A Data Mining Approach for Business Hamid Rastegari, Mohd Noor Md. Sap Faculty of Computer Science and Information Systems University Technology Malaysia [email protected], [email protected] Abstract: various attributes of data and between several Modern business is rushing toward electronic snapshots of data over a period of time. These hidden commerce. Electronic commerce has been put patterns have enormous potential in predictions and forward as a new method of implementing trade personalizations in e-commerce. Data mining has activities. If the transition is done properly, it enables been pursued as a research topic by at least three better management, new services, lower transaction communities: the statisticians, the artificial costs and better customer relations. However, intelligence researchers, and the database engineers successful business depends on speed and skill of (Raghavan, 2005). In this section present a brief knowledge discovery. Data mining has been overview of some of the features of each of these considered to be a tool of business for knowledge approaches. discovery. This paper discusses the important role of business based on data mining knowledge 2.1 Role of statistics in data mining development to detection the relation of data mining and electronic commerce. Extracting causal information from data is often Keywords: business, e-commerce, data mining. one of the principal goals of data mining and more generally of statistical inference. Statisticians have 1. Introduction done aggregate data analyses on data for decades; thus data mining has actually existed from the time Electronic commerce has changed the face of large scale statistical modeling has been made business. It allows better customer management, new possible (Carbone 2000). strategies for marketing, an expanded range of Statisticians consider the causal relationship products and more efficient operations. A key enabler between the dependent variables and independent of this change is using of increasingly sophisticated variables as proposed by the user (usually the domain data mining tools. expert), and attempt to capture the degree and nature Data mining tools generate new information for of dependence between the variables. Modeling decision makers from very large databases. The methods include simple linear regression, multiple various mechanisms of this generation include regressions, and nonlinear regression. Such models abstractions, aggregations, summarizations, and are often parameter driven and are arrived at after characterizations of data (Carbone 2000). These solving attendant optimization models. forms, in turn, are the result of applying sophisticated The regression methods may be considered modeling techniques from the diverse fields of analogous to the association rules in data mining. In statistics, artificial intelligence, database management the latter case, rule-mining algorithms propose the and computer graphics. correlation of item sets in a database, across various Although much work has been done to date, more attributes of the transactions. studies need to be conducted to as various subjects in Time series modeling on the other hand, is more a variety of e-commerce problems. The purpose of relevant in sequential mining. This is used to unearth this paper is a present of data mining methods and correlations and patterns in temporally ordered data. expression application of data mining in business. It For a more detailed overview of time series methods, is a briefing of works that have been done in this the reader may refer to Box et al (1994). area. This study can be useful for future work. 2.2 Role of AI in data mining 2. Data Mining Methods Artificial intelligence, on the other hand, has Having a huge amount of data, make some provided a number of useful methods for data problems for detection of hidden relationships among mining. Decision tree is most useful in classification problems. With this technique, map data into several assumptions, such as normality or independent error classes and a tree is constructed to model the terms. classification process. Machine learning is a set of Artificial intelligence based methods using neural methods that enable a computer to learn relations networks are used in clustering and classification from the given data sets. With minimal or no methods of data mining. They can also be used in hypothesis from the user, learning algorithms do sequential mining (Park 2000). For instance, market come up with meaningful relations and also explain basket analysis which concerns itself with identifying them well. Some of the most popular learning hidden customer segments could be solved using systems include the neural networks and support neural networks with unsupervised learning. An vector machines. We briefly present the relevant online web store may want to provide different issues below. grades of service to its users depending on the Neural networks are predominantly used to learn frequency of customers‟ visits to their websites. linear and nonlinear relationships between variables Identifying the basket of such customer segments of interest. The architecture, in general, consists of a could be done using clustering methods. Finally, if perceptron with input and output nodes with the web store wants to identify the factors weighted edges connecting the two nodes. A neural contributing to repeat customers, they could use the network with two layers is thus a bi-partite acyclic nonlinear regression expressions obtained using graph. The perceptron, which is the learning neural networks. machine, is „trained‟ in order to arrive at an optimal „weight vector‟. The output is then expressed as a 2.3 Role of database in data mining (weighted) linear combination of the inputs. Learning consists of solving an underlying optimization model Keeping in mind that data mining approaches rely which is solved using gradient descent based heavily on the availability of high quality data sets, methods. the database community has invented an array of Neural networks are also useful in clustering data relevant methods and mechanisms that need to be sets. The most popular method available to cluster used prior to any data mining exercise. Extract, data sets is the K-means algorithm. Given an M- transform and load (ETL) applications are worthy of dimensional data set, the idea is to try and locate the mention in this context. Given an enterprise system minimal number of centroids around which the data like an enterprise resource planning system (ERP), it set clusters itself. Thus the onus is to define an is likely that the number of transactions that happen appropriate distance vector that helps partition the by the minute could run into hundreds, if not data sets into as minimally overlapping sub-sets as thousands. Data mining can certainly not be run on possible. In general, Euclidean distance metrics are the transaction databases in their native state. It proposed for „optimally‟ partitioning a given data set. requires to be extracted at periodic intervals, The optimization is again based on minimizing the transformed into a form usable for analysis, and sum of squares of an appropriate error term, using loaded on to the servers and applications that work on classical gradient based methods. the transformed data. Today, software systems exist The advantages of neural networks over the in the form of data warehousing solutions that are conventional statistical analysis methods are as often bundled with the ERP system, to perform this follows (Park 2000). First, neural networks are good complex and important task. at modeling nonlinear relationships and interaction It is to be observed that data warehouses are while conventional statistical analysis in most cases essentially snapshots of transactional data aggregated assumes linear relationship between independent along various dimensions (including time, variables and dependent variables. Neural networks geographies, demographies, products etc.) In order to build their own models with the help of learning run data mining algorithms, it is common practice to process whether the relationships among variables use the data available in the data warehouse rather are linear or not. Secondly, neural networks perform than by running real time scripts to fetch well with missing or incomplete data. A single transactional data. This is for the simple reason that missing value in regression analysis leads to removal for practical purposes, it is sufficient to include of the entire observation or removal of the associated snapshots of data taken at say, weekly or monthly variable from all observations in the data set being basis, for analysis. Real-time data is not relevant for analyzed. However, neural networks update weights tactical decision making, which is where data mining between input, output, and intermediate nodes, so is used. Data warehousing is nevertheless fraught that even incomplete data can contribute to learning with technological challenges. and produce desired output results. Finally, neural When one has terabytes of data available, the goal networks do not require scale adjustment or statistical of database engineers in data mining is to create structures and mechanisms to efficiently read in data technology systems for the website include into memory and run algorithms like A priori sophisticated data mining tools that take care of (Agrawal & Srikant 1994). Such algorithms assume knowledge representation of customer profiles and