As means of data collection have become more capable, the need for non-linear, multi-variate modeling techniques has become more and more apparent. Data collection streams and the number of meaningful variables are broadening. Traditional data modeling methods simply were not designed to work with one hundred or more variables.

In answer to this, the last decade has seen the emergence of machine learning or artificial intelligence as a means of modeling complex patterns in data. Technologies such as neural networks, genetic algorithms and fuzzy logic have been very effective at finding interrelationships between multiple variables and modeling real-world, non-linear data. However, the business world has been reluctant to accept these methods due to computational intensity, and most of all, the inability to clearly trace and explain results.

With AFFM^TM, data mining may be performed by savvy business users rather than an analytical team. AFFM^TM may also be integrated in such a way that results are translated into domain specific language through a rule-driven knowledge-base. As well, comparing weights between any two or more variables is straight forward for convenient data visualization. This is particularly useful since the weights contain both vector correlation and pattern matching experience.  Some view AFFM^TM as the next significant step in adaptive pattern based modeling.

AFFM^TM ADVANTAGES

RESULTS: AFFM^TM is an innovative, synergistic technology which combines the benefits of several data mining methods, dramatically extending the ability to extract meaningful information from data.

AFFM^TM is not only unique in its ability to detect rare patterns in data, but does so without requiring historical outcomes. Information may be clustered not just based on correlation between variables, but also due to similarities between patterns associated with the variables. Unlike neural networks which generalize and are prone to over-training, AFFM's fuzzy-pattern matching feature provides the unique capability of identifying novel transactions, particularly useful in areas such as fraud and fault detection.

EXPLAINABILITY: AFFM^TM overcomes a major drawback of classic "AI" techniques: the ability to translate its results back into real-world terms.

Although neural networks are very effective non-linear estimators, they have been prone to dismissal as valid modeling tools because of their "black box" nature. AFFM^TM overcomes that limitation since its weights can be de-scaled and explained through the automatic creation of knowledge bases, or visualized by comparing the resulting weights between two or more nodes to reveal the variables that are discriminatory.

SPEED: Unlike most computer-aided pattern discovery algorithms, AFFM^TM requires only a single pass through the training data set!

Genetic algorithms and most neural network paradigms are computationally intensive since they must iterate through training data until the desired level of accuracy is attained. Alternatively, AFFM^TM is a "one pass" method which can be used in real time and near real time systems. Even problems with high numbers of variables train very efficiently with AFFM^TM.

SCALABILITY: The "Adaptive Temporal Correlation Network" (ATCN^TM) technology extends the concepts in AFFM^TM to a large-scale pattern recognition and auto correlation system.

ATCN^TM is composed of a collection of interconnected AFFM^TM modules. ATCN^TM has use in any situation where real or near real time monitoring of a number of broad-banded data sets is required. AFFM's greatest benefit would be recognized by those with the need for globally understanding information drawn from a confluence of seemingly unrelated data streams. Drug traffic detection and network intrusion detection are strong applications for this technology.

FEATURES

• Fast - One pass training
• Gains experience over time
• Weights are understandable in real world terms
• Ideal for use with rule generators such as ID3
• Supports visualization methods directly
• Real time capable
• Vigilant - detects novel patterns
• Rare patterns retain identity

ADDITIONAL INFORMATION
To initiate a discussion regarding the potential role of AFFM^TM in your environment, please extend a call invitation, or send an Email request. 

* AFFM and ATCN are trademarks of American Heuristics Corporation