User:Ian Helmke/Sandbox

Machine Learning Additions

This is text I'd like to add to the machine learning page (pending heavy editing). This text would go underneath the "Issues in Training and Evaluation" section.

Overfitting

Overfitting occurs in classifiers when a classifier creates an extremely accurate model for classifying the example data used to crate the model, yet is inaccurate on other examples.

This can sometimes occur because a classifier looks at irrelevant data points in the training data. Classifiers often give more weight to features seen less commonly in the training data, and if data given to the classifier shares similar uncommon values, it groups those values accordingly. As a result, the classifier is extremely accurate when classifying the training data and appears to be useful, but when used to organize data outside of the model, it becomes inaccurate.

Machine learning techniques can prevent overfitting by imposing a penalty upon itself for complicated models. A simpler model is oftentimes more consistent with the trends of the data in question.

Imbalance of Data

Data within training sets can be imbalanced, where the training set has more examples of one category of data than another. Many classifying algorithms assume that the ratio of the different categories of training data that they receive are roughly equivalent to the ratio of real data that will belong in each category.

There may also be characteristics that differentiate members of the minority class that are missed when data is imbalanced. For example, a classifier training on examples of cancer patients may fail to differentiate between different types of cancer in the minority class.

One way to prevent this imbalance of data is to use active example selection. Active example selection builds the classifier model slowly, adding a few pieces of sample data of each class at a time. The model is tested at every stage, and documents which improve the accuracy of the model are kept in the classifier, while ones that do not change the model or make its performance worse are removed. This ensures that only meaningful pieces of data are used to train the classifier, and since the ratio of documents is closer to 1:1, an imbalance of data is less of an issue.

Evaluating Results

There are a number of techniques for evaluating the results of a machine learning algorithm. Some of these techniques are also used in natural language processing. Machine learning techniques are generally evaluated by their results, and some methods (such as neural networks) are considered "black box" forms of classification, since it is not easy to understand how or why the underlying implementation is sorting a particular way.

In some cases, particularly with classifiers, the outcome of the machine learning algorithm is compared to a set of data classified by experts. The machine learning algorithm creates a model based on a set of training data, and classifies a second sample set of data. A group of experts also annotate the second set of data, marking it according to how they believe it should be classified (not according to how they believe a machine would classify it). The training and evaluation data is generally a tiny subset of the data available. The results of the algorithm are compared to the results of the experts and arranged into two scores: precision and recall. Precision accounts for situations where the classifier put something in a category where it did not belong. Recall accounts for situations where the classifier did not put something in a category it should have. If a classifier is able to produce good results for a subset of data, it should also be successful at classifying a larger set of similar data.

Classification and clustering algorithms can also be measured against other algorithms. This is useful when an algorithm is attempting to improve performance (speedwise, for example) while providing similar levels of precision and recall relative to another algorithm. It can also be used to show that an algorithm is an improvement over a previous generation, or to show which algorithm is most useful for organizing data for a particular problem.

Scalability

Modern computers have the ability to multitask exceptionally well through the use of multiple cores or CPUs. Machine learning algorithms are often used to process large quantities of data, and optimizing them for multiple CPUs greatly improves their performance.

Biclustering

The following material would go on a new page of the above title (probably linked to the ML page)

Biclustering is an unsupervised machine learning method which searches for similarities in specific subsections of the input data. Biclustering is unique among machine learning methods because it searches for similarities in small parts of the data, instead of putting a piece of data into a single group.

Biclustering was first discovered in 1970. As of 2011, it is a commonly used technique in bioinformatics, particularly in the area of gene expression, or identifying groups of genes that are similar between different people.

Processing

Clustering algorithms take vectors as input. They sort the vectors according to how similar they are by comparing all of the features in the vector, creating several groups that each contain a bunch of (hopefully) similar vectors. Biclustering looks at all of the vectors as a single input matrix, and attempts to find regions of the input which look similar.

Biclustering is a useful technique for finding trends in data when each vector of data is large because it can spot trends in specific parts of data that clustering cannot. A normal clustering algorithm sorts pieces of data according to features that the majority of them share. Biclustering organizes data according to parts of them that seem similar.

Applications

Biclustering is particularly useful in the medical field, where it can be used, for example, to find genes related to a specific disease in a group of patients. If each vector represents how a person expresses traits, biclustering can be used to determine a set of genes which is associated with cancer. It can even be used to find similarities and differences between different varieties of cancers.