Although composed of unstructured texts, documents contained in textual archives such as public announcements, patient records and annual reports to shareholders often share an inherent though undocumented structure. In order to facilitate efficient, structure-based search in archives and to enable information integration of text collections with related data sources, this inherent structure should be made explicit as detailed as possible. Inferring a semantic and structured XML document type definition (DTD) for an archive and subsequently transforming the corresponding texts into XML documents is a successful method to achieve this objective. The main contribution of this paper is a new method to derive structured XML DTDs in order to extend previously derived flat DTDs. We use the DIAsDEM framework to derive a preliminary, unstructured XML DTD whose components are supported by a large number of documents. However, all XML tags contained in this preliminary DTD cannot a priori be assumed to be mandatory. Additionally, there is no fixed order of XML tags and automatically tagging an archive using a derived DTD always implicates tagging errors. Hence, we introduce the notion of probabilistic XML DTDs whose components are assigned probabilities of being semantically and structurally correct. Our method for establishing a probabilistic XML DTD is based on discovering associations between, resp. frequent sequences of XML tags.
Most organizations are not only “drowning” in data, they are
also “struggling” to cope with huge amounts of text
documents. Tan points out that up to 80% of a company’s
information is stored in unstructured textual documents [
Hence, capturing interesting and actionable knowledge from textual databases is a major challenge for the data mining community. Creating semantic markup is one form of providing explicit knowledge about text archives to facilitate searching and browsing or to enable information integration
The work of this author is funded by the German Research Society
(DFG grant no. SP 572/4-1).
with related data sources. Unfortunately, most users are not
willing to manually create metadata due to the efforts and
costs involved [
In this paper, we present the KDD approach pursued in the research project DIAsDEM whose German acronym stands for “Data Integration for Legacy Systems and SemiStructured Documents by Means of Data Mining Techniques”. Our goal is semantic tagging of textual content with meta-data to facilitate searching, querying, identification of and integration with associated texts and relational data. Hence, we aim at deriving a structured XML DTD that serves as a quasi-schema for the document collection and enables the provision of database-like querying services on textual data. DIAsDEM focuses on text collections with domain-specific vocabulary and syntax that frequently share an inherent, but undocumented structure.
The DIAsDEM framework for semantic tagging of
domainspecific texts was introduced in [
Given that all XML tags are derived by data mining techniques (i.e. iterative clustering as explained in section 3), they are not crisp due to tagging errors. Taking this critical fact into account, we introduce the notion of a probabilistic DTD that describes the most likely orderings of XML tags and that contains statistical properties for each tag. The structured DTD will be the basis for future information integration efforts that involve XML archives generated by the DIAsDEM Workbench. We introduce two algorithms for inferring a probabilistic DTD that utilize association rule discovery algorithms and sequence mining techniques.
The rest of this paper is organized as follows: The next <?xml version="1.0" encoding="ISO-8859-1"?> <!DOCTYPE CommercialRegisterEntry SYSTEM ’CommercialRegisterEntry.dtd’> <CommercialRegisterEntry> <BusinessPurpose> Der Betrieb von Spielhallen in Teltow und das Aufstellen von Geldspiel- und Unterhaltungsautomaten. </BusinessPurpose> <ShareCapital AmoutOfMoney="25000 EUR"> Stammkapital: 25.000 EUR. </ShareCapital> <LimitedLiabilityCompany> Gesellschaft mit beschränkter Haftung. </LimitedLiabilityCompany> <ConclusionArticles Date="12.11.1998; 19.04.1999"> Der Gesellschaftsvertrag ist am 12.11.1998 abgeschlossen und am 19.04.1999 abgeändert. </ConclusionArticles> (...) Einzelvertretungsbefugnis kann erteilt werden. <AppointmentManagingDirector Person="Balski; Pawel; Berlin; 14.04.1965"> Pawel Balski, 14.04.1965, Berlin, ist zum Geschäftsführer bestellt. </AppointmentManagingDirector> (...) <PublicationMedia> Nicht eingetragen: Die Bekanntmachungen der Gesellschaft erfolgen im Bundesanzeiger. </PublicationMedia> </CommercialRegisterEntry> section briefly discusses related work. Section 3 gives an overview of our framework for semantic tagging of domainspecific text collections. Section 4 introduces the notion of probabilistic DTDs for textual archives and develops two methods for deriving them. Finally, we conclude and give directions for future research in section 5.
Nahn and Mooney propose the combination of methods from
KDD and information extraction to perform text mining tasks
[
Our approach shares with this research thread the objective of extracting semantic concepts from texts. However, concepts to be extracted in DIAsDEM must be appropriate to serve as elements of the XML DTD. Among other implications, discovering a concept that is peculiar to a single text unit is not sufficient for our purposes, although it may perfectly reflect the corresponding content. In order to derive a DTD, we need to discover groups of text units that share some semantic concepts. Moreover, we concentrate on domain-specific texts, which significantly differ from average texts with respect to word frequency statistics. These collections can hardly be processed using standard text mining software because the integration of relevant domain knowledge is a prerequisite for successful knowledge discovery.
There are only a few research activities aiming at the
transformation of texts into semantically annotated XML
documents: Becker et al. introduce the search engine
GETESS that supports query processing on texts by deriving and
processing XML text abstracts [
In this paper, the notion of semantic tagging refers to the activity of annotating texts with domain-specific XML tags that might contain additional attributes. Rather than classifying entire documents or tagging single terms, we aim at semantically tagging text units such as sentences or paragraphs. Table 1 illustrates this concept of semantic tagging, whereas each sentence of this German Commercial Register entry is a text unit. In this example, the semantics of most sentences are made explicit by XML tags that partly contain additional attributes describing extracted named entities (e.g., names of persons and amounts of money). The XML document depicted in Table 1 was created by applying the DIAsDEM framework to a collection of 1,145 textual Commercial Register entries containing 10,785 text units. This collection includes all entries related to foundations of companies in the district of the German city Potsdam in 1999. In Germany, companies are obliged by law to submit various information about business affairs to local Commercial Registers. Although Commercial Registers are an important source of information in daily business transactions, their textual content can only be searched using full-text queries at the moment. Hence, semantically semi-structuring these textual archives provides the basis for information integration and creation of value-adding services related to information brokerage. XML query languages could be employed to submit both both content- and structure-based queries against semantically tagged XML archives.
Our framework pursues two objectives for a given archive
of text documents: All text documents should be
semantically tagged and an appropriate, preliminary flat XML DTD
should be derived for the archive. Semantic tagging in
DIAsDEM is a two-phase process. We have designed a knowledge
discovery in textual databases (KDT) process that constitutes
the first phase in order to build clusters of semantically
similar text units, to tag documents in XML according to the
results and to derive an XML DTD describing the archive.
The KDT process that was introduced in [
In DIAsDEM we concentrate on the semantic tagging of similar text documents originating from a common domain. Nevertheless, the DIAsDEM approach is appropriate for semantically tagging various kinds of archives such as public announcements of courts and administrative authorities, quarterly and annual reports to shareholders, textual patient records in health care applications as well as product and service descriptions published on electronic marketplaces. ======= ========== ========== ==============
======= ============== ============== ==================
============ ================================================== ============ ===================================== ======== ===================
=============== == ========== ================== === === ===
======== ======== ======== ======== In the remainder of this section, we briefly introduce the first phase of the DIAsDEM framework whose iterative and interactive KDT process is depicted in Figure 1. This process is termed “iterative” because the clustering algorithm is invoked repeatedly. Our notion of iterative clustering should not be confused with the fact that most clustering algorithms perform multiple passes over the data before converging.
This process is also “interactive”, because a knowledge engineer is consulted for cluster evaluation and final cluster naming decisions at the end of each iteration.
Besides the initial text documents to be tagged, the following domain knowledge constitutes input to our KDT process: A thesaurus containing a domain-specific taxonomy of terms and concepts, a preliminary UML schema of the domain and descriptions of specific named entities of importance, e.g. persons and companies. The UML schema reflects the semantics of named entities and the relationships among them, as they are initially conceived by application experts. This schema serves as a reference for the DTD to be derived from discovered semantic tags, but there is no guarantee that the > SupervisoryBoard | (...) | Owner | FoundationPartnership )* <!ELEMENT CommercialRegisterEntry ( #PCDATA | BusinessPurpose | ShareCapital | ModificationMainOffice | FullyLiablePartner | AppointmentManagingDirector | GeneralPartnership | InitialShareholders | NonCashCapitalContribution | LimitedLiabilityCompany | ConclusionArticles | ModificationRegisteredName | <?xml version="1.0" encoding="ISO-8859-1"?> <!ELEMENT BusinessPurpose (#PCDATA)> <!ELEMENT ShareCapital (#PCDATA)> (...) <!ELEMENT FoundationPartnership (#PCDATA)> final DTD will be contained in or will contain this schema. clusters containing approx. 85% of text units.
Similarly to a conventional KDD process, our process starts
with a preprocessing phase: After setting the level of
granularity by determining the size of text units to be tagged,
the Java- and Perl-based DIAsDEM Workbench performs
basic NLP preprocessing such as tokenization, normalization
and word stemming using TreeTagger [
In the pattern discovery phase, all text unit vectors contained in the initial archive are clustered based on similarity of their content. The objective is to discover dense and homogeneous text unit clusters. Clustering is performed in multiple iterations. Each iteration outputs a set of clusters, which the DIAsDEM Workbench partitions into ”acceptable” and ”unacceptable” ones according to our quality criteria. A cluster of text unit vectors is ”acceptable”, if and only if (i) its cardinality is large and the corresponding text units are (ii) homogeneous and (iii) can be semantically described by a small number of text unit descriptors. Members of “acceptable” cluster are subsequently removed from the dataset for later labeling, whereas the remaining text unit vectors are input data to the clustering algorithm in the next iteration. In each iteration, the cluster similarity threshold value is stepwise decreased such that “acceptable” clusters become progressively less specific in content. The KTD process is based on a plug-in concept that allows the execution of different clustering algorithms within the DIAsDEM Workbench. In the case study, we employed the demographic clustering function included in the IBM Intelligent Miner for Data that maximizes the value of Condorcet’s criterion. After three iterations, the DIAsDEM Workbench discovered altogether 73 “acceptable” The postmining phase consists of a labeling step, in which “acceptable” clusters are semi-automatically assigned a label. Ultimately, cluster labels are determined by the knowledge engineer. However, the DIAsDEM Workbench performs both a pre-selection and a ranking of candidate cluster labels for the expert to choose from. All default cluster labels are derived from feature space dimensions (i.e. from text unit descriptors) that are prevailing in each “acceptable” cluster. Cluster labels actually correspond to XML tags that are subsequently used to annotate cluster members. Finally, all original documents are tagged using valid XML tags. Additionally, XML tags are enhanced by attributes reflecting previously extracted named entities and their values. Table 2 contains an excerpt of the flat, unstructured XML DTD that was automatically derived from XML tags in the case study.
It coarsely describes the semantic structure of the resulting XML collection. Currently, named entities that serve as additional attributes of XML tags are not fully evaluated by the DIAsDEM Workbench.
The output of the DIAsDEM Workbench is a set of semantic XML tags which should be used as XML tags to describe the content of the archive documents. To reflect the content of the archive at an abstract level, it is essential to compose the tags into a DTD. Since the semantic annotations are derived with data mining techniques, they are not crisp. Thus, it is essential that the validity of each tag is expressed in quantitative terms and is estimated properly. Furthermore, an ordering should be imposed upon the tags. Hence, after deriving semantic XML tags, we combine them into a probabilistic DTD by (i) deriving the most likely ordering of the tags and (ii) computing the statistical properties of each tag inside the document type definition.
The reader may recall that a semantic annotation is actually the label of a cluster discovered by the DIAsDEM Workbench. The underlying clustering mechanism produces nonoverlapping clusters. This implies that a text unit belongs to exactly one cluster, to the effect that it can be annotated with the label of this cluster only. Hence, the tags/labels derived the DIAsDEM Workbench cannot be nested. An extension of the DIAsDEM Workbench by a hierarchical clustering algorithm would allow for the establishment of subclusters and thus for the nesting of (sub)cluster labels. However, this is planned as future work.
The objectives of the DTD establishment method are the specification of the most appropriate ordering of tags, the identification of correlated or mutually exclusive tags and the adornment of each tag and each correlation among tags with statistical properties. These properties form the basis for reliable query processing, because they determine the expected precision and recall of the query results. In the following, we first introduce the statistical properties we consider for the DTD tags and their associations and describe the methodology for computing these statistics. To model the complete statistical information pertinent in these tags and their relationships, we use a hypergraph structure. We then introduce a mechanism that derives a probabilistic DTD from this graph. TagSupport The tags of the DTD are cluster labels derived by a statistical approach. Thus, in terms of XML, they are observed as optional per se. In many application areas, a domain expert can provide suggestions as to which tags should be observed as mandatory. Despite this, there is no guarantee that the expert’s suggestions hold true in the archive: The text unit containing this information may have been misclassified by the DIAsDEM Workbench, or the information may be simply absent from the document. For example, although one would expect that each movie has a regisseur, there are movies whose regisseur is unknown or inapplicable, due to the nature of the movie. The property TagSupport offers an indicator of whether a tag may be considered as potentially mandatory. We define it as the ratio of documents where this tag appears to the total number of documents in the archive. of error type II text units is higher, indicating that some text units were not placed in the cluster they semantically belong to. With 0.95 confidence, the overall error rate in the entire dataset is in the interval [2.6%, 5.9%] which is a promising result.
GroupSupport In most of the above statistics, we juxtapose the frequence of appearance of a tag with the frequency of a group of tags, be it a set or a sequence. We use the term
The statistical properties of DTD tags are depicted in Table 3 and described in the following paragraphs. The first column contains the names of the properties. The second column reflects whether the property is peculiar to the whole set of tags as cluster labels (i.e. the whole ”model”), to each tag or to a group of associated tags. The last column names the mechanism to be applied to derive the value of each property for each tag.
Accuracy The DIAsDEM Workbench derives semantic XML tags as labels of clusters. These clusters constitute a model over the data, in the conventional statistical sense. In terms of data classification, such models are subject to misclassification errors. We identify two types of misclassification:
Error type I: A text unit is assigned to the wrong cluster, i.e. the cluster label does not reflect the content of the text unit.
Error type II: A text unit is not assigned to any cluster, although there is a cluster with a label reflecting the content of the text unit.
For the envisaged DTD, only the error type I is relevant. We use the term accuracy of the model as the probability that cluster labels reflect the content of cluster members. The accuracy value affects the DTD as a whole, it is not peculiar to individual tags. Therefore, we do not incorporate this value in the statistical adornment of the individual tags.
In order to evaluate the quality of out approach in absence of pre-tagged documents, we drew a random sample containing 5% out of 10,785 text units and asked a domain specialist to verify the annotations of these text units with respect to both error types. Within the sample, error type I (error type II) occured in 0.4% (3.6%) of text units. Hence, tagged text units are most likely to be correctly processed. The percentage Property Accuracy TagSupport AssociationConfidence AssociationLift LocationConfidence GroupSupport
Radius model tag set of tags set of tags sequence of tags set or sequence of tags y 1 : : : ! An edge represents a relationship of the form yn x; x however, we use the convention that is the source node and We represent the tags and their associations in a directed graph. Its nodes are individual tags, sequences of adjacent tags or sets of co-occuring tags. Each node is adorned with the statistical properties pertinent to a tag, resp. tag group. the group of nodes in the rule’s LHS is the target. Similarly to nodes, an edge is adorned with the statistics of the orderinsensitive or order-sensitive association it represents. GroupSupport as the ratio of the number of documents containing a group of tags to the total number of documents. In fact, for any set of at least two tags, this property assumes one value for the set and as many values as are the perturbations of set members. In the following subsection, we show how we model the statistical information pertinent to individual tags, to tag groups (i.e. sets or sequences) and to relationships among them in a seamless way.
Some of the values of the statistical properties depicted in Table 3 are already made available as part of the DIAsDEM Workbench output, while the remaining ones can be computed by data mining algorithms. To exploit these values for the establishment of a probabilistic DTD, we need a representation model and an algorithm that builds the DTD when processing this model. We introduce here a generic graph structure, in which all statistical information on tags, groups of tags and tag relationships is depicted. This structure is appropriate for the establishment of a DTD or an XMLschema with rich statistical adornments. In the next subsection, we discuss two algorithms for DTD establishment. Computation method DIAsDEM Workbench simple statistics association rule discovery association rule discovery (ARD) sequence mining (SeqM)
ARD/SeqM lists. Second, we annotate each list with a flag that indicates whether the group depicted by the list is order-sensitive or order-insensitive; in the latter case, the ordering of the list is irrelevant but must be unique. Third, we guarantee uniqueness, i.e. that all permutations of the same group of tags are mapped into the same order-insensitive list, by requiring that order-insensitive list are lexicographically ordered. V A signature: By this definition, a tag can only appear in the graph if its TagSupport is more than zero. This is consistent with the fact that XML tags are derived with a KDD method.
For the groups of tags, we must distinguish among ordersensitive and order-insensitive groups. To do so, we perform three steps. First, we model tag groups as ordered
V 0 where contains only those groups of annotations, for which the GroupSupport value is above a given threshold.
This threshold can be specified as input to the mining software, as is usual in KDD applications, or may be set as low as 0. Of course, the threshold value affects the size of the graph and the execution time of the algorithm that traverses it to build the DTD. X := (0; 1] [ fN U where LLg, with signature:
In this signature, the statistical properties refer to the edge’s source given the group of nodes in the edge’s target. If the target is a sequence of adjacent tags, then the location confidence is the only valid statistical property, while the association confidence and lift are inapplicable. If the target is a set of tags, then the location confidence is inapplicable. When a statistical property is inapplicable, it assumes the NULL value.
Graph Properties The components of our “DTDestablishment graph” are tags, groups of tags and relationships among them, all adorned with statistical values.
All tags discovered by the DIAsDEM Workbench are present in this graph. Which groups of tags are present depends on the threshold value for the group support. Conceivable are both a minimalistic approach with a high threshold, by which only very frequent groups are present, and a maximalistic approach with a zero-value threshold, by which all tag combinations occuring in the documents are present.
If we opt for the minimalistic approach, the graph will not be connected in the general case. It will contain only the groups of tags being more frequent than a threshold, and the frequent relationships among them. Certain tags may be isolated, because they only rarely appear in combination with other tags. Contrary to it, the maximalistic approach ensures that all combinations of tags appearing together in documents are depicted in the graph, and that the graph is connected, except of the unlikely case that some documents contain a single tag not occuring in any other documents.
The upper limit to the graph size indicates that threshold values for the statistical properties are essential for obtaining a manageable graph. On the other hand, each cutoff value implies an information loss. Therefore, we observe the DTDestablishment graph under the maximalistic approach as a reference structure and introduce two algorithms that derive a probabilistic DTD by constructing only a part of this graph. In the following, we present two algorithms that derive a DTD by constructing part of the DTD-establishment graph.
Each tag of this DTD is adorned by only two (derived) probabilistic values, one referring to the tag itself and one to its location inside the DTD. The algorithms are using different heuristics to derive this DTD: the first one concentrates on the pairs of tags appearing most frequently together, while the second one gives preference to maximal sequences of tags. The reader may recall that the computation of the statistics for the relationships among the tags require the activation of data mining software. Hence, each of the algorithms is backed by a miner that returns the desired statistics.
The DTD-establishment graph in its maximalistic version captures all relationships among the semantic tags found by the DIAsDEM Workbench. Similarly to the process of schema establishment for a conventional database application, the designer must decide which relationships among the real-world entities are worth capturing and which are not. In our context, “worth capturing” refers to statistical values, presuming that a DTD should reflect the relationships usually present in the documents rather than the rare ones. However, the DTD-establishment graph contains relationships among sets and among sequences of tags, each one adorned with different (and only partially comparable) statistics.
This algorithm observes a DTD as a set of alternative sequences and builds each sequence backwards, starting at each last tag and proceeding until the first one. Concretely, the algorithm builds “maximal” sequences, where maximality means that the first tag of the sequence is the first tag in most of the documents supporting the sequence. : : : ; The maximum among c0; cu determines the rest of the procedure: If c0 is maximum, is the first element of the sequence. In this case, the sequence is marked as “done” and as “maximal” according to the maximality criterion already mentioned. where c0 is the ratio of documents where has no predecessor divided by the total number of documents. k 1 confidences of the other tags is less than some small k k values. In other words, there are tags with u, such k k ". Then, all tags are acceptable alternatives, resulting to If there is a ci larger than the other elements, then xi is the predecessor of in the sequence. However, it can be the case that the maximum is only marginally larger than the other that (i) one of them has shows the maximum location confidence but (ii) the difference of this value from the location alternative subsequences. s = 1 : : : s a subsequence k, then is “done” but it must x number of documents and comparing this value to the tag 1 puting the ratio of documents starting with over the whole s and thus is maximal. Otherwise, documents starting with jx cj s 1 ", then most documents containing start with 1 mostly adhere to a different maximal sequence.
Identifying maximal tag-sequences. In each iteration, the algorithm considers longer frequent sequences returned by the sequence miner, namely those ending with each subsequence already built. If no frequent sequence is found for also be checked whether it is maximal. This implies comsupport of 1, say c. The comparison is performed across the same guidelines as for alternative tag predecessors: if s side each maximal sequence containing it:
Statistics of maximal tag-sequences. At a final step, the algorithm filters out all sequences that are done but are not maximal. It then assigns probability values to each tag ins this tag with respect to the subsequence of leading to it.
The TagConfidence is the tag’s TagSupport multiplied by the accuracy of the model output by the DIAsDEM Workbench.
The TagPositionConfidence is the location confidence of Backward versus forward sequence construction. The backward-sequence-construction method generates alternative sequences of DTD tags by pruning the frequent sequences of adjacent tags produced by a sequence miner. An equivalent method can be devised by forward-sequence construction. This would have the advantage of being appropriate for incorporation to a sequence miner’s core as well, since most miners of this category perform forward sequence construction: In that case, the mining kernel would be modified to expand a sequence by the most likely successor tag only.
This algorithm observes a DTD as a tree of alternative subsequences and adorns each tag with its support with respect to the subsequence leading to it inside the tree: this is the number of documents starting with this subsequence of tags. Similarly to the sequence-construction algorithm described above, a tag may appear in more than one subsequences, having different predecessors in each one.
Observing the DTD as a tree implies a common root. In
the general case, each document of the archive may start at
a different tag. We assume a dummy root, the children of
which are those tags that appear first in documents. In
general, a tree node refers to a tag , and its children refer to
the tags appearing after in the context of ’s own
predecessors. In a sense, the DTD as a tree of alternatives resembles
a DataGuide as proposed in [
The tree-of-alternatives differs from the sequenceconstruction algorithm in two ways: Firstly, it considers all sequences of tags that appear in documents instead of frequent ones only. Secondly, it only observes complete sequences, while a sequence miner returns arbitrary subsequences of tags.
The tree-of-alternatives method is realized by the
preprocessor module of the Web usage miner WUM [
Most of the knowledge hidden in electronic media of an organization is encapsulated in documents. Acquiring this knowledge implies effective querying of the documents as well as the combination of information pieces from different textual assets. This functionality is usually confined to databaselike query processors, while text search engines scan individual assets and return ranked results. In this study, we have presented a methodology that enables query processing and joining of text sources by structuring them. We propose the derivation of an XML DTD over a domain-specific text archive by means of data mining techniques.
The semantic characterization of text units is the core of our approach as well as the derivation of XML tags from these characterisations. This is undertaken by the DIAsDEM Workbench which is concisely described in the first part of this study. Our main emphasis is on combining these tags that reflect the semantics of many text units across the archive into a single DTD that reflects the semantics of the archive as a whole. We have shown that this DTD is a probabilistic ap(...)
proximation of the archive content and have derived a set of statistical properties that reflect the quality of this approximation, for the whole DTD, for tags inside the DTD and for relationships among these tags.
The statistical properties of tags and of their relationships form the basis for combining them into a complete DTD in the XML sense or even into an XMLschema. We use a graph structure to depict all statistics that can serve as a basis for this operation and propose two mechanisms that derive DTDs by employing a mining algorithm and a set of heuristic rules. We have tested our methodology on an archive of documents from a regional Commercial Register in Germany: We have derived a set of tags with the DIAsDEM Workbench and then implemented one of the proposed mechanisms to derive a DTD for it.
Our future work includes the implementation of the second mechanism for DTD derivement and the establishment of a framework for the comparison of derived DTDs in terms of expressiveness and accuracy. Of course, the ultimate goal of our work is the establishment of a full-fledged querying mechanism over the text archives. To this purpose, we intend to couple our DTD derivation methods with a query mechanism for semi-structured data. Since the DTDs we derive are of probabilistic nature, this implies also the design of a model that evaluates the quality of the query results.
We thank the German Research Society for funding the project DIAsDEM, the Bundesanzeiger Verlagsgesellschaft mbH for providing data and our project collaborators Evguenia Altareva and Stefan Conrad for helpful discussions. The IBM Intelligent Miner for Data is kindly provided by IBM in terms of the IBM DB2 Scholars Program.