Solving Data Mining Problems Through Pattern Recognition

Description

• Copyright 1998
• Dimensions: 7" x 9-1/4"
• Pages: 400
• Edition: 1st

• Book
• ISBN-10: 0-13-095083-1
• ISBN-13: 978-0-13-095083-3

9508C-2

Apply pattern recognition to find the hidden gems in your data!

Data mining technology is helping businesses everywhere to work smarter by revealing unknown patterns within existing archives. Applying the latest advances in pattern recognition software can give you a key competitive edge across all data mining applications. The tutorials and software package included in Solving Data Mining Problems through Pattern Recognition take advantage of machine learning techniques and neural networks to help you get the most out of your data. Besides explaining the most current theories, Solving Data Mining Problems through Pattern Recognition takes a practical approach to overall project development concerns.

The rigorous, multi-step method includes:

• Defining the pattern recognition problem
• Collection, preparation, and preprocessing of data
• Choosing the appropriate algorithm and tuning algorithm parameters Training, testing, and troubleshooting.

Pattern classification, estimation, and modeling are addressed using the following algorithms:

• Linear and logistic regression
• Unimodal Gaussian and Gaussian mixture
• Multilayered perceptron/backpropagation and radial basis function neural networks
• K nearest neighbors and nearest cluster
• K means clustering.

While some aspects of pattern recognition involve advanced mathematical principles, most successful projects rely on a strong element of human experience and intuition. Solving Data Mining Problems through Pattern Recognition provides a strong theoretical grounding for beginners, yet it also contains detailed models and insights into real-world problem-solving that will inspire more experienced users, be they database designers, modelers, or project leaders.

This book includes a free, 90-day trial copy of Pattern Recognition Workbench, a powerful, easy-to-use system that combines machine learning, neural networks, and statistical algorithms to help you apply pattern recognition to your data right now. The enclosed CD-ROM runs under Windows(r) 95 and Windows NT(tm).



Sample Content

Table of Contents

1. Introduction .

Pattern Recognition by Humans. Pattern Recognition by Computers. Data Mining and Pattern Recognition. Types of Pattern Recognition. Classification. Calculation in Classification. Uncertainty in Classification. Computer-Automated Classification.Estimation. Calculation in Estimation. Uncertainty in Estimation. Computer-Automated Estimation. Developing a Model. Fixed Models. Parametric Models. Nonparametric Models. Preprocessing. A Continuum of Methods. Biases Due to Prior Knowledge. The Purpose of this Book.

2. Key Concepts: Estimation.

Terminology and Notation. Characteristics of an Optimal Model. Sources of Error. Fixed Models. Parametric Models. Example: Linear Regression. Generalization. Shortcomings of Parametric Methods. Iteration through Parametric Forms. Nonparametric Models. The Underlying Modeling Problem. Heuristics in Nonparametric Modeling. Approximation Architectures. A Practical Nonparametric Approach. The Role of Preprocessing. Statistical Considerations.

3. Key Concepts: Classification.

Terminology and Notation. Characteristics of an Optimal Classifier. Types of Models. Decision-Region Boundaries. Probability Density Functions. Posterior Probabilities. Approaches to Modeling. Fixed Models. Parametric Models. Nonparametric Models. The Role of Preprocessing. The Importance of Multiple Techniques.

Appendix. Statistical Considerations.


4. Additional Application Areas.

Database Marketing. Response Modeling. Cross Selling. Time-Series Prediction. Detection. Probability Estimation. Information Compression. Sensitivity Analysis.

5. Overview of the Development Process.

Defining the Pattern Recognition Problem. Collecting Data. Preparing Data. Preprocessing. Selecting an Algorithm and Training Parameters. Training and Testing. Iterating Steps and Troubleshooting.

Appendix. Evaluating the Final Model.


6. Defining the Pattern Recognition Problem.

What Problems Are Suitable for Data-Driven Solutions? How Do You Evaluate Results? Is It a Classification or Estimation Problem?What Are the Inputs and Outputs?

Appendix. Defining the Problem in PRW.


7. Collecting Data.

What Data to Collect. How to Collect Data. How Much Data Is Enough. Using Simulated Data.

Appendix. Importing Data into PRW.


8. Preparing Data.

Transforming Data into Numerical Values. Inconsistent Data and Outliers.

Appendix. Preparing Data in PRW.

Handling Missing Data. Converting Non-Numeric Inputs. Handling Inconsistent Data or Outliers.

9. Data Preprocessing.

Why Should You Preprocess Your Data? Averaging Data Values. Thresholding Data. Reducing the Input Space. Normalizing Data. Why Normalize Data? Types of Normalization. Modifying Prior Probabilities. Other Considerations.

Appendix A: Preprocessing in PRW.

Averaging Time-Series Data. Thresholding and Replacing Input Values. Reducing the Input Space. Normalizing Data. Modifying Prior Input Probabilities.

10. Selecting Architectures and Training Parameters.

Types of Algorithms. How to Pick an Algorithm.Practical Constraints. Memory Usage. Training Times. Classification/Estimation Times. Algorithm Descriptions. Linear Regression. Logistic Regression. Unimodal Gaussian. Multilayered Perceptron/Backpropagation. Radial Basis Functions. K Nearest Neighbors. Gaussian Mixture. Nearest Cluster. K Means Clustering. Decision Trees. Other Nonparametric Architectures. Algorithm Comparison Summary.

Appendix A: Selecting Algorithms and Training Parameters in PRW.

Selecting an Algorithm in PRW. Setting Algorithm Parameters. Linear Regression. Logistic Regression. Unimodal Gaussian. Backpropagation/MLP. Radial Basis Functions. K Nearest Neighbors. Gaussian Mixture. Nearest Cluster. K Means Clustering.

11. Training and Testing.

Train, Test, and Evaluation Sets. Validation Techniques. Cross Validation. Bootstrap Validation. Sliding Window Validation.

Appendix A: Training, Testing, and Reporting in PRW.

The Experiment Manager. Running Experiments. Enabling and Disabling Experiments. Scheduling Experiments. Selecting Report Options. Viewing Different Reports. Cross Validation. Sliding Window Validation.

12. Iterating Steps and Trouble-Shooting.

Iterating to Improve Your Solution. Automated Searches.Input Variable Selection. Algorithm Parameter Searches. Trouble-Shooting. Training Error Is High. Test Error Is High. Classification Problem Performs Poorly on Some Classes. Problems with Production Accuracy. Decision Tree Works Best by Far. Backpropagation Does Not Converge. Backpropagation Finds a Local Minimum Solution. Matrix Inversion Problem. Unimodal Gaussian Has High Training Error. Gaussian Mixture Diverges. RBF Has High Training Error.

Appendix A: Iterating in PRW.

Overview of PRW Features. Creating Multiple Spreadsheets. Creating Multiple Experiment Managers. Using Multiple Work Sessions. Using Automated Searches. Preprocessing Data. Exporting Experiments and Reports. Re-Using Experiment Parameters. Building User Functions.

Appendix A: References and Suggested Reading.


Appendix B: Pattern Recognition Workbench.


Appendix C: Unica Technologies, Inc.

About Unica. Unica's Software Products.

Appendix D: Glossary.


Index.


Software License Agreement.


What's on this CD.

Preface

Preface

Data Mining

Data mining is a term usually applied to techniques that can be used to find underlying structure and relationships in large amounts of data. These techniques are drawn primarily from the related fields of neural networks, statistics, pattern classification, and machine learning. They are becoming more important as computer automation spreads and as the processing and storage capabilities of computers increase. Widely available, low-cost computer technology now makes it possible to both collect historical data and also to institute on-line analysis and controls for newly arriving data.

Applications

Data mining techniques are being successfully used for many diverse applications. These include paper and sheet metal production control, medical diagnosis and risk prediction, credit-card fraud detection, computer security break-in and misuse detection, computer user identity verification, aluminum and steel smelting control, oil refinery control, pollution control in power plants, fraudulent income tax return detection, automobile engine control and fault detection, electric motor fault and failure prediction, mass mailing and telemarketing, and simplifying world-wide-web usage by predicting useful sites from past user behavior. Benefits of Data Mining

Benefits in these and other applications include reduced costs due to more accurate control, more accurate future predictions, more effective fault detection and prediction, fraud detection and control, and automation of repetitive human tasks. In addition, services can be improved and extended due to a better understanding of underlying processes and human behavior. Outline of this Book

This book provides a concise introduction to some of the most important input-output mapping, prediction, pattern classification, and clustering algorithms useful for data mining. This introduction is based on many collective years of experience by the authors, which has led to a focus on practical issues that must be addressed to successfully solve data mining problems. The book provides a basic road map for experts who know much about a specific application, but little about neural networks, statistics, pattern classification, or machine learning.

This road map first helps potential users determine whether input-output mapping, prediction, pattern classification, or clustering algorithms are appropriate for a given application. It then helps users determine which measurements, attributes, or features might be useful as inputs to these algorithms and provides guidance in collecting and formatting this data for computer analysis. Guidelines are then presented for accurately accessing performance using separate training, evaluation, and test data partitions or cross-validation. Finally, each important algorithm is described and guidance is provided concerning settings for parameters used to control the many algorithms. Multi-Algorithm Approach

An important truism presented in this book is that data mining is an art and that there is no single simple approach that is best for all problems. Rather, there are many algorithms and data representations, and the best strategy is to interactively experiment to find an approach that works for a particular data set. This human interaction is greatly simplified by the availability of software toolkits which allow users to interactively explore many algorithms on a common data set using the same performance metrics. This book focuses on one comprehensive software toolkit (Pattern Recognition Workbench) that includes most of the algorithms described and has the capability of handling large data sets. Details concerning this software, however, are relegated to the Appendix and to sections at the ends of chapters. These details can thus be skipped or used as examples of the types of information required to apply the various algorithms. Intended Audience

This book is most useful for persons who have a specific application in mind, but who know little about data mining algorithms. They can use this book to determine whether the algorithms presented can be applied to their application, to learn terminology, and to provide guidance when they try out some of the recommended approaches using a software toolkit. More experienced users who want to understand the theory behind prediction, mapping, control, pattern classification, and clustering or who would like to read detailed descriptions concerning specific data mining applications should explore other more advanced texts.

Richard P. Lippmann



Updates

Submit Errata

