Ludmila Aleksejeva

Forecasting Product Life Cycle Phase Transition Points with Modular Neural Networks Based System

Management of the product life cycle and of the corresponding supply network largely depends on information in which specific phase of the life cycle one or another product currently is and when the phase will be changed. Finding a phase of the product life cycle can be interpreted as forecasting transition points between phases of life cycle of these products. This paper provides a formulation of the above mentioned task of forecasting the transition points and presents the structured data mining system for solving that task. The developed system is based on the analysis of historical demand for products and on information about transitions between phases in life cycles of those products. The experimental results with real data display information about the potential of the created system.

Influence of Membership Functions on Classification of Multi-Dimensional Data

Influence of Membership Functions on Classification of Multi-Dimensional Data The aim of this study is to explore whether the number of intervals for each attribute influences the classification result and whether a larger number of intervals provide better classification accuracy using the Fuzzy PRISM algorithm. The feature selection has been carried out using Fast correlation-based filter solution, and then the decreased data sets have been applied in experiments with preferences used in the previous experiment series. The article also provides conclusions about the obtained classification results and analyzes criteria of certain experiments and their impact on the final result. Also a series of experiments was carried out to assess how and whether the classification result is influenced by categorization of continuous data, which is one of the membership function construction steps; Fuzzy unordered rule induction algorithm was used. The experiments have been carried out using four real data sets - Golub leukemia, Singh prostate, as well as Gastric cancer and leukemia donor data sets of the Latvian Biomedical Research and Study Center. Piederības funkciju ietekme daudzatribūtu datu klasifikācijā Šajā rakstā pētīts tas, vai katra atribūta intervālu skaits ietekmē klasifikācijas rezultātu, kā arī tas, vai lielāks intervālu skaits nodrošina arī labāku klasifikācijas rezultātu. Eksperimentu veikšanai izmantots FuzzyPRISM algoritms. Eksperimentos izmantotas četras reālas datu kopas - Golub leukemia, Singh prostate, Leukemia II un Latvijas biomedicīnas pētījumu un studiju centra kuņga vēža pacientu un veselo pacientu datu kopas. Visām datu kopām ir ļoti liels atribūtu skaits (līdz pat 10 000 atribūtu) un salīdzinoši neliels ierakstu skaits. Pirmajā sērijā, kurā bija divpadsmit eksperimenti, netika veikta atribūtu atlase. Nākamajā sērijā veikta atribūtu atlase, izmantojot Fast Correlation Based Filter risinājumu, un atkārtoti eksperimenti ar iepriekšējā eksperimentu sērijā izmantotajiem uzstādījumiem. Var secināt, ka vairāk likumu iegūts atribūtu atlases eksperimentos. Papildus trim eksperimentiem apmācības kopā veikta atribūtu atlase, izmantojot Fast Correlation Based Filter ar desmitkārtīgo šķērsvalidāciju, lai pārliecinātos, par to kā šķērsvalidācija ietekmē gala rezultātu. Izdarīti secinājumi par iegūtajiem klasifikācijas rezultātiem, kā arī analizēti atsevišķi eksperimentu parametri un to ietekme uz gala rezultātu. Izmantojot algoritmu FURIA, veikta arī eksperimentu sērija, lai noskaidrotu kā un vai klasifikācijas rezultātu ietekmē skaitlisku datu pārveidošana par kategoriskiem, kas ir viens no piederības funkciju konstruēšanas soļiem. Salīdzinot klasifikācijas rezultātus, tika secināts, ka visaugstākos rezultātus uzrāda eksperimenti ar originālo datu kopu ar nepārtrauktām atribūtu vērtībām, tomēr iegūtie klasifikācijas rezultāti, sākot ar dalījumu 10 intervālos, tuvojas pilno datu kopu rezultātiem. Tāpēc izvēloties, cik intervālos dalīt atribūta vērtību, jābūt skaidrībai, kas ir galvenais - klasifikācijas precizitāte, interpretējamība vai skaitļošanas ilgums. Влияние функций принадлежности на классификацию данных со многими атрибутами Статья посвящена исследованию следующих вопросов: влияет ли число интервалов определения каждого атрибута на результат классификации, обеспечивает ли увеличение числа интервалов улучшение результата классификации. Для проведения экспериментов использован алгоритм FuzzyPRISM. В экспериментах использованы четыре реальных множества данных Golub leukemia, Singh prostate, Leukemia II и множество данных о здоровых и больных раком желудка пациентах Латвийского центра биомедицины. Для всех множеств данных характерно очень большое число атрибутов (до 10 000) и сравнительно небольшое число записей. В первой серии из двенадцати экспериментов отбор атрибутов не проводился. В следующей серии отбор атрибутов проводился с использованием алгоритма Fast Correlation Based Filter, и далее повторялись эксперименты с установками, используемыми в экспериментах предыдущей серии. Можно заключить, что больше правил получено в экспериментах, основанных на отборе атрибутов. Дополнительно в трех экспериментах на обучающем множестве производился отбор атрибутов по алгоритму Fast Correlation Based Filter, а также использовалась 10кратная кроссвалидация (для проверки ее влияния на конечный результат). Сделаны выводы о полученных результатах классификации, проанализированы параметры отдельных экспериментов и их влияние на конечный результат. С использованием алгоритма Fuzzy Unordered Rule Induction Algorithm проведена также серия экспериментов, позволяющая выяснить влияние преобразования численных данных в категорические (что является одним из этапов конструирования функций принадлежности) на результат классификации. Сравнивая результаты классификации, можно заключить, что наилучшие результаты получены в экспериментах с полным оригинальным множеством данных, которое характеризуется непрерывными оценками атрибутов; однако, начиная с деления оценок атрибутов на 10 интервалов, полученные результаты классификации приближаются к результатам на полном множестве данных. Поэтому при выборе числа интервалов, на которые нужно делить оценки атрибутов, целесообразно выяснить, что важнее точность классификации, интерпретация результатов или продолжительность расчетов.

A Study on the Behaviour of the Algorithm for Finding Relevant Attributes and Membership Functions

A Study on the Behaviour of the Algorithm for Finding Relevant Attributes and Membership Functions One of the most recent approaches in machine learning is fuzzy rules usage for solving classification problems. This paper describes the algorithm for finding relevant attributes and searching for membership functions. Experimental results are used to clarify - which data sets can be used to automatically gain primary membership functions from primary data. This quality - gaining of membership functions - is one of the pros of the algorithm, because it eases resolution of classification task. The ability to use it with fuzzy data is one more merit. As a result, there are obtained reliable fuzzy classification rules to separate classes. By reconstructing primary membership functions also the number of IF-THEN rules gained from decision tables is reduced up to three times. Four experiments are conducted with different training and testing data set sizes. Conclusions are made about the optimal size of the training and testing data set that is necessary for achieving better results as well as about the data this algorithm is appropriate for. Finally, possible directions for further research are outlined.

Database Concepts in a Domain Ontology

Abstract There are multiple approaches for mapping from a domain ontology to a database in the task of ontology-based data access. For that purpose, external mapping documents are most commonly used. These documents describe how the data necessary for the description of ontology individuals and other values, are to be obtained from the database. The present paper investigates the use of special database concepts. These concepts are not separated from the domain ontology; they are mixed with domain concepts to form a combined application ontology. By creating natural relationships between database concepts and domain concepts, mapping can be implemented more easily and with a specific purpose. The paper also investigates how the use of such database concepts in addition to domain concepts impacts ontology building and data retrieval.

Genetic Algorithm Based Random Selection-Rule Creation for Ontology Building

This paper investigates the possibility of creating ontology concepts from information contained in a database, by finding random queries with the help of a genetic algorithm. This is done, with the aim to help ontology building. Based on the structure of the database random chromosomes are created. Their genes describe possible selection criteria. By using a genetic algorithm, these selections are improved. Due to the size of the database, an approach for finding fitness from general characteristics, instead of an in-depth analysis of the data is considered. After the algorithm finished improving the chromosomes in the population, the best chromosomes are chosen. They are considered for implementation as ontology concepts. These ontology concepts can be used as descriptions of the information contained in the database. Because genetic algorithms are not usually used for ontology building, this paper investigates the feasibility of such an approach.

Decision Tree Creation Methodology Using Propositionalized Attributes

Abstract The aim of the article is to analyse and thoroughly research the methods of construction of the decision trees that use decision tree learning with statement propositionalized attributes. Classical decision tree learning algorithms, as well as decision tree learning with propositionalized attributes have been observed. The article provides the detailed analysis of one of the methodologies on the importance of using the decision trees in knowledge presentation. The concept of ontology use is offered to develop classification systems of decision trees. The application of the methodology would allow improving the classification accuracy.

Neural Network Modelling for Sports Performance Classification as a Complex Socio-Technical System

Abstract Extraction of meaningful information by using artificial neural networks, where the focus is upon developing new insights for sports performance and supporting decision making, is crucial to gain success. The aim of this article is to create a theoretical framework and structurally connect the sports and multi-layer artificial neural network domains through: (a) describing sports as a complex socio-technical system; (b) identification of pre-processing subsystem for classification; (c) feature selection by using data-driven valued tolerance ratio method; (d) design predictive system model of sports performance using a backpropagation neural network. This would allow identifying, classifying, and forecasting performance levels for an enlarged data set.