Search Engines Information Retrieval in Practice All slides Addison Wesley, 2008 Classification and Clustering Classification and clustering are classical pattern recognition / machine learning problems Classification Asks what class does this item belong to? Supervised learning task Clustering Asks how can I group this set of items? Unsupervised learning task Items can be documents, queries, emails, entities, images,

etc. Useful for a wide variety of search engine tasks Classification Classification is the task of automatically applying labels to items Useful for many search-related tasks Spam detection Sentiment classification Online advertising Two common approaches Probabilistic Geometric How to Classify? How do humans classify items?

For example, suppose you had to classify the healthiness of a food Identify set of features indicative of health fat, cholesterol, sugar, sodium, etc. Extract features from foods Read nutritional facts, chemical analysis, etc. Combine evidence from the features into a hypothesis Add health features together to get healthiness factor Finally, classify the item based on the evidence If healthiness factor is above a certain value, then deem it healthy Ontologies Ontology is a labeling or categorization scheme

Examples Binary (spam, not spam) Multi-valued (red, green, blue) Hierarchical (news/local/sports) Different classification tasks require different ontologies Nave Bayes Classifier Probabilistic classifier based on Bayes rule: C is a random variable corresponding to the class D is a random variable corresponding to the input (e.g. document) Probability 101: Random Variables

Random variables are non-deterministic Can be discrete (finite number of outcomes) or continues Model uncertainty in a variable P(X = x) means the probability that random variable X takes on value x Example: Let X be the outcome of a coin toss P(X = heads) = P(X = tails) = 0.5 Example: Y = 5 - 2X If X is random, then Y is random If X is deterministic then Y is also deterministic Note: Deterministic just means P(X = x) = 1.0! Nave Bayes Classifier Documents are classified according to:

Must estimate P(d | c) and P(c) P(c) is the probability of observing class c P(d | c) is the probability that document d is observed given the class is known to be c Estimating P(c) P(c) is the probability of observing class c Estimated as the proportion of training documents in class c: Nc is the number of training documents in class c N is the total number of training documents Estimating P(d | c) P(d | c) is the probability that document d is

observed given the class is known to be c Estimate depends on the event space used to represent the documents What is an event space? The set of all possible outcomes for a given random variable For a coin toss random variable the event space is S = {heads, tails} Multiple Bernoulli Event Space Documents are represented as binary vectors One entry for every word in the vocabulary Entry i = 1 if word i occurs in the document and is 0 otherwise Multiple Bernoulli distribution is a natural way to model distributions over binary vectors

Same event space as used in the classical probabilistic retrieval model Multiple Bernoulli Document Representation Multiple-Bernoulli: Estimating P(d | c) P(d | c) is computed as: Laplacian smoothed estimate: Collection smoothed estimate: Multinomial Event Space Documents are represented as vectors of term frequencies One entry for every word in the vocabulary Entry i = number of times that term i occurs in the

document Multinomial distribution is a natural way to model distributions over frequency vectors Same event space as used in the language modeling retrieval model Multinomial Document Representation Multinomial: Estimating P(d | c) P(d | c) is computed as: Laplacian smoothed estimate: Collection smoothed estimate: Support Vector Machines

Based on geometric principles Given a set of inputs labeled + and -, find the best hyperplane that separates the +s and -s Questions How is best defined? What if no hyperplane exists such that the +s and -s can be perfectly separated? Best Hyperplane? First, what is a hyperplane? A generalization of a line to higher dimensions Defined by a vector w With SVMs, the best hyperplane is the one with the maximum margin If x+ and x- are the closest + and - inputs to

the hyperplane, then the margin is: Support Vector Machines w x > 0 + Hy w per + x =

w + + x pla = ne 0 +

M ar gin 1 w + 1 +

x = + + + w x < 0

Best Hyperplane? It is typically assumed that , which does not change the solution to the problem Thus, to find the hyperplane with the largest margin, we must maximize . This is equivalent to minimizing . Separable vs. Non-Separable Data + + +

+ + + + + + Separable

+ + + +

+ + + + +

+ Non-Separable + Linear Separable Case In math: In English: Find the largest margin hyperplane that separates the +s and -s

Linearly Non-Separable Case In math: In English: i denotes how misclassified instance i is Find a hyperplane that has a large margin and lowest misclassification cost The Kernel Trick Linearly non-separable data may become linearly separable if transformed, or mapped, to a higher dimension space Computing vector math (i.e., dot products) in very high dimensional space is costly The kernel trick allows very high dimensional dot products to be computed efficiently

Allows inputs to be implicitly mapped to high (possibly infinite) dimensional space with little computational overhead Kernel Trick Example The following function maps 2-vectors to 3-vectors: Standard way to compute is to map the inputs and compute the dot product in the higher dimensional space However, the dot product can be done entirely in the original 2-dimensional space: Common Kernels The previous example is known as the polynomial kernel (with p = 2)

Most common kernels are linear, polynomial, and Gaussian Each kernel performs a dot product in a higher implicit dimensional space Non-Binary Classification with SVMs One versus all Train class c vs. not class c SVM for every class If there are K classes, must train K classifiers Classify items according to: One versus one Train a binary classifier for every pair of classes Must train K(K-1)/2 classifiers Computationally expensive for large values of K SVM Tools

Solving SVM optimization problem is not straightforward Many good software packages exist SVM-Light LIBSVM R library Matlab SVM Toolbox Evaluating Classifiers Common classification metrics Accuracy (precision at rank 1) Precision Recall F-measure ROC curve analysis Differences from IR metrics

Relevant replaced with classified correctly Microaveraging more commonly used Classes of Classifiers Types of classifiers Generative (Nave-Bayes) Discriminative (SVMs) Non-parametric (nearest neighbor) Types of learning Supervised (Nave-Bayes, SVMs) Semi-supervised (Rocchio, relevance models) Unsupervised (clustering) Generative vs. Discriminative Generative models Assumes documents and classes are drawn from joint

distribution P(d, c) Typically P(d, c) decomposed to P(d | c) P(c) Effectiveness depends on how P(d, c) is modeled Typically more effective when little training data exists Discriminative models Directly model class assignment problem Do not model document generation Effectiveness depends on amount and quality of training data Nave Bayes Generative Process Class 1 Class 2 Generate class according to P(c) Generate document

according to P(d|c) Class 3 Class 2 Nearest Neighbor Classification +

+ + + + +

+ + + +

+ + + +

+ + + + + +

+ + + + + + +

+ + + + +

Feature Selection Document classifiers can have a very large number of features Not all features are useful Excessive features can increase computational cost of training and testing Feature selection methods reduce the number of features by choosing the most useful features Information Gain Information gain is a commonly used feature selection measure based on information theory

It tells how much information is gained if we observe some feature Rank features by information gain and then train model using the top K (K is typically small) The information gain for a Multiple-Bernoulli Nave Bayes classifier is computed as: Classification Applications Classification is widely used to enhance search engines Example applications Spam detection Sentiment classification Semantic classification of advertisements Many others not covered here!

Spam, Spam, Spam Classification is widely used to detect various types of spam There are many types of spam Link spam Adding links to message boards Link exchange networks Link farming Term spam URL term spam Dumping

Phrase stitching Weaving Spam Example Spam Detection Useful features Unigrams Formatting (invisible text, flashing, etc.) Misspellings IP address Different features are useful for different spam detection tasks Email and web page spam are by far the most widely studied, well understood, and easily detected types of spam

Example Spam Assassin Output Sentiment Blogs, online reviews, and forum posts are often opinionated Sentiment classification attempts to automatically identify the polarity of the opinion Negative opinion Neutral opinion Positive opinion Sometimes the strength of the opinion is also important Two stars vs. four stars Weakly negative vs. strongly negative Classifying Sentiment

Useful features Unigrams Bigrams Part of speech tags Adjectives SVMs with unigram features have been shown to be outperform hand built rules Sentiment Classification Example Classifying Online Ads Unlike traditional search, online advertising goes beyond topical relevance A user searching for tropical fish may also be interested in pet stores, local aquariums, or even scuba diving lessons

These are semantically related, but not topically relevant! We can bridge the semantic gap by classifying ads and queries according to a semantic hierarchy Semantic Classification Semantic hierarchy ontology Example: Pets / Aquariums / Supplies Training data Large number of queries and ads are manually classified into the hierarchy Nearest neighbor classification has been shown to be effective for this task Hierarchical structure of classes can be used to improve classification accuracy

Semantic Classification Aquariums Aquariums Fish Fish Rainbow Rainbow Fish Fish Resources Resources Web Page

Supplies Supplies Discount Tropical Fish Food Feed your tropical fish a gourmet diet for just pennies a day! www.cheapfishfood.com Ad Clustering A set of unsupervised algorithms that attempt to find latent structure in a set of items Goal is to identify groups (clusters) of similar items Suppose I gave you the shape, color, vitamin C content, and price of various fruits and asked you to cluster them

What criteria would you use? How would you define similarity? Clustering is very sensitive to how items are represented and how similarity is defined! Clustering General outline of clustering algorithms 1. Decide how items will be represented (e.g., feature vectors) 2. Define similarity measure between pairs or groups of items (e.g., cosine similarity) 3. Determine what makes a good clustering 4. Iteratively construct clusters that are increasingly good 5. Stop after a local/global optimum clustering is found

Steps 3 and 4 differ the most across algorithms Hierarchical Clustering Constructs a hierarchy of clusters The top level of the hierarchy consists of a single cluster with all items in it The bottom level of the hierarchy consists of N (# items) singleton clusters Two types of hierarchical clustering Divisive (top down) Agglomerative (bottom up) Hierarchy can be visualized as a dendogram Example Dendrogram M

L K J I H A B C D E F

G Divisive and Agglomerative Hierarchical Clustering Divisive Start with a single cluster consisting of all of the items Until only singleton clusters exist Divide an existing cluster into two new clusters Agglomerative Start with N (# items) singleton clusters Until a single cluster exists Combine two existing cluster into a new cluster How do we know how to divide or combined clusters? Define a division or combination cost Perform the division or combination with the lowest cost

Divisive Hierarchical Clustering A A D E D G B G B

F C E F C A A D E

D G B C F E G B C F

Agglomerative Hierarchical Clustering A A D E D G B G B

F C E F C A A D E

D G B C F E G B C F

Clustering Costs Single linkage Complete linkage Average linkage Average group linkage Clustering Strategies Single Linkage A D E

D G B C Average Linkage D E C F

F Average Group Linkage G B E G B

F C A Complete Linkage A Agglomerative Clustering Algorithm K-Means Clustering

Hierarchical clustering constructs a hierarchy of clusters K-means always maintains exactly K clusters Clusters represented as centroids (center of mass) Basic algorithm: Step 0: Choose K cluster centroids Step 1: Assign points to closet centroid Step 2: Recompute cluster centroids Step 3: Goto 1 Tends to converge quickly Can be sensitive to choice of initial centroids Must choose K! K-Means Clustering Algorithm K-Nearest Neighbor Clustering

Hierarchical and K-Means clustering partition items into clusters Every item is in exactly one cluster K-Nearest neighbor clustering forms one cluster per item The cluster for item j consists of j and js K nearest neighbors Clusters now overlap 5-Nearest Neighbor Clustering A A A A A

A D D D C C C C B B B B B B C

C D D D Evaluating Clustering Evaluating clustering is challenging, since it is an unsupervised learning task If labels exist, can use standard IR metrics, such as precision and recall If not, then can use measures such as cluster precision, which is defined as: Another option is to evaluate clustering as part of an end-to-end system

How to Choose K? K-means and K-nearest neighbor clustering require us to choose K, the number of clusters No theoretically appealing way of choosing K Depends on the application and data Can use hierarchical clustering and choose the best level of the hierarchy to use Can use adaptive K for K-nearest neighbor clustering Define a ball around each item Difficult problem with no clear solution Adaptive Nearest Neighbor Clustering A A

D B B B B B B C C C C C Clustering and Search Cluster hypothesis Closely associated documents tend to be relevant to the

same requests van Rijsbergen 79 Tends to hold in practice, but not always Two retrieval modeling options Retrieve clusters according to P(Q | Ci) Smooth documents using K-NN clusters: Smoothing approach more effective Testing the Cluster Hypothesis