Here is a summary of five of the main methods utilized to uncover patterns when data mining:
Anomaly detection : in a large data set it is possible to get a picture of what the data tends to look like in a typical case. Statistics can be used to determine if something is notably different from this pattern. For instance, the IRS could model typical tax returns and use anomaly detection to identify specific returns that differ from this for review and audit.
Association learning: This is the type of data mining that drives the Amazon recommendation system. For instance, this might reveal that customers who bought a cocktail shaker and a cocktail recipe book also often buy martini glasses. These types of findings are often used for targeting coupons/deals or advertising. Similarly, this form of data mining (albeit a quite complex version) is behind Netflix movie recommendations.
Cluster detection: one type of pattern recognition that is particularly useful is recognizing distinct clusters or sub-categories within the data. Without data mining, an analyst would have to look at the data and decide on a set of categories which they believe captures the relevant distinctions between apparent groups in the data. This would risk missing important categories. With data mining it is possible to let the data itself determine the groups. This is one of the black-box type of algorithms that are hard to understand. But in a simple example – again with purchasing behavior – we can imagine that the purchasing habits of different hobbyists would look quite different from each other: gardeners, fishermen and model airplane enthusiasts would all be quite distinct. Machine learning algorithms can detect all of the different subgroups within a dataset that differ significantly from each other.
Classification: If an existing structure is already known, data mining can be used to classify new cases into these pre-determined categories. Learning from a large set of pre-classified examples, algorithms can detect persistent systemic differences between items in each group and apply these rules to new classification problems. Spam filters are a great example of this – large sets of emails that have been identified as spam have enabled filters to notice differences in word usage between legitimate and spam messages, and classify incoming messages according to these rules with a high degree of accuracy.
Regression: Data mining can be used to construct predictive models based on many variables. Facebook, for example, might be interested in predicting future engagement for a user based on past behavior. Factors like the amount of personal information shared, number of photos tagged, friend requests initiated or accepted, comments, likes etc. could all be included in such a model. Over time, this model could be honed to include or weight things differently as Facebook compares how the predictions differ from observed behavior. Ultimately these findings could be used to guide design in order to encourage more of the behaviors that seem to lead to increased engagement over time.
Several of these seem similar to methods commonly used by sociologists:
1. Anomaly detection seems like looking for outliers. On one hand, outliers can throw off basic measures of central tendency or dispersion. On the other hand, outliers can help prompt researchers to reassess their models and/or theories to account for the unusual cases.
2. Cluster detection and/or classification appear similar to factor analysis. This involves a statistical analysis of a set of variables to see which ones “hang together.” This can be helpful for finding categories and reducing the number of variables in an analysis to a lesser number of important concepts.
3. Regression is used all the time both for modeling and predictions.
This all reminds me of what I heard in graduate school about the difference between data mining and statistical research: data mining amounted to atheoretical analysis. In other words, you might find relationships between variables (or apparent relationships between variables – could always be a spurious association or there could be suppressor or distorter effects) but you wouldn’t have compelling explanations for these relationships. While you might be able to develop some explanations, this is a different process than hypothesis testing where you set out to look and test for relationships and patterns.