There is a ton of buzz around data mining and, as always, many new names being injected into the same topics showing a lack of study. While the buzz exists, knowing when to deploy an algorithm can be tricky. Based on a deep dive into the subject recently and a dire need to program these algorithms in Java, I present a brief overview of tools with benchmarks and examples to hopefully follow later. Basic concepts are presented first followed by some algorithms that I have not really honed yet (benchmarks are not feasible for now but will be incredibly soon).
To read this thoroughly, I highly recommend following the links. This is a starting point and is also a summary of what I have found to now. Basically, the goal is to help avoid the hours it takes to find information on each and do some stumbling by reading this document.
A word of caution, effective matrix based solutions use a large amount of data. Other fuzzy algorithms exist for discovering relate-ability between small sets of items. For strings, there is distance matching such as Jaro-Winkler or Levenshtein with rule based comparisons and lookup tables to minimize error (say between Adama and Obama). Statistics can enhance this process if there is a need to take the best rating. Train a model for discovering whether the hypothesis that two entities distances makes them the same as opposed to the null hypothesis that it does not after filtering out some common issues.
Linear algebra is a critical foundation of text mining. Different elements are thought of as equations When we have different documents or images, each document or image is often considered to form an equation. This equation can then be presented in a matrix, a simple and easy way to get rid of thinking in really complex terms.
You may have heard of the vaulted differential equation. If not, some reading is in order from a site I used in college when there was not time for the book. The basis of a large portion of differential equations can be written in a matrix. This is important due to the eigen vector and eigen value. To be sure these concepts are crucial for solving matrices to find equations that explain a set of models. Drexel Universities eigenfaces tutorial provides a fairly solid understanding of the way a matrix is used in most text mining. However, similarity ratings are used to compare documents rather than a co-variance matrix for most tasks.
The end result of studying these methods is the ability to look under the hood at today’s hottest text mining technologies.
The Vector Space and TF-IDF
Understanding vectors and vector operations is another crucial step to understanding the mining process. A basic vector is a set of points representing a position in these planes . Vectors can be added, subtracted,multiplied, and, most importantly, stuffed in a matrix where their points can be used with basic linear algebra to find relationships.
Vectors have magnitude and distance. The angles and distances can be compared. Note that, while data loss may be present in finding the right magnitude and distance, the units used should be the same (it would be a terrible idea to think in terms of say millisecond-meter-documents-ice cream cones), it provides a sane basis for considering data. It is up to the miner to choose the most representative points for use in the vector.
In text mining, term frequency-inverse document frequency rating is used in many commercial algorithms including search engines. If the name is not enough, it is basically a frequency ratio based on the ratio of individual document . It works best on more than one document and an offset of 0.5 for term frequency helps offset the effect of large documents slightly by bumping up the rating. Inverse document frequency utilizes a logarithm function to ensure that the rating remains between 0 and 1.
Multiply the following equations together to find the result as described by Wikepedia:
No matter what you do, similarity ratings are the key to making the process work. There are a several that can be used. If the data can be represented fairly well, co-variance is an option . However, text data is not that well suited to using co-variance. This is due to varying styles that represent the same object and, most importantly, issues with quantization. Natural language is naturally fuzzy. Therefore, cosines usually offers a much better solution.
This cosines equation takes the product of two vectors or the sum of two vectors and divides by the product of there normalized vectors or the sum of their normalized vectors. It follows from vector algebra. The result is an angle representing the ‘degree’ of similarity. This can be used for comparison.
Word Net, Disambiguation, and Stemming
The process of disambiguation and stemming are crucial to text mining. There are many sentence processing methods as NLTK shows. At their core is WordNet and other dictionaries. WordNet is a freely available graph of an english dictionary. Most tools work with WordNet for finding root words, disambiguation, and cleaning.
Part of Speech or POS tagging is involved in both disambiguation or stemming. Maximum entropy models are used to discover a part of speech based on common usage.
Disambiguation attempts to resolve words with multiple meanings to their most probable meaning. The worst algorithm is original Lesk but involves only the use of WordNet. Accuracy hovers around 50 percent. Simplified Lesk achieves better results. Lesk finds overlapping words and frequencies to determine the best synonym to replace an ambiguous word. The better algorithms try to use clustering bayes for word sense discovery. Cosines may be used to improve Lesk as well.
Stemming reduces words to their roots. Most WordNet tools use existing classifications with POS tagging to achieve this result.
A Note on Regression Models
Lets be certain, prediction is not well suited to categorization. Changes in word choice across a large number of documents and decisions on importance do not always mean the same thing. Therefore, regression models tend to work poorly. The data is not likely continuous as well. Think of writing like a magnetic field with eddy currents. Predicting the effect due to an encounter with these currents is really, really difficult. Basically, run into an eddy current and you are going to have a really, really bad day. That is not to say that an equation can be created that fits most of the data with respect to location of a point, basically a differential equation. It will likely not be generic and be incredibly difficult to find.
Regression works well on continuous and more natural events.
Classification Tree and a Random Forest
Another often poor performer in categorization of text data is the classification tree. They are as good as the number of rules you are willing to create. However, they may be combined with multinomial Bayes for writing that is uniform and professional (say a legal document) to achieve some success. They are particularly useful after filtering data using LSA/HDP or Multinomial Bayes with decisions that work like a bayesian model when thinking about the bigger picture.
Basically, a classification tree uses probabilities within groupings to ascertain an outcome and moving down to the appropriate left or right child node based on a yes or no response to the question do you belong?
This process works well with defined data when there is a good degree of knowledge about a subject (say gene mapping) but text mining often uses fuzzy data with multiple possible meanings and disambiguation is not entirely accurate, Lesks original algorithm only acheived 50 percent accuracy and an LSA model hovers between 80-90%. Improving the quality can be done with multiple trees or possibly by training off of an extremely large set using cosines instead of raw frequency.
There are multiple methods for tree creation, two are random forests and bagging. Bagging takes multiple trees and averages the probabilities for decisions, using this average for their respective nodes. Random forests find random subsets of features, find probabilities based on them, and select the stronger predictor for a node. The latter approach is best with a much larger set of known features. The number of forests is usually the square root of the total number of features.
Again, the features must be known and fairly homogeneous. Text data is often not.
Multinomial Bayesian Classifier
Multinomial Bayesian Classification is a method that classifies data based on the frequency of words in different categories and their probabilities of occurrence. It is fairly straightforward, find the frequencies or train a set of frequencies on a word by word or gram by gram (a gram being an n-pairing and thus n-gram of words), find probabilities by sentence, take the best one.
MNB works well when writing differs starkly, say with subject matters that differ greatly. It is good for tasks such as separating spam from policies and code in html data when large amounts of training data are present.
Clustering with LSA or HDP
Clustering works well when something is known about the data but categorization is not well done manually. Most algorithms avoid affinity propagation which usually uses the square root of total inputs as the number of clusters any way. Matrices are used heavily here as eigen values and eigen vectors derive an equation that can be used to find relate-ability between documents.
LSA uses raw frequencies or more effectively cosines in the same manner as eigen faces to compare vectors. The end result, however, is an equation representing a category. By matrix inversion and multiplication, all elements are compared. In this case each ij entry in the matrix is a cosine or frequency of a word in a document. HDP (hierarchical direchlet process) is similar but attempts to learn more about the results and improve on the process. It takes much longer than LSA and is experiemental.
If trying to discover new information about text or trying to find the best fit of a number of categories, these methods are useful.
Maximum Entropy Models
Maximum entropy models work well on heterogenous data in a manner similar to Bayes. Gensims sentence tagger classifies sentences from non-sentences in this way. The models find entropy using the maxent principle which uses frequencies and likelihood of occurrence to find outcomes. It works quite well with the correct training sets.
If conditional independence is not assumable and nothing is known about a set, this model is useful. Categories should be known beforehand.