh, et al.
- 1991
(Show Context)
Citation Context ...r would now be very large hence making adjusted gain(X) small. In section 7 we will discuss error measures. Especially we show that the mean squared error also successfully used to train neural nets (=-=[21]-=-) does not have the problems of the entropy measures. The author of this report seems actually to be the first one who proposed and used mean squared error in symbolic machine learning. 5.3 Generating... |

2069 |
Foundations of Logic Programming
- LLOYD
- 1987
(Show Context)
Citation Context ...bottom up evaluation strategy. We are however aware of the 2 RULE LANGUAGES 25 many other proposals defining a semantics of Datalog rules. Among those proposals we only name the Clark completion (see =-=[24]-=-). Suppose now the following rules and facts. has cancer(x) / person(x); :non smoker(x) (46) has cancer(x) / ancestor(y; x); has cancer(y) (47) person(x) / parent(x; y) (48) person(y) / parent(x; y) (... |

1401 | Depth-first search and linear graph algorithms - Tarjan - 1972 |

934 | Learning logical definitions from relations
- Quinlan
- 1990
(Show Context)
Citation Context ...the hypothesis space and hence easies and accelerates the rule generation tremendously. The following restrictions can be often found. 1. No constants may occur in the rules (e.g. in LINUS [10], FOIL =-=[32]-=- and ProbKB [50]). 2. For reasons discussed in section 2.2, a rule has to be allowed. 3. The set of possible variables occurring in a single rule is also restricted a priory to be a subset of a fixed ... |

890 | The CN2 induction algorithm - Clark, Niblett - 1989 |

855 |
UCI Repository of Machine Learning Databases
- Murphy, Aha
(Show Context)
Citation Context ...graph plotting the success rate against the number of training data. Note that the tendency of the curve is stabilizing. 8.3.2 Postoperative Patient Data Postoperative Patient data is from Uni Irvine =-=[64]-=-. The classification task of this database is to determine where patients in a postoperative recovery area should be sent to next. Because hypothermia is a significant concern after surgery, the attri... |

802 | A Predicate Logic as a Programming Language
- KOWALSKI
- 1974
(Show Context)
Citation Context ...literal in the rule body holds. We define the semantics of a given rule set and fact set. A formal treatment of this semantics is for instance given in [2] and [47]. These two descriptions generalize =-=[12] for that -=-also negation in rule bodies is taken into account. We give here the "standard" semantics of such rules in the database area by describing a concrete bottom up evaluation strategy. We are ho... |

709 | Efficient and effective clustering methods for spatial data mining
- Ng, Han
- 1994
(Show Context)
Citation Context ... bases involving large factual information. This approach also provides no solution for the partial dependency problem previously outlined. There are many other fine and noteworthy studies (e.g., see =-=[10, 66, 65]-=-). Some of these proposals allow very general modeling capabilities we do not include in our semantics. None of those works, CHAPTER 3. UNCERTAINTY 47 however, provides a probabilistic way of dealing ... |

681 |
Towards a theory of declarative knowledge
- Apt, Blair, et al.
- 1988
(Show Context)
Citation Context ...ay requires a kind of normalization at some intermediate steps during rule evaluation. This is mainly needed to handle queries involving duration. If h(1) is true during the set of intervals f[1; 3]; =-=[2; 4]g 4 TIME 49 then the-=- evaluation of I : duration(I) ? 2) h(1) should return [1; 4], so "meeting" and "overlapping" intervals must be merged together. A similar but more general approach to handle time ... |

489 |
Probabilistic logic
- Nilsson
- 1986
(Show Context)
Citation Context ... both have the interval [0:5; 0:5]. But if we have only the first rule and the same facts, then p gets the uncertainty [0:5; 0:5]. A way to attach a probability to an arbitrary formula is proposed in =-=[69]-=-. But Nilsson does not deal with rule-based reasoning, and this method suffers from computational intractability. The computation of an uncertainty value according to this theory needs exponential tim... |

367 |
The multipurpose incremental learning system AQ15 and its testing application to three medical domains
- Michalski, Mozetic, et al.
- 1986
(Show Context)
Citation Context ...tion), and FQL . The basic work on symbolic Machine learning has been carried out using the knowledge representation tool of a decision tree which is equivalent to propositional deduction rules (e.g. =-=[26, 33]-=-). More recent approaches use Datalog as knowledge representation language [27, 10]. Datalog is more expressive than propositional rules, however, unlike the later, Datalog is not able to handle uncer... |

334 |
Principles of Database System
- Ullman
- 1982
(Show Context)
Citation Context ...ires explicit training data and the goal (classify into play or don't play) to be given and known respectively. Note also that this learning algorithm can only be applied to one single flat relation (=-=[45]-=-). If we would like to take into account attributes of objects spread over several tables then we would first have to join these tables to yield one single big table and then to apply this algorithm o... |

331 | The Temporal Query Language TQuel
- Snodgrass
- 1987
(Show Context)
Citation Context ...and ffl how certain it is that C bought or will buy US dollars. Many sophisticated methods to handle time in databases have been proposed and a lot of progress has been achieved in the process (e.g., =-=[85, 88]-=-). Similar efforts have been made to establish a practical and expressive model of time in the artificial intelligence area (e.g., [6, 38, 87]). Whereas in the database area much has been achieved to ... |

308 | Automated learning of decision rules for text categorization
- Apte, Damerau, et al.
- 1994
(Show Context)
Citation Context ... of all those possibilities we choose the best performing one. Afterwards, the second atom is deleted or swapped out; and so on. The swapping is actually a variation of an idea due to the Swap system =-=[8]-=- 4 . This process generates the first conjunct, in our example, (bonds(t1) ENTAILS ft1g) AND (political news(t1) ENTAILS ft1g) Subsequently, a second conjunct is generated in the way described. The ad... |

300 |
The Functional Data Model and the Data Language DAPLEX.
- Shipman
- 1981
(Show Context)
Citation Context ...less expensive fuzzy and possibilistic semantics for instance. ffl We illustrate the concepts enhancing a functional and object-oriented query language proposed in [81] which extents the Duplex model =-=[84]-=- on recursion. It is, however, straightforward applying the same concepts to the relational model. It has been widely recognized that the handling of time can be categorized along the following main l... |

292 | Approximating probabilistic inference in Bayesian belief networks is NP-hard
- Dagum, Luby
- 1993
(Show Context)
Citation Context ... (TKB ), and lfp 0 (TKB ) is computable in polynomial time. However, approximation in Bayesian belief nets (a Bayesian belief net is equivalent to a set of propositional rules) can go arbitrarily bad =-=[24]-=-, hence, even though our approximation looks very promising for many cases, it can go arbitrarily bad. The precision of the approximation lfp 0 (TKB ) is specified by choosing a constant k. This const... |

285 |
Representing and reasoning with probabilistic knowledge: a logical approach to probabilities
- Bacchus
- 1990
(Show Context)
Citation Context ... bases involving large factual information. This approach also provides no solution for the partial dependency problem previously outlined. There are many other fine and noteworthy studies (e.g., see =-=[10, 66, 65]-=-). Some of these proposals allow very general modeling capabilities we do not include in our semantics. None of those works, CHAPTER 3. UNCERTAINTY 47 however, provides a probabilistic way of dealing ... |

234 |
The management of probabilistic data
- Barbara, Garcia-Molina, et al.
- 1992
(Show Context)
Citation Context ...odel of time in the artificial intelligence area (e.g., [6, 38, 87]). Whereas in the database area much has been achieved to handle time, the efforts to deal with uncertainty are not that many (e.g., =-=[12, 13, 99]-=-). On the other hand, in artificial intelligence there are also numerous proposals of how to deal with uncertainty (e.g., see [39, 71]). Interestingly, however, there seems to be no proposal trying to... |

213 | Efficient distribution‐free learning of probabilistic concepts
- Kearns, Schapire
- 1994
(Show Context)
Citation Context ...To guide the search down the hypothesis tree we have to know what the "best" or "right" solution is. For probabilistic concepts or rules we know that the right hypothesis has the l=-=east quadratic loss [51]-=-. Hence, having a new rule candidate s i generated as a direct specialization of r the gain of replacing r by s i is gain(r; s i ) = E(F; R+ r) \Gamma E(F; R+ s i ) (7.1) where E(KB) is the estimated ... |

190 |
Data-Driven Discovery of Quantitative Rules in Relational Databases,
- Han, Cai, et al.
- 1993
(Show Context)
Citation Context ...erful. 6. In contrast to DISCOVER, regression analysis can only deal with numbers, not with symbolic data. CHAPTER 9. LEARNING PROBABILISTIC FQL* 128 7. In contrast to other mining techniques such as =-=[43, 44, 65]-=-, DISCOVER can also deal with weighted information. From above list, the experimental results and the sophisticated functionality DISCOVER offers follows that DISCOVER is a promising data mining tool ... |

187 |
Structural Complexity I
- Balcázar, Dı́az, et al.
- 1988
(Show Context)
Citation Context ...such a way requires a kind of normalization at some intermediate steps during rule evaluation. This is mainly needed to handle queries involving duration. If h(1) is true during the set of intervals f=-=[1; 3]; [2; 4]g 4 TIME 49 -=-then the evaluation of I : duration(I) ? 2) h(1) should return [1; 4], so "meeting" and "overlapping" intervals must be merged together. A similar but more general approach to hand... |

172 | Generating production rules from decision trees
- Quinlan
- 1987
(Show Context)
Citation Context ...l rules which are equivalent in expressive power to decision trees. In this section we introduce basic symbolic machine learning techniques. More details on the topics introduced here can be found in =-=[26, 30, 33]-=-. The emphasis here is on learning decision trees, but since every learned decision tree can afterwards be transformed into a propositional rule representation having the advantages discussed in secti... |

156 | Probabilistic logic programming
- Ng, Subrahmanian
- 1992
(Show Context)
Citation Context ... bases involving large factual information. This approach also provides no solution for the partial dependency problem previously outlined. There are many other fine and noteworthy studies (e.g., see =-=[10, 66, 65]-=-). Some of these proposals allow very general modeling capabilities we do not include in our semantics. None of those works, CHAPTER 3. UNCERTAINTY 47 however, provides a probabilistic way of dealing ... |

125 |
Modern Probability Theory and Its Applications.
- Parzen
- 1960
(Show Context)
Citation Context ...en (2 S ; "; [) is a Boolean algebra where " and [ denote set intersection and set union respectively. We introduce that part of probability theory we really need and recommend the excellent=-= textbook [29]-=- for further studies in probability theory. Let (A; ; ) be a Boolean algebra and P : A ! R be a function from domain A to the real numbers. Function P is called a probability function if for each elem... |

118 | Systems for knowledge discovery in databases,
- Matheus, Chan, et al.
- 1993
(Show Context)
Citation Context ...ng phase. Chapter 8 Consistent Forecastings 8.1 Introduction In various application domains such as business, manufacturing and medicine, forecasting and discovery tools become increasingly important =-=[4, 59]-=-. Sales managers base forecasts on various product history variables; marketers consider the complex set of buyer preferences and preferred product options; financial analysts classify levels of credi... |

114 |
A Guide To Sql Standard.
- Date, Darwen
- 1996
(Show Context)
Citation Context ...each Datalog rule can be translated to an equivalent relational algebra expression. This means that we can - on top of an existing database system interfaced through relational algebra (probably SQL, =-=[9]-=-) - build a Datalog rule evaluator. The facts are then stored as tuples in tables and on top of this database the rules are evaluated by executing their equivalent SQL representation on the database s... |

105 | Quantitative deduction and its fixpoint theory
- Emden
- 1986
(Show Context)
Citation Context ...n and selection (see [45, pp 151-153]) to their fuzzy counterparts. This creates the field of "fuzzy databases". We will here however shortly discuss the application of fuzzy logic to extend=-= Datalog ([11]). This re-=-veals the strength of fuzzy logic in rule based systems, computational cheapness, and its weakness, "not precise enough calculation". We extend our toy database (46) to (55) introduced in se... |

93 | Supporting valid-time indeterminacy
- DYRESON, R
- 1998
(Show Context)
Citation Context ...at dealing with several dimensions of linear orderings and uncertainty is needed in two or three dimensional geographic applications. One of the few remarkable papers dealing with time uncertainty is =-=[26]-=-. Parts of these ideas also went in the TSQL2 language definition [32]. The former paper discusses how to deal with time uncertainty (called time indeterminacy). However, the paper does not provide a ... |

76 |
Data Models
- Tsichritzis, Lochovsky
- 1982
(Show Context)
Citation Context ... expressed by means of single real numbers, time representations seem to be more naturally done using intervals. Finally, where to attach is basically a decision imposed by the underlying data model (=-=[44]-=-). FQL is objectoriented, hence the weights will be attached on the attribute level although a simulation technique (see 4.3) can be used to also weight on the object level. Datalog is based on the re... |

74 |
Probabilistic Inference in Intelligent Systems. Networks of Plausible Inference
- Pearl
- 1988
(Show Context)
Citation Context ...the efforts to deal with uncertainty are not that many (e.g., [12, 13, 99]). On the other hand, in artificial intelligence there are also numerous proposals of how to deal with uncertainty (e.g., see =-=[39, 71]-=-). Interestingly, however, there seems to be no proposal trying to provide a uniform model of how to express time and uncertainty. [54] deals with either time or uncertainty but not with both simultan... |

73 |
Decision support and expert systems
- Turban
- 1990
(Show Context)
Citation Context ...h factors are the most influential for predicting the Deutsch Mark against the US dollar for instance. There are other approaches for producing forecastings in financial application domains (e.g. see =-=[50, 2, 31, 3, 91]-=- describing approaches by J.P. Morgan, New York, Olson & Associates in Zurich and many others). But all those techniques and approaches rely on merely quantifiable data. Hence,they do not try to take ... |

69 | Data mining: The search for knowledge in databases
- Holsheimer, Siebes
- 1994
(Show Context)
Citation Context ...earning capabilities. In the literature, also the term data mining is often used for what we call knowledge discovery. An attempt to give a definition for "data mining" is for instance the f=-=ollowing ([22]-=-). Data mining is the search for relationships and global patterns that exist in large databases, but are `hidden' among the vast amounts of data, such as a relationship between patient data and their... |

65 | A semantical framework for supporting subjective and conditional probabilities in deductive databases
- Ng, Subrahmanian
- 1993
(Show Context)
Citation Context ...applying the reasoning proposed in [41] whereas our method works for this case as well. Additionally, we are considering rule-based reasoning even involving recursion which is not considered in [41]. =-=[67]-=- is another approach to uncertainty reasoning. It uses the same uncertainty intervals as [68], but the rule heads or facts may be conjunctions or disjunctions of atoms. Yet every step of the proposed ... |

61 | A probabilistic framework for vague queries and imprecise information in databases
- Fuhr
- 1990
(Show Context)
Citation Context ...xt extend Datalog and FQL in a probabilistic way. 3.3 Probablistic Datalog Many real world problems cannot be described or solved by deterministic information because of inherent vagueness (e.g., see =-=[12, 36]-=-). We demonstrate the truth of this statement in a real-world application to which we have successfully applied the rule-based semantics described and defined in this paper. Assume a database containi... |

57 |
Logic programs with uncertainties: A tool for implementing rulebased systems.
- Shapiro
- 1983
(Show Context)
Citation Context ...nagement systems we would like to have a means to handle rules involving variables and recursion, neither of which are provided by the Mycin approach. Our proposal is closest to the excellent studies =-=[83]-=- and [53]. The first difference to our approach is that both of these studies are restricted to negation-free rules. In [83] a model theoretic semantics is given without providing a fixpoint semantics... |

56 |
Theory of database queries
- Chandra, A
- 1988
(Show Context)
Citation Context ...emantics making it tractable in real-world applications. It has been argued that the design of a knowledge representation language is a trade-off between expressiveness and computational tractability =-=[9, 17, 82]-=-. Similarly, we think that the design of a calculus to deal with uncertainty is also a compromise between precision and computational complexity. CHAPTER 3. UNCERTAINTY 43 We present here new techniqu... |

54 |
On the semantics of rule-based expert systems with uncertainty
- Kifer, Li
- 1988
(Show Context)
Citation Context ...systems we would like to have a means to handle rules involving variables and recursion, neither of which are provided by the Mycin approach. Our proposal is closest to the excellent studies [83] and =-=[53]-=-. The first difference to our approach is that both of these studies are restricted to negation-free rules. In [83] a model theoretic semantics is given without providing a fixpoint semantics. This fi... |

52 |
Multiple Predicate Learning. In
- DeRaedt, Lavrac, et al.
- 1993
(Show Context)
Citation Context ...ing just one goal predicate and training examples about one goal predicate one 6 LEARNING DATALOG RULES 79 has several goals and training data on all these goals. Let us take the example presented in =-=[34]-=-. The background information consists of facts such as male(prudent); male(willem); male(etienne); male(leon); ::: female(laura); female(esther); female(rose); female(alice); ::: parent(bart; stijn); ... |

51 | Stable semantics for probabilistic deductive databases - Ng, Subrahmanian - 1994 |

48 |
The IRIS architecture and implementation
- Wilkinson, Lyngbaek, et al.
- 1990
(Show Context)
Citation Context ...ribe the semantics of such functions. We do this by following the original paper [38, pp 24-33] describing this language. FQL is actually an extension of the Iris database system query language OSQL (=-=[46]-=-) developed at HP labs on recursive function definitions. FQL stands for functional query language and the star denotes the power of computing transitive closures (i.e. recursive functions). A shorten... |

46 |
Technical analysis explained.
- Pring
- 2002
(Show Context)
Citation Context ...d to combine the forecastings of thirty different traders so as to make an overall and more reliable prediction. But as it turned out, also this combination could not improve the prediction accuracy. =-=[73]-=- is a nice textbook giving an overview on recent popular technical analysis methods. Technical analysis is concerned with forecasting reversals in trends from charts. Such methods include the localiza... |

43 |
On Completeness of Historical Relational Query Languages
- Clifford, Croker, et al.
- 1994
(Show Context)
Citation Context ...the data) and valid time (the time for which the data is believed to be correct). 3. In the relational and deductive model, time stamps can either be provided at the tuple or the attribute level (see =-=[23]-=-). The equivalent choice in object-oriented languages is whether functions or objects are time stamped. Obviously, uncertainty can be categorized along the same lines. 1. We can store uncertainty inte... |

37 |
Temporal Logics and their Applications
- Galton
- 1987
(Show Context)
Citation Context ...ed at either end ([ and ] respectively) or open ( ( and ) respectively). We tacitly took the so called first-order approach to time. In some application domains, however, the modal approach is taken (=-=[16]-=-). The notion Fp (for `it will be the case that p'), Gp (for `it will always be the case that p'), Pp (for `it has been the case that p), and Hp (for `it has always been the case that p) are used. Not... |

37 |
An analysis of four uncertainty calculi
- Henkind, Harrison
- 1988
(Show Context)
Citation Context ...stic and the fuzzy approach to dealing with uncertainty. Among others we mention Dempster-Shafer theory, Rough Set approach and so on. A discussion of these approaches can be found in sources such as =-=[19, 17, 20]-=-. In our view, [18] presents an outstanding but seldomly take note of approach developing the model logic LL, a logic for likelihood reasoning. In this language, there are formulas built from proposit... |

36 |
A logic to reason about likelihood
- Halpern, Rabin
(Show Context)
Citation Context ... to dealing with uncertainty. Among others we mention Dempster-Shafer theory, Rough Set approach and so on. A discussion of these approaches can be found in sources such as [19, 17, 20]. In our view, =-=[18]-=- presents an outstanding but seldomly take note of approach developing the model logic LL, a logic for likelihood reasoning. In this language, there are formulas built from propositional predicates an... |

35 |
Y.C.: Incomplete information in relational temporal databases
- Gadia, Nair, et al.
- 1992
(Show Context)
Citation Context ...his expresses uncertainty about (i) the time and/or (ii) the occurrence of the event. Such a situation needs the management of time and uncertainty discussed in this paper. Another remarkable work is =-=[79]-=-. However, this model of uncertainty is too restrictive for applications such as forecasting currency exchange rates. [79] deals with only three truth values instead two. A strength of our proposal is... |

34 |
A probabilistic relational data model
- Barbara, Garcia-Molina, et al.
- 1990
(Show Context)
Citation Context ...e (i.e. a boolean algebra together with a probability function) are appropriately translated into LL, then any deduction drawn from the statements in LL can also be drawn from that probability space. =-=[4]-=- presents an approach to handle uncertainty in relational databases. In contrast to our discussion on probabilistic Datalog where each fact or tuple has attached a weight, they provide a 3 UNCERTAINTY... |

32 |
A foundation of evolution from relational to object databases
- Beech
- 1988
(Show Context)
Citation Context ...age, but restrict ourselves to the fundamental concepts. We ignore issues such as object deletion and addition as well as inheritance (see for more information [23, 39]). Unlike OSQL (object SQL, see =-=[5]-=-) functions, FQL functions can be recursive or even mutually recursive. Like Datalog, FQL has a fixpoint semantics that can be computed within finite 2 RULE LANGUAGES 32 time or more precisely within ... |

30 |
SPIDER: A multiuser information retrieval system for semistructured and dynamic data
- Schäuble
- 1993
(Show Context)
Citation Context ... shortened version of the definition of FQL can be found in [39]. This language is currently implemented at ETH Zurich and serves as the basis for the database and information retrieval system SPIDER =-=[23, 37]-=- 4 . The 4 ETH Zurich and HKUST are jointly working on extending this system. ETH concentrates on efficiency and implementation, whereas HKUST on finding new fundamental concepts for solving discovery... |

29 |
Probability logic. Notre Dame
- Hailperin
- 1984
(Show Context)
Citation Context ...ial, not in the size of the knowledge base - much worse - exponential in the number of constant symbols assumed to be in the underlying language. So we end up again with computational intractability. =-=[41]-=- links probability theory to logic. The problem considered is: given a set of variables and an assignment of real numbers in the interval [0,1] to the conjunctions of atoms over these variables, how c... |

27 |
Managing Uncertainty in Expert Systems
- Grzymala-Busse
- 1991
(Show Context)
Citation Context ...nformation. 3 UNCERTAINTY 37 3.2 Other Approaches to Uncertainty We briefly give an overview on fuzzy sets whose roots go back to the paper [55]. A similar, a bit longer such overview can be found in =-=[17]-=-. Given is an universe of discourse U , a set of elements. Each subset S of U is characterized by a membership functionsS : U ! [0; 1]. S is usually called a fuzzy set. Two fuzzy sets S and S 0 are eq... |

25 |
Inductive Learning in Deductive Databases.
- Dzeroski, Lavrac
- 1993
(Show Context)
Citation Context ...sing the knowledge representation tool of a decision tree which is equivalent to propositional deduction rules (e.g. [26, 33]). More recent approaches use Datalog as knowledge representation language =-=[27, 10]-=-. Datalog is more expressive than propositional rules, however, unlike the later, Datalog is not able to handle uncertain or probabilistic information, can not deal with function symbols and has limit... |

24 | New directions for uncertainty reasoning in deductive databases
- GUNTZER, KIESSLING, et al.
- 1991
(Show Context)
Citation Context ...f an uncertainty value according to this theory needs exponential time in the size of the data and there is no general polynomial algorithm to approximate the proposed reasoning. Also the proposal of =-=[40]-=- which deals with propositional rules and uncertainties is intractable. To overcome the problem of inefficiency, the designers of Mycin, a medical expert system, have approximated basic probability th... |

23 | Polynomial time query processing in temporal deductive databases
- Chomicki
- 1990
(Show Context)
Citation Context ...e in rule based systems. We will necessarily be more precise and hence also more formal when extending our powerful rule language (P)FQL to cope with time. 4 TIME 48 A recently investigated approach (=-=[7]-=-) is to add an additional argument to each predicate and to adopt the linearly ordered time points approach. The rule base f light(x; y; T + 7) / f light(x; y; T ) (87) f light(Munich; HK; 28=9=93) (8... |

23 | Probabilistic knowledge bases
- Wüthrich
- 1995
(Show Context)
Citation Context ...o handle uncertain or probabilistic information, can not deal with function symbols and has limited modeling capabilities for database applications. The first deficiency has been overcome by the work =-=[48, 49]-=-. Based on these studies the latest development in symbolic Machine Learning was the development of techniques discoverying probabilistic Datalog rules [51, 50]. More expressive than Datalog, because ... |

17 |
On the expressive power of query languages
- Schauble, Wuthrich
- 1994
(Show Context)
Citation Context ...g on extending this system. ETH concentrates on efficiency and implementation, whereas HKUST on finding new fundamental concepts for solving discovery problems. 2 RULE LANGUAGES 33 discussion follows =-=[38, 39]-=- 5 . 6 5 For the written final exam, you need to know at most what we discussed in class and not the full version of this paper nor any proofs therein. 6 It could be that you can have a live demo of a... |

16 |
A Formal Definition for Expressive Power of Knowledge Representation Languages
- Baader
- 1990
(Show Context)
Citation Context ...emantics making it tractable in real-world applications. It has been argued that the design of a knowledge representation language is a trade-off between expressiveness and computational tractability =-=[9, 17, 82]-=-. Similarly, we think that the design of a calculus to deal with uncertainty is also a compromise between precision and computational complexity. CHAPTER 3. UNCERTAINTY 43 We present here new techniqu... |

13 |
1991] "Temporal Reasoning in Deductive Databases
- Sripada
(Show Context)
Citation Context .... In contrast to valid time we can also record transaction time. Transaction time tells us when the data was persistently stored in the database. ffl In some approaches time may be arbitrarily nested =-=[40]-=-. For example, at time t 1 we store that we believed at time t 2 that the price of a share was x at time t 3 . In this example, t 1 is a transaction time whereas t 2 and t 3 are believe or valid times... |

13 |
Relating Dempster-Shafer theory to stable semantics
- Ng, Subrahmanian
- 1991
(Show Context)
Citation Context ...ompetitiveness to a degree as high as is possible using our semantics. The study [54] gives a basis for handling uncertainty and temporal aspects in rule-based systems and extends the fuzzy approach. =-=[68]-=- integrates Dempster-Shafer theory into logic programs. But Dempster-Shafer theory seems also in-adequate for the applications we have in mind. For example, the two rules p : [x; y] / q : [x; y] and p... |

12 |
Statistical and Scientific Databases
- Michalewicz
- 1992
(Show Context)
Citation Context ...e answer can be obtained using either a top down or bottom up evaluation. A second example is the following. Holds(r(x) / p(x); :q(x); [50; 110]) Holds(p(A); [20; 150]) CHAPTER 4. TIME 62 Holds(p(B); =-=[60; 300]-=-) Holds(q(A); [70; 100]) Holds(q(A); [120; 170]) Now we can use special evaluation techniques to obtain the result T = f[50; 70]; [100; 110]g on the query / Holds(r(A); T ). We got a flavor of the pro... |

10 | An overview of database mining techniques
- Kero, Russell, et al.
- 1995
(Show Context)
Citation Context ...experimented with all kinds of techniques and fine-tunings. In the more general field of data mining or knowledge discovery many promising techniques have been developed for a variety of applications =-=[4, 43, 44, 65, 52]-=-. The main difference of these results to our work here is that we are specifically and only concerned with one single application, whereas the before mentioned techniques are fine-tuned to many other... |

10 |
On the Expressive Power of Annotated Logic Programs
- Kifer, Subrahmanian
- 1989
(Show Context)
Citation Context ...he application mentioned at the beginning, but with fuzzy interpretation we found no rule set which could express the competitiveness to a degree as high as is possible using our semantics. The study =-=[54]-=- gives a basis for handling uncertainty and temporal aspects in rule-based systems and extends the fuzzy approach. [68] integrates Dempster-Shafer theory into logic programs. But Dempster-Shafer theor... |

9 | On the Learning of Rule Uncertainties and their Integration into Probabilistic Knowledge Bases.
- Wuthrich
- 1993
(Show Context)
Citation Context ...and is therefore naturally tuple weighted. We next extend Datalog and FQL in a probabilistic way. 3 UNCERTAINTY 43 3.3 Probablistic Datalog Our discussion follows [48, 49] and takes some parts out of =-=[52]-=-. 7 8 3.4 Probablistic FQL Our discussion follows [53] [52]. 9 . 7 For the written final exam, you need to know at most what we discussed in class and not the full versions of these papers nor any pro... |

8 |
Large Deductive Databases with Constraints
- Wuthrich
- 1991
(Show Context)
Citation Context ...er, if we assume that also b is a legal constant in our universe then q(a) becomes deducible. Hence the rule is domain dependent. It holds that if a rule is allowed then it is domain independent (see =-=[47]-=- for an exhaustive discussion of this issue). 2 RULE LANGUAGES 24 An attractive property of Datalog is that each Datalog rule can be translated to an equivalent relational algebra expression. This mea... |

8 | Discovering probabilistic decision rules
- Wuthrich
- 1997
(Show Context)
Citation Context ...eliability of the discovered relationships. We next present in section 1.3 a short case study of a knowledge discovery application in a simple but non trivial domain (the full report of this study is =-=[50]-=-). This study relies on the query language probabilistic Datalog (Datalog extended to handle probabilistic rules and facts). The aim is to give an impression of knowledge discovery before going into m... |

8 |
Learning Probabilistic Rules
- Wuthrich
- 1993
(Show Context)
Citation Context ...es to learn probabilistic concepts and make also a comparison to neural net learning approaches. 7 LEARNING PROBABILISTIC KNOWLEDGE 80 7 Learning Probabilistic Knowledge The discussion follows mainly =-=[51]-=- and [50]. We show then that the introduced concepts can also be used to learn probabilistic FQL . We basically only need satisfactory answers to the questions 30 and 31. Question 30 Suppose to have a... |

7 |
A Temporal Approach to Belief Revision in Knowledge Bases
- Sripada
- 1993
(Show Context)
Citation Context ... share was x at time t 3 . In this example, t 1 is a transaction time whereas t 2 and t 3 are believe or valid times. Such approaches can lead to sophisticated reasoning about time in knowledge bases =-=[40, 41]-=-. ffl We can keep track of linear time or conceptual branching time. When adopting branching time then we deal with different versions of data. For example, we can store that at time t:t1 we believed ... |

6 |
et al. A TSQL2 Tutorial
- Snodgrass, Ahn, et al.
- 1994
(Show Context)
Citation Context ... is needed in two or three dimensional geographic applications. One of the few remarkable papers dealing with time uncertainty is [26]. Parts of these ideas also went in the TSQL2 language definition =-=[32]-=-. The former paper discusses how to deal with time uncertainty (called time indeterminacy). However, the paper does not provide a solution to deal with time and uncertainty. If an event occurred but w... |

5 |
guide to statistical and mathematical analysis, version 3.3
- S-PLUS
- 1995
(Show Context)
Citation Context ... predict the competitiveness from the premises. The expert classified each premise of a company into one of seven columns (very bad to very good, see figure 1.2). We simply divide the closed interval =-=[0; 1]-=- equally by the seven numbers 0; 0:17; 0:33; 0:5; 0:67; 0:83; 1 and attach these numbers as weights to the premise facts instead. The McKinsey matrix has 3 columns, so that a firm can be seen as nonco... |

4 |
Knowledge base refinement by backpropagation
- Fu
- 1991
(Show Context)
Citation Context ...abilistic formalism and the choice of an appropriate error measure make the heuristic pruning unnecessary. The only other work we are aware of using mean squared error in symbolic machine learning is =-=[15]-=-. Fu uses backpropagation algorithm which is based on mean squared error to deduce the probabilities of a given set of propositional rules. However, what the author proposed is much more general. Name... |

4 |
Probabilistic interpretations for MYCIN’s certainty factors.
- Heckermann
- 1986
(Show Context)
Citation Context ...stic and the fuzzy approach to dealing with uncertainty. Among others we mention Dempster-Shafer theory, Rough Set approach and so on. A discussion of these approaches can be found in sources such as =-=[19, 17, 20]-=-. In our view, [18] presents an outstanding but seldomly take note of approach developing the model logic LL, a logic for likelihood reasoning. In this language, there are formulas built from proposit... |

4 |
FQL : The Query Language of the Information Retrieval System SPIDER
- Knaus
- 1994
(Show Context)
Citation Context ...(y,5)) We do not describe the full language, but restrict ourselves to the fundamental concepts. We ignore issues such as object deletion and addition as well as inheritance (see for more information =-=[23, 39]-=-). Unlike OSQL (object SQL, see [5]) functions, FQL functions can be recursive or even mutually recursive. Like Datalog, FQL has a fixpoint semantics that can be computed within finite 2 RULE LANGUAGE... |

4 | A Probabilistic Query Language
- Wuthrich
- 1994
(Show Context)
Citation Context ...xperts are unwilling and unable to make numerical estimates of weights. This is basically correct, but if we are able to automatically "learn" such numerical values then they are no longer a=-= problem ([53]-=- discusses the question where weights could come from in a numerical or quantitative approach to uncertainty; note that fuzzy logic and probability theory are both numerical.). They introduce then two... |

3 |
A New Approach to Rule Induction and Pruning
- Fensel, Klein
- 1991
(Show Context)
Citation Context ...en specializing it (specialization is used in FOIL [32]), or by taking the most specific rule and then generalizing it by deleting for instance literals from its body (generalization is used in RELAX =-=[13]-=-). We need therefore a notion of specialization and generalization. In mathematical terms, all Datalog rule generators assume a direct specialization relationship ! d on the set of admissible rules H.... |

3 | Discovery of Data Evolution Regularities in Large Databases
- Han, Cai, et al.
- 1995
(Show Context)
Citation Context ...erful. 6. In contrast to DISCOVER, regression analysis can only deal with numbers, not with symbolic data. CHAPTER 9. LEARNING PROBABILISTIC FQL* 128 7. In contrast to other mining techniques such as =-=[43, 44, 65]-=-, DISCOVER can also deal with weighted information. From above list, the experimental results and the sophisticated functionality DISCOVER offers follows that DISCOVER is a promising data mining tool ... |

3 | A Probabilstic NF2 Model for Imprecision in Databases - Fuhr, Rolleke - 1994 |

2 |
Support for time in deductive databases
- Bohlen, Marti
- 1993
(Show Context)
Citation Context ...uring a set of such intervals). A rule such as H / B can also be stamped with a restricting time interval [t; t 0 ] to become H / B : [t; t 0 ]. The implementation of such an approach is presented in =-=[6]. The sema-=-ntics of such a rule base is defined by means of definitions of logical and ("," in Datalog), logical or (two rules with the same head in Datalog) and negation. Informally, having the rule b... |

2 |
Implementing Fuzzy Logic for a Deductive Database Manager
- Marti, Wieland
- 1990
(Show Context)
Citation Context ... 2 ; a 3 ); has cancer(a 2 ) : 0:5 3 UNCERTAINTY 40 to yield 0:5 = max(fmin(f1; 0:2; 0:5g); min(f1; 0:5; 0:5g); min(f1; 0:5; 0:5g)g). The implementation of such a fuzzy Datalog system is presented in =-=[25]-=-. The main advantage of the fuzzy way of combining weights comes from the fact that at each stratification level and at each single iteration level on a particular stratification level we can forget h... |

2 |
Temporal Logics and their Applications, chapter 4
- Sadri
- 1987
(Show Context)
Citation Context ... and being less or equal to t 2 then we assume that y is still valid at time t 2 . Similar default persistence is also used in recent temporal reasoning approaches (e.g. Kowalskis event calculus, see =-=[35]-=- and [40, pp 25-27]). Since we are time stamping on the attribute level we assume that if the value of an attribute is changed then the whole attribute (which is a bag) is changed. For example, let ba... |

2 |
Consistent Predictions for Categorical Data
- Cho, Wuthrich
- 1996
(Show Context)
Citation Context ...rameter settings for finding a fitting have been tried out. Finally, the best parameter setting has been picked. Once given the best parameter setting, finding the fitting took less than 1 minute. 3. =-=[18]-=- describes techniques for finally deciding in what class patient p is. The problem is as follows. Given the independently produced results a(p) = (p; 0:8); i(p) = (p; 0:3) and i(p) = (p:7), decide in ... |

2 |
Treasury Operations and The Foreign Exchange Challenge
- Chorafas
- 1992
(Show Context)
Citation Context ...tations such as Holds(Holds(Clause; T 1 ); T 2 ). Consider the following rule base. Holds(friend(x; y) / classmate(x; y); likes(x; y); [10; 180]) Holds(classmate(x; y) / attends(x; z); attends(y; z); =-=[20; 900]) Hol-=-ds(attends(Bob; Logic); [50; 200]) Holds(attends(John; Logic); [100; 150]) Holds(attends(Mary; Logic); [80; 240]) Holds(likes(Mary; Bob); [40; 1]) In the above database, the query "for what perio... |

2 |
Dacorogna et al. Heterogenous Real-time Trading Strategies in the Foreign Exchange Market
- M
- 1994
(Show Context)
Citation Context ...h factors are the most influential for predicting the Deutsch Mark against the US dollar for instance. There are other approaches for producing forecastings in financial application domains (e.g. see =-=[50, 2, 31, 3, 91]-=- describing approaches by J.P. Morgan, New York, Olson & Associates in Zurich and many others). But all those techniques and approaches rely on merely quantifiable data. Hence,they do not try to take ... |

2 |
Currency Options, Hedging and Trading
- Jr, Lombard, et al.
- 1992
(Show Context)
Citation Context ...1 ); T 2 ). Consider the following rule base. Holds(friend(x; y) / classmate(x; y); likes(x; y); [10; 180]) Holds(classmate(x; y) / attends(x; z); attends(y; z); [20; 900]) Holds(attends(Bob; Logic); =-=[50; 200]) Hol-=-ds(attends(John; Logic); [100; 150]) Holds(attends(Mary; Logic); [80; 240]) Holds(likes(Mary; Bob); [40; 1]) In the above database, the query "for what period or time interval T are Mary and Bob ... |

2 |
Exchange Rates, Prices and World Trade
- Manzur
- 1993
(Show Context)
Citation Context ...market closes. But before describing our decision-support system and the experimental results achieved, we review some work done previously to predict short-term forecasts of currency exchange rates. =-=[57]-=- discusses empirical tests on the ability of foreign exchange market participants to forecast the future value of the Australian dollar for one- and four-week horizons. The Australian dollar is known ... |

2 |
Objects' Relationships
- Pintado
- 1994
(Show Context)
Citation Context ..., New York, Olson & Associates in Zurich and many others). But all those techniques and approaches rely on merely quantifiable data. Hence,they do not try to take into account "unmeasurable"=-= factors. [72]-=- presents nice graphical techniques and tools to visualize financial data. These tools can not directly be used for predictions but rather can be used to visualize the risk/return field of portfolios.... |

2 | An Object-Oriented Data Model for a Time - Dreyer, Dittrich, et al. - 1994 |

2 | Managing Time and Uncertainty - W��uthrich - 1994 |

1 |
Database Mining: A Prformance Perspective
- Agrawal, Imielinski, et al.
- 1993
(Show Context)
Citation Context ...ay requires a kind of normalization at some intermediate steps during rule evaluation. This is mainly needed to handle queries involving duration. If h(1) is true during the set of intervals f[1; 3]; =-=[2; 4]g then the evaluatio-=-n of I : duration(I) ? 2) h(1) should return [1; 4], so "meeting" and "overlapping" intervals must be merged together. A similar but more general approach to handle time in deducti... |

1 |
Automatically Generating FQL Derived Function Definitions. Dec 94
- Chung
- 1994
(Show Context)
Citation Context ...usly speeds up the learning process. The definition of most general derived function definition, as well as the definition of the set of specializations of a given function definition can be found in =-=[21]-=- and appendix A of this paper. We exemplify the generator module. Suppose that a user has defined the typed variables x : X; y11 : Y 1; y12 : Y 1; y21 : Y 2; y22 : Y 2; y31 : Y 3; y32 : Y 3; x4 : X4 (... |

1 |
Inductive Aquisition of Expert Knowledge
- Muggelton
- 1990
(Show Context)
Citation Context ... of some other techniques. These techniques include ffl decision trees (DT) [16, 77] ffl regression analysis (RA) [48] CHAPTER 9. LEARNING PROBABILISTIC FQL* 119 ffl Inductive Logic Programming (ILP) =-=[62, 27]-=- ffl neural nets (NN) [35, 47] Figure 9.1: Starting DISCOVER. 9.1 Mining Techniques 9.1.1 Overview Discover is a front-end to the commercially available database management systems Oracle and Sybase r... |

1 |
Some Proofs about Probabilistic Knowledge Bases
- Wuthrich
- 1994
(Show Context)
Citation Context ...sq is unsatisfiable. 3. IKB (p) = 1 if p is the certain event, i.e. a tautology. 4. IKB (p)sIKB (q) if p logically implies q. The proof of proposition 1 as well as proposition 2 below can be found in =-=[94]-=-. Proposition 1 reveals the full correspondence to probability theory. Even though this correspondence does not prove that our calculus is the right one, we dare to claim that this is at least a stron... |