Back to search

University of Patras

Αποδοτική οργάνωση και διαχείριση πολυδιάστατων αντικειμένων για την ανακάλυψη γνώσης

Abstract

dc:description

The subject of this dissertation is the invention of methods which assure effective organization and management of multi-dimensional objects in order to achieve knowledge discovery. The initial target behind this study was the needs of a demanding application intending to map the human brain in order to help the localization of epileptic foci. During the corresponding research, the Representation and Management needs of human brain data raised two core research problems:  The representation peculiarity of the composite, non-uniform, network structured three-dimensional objects(brain objects), and  The needs for effective management-use of known and derived data and knowledge dependencies, which can upgrade the application performance and dynamics. The most important part of our relative research, leaded to the: o Investigation of similarity search aspects. As this research area has great application and open problem width, it constitutes a great part of this dissertation. Taking into account that the certain geometrical and knowledge dependency features of human brain data are common – all or part of them - in many modern multimedia applications, the above problems are included in the basic Data Base research problems. Focusing our research in the above problems, we lead up to the:  Definition of new flexible data types, concepts, models, tools and data and knowledge ordering methods (Data Base BDB and models 3D-IFO and MITOS) which organize our data more flexibly and effectively, using methods that not only assure easier data access but also exploit their ‘hidden’ relationships and dependencies for more knowledge discovery.  Definition of new search trees and methods for: o Effective detection of partial similarity among multi-dimensional objects ( Lui k-n-match και INTESIS). o Obliteration of the high performance fall which occurs in similarity trees as dimensionality increases (‘dimensionality curse’) (Digenis structure). o Detection of object features/attributes/properties (dimensions) which have similar course in the time course – for multi-dimensional objects mostly – aiming at the study and detection of possible interaction among them (Chiron proposal). Generally, this dissertation consists of two basic parts, which refer to two research areas with great interaction: • The Multi-Dimensional Data Base Modelling • The Similarity Search among Multi-Dimensional objects. Ιn the first chapter, the problem of human brain mapping for the localization of epileptic foci is discussed. This problem raises issues related to the peculiarities of the representation and the organization of three dimensional objects with complex structures/shapes and functional dependencies and relationships among them (brain objects). In the beginning, the logical model BDB (Brain Data Base) is proposed as a first approach, introducing new entity types. In the corresponding study, a very interesting proposal is the introduction of hierarchical ordering in the Relational Model in order to organize the brain areas according to their frequencies of epileptic foci presence, improving statistically the corresponding response times. In the following, the needs of the application are studied in the basis of a Semantic – IFO model - and of an Object-oriented Model, resulting in the definition of the 3D-IFO and the MITOS (Model for the Intelligent Three-dimensional Object Support) model, respectively. In the framework of 3D-IFO model, new data types and new operators have been introduced, in order to achieve effective representation and better management of the complex brain objects. Additionally, a new constructor and the suitable attribute for its support have been introduced, in order to effectively represent the ordering among brain parts, based on a certain criterion, thus facilitating combined data retrieval. In the end, the object-oriented model MITOS, introduces a new data model (MITOS Data Model – MDM) which cooperates with an intelligent knowledge base approach (MITOS Query Language – MQL). MITOS model introduces many novelties which serve a more expressive and intelligent representation and management of multi-dimensional data and knowledge. One of these novelties constitutes the definition of one more basic object characteristic (in object-oriented theory), the relationship with the environment, releasing it from the situation or the behaviour, where its concept and representation weakens. The second MITOS novelty concerns MQL and is related to the introduction of the concept of ‘key’ in the rules area. The extension of this potentiality – the idea comes from Data Base area – leads in fact to a kind of a key, with a meaning that it could have in Knowledge Bases and can not be exactly the same with that in Data Bases, because of the specific distinctions of these two Bases. The subject of the second chapter is the detection of a least investigated similarity kind among multi-dimensional objects, the partial similarity. Partial similarity refers to similarities which are not full but they really exist. It is difficult to capture them using common techniques based on similarity functions (e.g. Euclidian distance) because these functions are affected by the whole set of object dimensions. Thus, when the objects are similar but ‘very different’ in few dimensions (e.g. very different colour and size) then the corresponding calculated functions (distances) will have very high values because of these few high dissimilarities and the similarity result will be negative while the objects will actually be similar. On the other hand, when between two objects there are low dissimilarities in most dimensions, they are actually dissimilar but the resultant function will have low value, so the dissimilar objects will be discerned as similar. In both cases, the common full similarity detection methods are not reliable. The few studies that have investigated partial similarity, have mostly focused on geometric data. The study which is extended to multi-dimensional objects in general, by Koudas et al., introduces the effective methods k-n-match and frequent k-n-match, which result in k objects being similar to the given ones not in all their dimensions but at least in n ones, avoiding in this way those few very dissimilar dimensions –if any- which lead to false results. Nonetheless, these methods have some weaknesses which finally result either in full similarity (when finally, in frequent k-n-match, all n’s are taken into account) or in an occasional partial similarity detection (with the suitable n’s, which should not be very high or very low, without having however any type or method to calculate the ‘best’ n’s). Based on these methods, we propose two techniques which can provably detect real partial similarities. The first of them, Lui k-n-match, succeeds in the approximate specification of the suitable n’s for the k-n-matches, based on human-computer interaction and on the suitable checks of the similar objects that k-n-matches propose. More precisely, using k-n-match, for each n a set with objects possibly similar to the given one (query object), is proposed . The user filters this set and decides which objects of the proposed set are really similar to the given one. This procedure continues until the point where, while n has become larger than d/2* , the user does not select any object as similar from the proposed object set. In this way, the results are more reliable and valid (because of human-computer interaction) while in parallel the number of the executed k-n-matches are remarkably reduced. The second partial similarity detection method (INTESIS) is based on the following observation: when two objects are similar, it usually means that they are similar in most of their characteristics. In data bases, each of object characteristic is represented by a set (usually small) of features-attributes(dimensions). Thus, if this correspondence between a characteristic and a set of attributes is defined by the developer of each application - creating dimension subsets – then the following can be successively done: a) A full similarity detection for each dimension subset b) Organization of these subsets in the corresponding hierarchical trees for their easy and effective management. The additional simplification of this choice derives from the fact that as long as the dimension subsets are small, the selection of the corresponding tree will be a very easy task, while almost all hierarchical trees have high performance for low dimensionalities. Consequently, the developer of each application can use the hierarchical tree that he/she considers as best (our proposition is R*-tree). Finally, in order to complete the procedure, the application developer has to define which is the minimum number of the requisite similar characteristics that indicate partial similarity, for the particular application. For the evaluation of the method, first of all, it is necessary to mention that it refers to a total number of dimensions less than 100 and consequently to a relatively small number of trees. As it is obvious, the final response time in a uniprocessor system is the sum of the response times of each tree. Taking into account that the number of dimensions which correspond to each tree is small, these trees have very good response times and consequently the total response time is low enough. While the use of each tree does not presuppose the use of another tree before or after it, the search in each tree can be performed in parallel. Therefore, in a multi-processing system, the total response time can be considerably reduced, achieving to reach the time needed for only one tree (when the number of processors is equal to the number of trees). Furthermore, bearing in mind that partial similarity forms a very demanding similarity search kind, not only the response times in multi-processing systems but those times in a uniprocessor system constitute satisfying performances. The third chapter studies the potentiality of defining a new structure which does not ‘suffer’ from ‘dimensional curse’, while it assures good performance for low dimensionalities too. The latest studies have resulted in the following: Although the known similarity trees (either based on space partitioning or on data partitioning perform effectively in cases of low dimensionality, their performance generally degrades as dimensionality increases (especially for more than 10 dimensions). On the other hand, VA-File constitutes a simple approximate method (it is a simple array-file of object geometric approximations) which manages to outperform any other similarity search method at high dimensionality but it has low performance for low dimensionality. In order to overcome this determinant dependence between the performance and the dimensionality of a data-object set, we propose the new hybrid structure called Digenis , which marries the logic of similarity trees with VA-Files logic. More precisely, a static parametric tree (Digenis tree) is defined in conceptual level while the Digenis file, based on Digenis tree, is used in physical level. Using this correlation, a) the similarity search procedure is located in a small part of the file, excluding most dissimilar objects from the search and b) the method is used effectively for both low and high dimensional objects, preserving generality and flexibility. The first necessary definition for Digenis proposal is related to the object families. They consist of objects having a small distance among them (based on a certain limit fl defined from the analyst, in each case) and they are represented by the ‘mean’ object of the family (if it does not exist, it is created just for this role). Each object family is hosted in a node of Digenis tree. The distance which is used is the most spread one, the Euclidian distance, for which the triangle inequality – where the method is mainly based - stands. Additionally, a second distance limit (Lt) is defined – from the analyst- which forms the limit used to conclude if two objects are similar or not. Finally, the root of the Digenis tree is the most ‘centered’ family in the total object area and the nodes being the children of it are its ch nearest families-nodes. The children of each of them are its ch nearest families, and so on. The triangle inequality which stands among the Euclidian distances of the object-families, is proved to be a very useful Theorem for the safe check exclusion of a great part of the tree , leading to a final check in a small tree area. The search analysis of the tree is very helpful for the use of Digenis file, where the real search is performed (physical level). The corresponding Digenis file in the physical level is created if each tree node composes a record of the file, beginning from the tree root and passing from each level, from left to right. Using this correspondence, the proved Digenis tree techniques of easy, safe and quick exclusion of Digenis record areas can be used. The (theoretical and experimental) evaluation of the method results in the detection of certain advantages and disadvantages of it. Advantages:  The file has very good performance for high dimensionalities. This was expected because the file works as a kind of VA-File, where the records are compact geometric approximations. This matters because both the use of object families achieves a first data ‘compression’ and the pre-arrangement of the objects via the conceptual use of the tree lead to a kind of grouping of neighboring objects in neighboring areas.  The file has also good performance for low dimensionality, because in comparison to VA-File, it is expectably better while it is based on a tree structure. For the same reason, Digenis file is competitive to the classic hierarchical similarity trees. Drawbacks:  The fact that the number of children for each node is statically defined as ch in each application is a disadvantage for the construction of the tree, because usually the really similar families may be more or less than ch. Confrontation: If the similar families of a node are more than ch, then only the ch closest to the family are placed as its children, in the next level. If they are less than ch, then a limit of children and full nodes in the tree is defined. When this limit is overcome, the nodes-children are normally placed in the tree and the rest nodes –until ch-th one – remain empty. When however the number of the children of a family and of the full nodes in the tree, are less than this limit, the corresponding subtree is separated, creating a new smaller tree – with smaller ch – while the initial tree is reorganized. Consequently, the final application may include more than one Digenis tree, which are accessed from the bigger to the smaller, until the similarity is found (if any).  Perhaps there are remote areas of object families – without any connection with other families – which can not be included in any other tree. Confrontation: A file including sequentially the remote families (called ‘remote’ file’) is created. During the similarity search, this file is the first which is accessed because while it usually hosts a few families, the search will be quick enough. If a similarity is detected (among the query object and a family in the file), then the search in the trees will be avoided while if no similarity exists, the time overhead of the file search is almost negligible, because of its size. In the last chapter, a new kind of dynamic similarity search is investigated. It is related with the time streams not of the objects themselves but of their properties/attributes/dimensions. In other words, what is detected is whether the courses of two or more properties resemble. This kind of information can be very useful for several kinds of applications (medical, financial, scientific in general, e.t.c). Using medical data related to hormonal tests as an example, we prove that, based on our method Chiron , the hormones which are developed in the same way are accurately and effectively detected. More precisely, new objects (property course objects or Chiron objects) which encode the variations of each property in certain time intervals, are defined and organized in a tree (Chiron tree). The way these objects are defined, their differences and the Chiron tree itself make its navigation and the detection of similar Chiron objects – and consequently of properties which are developed in a similar way - a quick and easy procedure. This is achieved via the distribution of the Chiron objects in the Chiron tree according to the number of the different digits that exist among them. In this way, when we search in the Chiron tree for objects similar to a given one, a simple and compact algorithm is used, which avoids a vast amount of useless checks among very different objects. Generally, the method is promising enough because it poses new problems for investigation, like the statistical analysis of its results, the search for objects that are developed in a reverse way, the management of time shifts among the property course objects and the Chiron tree optimization.

Degree

thesis:*
Grantor dc:publisher
University of Patras
Year dc:date
2010

Author and committee

dc:creator, dc:contributor.*
Authors dc:creator
  • Κροτοπούλου, Αικατερίνη
  • Krotopoulou, Aikaterini

Subjects

dc:subject × 19

Rights

Language dc:language
gre

Identifiers

dc:identifier.*
Identifier
10.12681/eadd/26174

Chain of custody

source
Harvested from
Greek National Archive of PhD Theses
Base URL
phdtheses.ekt.gr/eadd_oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Κροτοπούλου, Αικατερίνη; Krotopoulou, Aikaterini. Αποδοτική οργάνωση και διαχείριση πολυδιάστατων αντικειμένων για την ανακάλυψη γνώσης. University of Patras, 2010. http://hdl.handle.net/10442/hedi/26174