Attempt all the questions.
Section-A
Question1) What do you mean by decision trees? How do decision trees work? Describe the CART algorithm.
Question2) What do you understand by Neural Networks? How it works? How neural network techniques are applied in data mining?
Question3) What is web mining? How does data mining work? How it is used in business.
Question4) Describe in brief how information from a data warehouse promotes CRM (Customer Relationship Management)
Section-B
Case Study
Neural Networks have probably seen their greatest acceptance and application in the financial industry. Applications like credit card fraud, default (personal bankruptcy), and even customer attrition have all shown successful application with neural networks. For fraud alone, the dollar amounts to be saved by viable predictive models are staggering. In 1995 combined losses from credit card fraud and counterfeiting was $1.3 billion. Visa member banks alone lost more than $148 million to counterfeiters in 1994. The good news is that neural network system have been introduced that reduced that loss by 16 % to $124 million just one year later.
In spite of successes like these in fraud and other applications, holy grail of the financial applications is still time series prediction: being able to say what is going to happen next- whether it is predicting closing price of a stock, market, or even general overall shifts in the market. One particularly difficult prediction is foreign exchange rates between various currencies. As there are multiple players, trying to exploit small niches to make money. They create market and its behaviour. Since of many people playing with several information, the market is commonly thought to be “efficient”. Here “efficient market hypothesis” is based on the premise that there is little opportunity to exploit market changes because as soon as a small one occurs that may be predictable, someone probably already beat you to it. It also implies that price of anything in the market has been efficiently set to correct value for the present time on the basis of future risk and future possibilities for profit. Because of this it is generally believed that it is difficult to predict future market behaviour based on historical information in any better manner than basically at random. In the case presented here, 4 years of exchange rate data between Swiss franc and U.S. dollar were obtained for the years1985-1988. A standard feedforward neural network algorithm (no recurrent paths between nodes of different layers) was used with backpropogation learning algorithm. Two network architectures were tried. The first had seven input units, seven hidden units, and one output unit.; second had an input layer of seven units, two hidden layers (five units & two units) and a single output unit. Both networks individually, as well as the average of their outputs were tried. This averaging seemed to improve the performance of either network alone and was tried in an effort to remove any bias inherent in using one particular architecture. The filtering function used at each hidden node and output node took the summed inputs of the node and output a value between -1.0 & 1.0.
Output to be predicted was the direction of the exchange rate (1.0 for an upswing and -1.0 for a downturn)
Question5) Case Questions:
(i) What are the benefits of using neural network algorithm?
(ii) How does backpropogation learning algorithm works?
(iii) Why two hidden layers are used?
(iv) Explain the weakness of this model if any?