310 Frank H ¨ oppner J-measure compares the a priori distribution of X (binary variable, either the an- tecedent holds (X = x) or not (X = x)) with the a posteriori distribution of X given that Y = y. The relative information j(X |Y = y)= ∑ z∈{x,x} P(X = z|Y = y) log 2  P(X = z|Y = y) P(X = z)  yields the instantaneous information that Y = y provides about X ( j is also known as the Kullbach-Leibler distance). When applying the rule multiple times, on average we have the information J(X|Y=y) = P(Y=y)j(X|Y=y), which is the J-value of the rule and is bounded by 0.53 bit. The drawback is, however, that highly infrequent rules do not carry much information on average (due to the factor P(Y = y)), such that highly interesting but rarely occurring associations may not appear under the top-ranked rules. Other measures are conviction (a “directed”, asymmetric lift) (Brin et al., 1997B), certainty factors from MYCIN (Berzal et al., 2001), correlation coefficients from statistics (Tan and Kumar, 2002), Laplace or Gini from rule induction (Clark and Boswell, 1991) or decision tree induction (Breiman, 1996). For a comparison of var- ious measures of interestingness the reader is referred to (Hilderman and Hamilton, 2001), where also general properties rule measures should have are discussed. In (Ba- yardo and Agrawal, 1999) it is outlined that, given a fixed consequent, the ordering of rules obtained from confidence is identical to those obtained by lift or conviction (which is further generalized in (Bayardo et al., 1999)). 15.4.2 Interactive or Knowledge-Based Filtering Whatever the rule evaluation measure may propose, the final judgment about the interestingness and usefulness of a rule is made by the human expert or user. For instance, many measures consistently return those rules as most interesting that con- sists of a single item in the consequent, because in this case confidence is maximized (see Section 60.2.1). But the user may be interested in different items or item combi- nations in the consequent, therefore the subjective interestingness of these rules may be low in some applications. Indeed, all measures of interestingness rely on statisti- cal properties of the items and do not take background information into account. But background knowledge about the presence or absence of correlation between items may alert a human expert at some rules, which are undistinguishable from others if looked at the interestingness rate provided by statistical measures alone. Given the large number of rules, as a first step a visualization and navigation tool may help to quickly find interesting rules. In (Klemettinen et al., 1994) rule templates are proposed, allowing the user to restrict the set of rules syntactically. Some visu- alization techniques are also presented, such as visualizing rules in a graph, where the items represent nodes and rules are represented by edges that lead from the an- tecedent to consequent attributes. The thickness of an edge may illustrate the rating of the rule. The idea of (Dong and Li, 1998) is to compare the performance of a rule against the performance of similar rules and flag it only as interesting, if it deviates clearly 15 Association Rules 311 from them. A distance measure for rules is used to define the neighborhood of a rule (containing all rules within a certain distance according to some distance measure). Several possibilities to flag a rule as interesting are discussed: it may qualify as in- teresting by an unexpectedly high confidence value, if its confidence deviates clearly from the average confidence in its neighborhood, or by an unexpectedly sparse neigh- borhood, if the number of mined rules is small compared to the number of possible rules in the neighborhood. Another rule filtering approach is proposed in (Liu et al., 1999), where statistical correlation is used to define the direction of a rule: Basically, a rule has direction 1 / 0 / -1, if antecedent and consequent are positively correlated / uncorrelated / negatively correlated. From the set of predecessors of a rule, an ex- pected direction can be derived (e.g. if all subrules have direction 0, we may expected an extended rule to have also direction 0), and a rule is flagged as interesting if its di- rection deviates from this expected direction (closely related to (Brin et al., 1997A), see section 15.4.5). A completely different approach is to let the expert formalize her or his domain knowledge in advance, which then can be used to test the newly discovered rules against the expert’s belief. The Apriori algorithm is extended to find rules that con- tradict predefined rules in (Padmanabhan, 1998). Only contradictory rules are then reported, because they represent potentially new information to the expert. The first interactive approaches were designed as post processing systems, which generate all rules first and then allow for a fast navigation through the rules. The idea of rule template matching can be extended to a rule query language that is supported by the mining algorithm (Srikant et al., 1997). In contrast to the post-processing ap- proach, such an integrated querying will be faster if only a few queries are posted and may succeed in occasions where a complete enumeration of all rules fails, e.g. if the minimum support threshold is very low but many other constraints can be exploited to limit the search space. The possibility of posing additional constraints is therefore crucial for successful interactive querying (see Section 15.4.4). In (Goethals and Van den Bussche, 2000) the user specifies items that must or must not appear in the an- tecedent or consequent of a rule. From this set of constraints, a reduced database (reduced in the size and number of transactions) is constructed, which is used for faster itemset enumeration. The generation of a working copy of the database is ex- pensive, but may pay off for repeated queries, especially since intermediate results of earlier queries are reused. The technique is extended to process full-fledged Boolean expressions. 15.4.3 Compressed Representations Possible motivations for using compressed or condensed representations of rules in- clude • Reduction of storage needs and, if possible, computational costs • Derivation of new rules by condensing items to meta-items • Reduction of the number of rules to be evaluated by an expert 312 Frank H ¨ oppner Using a compressed rule set is motivated by problems occurring with dense databases (see Section 15.4.4): Suppose we have found a frequent itemset r containing an item i. If we include additional n items in our database that are perfectly correlated to item i,2 n −1 variations of itemset r will be found. As another example, suppose the database consists of a single transaction {a 1 , ,a n } and min supp = 1, then 2 n −1 frequent subsets will be generated to discover {a 1 , , a n } as a frequent itemset. In such cases, a condensed representation (Mannila and Toivonen, 1996) can help to reduce the size of the rule set. An itemset X is called closed, if there is no superset X  that is contained in every transaction containing X. In the latter example, only a single closed frequent itemset would be discovered. The key to finding closed itemsets is the fact that the support value remains constant, regardless of which subset of the n items is contained in the itemset. Therefore, from the frequent closed itemsets, all frequent itemsets can be reconstructed including their support values, which makes closed itemsets a lossless representation. Algorithms that compute closed frequent itemsets are, to name just a few, Close (Pasquier et al., 1999), Charm (Zaki, 2000), and Closet (Wang, 2003). An overview of lossless representations can be found in (Kryszkiewicz, 2001). Secondly, for a rule to become more meaningful, it may be helpful to consider several items in combination. An example is the use of product taxonomies in mar- ket basket analysis, where many rules that differ only in one item, say apple-juice, orange-juice, and cherry-juice, may be outperformed (and thus replaced) by a single rule using the generalized meta-item fruit-juice. Such meta-items can be obtained from an a-priori known product taxonomy (Han et al., 2000, Srikant and Agrawal, 1995). If such a taxonomy is not given, one may want to discover disjunctive combi- nation of items that optimize the rule evaluation automatically (see (Zelenko, 1999) for disjunctive combination of items, or (H ¨ oppner, 2002) for generalization in the context of sequences). Finally, the third motivation addresses the combinatorial explosion of rules that can be generated from a single (closed) frequent itemset. Since the expert herself knows best which items are interesting, why not present associations – as a com- pressed representation of many rules – rather than individual rules? In particular when sequences are mined, the order of items is fixed in the sequence, such that we have only two degrees of freedom: (1) Remove individual items before building a rule and (2) select the location which separates antecedent from consequent items. In (H ¨ oppner, 2002) every location in a frequent sequence is attributed with the J- value that can at least be achieved in predicting an arbitrary subset of the consequent. This series of J-values characterizes the predictive power of the whole sequence and thus represents a condensed representation of many associated rules. The tuple of J-values can also be used to define an order on sequences, allowing for a ranking of sequences rather than just rules. 15.4.4 Additional Constraints for Dense Databases Suppose we have extracted a set of rules R from database D. Now we add a very frequent item f randomly to our transactions in Dand derive rule set R  . Due to the 15 Association Rules 313 randomness of f, the support and confidence values of the rules in R do not change substantially. Besides the rules in R, we will also obtain rules that contain f either in the antecedent or the consequent (with very similar support/confidence values). Even worse, since f is very frequent, almost every subset of I will be a good predictor of f , adding another 2 |I| rules to R. This shows that the inclusion of f increases the size of R  dramatically, while none of these rules are actually worth investigating. In the market basket application, such items f usually do not exist, but come easily into play if customer data is included in the analysis. For dense databases the bottom- up approach of enumerating all frequent itemsets quickly turns out to be infeasible. In this section some approaches that limit the search space further to reduce time and space requirements are briefly reviewed. (In contrast to the methods in the next section, the minimum support threshold is not necessarily released.) In (Webb, 2000) it is argued that a user is not able to investigate many thousand rules and therefore will be satisfied if, say, only the top 1000 rules are returned. Then, optimistic pruning can be used to prune those parts of the search space that cannot contribute to the result since even under optimistic conditions there is no chance of getting better than the best 1000 rules found so far. This significantly limits the size of the space that has actually to be explored. The method requires, however, to run completely in main memory and may therefore be not applicable to very large databases. A modification to the case of numerical attributes (called impact rules) is proposed in (Webb, 2001). In (Bayardo et al., 1999) the set of rules is reduced by listing specializations of a rule only if they improve the confidence of a common base rule by more than a threshold min imp . The improvement of a rule is defined as imp(A→C) = min { A’ ⊂ A | conf(A→C) – conf(A’→C) } These additional constraints must be exploited during rule mining, since enumer- ating all frequent itemsets first and apply constraints thereafter is not feasible. The key to efficiency of DenseMiner is the derivation of upper bounds for confidence, improvement, and support, which are then used for optimistic pruning of itemset blocks. This saves the algorithm from determining the support of many frequent itemsets, whose rules will miss the confidence or improvement constraints anyway. This work is enhanced in (Bayardo and Agrawal, 1999) to mine only those rules that are the optimal rules according to a partial order of rules based on support and confi- dence. It is then shown that the optimal rules under various interestingness measures are included in this set. In practice, while the set of obtained rules is quite man- ageable, it characterizes only a specific subset of the database records (Bayardo and Agrawal, 1999). An alternative partial order is proposed (based on the subsumption of the transactions supported by rules) that better characterizes the database, but also leads to much larger rule sets. One of the appealing properties of the Apriori algorithm is that it is complete in the sense that all association rules will be discovered. We have seen that, on the other 314 Frank H ¨ oppner hand, this is at the same time one of its drawbacks, since it leads to an unmanageably large rule set. If not completeness but quality of the rules has priority, many of the rule induction techniques from statistics and machine learning may be also applied to discover association rules. As an example, in (Friedman, 1997) local regions in the data (corresponding to associations) are sought where the distribution deviates from the distribution expected under the independence assumption. Starting with a box containing the whole data set, it is iteratively shrunk by limiting the set of admis- sible values for each variable (removing items from categorical attributes or shifting the borders of the interval of admissible values for numerical attributes). Among all possible ways to shrink the box, the one that exhibits the largest deviation from the expected distribution is selected. The refinement is stopped as soon as the support falls below a minimum support threshold. Having arrived at a box with a large de- viation, a second phase tries to maximize the support of the box by enlarging the range of admissible values again, without compromising the high deviation found so far. From a box found in this way, association rules may be derived by exploring the correlation among the variables. The whole approach is not restricted to discovering associations, but finds local maxima of some objective function, such that it could also be applied to rule induction for numerical variables. Compared to partitioning approaches (such as decision or regression trees), the rules obtained from trees are much more complex and usually have much smaller support. Many of the approaches in the next section can also be used for dense databases, because they do not enumerate all frequent itemsets either. Similarly, for some tech- niques in this section the minimum support threshold may be dropped with some databases, but in (Webb, 2000, Bayardo and Agrawal, 1999) and (Friedman, 1997) minimum support thresholds have been exploited. 15.4.5 Rules without Minimum Support For some applications, the concept of a minimum support threshold is a rather ar- tificial constraint primarily required for the efficient computation of large frequent itemsets. Even in market basket analysis one may miss interesting associations, e.g. between caviar and vodka (Cohen et al. (2007)). Therefore alternatives to Apriori have been developed that do not require a minimum support threshold for association rule enumeration. We discuss some of these approaches briefly in this section. The observation in section 15.2.1 characterizes support as being downward closed: If an itemset has minimum support, all subsets also have minimum support. We have seen that the properties of a closure led to an efficient algorithm since it provided powerful pruning capabilities. In (Brin et al., 1997A) the upward closure of corre- lation (measured by the chi-square test) is proven and exploited for mining. Since correlation is upward closed, to utilize it in a level-wise algorithm the inverse prop- erty is used: For a (k + 1)-itemset to be uncorrelated, all of its k-subsets must be uncorrelated. The uncorrelated itemsets are then generated and verified in a similar fashion (“has minimum support” is substituted by “is uncorrelated”) as in the Apriori algorithm (Figure 15.2). The interesting output, however, is not the set of uncorre- lated items, but the set of correlated items. The border between both sets is identified 15 Association Rules 315 during the database pass: whenever a candidate k-itemset has turned out to be corre- lated it is stored in a set of minimally correlated itemsets (otherwise it is stored in a set of uncorrelated k-itemsets and used for pruning in the next stage). From the set of minimally correlated itemsets we know that all supersets are also correlated (upward closure), so the partitioning into correlated and uncorrelated itemsets is complete. Note that in contrast to Apriori and derivatives, no rules but only sets of correlated attributes are provided, and that no ranking for the correlation can guide the user in manual investigation (see Section 60.2.2). An interesting subproblem is that of finding all rules with a very high confidence (but no constraints on support), as it is proposed in (Li and Fang, 1989). The key idea is to, given a consequent, subdivide the database D into two parts, one, D 1 , containing all transactions that contain the consequent and the other, D 2 , containing those that do not. Itemsets that occur in D 1 but not at all in D 2 can be used as antecedents for 100%-confidence rules. A variation for rules with high confidence is also proposed. Other proposals also address the enumeration of associations (or association rules), but have almost nothing in common with the Apriori algorithm. Such an ap- proach is presented in (Cohen et al. (2007)), which heavily relies on the power of randomization. Associations are identified via correlation or similarity between at- tributes X and Y : S(X,Y) = supp(X ∩ Y) / supp (X ∪Y ). The approach is restricted to the identification of pairs of similar attributes, from which groups of similar vari- ables may be identified (via transitivity). Since only 2-itemsets are considered and due to the absence of a minimum support threshold, considerable effort is under- taken to prune the set of candidate 2-patterns of size |I| 2 . (In (Brin et al., 1997A) this problem is attacked by additionally introducing a support threshold guided by the chi-square test requirements.) The approach uses so-called signatures for each col- umn, which are calculated in a first database scan. Several proposals for signatures are made, the most simple defines a random order on the rows and the signature it- self is simply the first row index (under the ordering) in which the column has a 1. The probability that two columns X and Y have the same signature is proportional to their similarity S(X,Y ). To estimate this probability, k independently drawn orders are used to calculate k individual signatures. The estimated probabilities are used for pruning and exact similarities are calculated during a second pass. The idea of Mambo (Castelo and Giudici, 2003) is that only those attributes should be considered in the antecedent of a rule, that directly influence the conse- quent attribute, which is reflected by conditional independencies: If X and Y are conditionally independent given Z (or I(X,Y|Z)), then once we know Z, Y (or X) does not tell us anything of importance about X (or Y ): I(X,Y|Z) ⇔ P(X,Y | Z) = P(X | Z) P(Y | Z) As an example from automobile manufacturing (Rokach and Maimon, 2006), consider the variables “car model”, “top”, and “color”. If, for a certain car type, a removable top is available, the top’s color may be restricted to only a few possibili- ties, which consequently restricts the available colors for the car as a whole. The car model influences the color only indirectly, therefore we do not want to see rules lead- ing from car model to color, but only from top to color. If conditional independencies are known, then they can be used to prevent a rule miner from listing such rules. The 316 Frank H ¨ oppner idea of Mambo is to find these independencies and use them for pruning. A mini- mal set of variables MB for a variable X such that for all variables Y /∈ MB∪{X} we have I(X,Y|MB), is called a Markov Blanket of X. The blanket shields X from the remaining variables. The difficulty of this approach lies in the estimation of the Markov Blankets, which are obtained from a Markov Chain Monte Carlo method. In (Aumann & Lindell, 1999) an approach for mining rules from one numeri- cal attribute to another numerical attribute is proposed, that also does not require a minimum support threshold. The numerical attribute in the antecedent is restricted by an interval and the consequent characterizes the population mean in this case. A rule is generated only if the mean of the selected subset of the population differs significantly from the overall mean (according to a Z-test), and a “minimum differ- ence” threshold is used to avoid enumerating rules which are significant but only marginally different. The idea is to sort the database according to the attribute in the antecedent (which is, however, a costly operation). Then any set of consecutive transactions in the sorted database, whose consequent values are above or below av- erage, gives a new rule. They also make use of a closure property: if two neighboring intervals [a,b] and [b, c] lead to a significant change in the consequent variable, then so does the union [a,c]. This property is exploited to list only rule with maximal intervals in the antecedent. 15.5 Conclusions We have discussed the common basis for many approaches to association rule min- ing, the Apriori algorithm, which gained its attractivity and popularity from its sim- plicity. Simplicity, on the other hand, always implies some insufficiencies and opens space for various (no longer simple) improvements. Since the first papers about as- sociation rule mining have been published, the number of papers in this area has exploded and it is almost impossible to keep track of all different proposals. There- fore, the overview provided here is necessarily incomplete. Over time, the focus has shifted from sparse (market basket) data to (general purpose) dense data, and reasonably large itemsets (promoting less than 0.1% of the customers in a supermarket probably is not worth the effort) to small patterns, which represent deviations from the mean population (small set of most profitable customers, which shall be mailed directly). Then it may be desirable to get rid of constraints such as minimum support, which possibly hide some interesting patterns in the data. But on the other hand, the smaller the support the higher the probability of observing an incidental rather than a meaningful deviation from the average pop- ulation, especially when taking the size of the database into account (Bolton et al., 2002). A major issue for the upcoming research is therefore to limit the findings to substantive “real” patterns. 15 Association Rules 317 References Agrawal R. and Srikant R. Fast Algorithms for mining association rules. Proc. Int. Conf. on Very Large Databases, 487-499, 1994 Agrawal R. and Srikant R. Mining Sequential Patterns. Proc. Int. Conf. on Data Engineering, 3-14, 1995. Agrawal R., Mannila H., Srikant R., Toivonen H., Verkamo A.I. Fast Discovery of Asso- ciation Rules. In: Advances in Knowledge Discovery and Data Mining, Fayyad U.M., Piatetsky-Shapiro G., Smyth P., Uthurusamy R. (eds)., AAAI Press / The MIT Press, 307-328, 1996 Aumann Y., Lindell, Y. A Statistical Theory for Quantitative Association Rules, Journal of Intelligent Information Systems, 20(3):255-283, 2003 Bayardo R.J., Agrawal R. Mining the Most Interesting Rules. Proc. ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 145-154, 1999 Bayardo R.J., Agrawal R., Gunopulos D. Constrained-Based Rule Mining in Large, Dense Databases. Proc. 15 th Int. Conf. on Data Engineering, 188-197, 1999 Berzal F., Blanco I., Sanchez D., Vila M.A. A New Framework to Assess Association Rules. Proc. Symp. on Intelligent Data Analysis, LNCS 2189, 95-104, Springer, 2001 Bolton R., Hand D.J., Adams, N.M. Determining Hit Rate in Pattern Search, Proc. Pattern Detection and Discovery, LNAI 2447, 36-48, Springer, 2002 Borgelt C., Berthold M. Mining Molecular Fragments: Finding Relevant Substructures of Molecules, Proc. Int. Conf. on Data Mining, 51-58, 2002 Breiman L., Friedman J., Olshen R., Stone C. Classification and Regression Trees. Chapman & Hall, New York, 1984. Brijs T., Swinnen G., Vanhoof K. and Wets G. Using Association Rules for Product Assort- ment Decisions: A Case Study. Proc. ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 254-260, 1999 Brin S., Motwani R., and Silverstein C. Beyond market baskets: Generalizing association rules to correlations. Proc. ACM SIGMOD Int. Conf. Management of Data, 265-276, 1997 Brin S., Motwani R., Ullman, J.D., Tsur, S. Dynamic itemset counting and implication rules for market basket data. SIGMOD Record 26(2):255-264, 1997 Castelo R., Feelders A., Siebes A. Mambo: Discovering Association Rules. Proc. Symp. on Intelligent Data Analysis. LNCS 2189, 289-298, Springer, 2001 Clark P., Boswell R. Rule Induction with CN2: Some recent improvements. In Proc. Euro- pean Working Session on Learning EWSL-91, 151-163, 1991 Cohen E., Datar M., Fujiwara S., Gionis A., Indyk P., Motwani R., Ullman J., Yang C. Find- ing Interesting Associations without Support Pruning, IEEE Transaction on Knowledge Discovery 13(1):64-78, 2001 Dehaspe L., Toivonen H. Discovery of Frequent Datalog Patterns. Data Mining and Knowl- edge Discovery, 3(1):7-38,1999 Delgado M., Martin-Bautista M.J., Sanchez D., Vila M.A. Mining Text Data: Features and Patterns. Proc. Pattern Detection and Discovery, LNAI 2447, 140-153, Springer, 2002 Dong G., Li J. Interestingness of Discovered Association Rules in Terms of Neighbourhood- Based Unexpectedness. Proc. Pacific Asia Conf. on Knowledge Discovery in Databases, LNAI 1394, 72-86, 1998 Friedman J.H., Fisher N.I. Bump Hunting in High-Dimensional Data. Statistics and Com- puting 9(2), 123-143, 1999 318 Frank H ¨ oppner Gamberger D., Lavrac N., Jovanoski, V. High Confidence Association Rules for Medical Diagnosis, Proc. of Intelligent Data Analysis in Medical Applications (IDAMAP), 42- 51, 1999 Goethals B., Van den Bussche, J. On Supporting Interactive Association Rule Mining, Proc. Int. Conf. Data Warehousing and Knowledge Discovery, LNCS 1874, 307-316, Springer, 2000 Han J., Fu Y. Discovery of Multiple-Level Association Rules from Large Databases Proc. Int. Conf. on Very Large Databases, 420-431, 1995 Hilderman R.J. and Hamilton H.J. Knowledge Discovery and Measures of Interest. Kluwer Academic Publishers, 2001 H ¨ oppner F., Klawonn F. Finding Informative Rules in Interval Sequences. Intelligent Data Analysis, 6(3):237-256, 2002 H ¨ oppner F. Discovery of Core Episodes from Sequences. Proc. Pattern Detection and Dis- covery, LNAI 2447, 199-213, Springer, 2002 Klemettinen M., Mannila H., Ronkainen P., Toivonen, H. and Verkamo A.I. Finding Inter- esting Rules from Large Sets of Discovered Association Rules. Proc. Int. Conf. on In- formation and Knowledge Managament, 401-407, 1994. Koperski K., Han J. Discovery of Spatial Association Rules in Geographic Information Databases. Proc. Int. Symp Advances in Spatial Databases, LNCS 951, 47-66, 1995 Kryszkiewicz M. Concise Representation of Frequent Patterns based on Disjunction-free Generators. Proc. Int. Conf. on Data Mining, 305-312, 2001. Lee C.H., Lin C.R., Chen M.S. On Mining General Temporal Association Rules in a Publi- cation Database. Proc. Int. Conf. on Data Mining, 337-344, 2001 Li J., Zhang X., Dong G., Ramamohanarao K., Sun Q. Efficient Mining of High Confidence Association Rules without Support Thresholds. Proc. Principles of Data Mining and Knowledge Discovery, 406-411, 1999 Liu B., Hsu W., Ma Y. Pruning and Summarizing the Discovered Associations. Proc. ACM SIGKDD Conf. Knowledge Discovery and Data Mining, 125-134, 1999 Mannila H., Toivonen H. Multiple uses of frequent sets and condensed representations. Proc. ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 189-194, 1996 Mannila H., Toivonen H., Verkamo A.I. Discovery of Frequent Episodes in Event Sequences. Data Mining and Knowledge Discovery, 1(3):259-289, 1997 Miller R.J., Yang Y. Association Rules over Interval Data. Proc. Int. Conf. on Management of Data, 452-461, 1997 Mitchell T., Machine Learning, McGraw-Hill, 1997 Mobasher B., Dai H., Luo T., Nakagawa, M. Discovery and Evaluation of Aggregate Usage Profile for Web Personalization, Data Mining and Knowledge Discovery 6:61-82, 2002 Padmanabhan B., Tuzhilin A. A Belief-Driven Method for Discovering Unexpected Patterns. Int. Conf. on Knowledge Discovery and Data Mining, 94-100, 1998 Pasquier N., Bastide Y., Taouil R., Lakhal L. Efficient Mining of association rules using closed itemset lattices. Information Systems, 24(1):25-46, 1999 Piatetsky-Shapiro, G. Discovery, Analysis, and Presentation of Strong Rules. Proc. Knowl- edge Discovery in Databases, 229-248, 1991 Rokach, L., Averbuch, M. and Maimon, O., Information retrieval system for medical narra- tive reports. Lecture notes in artificial intelligence, 3055. pp. 217-228, Springer-Verlag (2004). 15 Association Rules 319 Rokach L. and Maimon O., Data mining for improving the quality of manufacturing: A feature set decomposition approach. Journal of Intelligent Manufacturing 17(3): 285299, 2006. Smyth P. and Goodman, R.M. An Information Theoretic Approach to Rule Induction from Databases. IEEE Trans. Knowledge Discovery and Data Engineering, 4(4):301-316, 1992 Srikant R., Agrawal R. Mining Generalized Association Rules. Proc. Int. Conf. on Very Large Databases, 407-419, 1995 Srikant R., Agrawal R. Mining Quantitative Association Rules in Large Relational Tables. Proc. ACM SIGMOD Conf. on Management of Data, 1-12, 1996 Srikant R., Vu Q., Agrawal R. Mining Association Rules with Constraints. Proc. Int. Conf. Knowledge Discovery and Data Mining, 66-73, 1997 Tan P.N., Kumar V. Selecting the Right Interestingness Measure for Association Patterns, Proc. ACM SIGKDD Conf. Knowledge Discovery and Data Mining, 32-41, 2002 Tsoukatos, I. and Gunopulos, D. Efficient Mining of Spatiotemporal Patterns. Proc. Int. Symp. Spatial and Temporal Databases, LNCS 2121, pages 425-442, 2001 Wang J., Han J., Pei J. CLOSET+: Searching for the Best Strategies for mining Frequent Closed Itemsets, Proc. ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 236-245, 2003 Webb G.I. Efficient Search for Association Rules. Proc. ACM SIGKDD Int. Conf. on Knowl- edge Discovery and Data Mining, 99-107, 2000 Webb G.I. Discovering Associations with numeric variables. Proc. ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 383-388, 2001 Zaki M.J. Generating non-redundant Association Rules. In Proc. ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 34-43, 2000 Zaki M.J. SPADE: An Efficient Algorithm for Mining Frequent Sequences. Machine Learn- ing 42(1):31-60, 2001 Zelenko D. Optimizing Disjunctive Association Rules. Proc. of Int. Conf. on Principles of Data Mining and Knowledge Discovery, LNAI 1704, 204-213, 1999 . Conf. Knowledge Discovery and Data Mining, 32- 41, 20 02 Tsoukatos, I. and Gunopulos, D. Efficient Mining of Spatiotemporal Patterns. Proc. Int. Symp. Spatial and Temporal Databases, LNCS 21 21, pages. Personalization, Data Mining and Knowledge Discovery 6:61- 82, 20 02 Padmanabhan B., Tuzhilin A. A Belief-Driven Method for Discovering Unexpected Patterns. Int. Conf. on Knowledge Discovery and Data Mining, . Conf. on Knowledge Discovery and Data Mining, 383-388, 20 01 Zaki M.J. Generating non-redundant Association Rules. In Proc. ACM SIGKDD Int. Conf. on Knowledge Discovery and Data Mining, 34- 43, 20 00 Zaki