What Kinds of Patterns Can Data Mining Discover?
Using the most relevant data (which may come from organizational databases or may be obtained from outside sources), data mining builds models to identify patterns among the attributes (i.e., variables or characteristics) that exist in a data set. Models are usually the mathematical representations (simple linear correlations and/or complex highly nonlinear relationships) that identify the relationships among the attributes of the objects (e.g., customers) described in the data set. Some of these patterns are explanatory (explaining the interrelationships and affinities among the attributes), whereas others are predictive (projecting future values of certain attributes). In general, data mining seeks to identify three major types of patterns:
Associations find commonly co-occurring groupings of things, such as “beers and diapers” or “bread and butter” commonly purchased and observed together in a shopping cart (i.e., market-basket analysis). Another type of association pattern captures the sequences of things. These sequential relationships can discover time-ordered events, such as predicting that an existing banking customer who already has a checking account will open a savings account followed by an investment account within a year.
Predictions tell the nature of future occurrences of certain events based on what has happened in the past, such as predicting the winner of the Super Bowl or forecasting the absolute temperature on a particular day.
Clusters identify natural groupings of things based on their known characteristics, such as assigning customers in different segments based on their demographics and past purchase behaviors.
These types of patterns have been manually extracted from data by humans for centuries, but the increasing volume of data in modern times has created a need for more automatic approaches. As data sets have grown in size and complexity, direct manual data analysis has increasingly been augmented with indirect, automatic data processing tools that use sophisticated methodologies, methods, and algorithms. The manifestation of such evolution of automated and semi-automated means of processing large data sets is now commonly referred to as data mining.
As mentioned earlier, generally speaking, data mining tasks and patterns can be classified into three main categories: prediction, association, and clustering. Based on the way in which the patterns are extracted from the historical data, the learning algorithms of data mining methods can be classified as either supervised or unsupervised. With supervised learning algorithms, the training data includes both the descriptive attributes (i.e., independent variables or decision variables) and the class attribute (i.e., output variable or result variable). In contrast, with unsupervised learning, the training data includes only the descriptive attributes. Figure 2.3 shows a simple taxonomy for data mining tasks, along with the learning methods and popular algorithms for each of the data mining tasks. Out of the three main categories of tasks, prediction patterns/models can be classified as the outcome of a supervised learning procedure, while association and clustering patterns/models can be classified as the outcome of unsupervised learning procedures.
Prediction is commonly used to indicate telling about the future. It differs from simple guessing by taking into account the experiences, opinions, and other relevant information in conducting the task of foretelling. A term that is commonly associated with prediction is forecasting. Even though many people use these two terms synonymously, there is a subtle difference between them. Whereas prediction is largely experience and opinion based, forecasting is data and model based. That is, in the order of increasing reliability, one might list the relevant terms as guessing, predicting, and forecasting. In data mining terminology, prediction and forecasting are used synonymously, and the term prediction is used as the common representation of the act. Depending on the nature of what is being predicted, prediction can be named more specifically as classification (where the predicted thing, such as tomorrow’s forecast, is a class label such as “rainy” or “sunny”) or regression (where the predicted thing, such as tomorrow’s temperature, is a real number, such as “65 degrees”).
Classification, or supervised induction, is perhaps the most common of all data mining tasks. The objective of classification is to analyze the historical data stored in a database and automatically generate a model that can predict future behavior. This induced model consists of generalizations over the records of a training data set, which help distinguish predefined classes. The hope is that the model can then be used to predict the classes of other unclassified records and, more importantly, to accurately predict actual future events.
Common classification tools include neural networks and decision trees (from machine learning), logistic regression and discriminant analysis (from traditional statistics), and emerging tools such as rough sets, support vector machines, and genetic algorithms. Statistics-based classification techniques (e.g., logistic regression, discriminant analysis) have been criticized as making unrealistic assumptions about the data, such as independence and normality, which limit their use in classification-type data mining projects.
Neural networks involve the development of mathematical structures (somewhat resembling the biological neural networks in the human brain) that have the capability to learn from past experiences, presented in the form of well-structured data sets. They tend to be more effective when the number of variables involved is rather large and the relationships among them are complex and imprecise. Neural networks have disadvantages as well as advantages. For example, it is usually very difficult to provide a good rationale for the predictions made by a neural network. Also, neural networks tend to need considerable training. Unfortunately, the time needed for training tends to increase exponentially as the volume of data increases, and, in general, neural networks cannot be trained on very large databases. These and other factors have limited the applicability of neural networks in data-rich domains. (See Chapter 5, “Algorithms for Predictive Analytics,” for more detailed coverage of neural networks.)
Decision trees classify data into a finite number of classes, based on the values of the input variables. Decision trees are essentially a hierarchy of if–then statements and are thus significantly faster than neural networks. They are most appropriate for categorical and interval data. Therefore, incorporating continuous variables into a decision tree framework requires discretization—that is, the conversion of continuous valued numeric variables to ranges and categories.
A related category of classification tools is rule induction. Unlike with a decision tree, with rule induction, the if–then statements are induced from the training data directly, and they need not be hierarchical in nature. Other, more recent techniques such as SVM, rough sets, and genetic algorithms are gradually finding their way into the arsenal of classification algorithms and are covered in more detail in Chapter 5 as part of the discussion on data mining algorithms.
Using associations—which are commonly called association rules in data mining—is a popular and well-researched technique for discovering interesting relationships among variables in large databases. Thanks to automated data-gathering technologies such as use of bar code scanners, the use of association rules for discovering regularities among products in large-scale transactions recorded by point-of-sale systems in supermarkets has become a common knowledge-discovery task in the retail industry. In the context of the retail industry, association rule mining is often called market-basket analysis.
Two commonly used derivatives of association rule mining are link analysis and sequence mining. With link analysis, the links among many objects of interest are discovered automatically, such as the link between web pages and referential relationships among groups of academic publication authors. With sequence mining, relationships are examined in terms of their order of occurrence to identify associations over time. Algorithms used in association rule mining include the popular Apriori (where frequent item sets are identified), FP-Growth, OneR, ZeroR, and Eclat algorithms. Chapter 4, “Data and Methods for Predictive Analytics,” provides an explanation of Apriori.
Clustering involves partitioning a collection of things (e.g., objects, events, etc., presented in a structured data set) into segments (or natural groupings) whose members share similar characteristics. Unlike in classification, in clustering, the class labels are unknown. As the selected algorithm goes through the data set, identifying the commonalities of things based on their characteristics, the clusters are established. Because the clusters are determined using a heuristic-type algorithm, and because different algorithms may end up with different sets of clusters for the same data set, before the results of clustering techniques are put into use, it may be necessary for an expert to interpret and potentially modify the suggested clusters. After reasonable clusters have been identified, they can be used to classify and interpret new data.
Not surprisingly, clustering techniques include optimization. The goal of clustering is to create groups so that the members within each group have maximum similarity and the members across groups have minimum similarity. The most commonly used clustering techniques include k-means (from statistics) and self-organizing maps (from machine learning), which is a unique neural network architecture developed by Kohonen (1982).
Firms often effectively use their data mining systems to perform market segmentation with cluster analysis. Cluster analysis is a means of identifying classes of items so that items in a cluster have more in common with each other than with items in other clusters. This type of analysis can be used in segmenting customers and directing appropriate marketing products to the segments at the right time in the right format at the right price. Cluster analysis is also used to identify natural groupings of events or objects so that a common set of characteristics of these groups can be identified to describe them.
Two techniques often associated with data mining are visualization and time-series forecasting. Visualization can be used in conjunction with other data mining techniques to gain a clearer understanding of underlying relationships. As the importance of visualization has increased in recent years, the term visual analytics has emerged. The idea is to combine analytics and visualization in a single environment for easier and faster knowledge creation. Visual analytics is covered in detail in Chapter 4. In time-series forecasting, the data consists of values of the same variable that is captured and stored over time, at regular intervals. This data is then used to develop forecasting models to extrapolate the future values of the same variable.