What are Associations in Machine Learning?
Associations are the specific measurable constraints on interestingness used in association rule learning. Regardless of the rules being employed to classify new data, the associations need to be defined by constraints to determine what is both interesting and relevant. While many variables could be selected as an association constraint, the most commonly used are:
Support – How frequently the pattern/items occur in the dataset.
Confidence – How often the rule being used has been true (conditional probability).
Lift – Actual success rate of the target model (rule) over the expected success from random chance.
Conviction – Actual incorrect predication rate over the expected failure rate from random chance.
How Does Association Learning Work?
Along with cluster analysis and anomaly detection, association rules are the most widely used unsupervised learning techniques. Association learning is a machine learning and data mining technique that creates rules for finding interesting relations between variables. Unlike conventional association algorithms measuring degrees of similarity, association rule learning identifies hidden correlations in databases by applying some measure of interestingness to generate an association rule for new searches.
Association rule algorithms count the frequency of complimentary occurrences, or associations, across a large collection of items or actions. The goal is to find associations that take place together far more often than you would find in a random sampling of possibilities. This rule-based approach is a fast and powerful tool for mining categorized, non-numeric databases.
A classic example of this system in practice is analyzing retail sales to find the best way to place items in a store. In a store with a million transactions a year, 10,000 sales might include newborn baby diapers and 100,000 include razor blades. At first glance, newborn diapers and razors seem statistically independent, with no apparent correlation. But rule mining would dig deeper into the transaction frequency and find out that 5,000 sales include both items.
So instead of simply learning that 1% of shoppers buy diapers and 10% buy razor blades, the association system generates a new rule that 50% of all shoppers purchasing newborn diapers will also buy razor blades. A much more useful bit of information for retailers.
Just as important, the rule-based approach enhances performance and generates new rules as it analyzes more data. With a large enough dataset, this allows the machine to mimic the human brain’s feature extraction and abstract association capabilities from raw data.The same basic technique has countless other applications as well