Topic 5D – Data Mining

Concepts

1. The counselor's activity is database querying: she already knows the question ("Who are the juniors below 2.5 GPA?") and retrieves specific records that match a known criterion. The researcher's activity is data mining: she does not start with a specific question but instead searches large historical data for previously unknown patterns — the discovery that a specific multi-year trajectory predicts low junior-year GPA was not known before the analysis. The key difference is whether you are retrieving known information or discovering unknown patterns.
2. Two reasons: (1) Performance — data mining queries are computationally intensive, scanning and analyzing enormous volumes of historical data. Running these queries on a live operational database would slow down or disrupt the normal transactional operations (recording grades, scheduling classes, processing payments) that depend on that database. (2) Data integration and cleaning — a data warehouse consolidates data from multiple operational systems (student information system, attendance system, assessment platform) into a single cleaned, consistent structure optimized for analysis. Operational databases are designed for fast individual transactions, not for complex analytical queries that span years of records across multiple systems.
3. Not necessarily. A correlation between two behaviors does not prove that one causes the other, or that both are driven by any single meaningful underlying factor. Before acting on this finding, you would want to ask: Is the sample large enough for the pattern to be statistically significant? Could a third variable (income level, age, geography) explain both behaviors independently? Is the pattern consistent across time periods and subgroups, or is it a coincidence of this particular dataset? What would acting on the pattern even mean practically? Patterns found by data mining are hypotheses to investigate, not conclusions to act on directly.

The Six Techniques

4. Cluster analysis finds natural groupings in data without any predefined categories — the algorithm divides the data into groups based on similarity, and the groups emerge from the data itself. Class description characterizes a group that already exists and is already labeled — you know which records belong to the group, and you want to know what features they have in common. You cannot use class description when you do not know your groups because class description requires pre-labeled data as input. If no one has identified and labeled the groups yet, there is nothing to describe — cluster analysis must come first to discover whether natural groupings exist.
5. • Books borrowed together: Association analysis. Association analysis finds items that frequently co-occur — "customers who buy X also buy Y" or, in this case, "patrons who borrow X also borrow Y."
   • What distinguishes early college completers from non-completers: Class discrimination. Two predefined groups exist (completed vs. did not complete), and the goal is to find the features that differentiate between them.
   • Four distinct patient profiles, no predetermined groups: Cluster analysis. The hospital has no predefined groupings; the algorithm discovers them from the data itself.
   • Unusual claims compared to typical ones: Outlier analysis. The goal is to identify records that deviate significantly from the norm — the definition of outlier analysis.

Ethics and Implications

6. Bias concern: the historical data used to build the model reflects past inequities. If students from certain demographic groups (by race, income, English-learner status) were historically less likely to graduate due to systemic factors — underfunded schools, fewer support resources — the model will learn to associate those demographic patterns with low graduation likelihood. The risk score may then assign high risk to students from those groups not because of anything specific to those individual students, but because the model has encoded past discrimination. The system could create a self-fulfilling prophecy by concentrating resources on students who fit a historical pattern rather than addressing the underlying structural factors. Correlation limit concern: even if the pattern is statistically valid, a risk score predicts group-level probabilities, not individual outcomes. A student assigned a high risk score may graduate with flying colors; one assigned a low risk score may drop out. Treating the score as a definitive individual prediction — rather than one input among many — oversimplifies what correlation-based models can actually tell us about individuals.
7. Important questions include: What student data is collected, and for how long is it retained? Who has access to the data — does it leave the district's control, and is it shared with the platform vendor? How is the data secured, and what happens in a breach? Does FERPA apply, and does the platform comply? Is the student's individual interaction data used to build models sold or applied to other districts or commercial products? Can parents opt out? What is the district's ability to audit the platform's use of the data? Are students and families informed about what is being collected?