Chapter 3 Data Mining and Data Visualization Modern

Chapter 3: Data Mining and Data Visualization Modern Data Warehousing, Mining, and Visualization: Core Concepts by George M. Marakas © 2003, Prentice-Hall Chapter 3 - 1

3 -1: A Picture is Worth a Thousand Words Data mining is the set of activities used to find new, hidden, or unexpected patterns in data. These techniques are often called knowledge data discovery (KDD), and include statistical analysis, neural or fuzzy logic, intelligent agents or data visualization. The KDD techniques not only discover useful patterns in the data, but also can be used to develop predictive models. © 2003, Prentice-Hall 2

Verification Versus Discovery In the past, decision support activities were primarily based on the concept of verification. This required a great deal of prior knowledge on the decision-maker’s part in order to verify a suspected relationship. With the advance of technology, the concept of verification began to turn into discovery. © 2003, Prentice-Hall 3

Data Mining’s Growth in Popularity One reason is that we keep getting more and more data all the time and need tools to understand it. We also are aware that the human brain has trouble processing multidimensional data. A third reason is that machine learning techniques are becoming more affordable and more refined at the same time. © 2003, Prentice-Hall 4

Making Accurate Predictions with Data Mining Although the literature contains statements such as “data mining will allow us to predict who will buy a particular product, ” that is against human nature. In situations where data mining is used to predict response to a marketing campaign, only about 5% of the people selected as “likely respondents” actually do respond. © 2003, Prentice-Hall 5

Making Accurate Predictions with Data Mining (cont. ) Although the accuracy of predicting individual behavior is not so good, it is better than it seems, since direct marketing efforts often have “hit rates” of only about 1% without data mining. © 2003, Prentice-Hall 6

3 -2: Online Analytical Processing (OLAP) Codd developed a set of 12 rules for the development of multidimensional databases: 1. 2. 3. 4. 5. 6. Multidimensional view Transparent to user Accessible Consistent reporting Client-server architecture Generic dimensionality © 2003, Prentice-Hall 7. 8. 9. 10. 11. 12. Dynamic sparse matrix handling Multiuser support Cross-dimensional ops Intuitive manipulation Flexible reporting Unlimited dimension and aggregation 7

OLAP as Implemented To date, it does not appear that any implementation exists that satisfies all 12 rules. Some people argue it might not even be possible to attain all of them. More recently, the term OLAP has come to represent the broad category of software technology that enables multidimensional analysis of enterprise data. © 2003, Prentice-Hall 8

Multidimensional OLAP (MOLAP) Data can be viewed across several dimensions. Here sales are arrayed by region and product. A fourth dimension could be added by using several graphs -- perhaps at different time points. Most analyses have many more dimensions than this. MOLAP handles data as an n-dimensional hypercube. © 2003, Prentice-Hall 9

Relational OLAP (ROLAP) A large relational database server replaces the multidimensional one. The database contains both detailed and summarized data, allowing “drill down” techniques to be applied. SQL interfaces allow vendors to build tools, both portable and scalable. This does require databases with many relational tables which may lead to substantial processor overhead on complex joins. © 2003, Prentice-Hall 10

A Typical Relational Schema © 2003, Prentice-Hall 11

3 -3: Techniques Used to Mine the Data Paralleling the popularity of data mining itself, the development of new techniques is exploding as well. Many innovations are vendor-specific, which sometimes does little to advance the state of the art. Regardless, data-mining techniques tend to fall into four major categories: 1. classification 2. association 3. sequencing 4. clustering © 2003, Prentice-Hall 12

Classification methods The goal is to discover rules that define whether an item belongs to a particular subset or class of data. For example, if we are trying to determine which households will respond to a direct mail campaign, we will want rules that separate the “probables” from the not probables. These IF-THEN rules often are portrayed in a tree-like structure. © 2003, Prentice-Hall 13

Association Methods These techniques search all transactions from a system for patterns of occurrence. A common method is market basket analysis, in which the set of products purchased by thousands of consumers are examined. Results are then portrayed as percentages; for example, “ 30% of the people that buy steaks also buy charcoal”. © 2003, Prentice-Hall 14

Sequencing Methods These methods are applied to time series data in an attempt to find hidden trends. If found, these can be useful predictors of future events. For example, customer groups that tend to purchase products tied-in with hit movies would be targeted with promotional campaigns timed to release dates. © 2003, Prentice-Hall 15

Clustering Techniques Clustering techniques attempt to create partitions in the data according to some distance metric. The clusters formed are data grouped together simply by their similarity to their neighbors. By examining the characteristics of each cluster, it may be possible to establish rules for classification. © 2003, Prentice-Hall 16

Data Mining Technologies Statistics – the most mature data mining technologies, but are often not applicable because they need clean data. In addition, many statistical procedures assume linear relationships, which limits their use. Neural networks, genetic algorithms, fuzzy logic – these technologies are able to work with complicated and imprecise data. Their broad applicability has made them popular in the field. © 2003, Prentice-Hall 17

Data Mining Technologies (cont. ) Decision trees – these technologies are conceptually simple and have gained in popularity as better tree growing software was introduced. Because of the way they are used, they are perhaps better called “classification” trees. © 2003, Prentice-Hall 18

The Knowledge Discovery Search Process Table 3 -2 contains a more detailed outline of the process, but the major steps are: l Define the business problem and obtain the data to study it. l Use data mining software to model the problem. l Mine the data to search for patterns of interest. © 2003, Prentice-Hall 19

The Knowledge Discovery Search Process (cont. ) l Review the mining results and refine them by respecifying the model. l Once validated, make the model available to other users of the DW. © 2003, Prentice-Hall 20

Creating a Data-Mining Model Although syntax differs from vendor to vendor, building a model on top of a database is much like creating a table: CREATE MODEL mail_list Income character input, Age integer input, Respond character input To populate it with data, use an SQL INSERT: INSERT INTO mail_list SELECT income, age, respond FROM client_list WHERE region = ‘Southeast” © 2003, Prentice-Hall 21

Creating a Data-Mining Model (cont. ) The process automatically created additional views of the model (mail_list_UNDERSTAND and mail_list_PREDICT). These can be examined: SELECT * FROM mail_list_UNDERSTAND WHERE input_column_name = ‘income” and input_column_value = “high” and output_column_name = “respond” and output_column_value = ‘yes” Once these are created, they are treated as tables in the database so they can be viewed and joined by other users. © 2003, Prentice-Hall 22

New Applications for Data Mining As the technology matures, new applications emerge, especially in two new categories, text mining and web mining. Some text mining examples are: l Distilling the meaning of a text l Accurate summarization of a text l Explication of the text theme structure l Clustering of texts © 2003, Prentice-Hall 23

Web mining is a special case of text mining where the mining occurs over a website. It enhances the website with intelligent behavior, such as suggesting related links or recommending new products. It allows you to unobtrusively learn the interests of the visitors and modify their user profiles in real time. They also allow you to match resources to the interests of the visitor. © 2003, Prentice-Hall 24

3 -4: Market Basket Analysis: The King of Algorithms This is the most widely used and, in many ways, most successful data mining algorithm. It essentially determines what products people purchase together. Stores can use this information to place these products in the same area. Direct marketers can use this information to determine which new products to offer to their current customers. Inventory policies can be improved if reorder points reflect the demand for the complementary products. © 2003, Prentice-Hall 25

Association Rules for Market Basket Analysis Rules are written in the form “left-hand side implies right-hand side” and an example is: Yellow Peppers IMPLIES Red Peppers, Bananas, Bakery To make effective use of a rule, three numeric measures about that rule must be considered: (1) support, (2) confidence and (3) lift © 2003, Prentice-Hall 26

Measures of Predictive Ability Support refers to the percentage of baskets where the rule was true (both left and right side products were present). 2. Confidence measures what percentage of baskets that contained the left-hand product also contained the right. 3. Lift measures how much more frequently the left-hand item is found with the right than without the right. 1. © 2003, Prentice-Hall 27

An Example Rule: Green Peppers IMPLIES Bananas Red Peppers IMPLIES Bananas Yellow Peppers IMPLIES Bananas Lift 1. 37 1. 43 1. 17 Support 3. 77 8. 58 22. 12 Confidence 85. 96 89. 47 73. 09 The confidence suggests people buying any kind of pepper also buy bananas. Green peppers sell in about the same quantities as red or yellow, but are not as predctive. © 2003, Prentice-Hall 28

Market Basket Analysis Methodology We first need a list of transactions and what was purchased. This is pretty easily obtained these days from scanning cash registers. Next, we choose a list of products to analyze, and tabulate how many times each was purchased with the others. The diagonals of the table shows how often a product is purchased in any combination, and the off-diagonals show which combinations were bought. © 2003, Prentice-Hall 29

A Convenience Store Example (5 transactions) Consider the following simple example about five transactions at a convenience store: Transaction 1: Transaction 2: Transaction 3: Transaction 4: Transaction 5: Frozen pizza, cola, milk Milk, potato chips Cola, frozen pizza Milk, pretzels Cola, pretzels These need to be cross tabulated and displayed in a table. © 2003, Prentice-Hall 30

A Convenience Store Example (5 transactions) Product Bought Pizza also Milk also Cola also Chips also Pretzels also Pizza 2 1 2 0 0 Milk 1 3 1 1 1 Cola 2 1 3 0 1 Chips 0 1 0 Pretzels 0 1 1 0 2 Pizza and Cola sell together more often than any other combo; a cross-marketing opportunity? Milk sells well with everything – people probably come here specifically to buy it. © 2003, Prentice-Hall 31

Using the Results The tabulations can immediately be translated into association rules and the numerical measures computed. Comparing this week’s table to last week’s table can immediately show the effect of this week’s promotional activities. Some rules are going to be trivial (hot dogs and buns sell together) or inexplicable (toilet rings sell only when a new hardware store is opened). © 2003, Prentice-Hall 32

Limitations to Market Basket Analysis A large number of real transactions are needed to do an effective basket analysis, but the data’s accuracy is compromised if all the products do not occur with similar frequency. The analysis can sometimes capture results that were due to the success of previous marketing campaigns (and not natural tendencies of customers). © 2003, Prentice-Hall 33

Performing Analysis with Virtual Items The sales data can be augmented with the addition of virtual items. For example, we could record that the customer was new to us, or had children. The transaction record might look like: Item 1: Sweater Item 2: Jacket Item 3: New This might allow us to see what patterns new customers have versus old customers. © 2003, Prentice-Hall 34

Computing Measures of Association Pizza Milk Cola Chips Pretzels Pizza 2 1 2 0 0 Milk 1 3 1 1 1 Cola 2 1 3 0 1 Chips 0 1 0 Pretzels 0 1 1 0 2 Let’s do some of the textbook’s example computations here …… © 2003, Prentice-Hall 35

Taxonomies The presence of items not purchased very frequently is an obstacle to a good market basket analysis. One way to deal with this is to eliminate products that occur with a frequency less than some threshold. A better idea would be to try to form groups of products that fall below the threshold. Four flavors of popsicle occur 9% of the time all together, but no more than 3% individually. © 2003, Prentice-Hall 36

Multidimensional Market Basket Analysis Rules can involve more than two items, for example Plant and Clay Pot IMPLIES Soil. These rules are built iteratively. First, pairs are found, then relevant sets of three or four. These are then pruned by removing those that occur infrequently. In an environment like a grocery store, where customers commonly buy over 100 items, rules could involve as many as 10 items. © 2003, Prentice-Hall 37

3 -5: Current Limitations and Challenges to Data Mining Despite the potential power and value, data mining is still a new field. Some things that thus far have limited advancement are: l Identification of missing information – not all knowledge gets stored in a database l Data noise and missing values – future systems need better ways to handle this l Large databases and high dimensionality – future applications need ways to partition data into more manageable chunks © 2003, Prentice-Hall 38

3 -6: Data Visualization: “Seeing” the Data © 2003, Prentice-Hall 39

Visual Presentation For any kind of high dimensional data set, displaying predictive relationships is a challenge. The picture on the previous slide uses 3 -D graphics to portray the weather balloon data numbers in text Table 11 -4. We learn very little from just examining the numbers. Shading is used to represent relative degrees of thunderstorm activity, with the darkest regions the heaviest activity. © 2003, Prentice-Hall 40

A Bit of History An early effort used sequences of twodimensional graphs to add depth. Current virtual reality programs allow the user to step through a data set. Try going to a realtor’s website and taking a tour of a house up for sale. © 2003, Prentice-Hall 41

Human Visual Perception and Data Visualization Data visualization is so powerful because the human visual cortex converts objects into information so quickly. The next three slides show (1) usage of global private networks, (2) flow through natural gas pipelines, and (3) a risk analysis report that permits the user to draw an interactive yield curve. All three use height or shading to additional dimensions to the figure. © 2003, Prentice-Hall 42

Global Private Network Activity High Activity Low Activity © 2003, Prentice-Hall 43

Natural Gas Pipeline Analysis Note: Height shows total flow through compressor stations. © 2003, Prentice-Hall 44

An “Enlivened” Risk Analysis Report © 2003, Prentice-Hall 45

Geographical Information Systems A GIS is a special purpose database that contains a spatial coordinate system. A comprehensive GIS requires: 1. Data input from maps, aerial photos, etc. 2. Data storage, retrieval and query 3. Data transformation and modeling 4. Data reporting (maps, reports and plans) © 2003, Prentice-Hall 46

The Special Capabilities of a GIS In general, a GIS contains two types of data: Spatial data: these elements correspond to a uniquely-defined location on earth. They could be in point, line or polygon form. Attribute data: These are the data that will be portrayed at the geographic references established by spatial data. Example: Data from an opinion poll is displayed for multiple regions in the United States. Clicking on an area allows the user to drill down to the results for smaller areas. © 2003, Prentice-Hall 47

Telephone Polling Results Note: On the “live” map, clicking on an area allows the user to drill down and see results for smaller areas. © 2003, Prentice-Hall 48

3 -7: Siftware Technologies Although data visualization product vendors seem to enter or leave the market with great frequency, several firms are beginning to develop significant brand loyalty. Red Brick – Helped category managers at H. E. B. in San Antonio to determine which products to put in which stores. Another application was the consolidation of three old data warehouses at Hewlett-Packard. © 2003, Prentice-Hall 49

Siftware -- Continued Oracle – A large suite of connectivity products allows transparent access to mainframe databases. Some major customers include John Alden Insurance, Shop. Ko Stores and Pacific Bell. Informix – Associated Grocers uses Informix data warehousing products at the heart of its three-tier client-server system. © 2003, Prentice-Hall 50

Siftware -- Continued Sybase – Sybase Warehouse WORKS is an integrated system designed around the four key functions in data warehousing. Silicon Graphics – Data mining software is mated to 3 -D visualization tools to allow users to fly through data. IBM – provides a number of decision support tools in its Information Warehouse Solutions. © 2003, Prentice-Hall 51