CAP4770: Data Mining

Course Description

CAP4770, Data Mining, teaches the discovery of patterns in large datasets: how to prepare messy real data for analysis, how the principal algorithm families work, when each is appropriate, and — the part students most often underestimate — how to tell whether a discovered pattern is real or an artefact of the search.

The University of West Florida describes a course that "exposes students to data mining concepts and techniques and different data mining software," covering "data pre-processing and cleaning, concept hierarchy generation, attribute relevance analysis, association rule mining, classification algorithms, and cluster analysis." Florida Gulf Coast University, which titles it Knowledge Discovery and Data Mining, covers "principles, concepts, and methods in the fields of data mining and knowledge discovery," with "algorithm development, current tools, and real-world applications," including "data visualization, exploration, clustering, classification, association rule mining, and anomaly detection."

The two descriptions agree closely, and both put pre-processing first — which reflects practice. Practitioners consistently report that most of the work in a data mining project is cleaning, integrating and transforming data rather than running algorithms, and a course that begins with the algorithms teaches a distorted picture of the job.

The term "data mining" predates "data science" and sits inside it. The distinction most instructors draw is that data mining emphasises discovery — finding structure that was not hypothesised in advance — where classical statistics emphasises testing hypotheses formed beforehand. That difference is what creates the course's central methodological caution: search hard enough through enough variables and you will find patterns in pure noise, so validation discipline is not an afterthought but the thing that separates a result from a coincidence.

CAP4770 is offered at approximately 9 Florida institutions and carries 3 credits with roughly 45 contact hours. It is a 4000-level course taken in the junior or senior year, normally as an elective in computer science, software engineering, information technology or data science programmes.

⚠ Two different prerequisite routes — and they signal different courses

• University of West Florida: prerequisite COP4710 — database systems. The course is positioned downstream of databases: you have data at scale, now mine it.
• Florida Gulf Coast University: prerequisite COP3003 — object-oriented programming. The course is positioned downstream of programming, and FGCU's description mentions "algorithm development" explicitly.

Both are legitimate, and the difference predicts emphasis. A database-gated section tends toward data preparation, SQL-scale data handling, warehousing and tool use; a programming-gated section tends toward implementing algorithms and understanding their mechanics. Neither institution lists statistics or linear algebra as a prerequisite, which is worth knowing: students without a statistics background can enrol, and will find the model-evaluation material harder than the programming. Check which route your section takes before assuming what the workload will be.

Learning Outcomes

Required Outcomes

• Explain the knowledge discovery in databases (KDD) process and situate data mining within it, from selection through pre-processing, transformation, mining and interpretation.
• Characterise data: attribute types, similarity and distance measures, and the statistical properties of a dataset relevant to method selection.
• Perform data cleaning: handle missing values, detect and treat outliers, resolve inconsistencies, and document the decisions made.
• Perform data integration and transformation: normalisation and standardisation, discretisation, concept hierarchy generation, aggregation, and feature construction.
• Perform dimensionality reduction and attribute relevance analysis, including feature selection methods and principal component analysis.
• Conduct exploratory data analysis and produce visualisations that reveal structure rather than decorate it.
• Explain and apply association rule mining, including support, confidence and lift, and the Apriori and FP-Growth algorithms.
• Interpret association rules critically, distinguishing statistically frequent co-occurrence from actionable relationship.
• Explain and apply the principal classification algorithms: decision trees, naive Bayes, k-nearest neighbours, logistic regression, support vector machines, and ensemble methods including random forests and boosting.
• Explain the bias-variance trade-off and recognise overfitting and underfitting from training and validation behaviour.
• Evaluate classifiers correctly: train/validation/test splits, cross-validation, confusion matrices, accuracy, precision, recall, F1, ROC curves and AUC.
• Explain why accuracy is misleading on imbalanced data and select an appropriate metric and sampling strategy for a skewed problem.
• Explain and apply clustering: k-means, hierarchical clustering and density-based methods such as DBSCAN; select k and evaluate cluster quality.
• Explain and apply anomaly and outlier detection methods, and explain their relationship to classification on rare classes.
• Use at least one data mining toolset competently — Python with scikit-learn and pandas, R, Weka, KNIME, or RapidMiner.
• Complete an end-to-end project on a real dataset: framing, preparation, modelling, evaluation and communication of findings.
• Explain the ethical and legal dimensions of mining data about people: privacy, consent, re-identification, algorithmic bias and disparate impact.

Optional Outcomes

• Explain data warehousing, OLAP and the multidimensional data model — commonly included where the course is database-gated.
• Apply text mining and natural language processing: tokenisation, TF-IDF, topic modelling, sentiment analysis.
• Apply neural networks and introductory deep learning.
• Mine sequential, time series or stream data.
• Apply graph and network mining, including community detection.
• Build recommender systems using collaborative and content-based filtering.
• Work with distributed processing frameworks such as Spark for datasets exceeding memory.
• Apply model interpretability methods (feature importance, SHAP, LIME) and explain why interpretability matters in regulated domains.
• Apply spatial data mining, or domain-specific mining in health care, finance or security.
• Deploy a model and explain the operational concerns: monitoring, drift and retraining.

Major Topics

Required Topics

• Introduction: what data mining is, the KDD process, the relationship to statistics, machine learning and data science, and the kinds of question it answers
• Data types and characterisation: nominal, ordinal, interval and ratio attributes; similarity and distance measures; data quality
• Data pre-processing: cleaning, missing value strategies, noise and outliers, integration and entity resolution
• Data transformation: normalisation, standardisation, discretisation, binning, concept hierarchy generation, feature engineering
• Dimensionality reduction: feature selection (filter, wrapper, embedded), principal component analysis, the curse of dimensionality
• Exploratory analysis and visualisation for discovery
• Association analysis: frequent itemsets, support, confidence, lift; Apriori; FP-Growth; rule generation and pruning; the multiple-comparisons problem in rule mining
• Classification: the general framework, training and generalisation
• Decision trees: information gain and Gini, splitting, pruning, interpretability
• Bayesian classification: naive Bayes and its independence assumption
• Instance-based learning: k-nearest neighbours and the effect of distance metric and scaling
• Linear classifiers: logistic regression and support vector machines; kernels in outline
• Ensemble methods: bagging, random forests, boosting
• Model evaluation: holdout, cross-validation, confusion matrix, precision and recall, F1, ROC and AUC; class imbalance and resampling
• Overfitting, the bias-variance trade-off, regularisation, and the necessity of a held-out test set
• Cluster analysis: partitioning (k-means, k-medoids), hierarchical (agglomerative, divisive), density-based (DBSCAN); choosing k; internal and external validation
• Anomaly and outlier detection: statistical, distance-based, density-based and isolation approaches
• Data mining tools and software in practice
• Ethics, privacy and bias: re-identification of "anonymised" data, consent, proxy discrimination, disparate impact, and the regulatory landscape
• The end-to-end project: framing a question, preparing data, modelling, validating and communicating

Optional Topics

• Data warehousing, OLAP and multidimensional modelling
• Text mining, TF-IDF, topic modelling and sentiment analysis
• Neural networks and introductory deep learning
• Time series and stream mining; concept drift
• Graph and social network mining
• Recommender systems
• Big data frameworks: Hadoop, Spark, distributed processing
• Model interpretability and explainable AI
• Spatial data mining and GIS integration
• Domain applications: fraud detection, health care analytics, cybersecurity, marketing
• Model deployment, monitoring and MLOps in outline

Resources & Tools

• Introduction to Data Mining (Tan, Steinbach, Karpatne & Kumar, Pearson) is the most widely adopted text for this course in Florida and nationally, and is the reference most syllabi follow chapter by chapter.
• Data Mining: Concepts and Techniques (Han, Kamber & Pei, Morgan Kaufmann) is the other standard — the source of the KDD-process framing and of the concept-hierarchy and attribute-relevance material UWF's description names — and is stronger on the database side.
• Data Mining: Practical Machine Learning Tools and Techniques (Witten, Frank, Hall & Pal) is the companion to Weka and is common where the course is tool-led.
• An Introduction to Statistical Learning (James, Witten, Hastie & Tibshirani) is free online in both R and Python editions and is the best available bridge for students who need the statistical foundation the prerequisites do not supply. Its heavier sibling, The Elements of Statistical Learning, is also free.
• Software: Python with pandas, scikit-learn, NumPy, Matplotlib and Seaborn is now the dominant teaching stack; R with tidyverse and caret remains common; Weka, KNIME and RapidMiner are used where the course emphasises tools over code; Jupyter notebooks and Google Colab are the usual working environments, and Colab is free and needs no local setup.
• Datasets for projects: the UCI Machine Learning Repository, Kaggle, data.gov, and Google Dataset Search. Florida-specific open data makes for stronger projects than generic Kaggle sets: the Florida Department of Law Enforcement crime data, AHCA hospital data via FloridaHealthFinder.gov, Florida Department of Education school data, the Florida Department of Transportation crash data, and county property appraiser records — all public under Florida's records law.
• Professional and ethical context: the ACM Code of Ethics, the ACM SIGKDD community, and the NIST AI Risk Management Framework for the governance material.

Career Pathways

• Data Scientist — SOC 15-2051. The headline destination. Entry-level data science roles are competitive and often expect a graduate degree, but the analyst roles below are genuinely reachable from a bachelor's with a strong project portfolio.
• Data Analyst and Business Intelligence Analyst — SOC 15-2051 and 13-1161. The most accessible entry point, and a common route into data science after two or three years.
• Machine Learning Engineer — SOC 15-1252 and 15-2051. Weighted toward software engineering skill as much as modelling.
• Data Engineer — SOC 15-1243 and 15-1252. ⚠ Worth knowing: data engineering is in higher demand and less crowded than data science, and the pre-processing and pipeline material in this course points directly at it.
• Database Administrator and Architect — SOC 15-1242 and 15-1243.
• Fraud and Risk Analyst — SOC 13-2054 and 13-2011. Anomaly detection is the direct application; Florida's insurance, banking and health care sectors employ heavily.
• Marketing Analyst and Customer Insights Analyst — SOC 13-1161. Association rules and segmentation are the working methods.
• Health Care Data Analyst / Clinical Informatics Analyst — SOC 15-2051 and 15-1211; see also HSA4191.
• Cybersecurity Analyst — SOC 15-1212. Intrusion and anomaly detection use these techniques directly.
• Bioinformatics and Research Data Analyst — SOC 15-2051 and 19-1029; see also BSC4434C.
• Florida employers of note: the financial services concentration in Tampa, Jacksonville, Miami and West Palm Beach (Citi, JPMorgan Chase, Fidelity, Raymond James, Truist); the insurance sector, which in Florida is unusually large and unusually analytics-driven because of catastrophe modelling; the health systems and payers (AdventHealth, Orlando Health, BayCare, Florida Blue, Humana); the theme park and hospitality operators, which run substantial demand-forecasting and pricing analytics; the cruise lines in Miami; the aerospace and defence contractors on the Space Coast; the simulation and training cluster in Orlando; Publix and the retail and logistics sector; and state agencies including FDLE, AHCA and the Department of Education.

Special Information

Position in the curriculum

CAP4770 is a junior- or senior-year elective. It follows the programming sequence and, at institutions that gate it on databases, the database course. It pairs naturally with machine learning, artificial intelligence, database systems and statistics courses, and it is frequently the course where a computer science student decides whether to pursue data science. At institutions with a data science or analytics concentration it may be a required core course rather than an elective.

⚠ The statistics gap — the thing to know before enrolling

Neither UWF nor FGCU lists statistics or linear algebra as a prerequisite, and that is typical for this course. It means students can and do enrol without a statistics background, and it produces a predictable pattern: the programming is manageable and the model evaluation is not. Understanding why cross-validation is necessary, what a p-value would and would not tell you here, why accuracy is the wrong metric on imbalanced data, and what overfitting actually is — all of this rests on statistical intuition that the prerequisites do not build.

The practical advice: take a statistics course before or alongside this one if you can, and if you cannot, work through An Introduction to Statistical Learning, which is free and written for exactly this gap. Students who skip this find the course superficially easy — the libraries do the work in three lines — and then cannot tell a good model from a bad one, which is the entire professional skill.

Course format and workload

Three credits, approximately 45 contact hours. Assessment is project- and assignment-driven: programming assignments implementing or applying algorithms, a term project on a real dataset, and examinations covering algorithm mechanics and evaluation. Expect eight to twelve hours a week outside class, unevenly distributed — data cleaning consumes more time than anything else and always more than planned.

Keep the project. A well-documented end-to-end project on a real dataset, with the preparation decisions explained and the evaluation done honestly, is the single most useful artefact this course produces. Hiring in data roles is portfolio-driven, and a GitHub repository with a clear README does more than the transcript line.

Transfer and articulation

CAP4770 is a 4000-level SCNS course: the number is recognised statewide, but upper-division credit is not covered by the A.A. transfer guarantee and applicability inside the major is the receiving department's decision. The course is not available before transfer from a Florida College System A.A.; the lower-division path is the programming sequence (COP2220/COP3014 and the data structures course), calculus and physics or science as required by the computing programme. The specific caution here is the prerequisite-route difference — a database-gated section and a programming-gated section cover overlapping but not identical ground, which does not appear on a transcript. Carry a syllabus if a receiving programme has a specific expectation.

Course-code variations across Florida

The CAP prefix is computer applications and covers the applied artificial intelligence and analytics area. Relevant numbers: CAP4770 (this course, as Data Mining or Knowledge Discovery and Data Mining), CAP4771-range variants, CAP4774 (Databases for Data Science at UWF — note it sits in a different department there, Cybersecurity and Information Technology, not Computer Science), CAP4630 and CAP4611-range artificial intelligence, CAP4641-range natural language processing, and CAP5xxx graduate versions. Adjacent prefixes: COP for programming and databases (COP3003 object-oriented programming, COP4710 database systems — the two prerequisite routes), CDA for computer architecture, CIS for information systems and cybersecurity, STA for statistics, and ISM for business analytics taught in the business school. ⚠ An ISM-prefixed business analytics course is not equivalent to CAP4770 for a computing major — it is tool-oriented rather than algorithm-oriented — and the substitution generally does not run in either direction.

AI Integration

This course is about machine learning, so the question is not whether AI belongs in it but how to be honest about what the tools have changed.

What has genuinely changed. Writing modelling code is no longer a barrier. A language model will produce a working scikit-learn pipeline — load, split, fit, evaluate — faster than a student can type it, and it will usually be correct. AutoML systems will search model families and hyperparameters automatically. This is real, and pretending otherwise does students no favours.

What has not changed, and is now more of the job than before. Four things, and they are precisely the course's content. Problem framing: deciding what to predict, what the unit of analysis is, and whether the available data can answer the question at all. Data understanding: knowing what a field actually means, which values are placeholders, where the collection process introduced bias, and whether a feature leaks the target. Evaluation judgement: choosing a metric that reflects the real cost of each error type, constructing a validation split that respects time or grouping structure, and recognising when a good score is an artefact. Consequence assessment: what happens when the model is wrong about a person.

The failure mode to internalise. A model can achieve an excellent validation score and be completely useless or actively harmful. Target leakage — a feature that encodes the outcome — produces near-perfect scores and collapses in production; it is the single most common serious error in student projects and in professional work. A validation split that ignores time lets a model see the future. Imbalanced data makes 99% accuracy achievable by predicting the majority class always. AI assistants do not catch any of these, because catching them requires knowing what the data means. That is the reason to learn the material rather than to generate it.

Where AI helps a student here. Explaining an algorithm a second way, debugging code, suggesting approaches, and drafting documentation. Used well it removes the syntax barrier and leaves more time for the parts that matter.

Where AI fails. Models produce code that runs and is subtly wrong — fitting the scaler on the full dataset before splitting, evaluating on training data, using the wrong cross-validation scheme for grouped or temporal data. They hallucinate library functions and parameters. And they will confidently interpret a result without knowing whether the data supports the interpretation.

Ethics, which is course content and not a footnote. Mining data about people carries obligations the code does not enforce. "Anonymised" datasets are frequently re-identifiable from a handful of attributes. Models trained on historical decisions reproduce the biases in those decisions — the documented case of a health system algorithm that used prior cost as a proxy for need, and consequently under-referred Black patients at equal illness severity, is the standard teaching example and is a data-mining error, not merely a policy one. Proxy variables reintroduce protected characteristics that were deliberately excluded. And in regulated domains — credit, employment, housing, health care, criminal justice — disparate impact is a legal exposure regardless of intent. The professional obligation is to ask what the model is for, who it affects, and what happens when it is wrong, before deploying it.

Academic integrity. Instructor policies vary widely in computing courses and are often permissive on code while strict on analysis and interpretation, which are the assessed skills. Read the syllabus. And note the self-interested reason to do the work: an interviewer will ask you to explain your project's evaluation choices, and a portfolio you cannot defend is worse than none.