Pattern Discovery in Time-oriented Data

We present a data mining system, EasyMiner which has been developed for interactive mining of interesting patterns in time-oriented databases. This system implements a wide spectrum of data mining functions, including generalisation, characterisation, classification, association and relevant analysis. By enhancing several interesting data mining techniques, including attribute induction and association rule mining to handle time-oriented data the system provide a user friendly, interactive data mining environment with good performance. These algorithms were tested on time-oriented medical data and experimental results show that the algorithms are efficient and effective for discovery of pattern in databases. INTRODUCTION Knowledge Discovery in Databases (KDD) is the effort to understand, analyse, and eventually make use of the huge volume of data available. According to Fayyad et al. [1] KDD is the non trivial process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data. In their opinion, there are usually many steps in a KDD process including selection, pre-processing, transformation, data mining, and interpretation/ evaluation of the results as shown in Figure 1. As seen in the Figure data mining is only one step of the process, involving the application of discovery tools to find interesting patterns from targeted data, but in the research community most often the term data mining and KDD have been used interchangeably. Finding patterns in databases is the fundamental operation behind several common data mining tasks, including association rule [2] and sequential pattern mining[3]. Data mining is the process of applying machine learning and other techniques to classical databases in order to extract implicit, previously unknown and potential useful patterns from database [4]. Time is an important aspect of all real-world phenomena. Conventional databases model an enterprise as it changes dynamically by a snapshot at a particular point in time. As information is updated in a conventional database, its old, out-of-date data is discarded forever, its changes over time are thus lost. But in many situations, this snapshot-type of database is inadequate. They cannot handle queries related to any historical data. For many applications such as accounting, banking, GIS systems and medical data the changes made their databases over time are a valuable source of information which can direct their future operation. The pattern discovered from conventional databases has limited value since the temporal nature of data is not taken into account but only the current or latest snapshot. In the rest of the paper the data mining techniques used by Easy Miner for discovering interesting patterns from time-oriented database are described. In section 2 relevance analysis method is presented with examples from medical domain. In section 3 association rules mining technique is presented and section 4 describes Easy Miner approach to pattern discovery by mining classification rules. Section 5 concludes with a summary of the paper and outline of the future work. Figure 1. The knowledge discovery process 1 Relevance Analysis As we know from real life, several facts are relevant with each other and there is a strong dependence between them, for example Age is relevant to Date of Birth, Title(Mr, Mrs, Miss) is relevant to Sex and Marital Status. This kind of knowledge is qualitative and it is quite useful to mine it from large databases that hold information about many objects(fields). For example a bank could look in its data and identify which are the factors that it should take in mind in order to give a Credit Limit to a customer. As we found by using EasyMiner Credit Limit is relevant to Account Status, Monthly Expenses, Marital Status, Monthly Income, Sex and etc. A number of statistical and machine learning techniques for relevance analysis have been proposed until now. The most popular and acceptable approach in the data mining community, is the method of measuring the uncertainty coefficient[4]. Let us suppose that the generated set from the collection of task relevant data is a set P of p data records. Suppose also, that the interesting attribute has m distinct values defining by this way m Knowledge DATA Selection Target data Preprocessing Data Mining Transformation PATTERNS Interpretation/ Evaluation Preprocessed data