The subject of this book is automated learning, or, as we will more often call it, Machine Learning (ML). That is, we wish to program computers so that they can “learn” from input available to them. Roughly speaking, learning is the process of converting experience into expertise or knowledge. The input to a learning algorithm is training data, representing experience, and the output is some expertise, which usually takes the form of another computer program that can perform some task. Seeking a formal-mathematical understanding of this concept, we'll have to be more explicit about what we mean by each of the involved terms: What is the training data our programs will access? How can the process of learning be automated? How can we evaluate the success of such a process (namely, the quality of the output of a learning program)?

WHAT IS LEARNING?

Let us begin by considering a couple of examples from naturally occurring animal learning. Some of the most fundamental issues in ML arise already in that context, which we are all familiar with.

Bait Shyness – Rats Learning to Avoid Poisonous Baits: When rats encounter food items with novel look or smell, they will first eat very small amounts, and subsequent feeding will depend on the flavor of the food and its physiological effect. If the food produces an ill effect, the novel food will often be associated with the illness, and subsequently, the rats will not eat it.

Let us begin our mathematical analysis by showing how successful learning can be achieved in a relatively simplified setting. Imagine you have just arrived in some small Pacific island. You soon find out that papayas are a significant ingredient in the local diet. However, you have never before tasted papayas. You have to learn how to predict whether a papaya you see in the market is tasty or not. First, you need to decide which features of a papaya your prediction should be based on. On the basis of your previous experience with other fruits, you decide to use two features: the papaya's color, ranging from dark green, through orange and red to dark brown, and the papaya's softness, ranging from rock hard to mushy. Your input for figuring out your prediction rule is a sample of papayas that you have examined for color and softness and then tasted and found out whether they were tasty or not. Let us analyze this task as a demonstration of the considerations involved in learning problems.

Our first step is to describe a formal model aimed to capture such learning tasks.

A FORMAL MODEL – THE STATISTICAL LEARNING FRAMEWORK

The learner's input: In the basic statistical learning setting, the learner has access to the following:

Domain set: An arbitrary set, χ. This is the set of objects that we may wish to label.

In the previous chapter we introduced the families of convex-Lipschitz-bounded and convex-smooth-bounded learning problems. In this section we show that all learning problems in these two families are learnable. For some learning problems of this type it is possible to show that uniform convergence holds; hence they are learnable using the ERM rule. However, this is not true for all learning problems of this type. Yet, we will introduce another learning rule and will show that it learns all convex-Lipschitz-bounded and convex-smooth-bounded learning problems.

The new learning paradigm we introduce in this chapter is called Regularized Loss Minimization, or RLM for short. In RLM we minimize the sum of the empirical risk and a regularization function. Intuitively, the regularization function measures the complexity of hypotheses. Indeed, one interpretation of the regularization function is the structural risk minimization paradigm we discussed in Chapter 7. Another view of regularization is as a stabilizer of the learning algorithm. An algorithm is considered stable if a slight change of its input does not change its output much. We will formally define the notion of stability (what we mean by “slight change of input” and by “does not change much the output”) and prove its close relation to learnability. Finally, we will show that using the squared l2 norm as a regularization function stabilizes all convex-Lipschitz or convex-smooth learning problems. Hence, RLM can be used as a general learning rule for these families of learning problems.

In this chapter we discuss how to assess the significance of the mined frequent patterns, as well as the association rules derived from them. Ideally, the mined patterns and rules should satisfy desirable properties such as conciseness, novelty, utility, and so on. We outline several rule and pattern assessment measures that aim to quantify different properties of the mined results. Typically, the question of whether a pattern or rule is interesting is to a large extent a subjective one. However, we can certainly try to eliminate rules and patterns that are not statistically significant. Methods to test for the statistical significance and to obtain confidence bounds on the test statistic value are also considered in this chapter.

RULE AND PATTERN ASSESSMENT MEASURES

Let I be a set of items and T a set of tids, and let D ⊆ T × I be a binary database. Recall that an association rule is an expression X → Y, where X and Y are itemsets, i.e., X, Y ⊆ I, and X ∩ Y = ∅. We call X the antecedent of the rule and Y the consequent.

The search space for frequent itemsets is usually very large and it grows exponentially with the number of items. In particular, a low minimum support value may result in an intractable number of frequent itemsets. An alternative approach, studied in this chapter, is to determine condensed representations of the frequent itemsets that summarize their essential characteristics. The use of condensed representations can not only reduce the computational and storage demands, but it can also make it easier to analyze the mined patterns. In this chapter we discuss three of these representations: closed, maximal, and nonderivable itemsets.

MAXIMAL AND CLOSED FREQUENT ITEMSETS

Given a binary database D ⊆ T × I, over the tids T and items I, let F denote the set of all frequent itemsets, that is,

F = {X | X ⊆ I and sup(X) ≥ minsup}

Maximal Frequent Itemsets

A frequent itemset X ∈ F is called maximal if it has no frequent supersets. Let M be the set of all maximal frequent itemsets, given as

M= {X | X ∈ F and 6 ∃Y ⊃ X, such that Y ∈ F}

The set M is a condensed representation of the set of all frequent itemset F, because we can determine whether any itemset X is frequent or not using M.

Before we can mine data, it is important to first find a suitable data representation that facilitates data analysis. For example, for complex data such as text, sequences, images, and so on, we must typically extract or construct a set of attributes or features, so that we can represent the data instances as multivariate vectors. That is, given a data instance x (e.g., a sequence), we need to find a mapping φ, so that φ(x) is the vector representation of x. Even when the input data is a numeric data matrix, if we wish to discover nonlinear relationships among the attributes, then a nonlinear mapping φ may be used, so that φ(x) represents a vector in the corresponding high-dimensional space comprising nonlinear attributes. We use the term input space to refer to the data space for the input data x and feature space to refer to the space of mapped vectors φ(x). Thus, given a set of data objects or instances xi, and given a mapping function φ, we can transform them into feature vectors φ(xi), which then allows us to analyze complex data instances via numeric analysis methods.

There exist many different clustering methods, depending on the type of clusters sought and on the inherent data characteristics. Given the diversity of clustering algorithms and their parameters it is important to develop objective approaches to assess clustering results. Cluster validation and assessment encompasses three main tasks: clustering evaluation seeks to assess the goodness or quality of the clustering, clustering stability seeks to understand the sensitivity of the clustering result to various algorithmic parameters, for example, the number of clusters, and clustering tendency assesses the suitability of applying clustering in the first place, that is, whether the data has any inherent grouping structure. There are a number of validity measures and statistics that have been proposed for each of the aforementioned tasks, which can be divided into three main types:

External: External validation measures employ criteria that are not inherent to the dataset. This can be in form of prior or expert-specified knowledge about the clusters, for example, class labels for each point.

Internal: Internal validation measures employ criteria that are derived from the data itself. For instance, we can use intracluster and intercluster distances to obtain measures of cluster compactness (e.g., how similar are the points in the same cluster?) and separation (e.g., how far apart are the points in different clusters?).

Many real-world applications such as bioinformatics, Web mining, and text mining have to deal with sequential and temporal data. Sequence mining helps discover patterns across time or positions in a given dataset. In this chapter we consider methods to mine frequent sequences, which allow gaps between elements, as well as methods to mine frequent substrings, which do not allow gaps between consecutive elements.

FREQUENT SEQUENCES

Let ∑ denote an alphabet, defined as a finite set of characters or symbols, and let |∑| denote its cardinality. A sequence or a string is defined as an ordered list of symbols, and is written as s = s1s2 …sk, where si ∈ ∑ is a symbol at position i, also denoted as s[i]. Here |s| = k denotes the length of the sequence. A sequence with length k is also called a k-sequence. We use the notation s[i : j] = sisi+1 … sj−1sj to denote the substring or sequence of consecutive symbols in positions i through j, where j > i. Define the prefix of a sequence s as any substring of the form s[1 : i] = s1s2 …si, with 0 ≤ i ≤ n. Also, define the suffix of s as any substring of the form s[i : n] = sisi+1 …sn, with 1 ≤ i ≤ n+1. Note that s[1 : 0] is the empty prefix, and s[n + 1 : n] is the empty suffix.

We have seen different classifiers in the preceding chapters, such as decision trees, full and naive Bayes classifiers, nearest neighbors classifier, support vector machines, and so on. In general, we may think of the classifier as a model or function M that predicts the class label ŷ for a given input example x:

ŷ = M(x)

where x = (x1, x2, …, xd)T ∈ Rd is a point in d-dimensional space and ŷ ∈ {c1, c2, …, ck} is its predicted class.

To build the classification model M we need a training set of points along with their known classes. Different classifiers are obtained depending on the assumptions used to build the model M. For instance, support vector machines use the maximum margin hyperplane to construct M. On the other hand, the Bayes classifier directly computes the posterior probability P(cj|x) for each class cj, and predicts the class of x as the one with the maximum posterior probability, ŷ = argmaxcj {P(cj|x)}. Once the model M has been trained, we assess its performance over a separate testing set of points for which we know the true classes. Finally, the model can be deployed to predict the class for future points whose class we typically do not know.

In many applications one is interested in how often two or more objects of interest co-occur. For example, consider a popular website, which logs all incoming traffic to its site in the form of weblogs. Weblogs typically record the source and destination pages requested by some user, as well as the time, return code whether the request was successful or not, and so on. Given such weblogs, one might be interested in finding if there are sets of web pages that many users tend to browse whenever they visit the website. Such “frequent” sets of web pages give clues to user browsing behavior and can be used for improving the browsing experience.

The quest to mine frequent patterns appears in many other domains. The prototypical application is market basket analysis, that is, to mine the sets of items that are frequently bought together at a supermarket by analyzing the customer shopping carts (the so-called “market baskets”). Once we mine the frequent sets, they allow us to extract association rules among the item sets, where we make some statement about how likely are two sets of items to co-occur or to conditionally occur. For example, in the weblog scenario frequent sets allow us to extract rules like, “Users who visit the sets of pages main, laptops and rebates also visit the pages shopping-cart and checkout”, indicating, perhaps, that the special rebate offer is resulting in more laptop sales.