Datafication, Exploratory Data Analysis, and Classification Models Study Guide

Course Overview, Logistics, and Learning Environment

• Course Schedule and Structure:

  • Lectures run for $2$ hours per week with a $10$-minute mid-lecture break (typically taken around $5$ to $12$ minutes prior to the second hour).

  • Module $1$ encompasses the first $4$ weeks of the semester and covers all content evaluated on the midterm test.

  • Lecture design focuses on active, real-time student engagement rather than passive slide reading.

• Engagement and Stats Apps:

  • Accessible on Canvas via the left-hand navigation menu under "Stats Apps" (toggle the menu display if hidden on mobile or small screens).

  • Allows anonymous live student input while attributing entries to Canvas profiles to record engagement marks.

  • Grace Period: Prep work contributions submitted up until Monday morning at $8:00\text{ AM}$ receive engagement credit, though completion prior to lectures is strongly recommended.

• Communication Protocols and Staff Access:

  • Instructor Office: Room $320$, Building $303$.

  • Help Center / Drop-in Location: Room $170$, Building $302$ (fully refurbished).

  • Drop-in Hours: Available for $1$ hour directly following lectures, Fridays all afternoon, and Mondays from $12:00\text{ PM}$ to $4:00\text{ PM}$.

  • Email Policy: Direct email is mandatory for contacting staff (addresses listed on the "Meet the Team" Canvas page). Canvas direct messaging must not be used due to lost messages.

  • Ed Discussion: Online platform on Canvas for course-related queries. Existing threads must be searched before creating new posts.

  • Classroom Conduct: In-class chatter is prohibited due to hall acoustics that carry background voice audio across the room.

• Course Assessments:

  • Quizzes: Located under the "Quizzes" tab on Canvas. Unlimited submission attempts are permitted prior to the deadline. Due every Friday at end of day (late penalties apply thereafter).

  • Weekly Tasks: Located in the course book under each chapter. Due every Monday at $10:00\text{ PM}$. Tasks utilize unique individualized datasets; final responses must be downloaded as an HTML document and uploaded to Canvas under "Assignments" -> "Task 1".

  • Open Resource Policy: Prep work course notes and personal study guides may be brought directly into tests and examinations.

• Class Representative Application Requirements:

  • Duties: Attend $2$ mandatory meeting times during the semester and gather student feedback.

  • Application Format: Send an email to the instructor containing exactly $2$ sentences ($1$ sentence detailing motivation for the role, $1$ sentence describing field of study and background).

  • Deadline: End of business on Friday.

• Technical Requirements and Course Tools:

  • Recommended Browser: Google Chrome is strongly advised (especially on macOS/Apple devices) for optimal compatibility with course web apps and scripts.

  • Insight Lite: Web application located under "Tools" in the course book, used for statistical data visualization, data import, and numerical analysis.

  • Panopto Video: Canvas folder housing all recorded lecture videos.

  • Software Guides: Available in the course book covering functions for Google Sheets and Insight Lite.

Datafication and Variables in Statistics

• Concept of Datafication:

  • Datafication is the algorithmic process of transforming raw unstructured data or existing variables into structured quantitative or categorical variables using formulas or code.

  • Examples include converting raw unstructured text strings into explicit character counts, word counts, or logical indicators.

• Sentiment Analysis:

  • Sentiment analysis converts qualitative text into measurable categories reflecting emotional tone (e.g., positive, negative, neutral).

  • Practical Applications: Used in social media monitoring, such as tracking consumer sentiment on Twitter following product releases (e.g., evaluating reactions to a novel licorice chocolate mango ice cream flavor).

  • Dictionary-Based Sentiment Analysis: Uses pre-defined word libraries to classify text based on target vocabulary (e.g., mapping "nervous" to negative sentiment or "excited" to positive sentiment).

• Structure of a Rectangular Data Set:

  • A rectangular data set is organized such that every column represents a variable and every row (excluding header row $1$) represents an individual case or entity.

  • Variable Types in Data Collection Spreadsheets:

  • timestamp: Date and time of submission.

  • writing ID: Unique cryptographic hash assigned to the entry.

  • device type: Categorical variable indicating hardware type (e.g., desktop/laptop vs. mobile).

  • screen width: Numeric variable recording screen resolution in pixels.

  • comment: Raw unstructured text string.

  • time taken in seconds: Continuous numeric variable representing completion duration.

  • user ID: Unique Canvas profile identifier mapped for engagement credit.

• Classification of Variables:

  • Categorical Variable: Takes discrete non-numeric values, labels, or factor levels (e.g., device type, sentiment category, prior statistics experience level).

  • Numeric Variable: Takes continuous or discrete numerical values representing direct counts or measurements (e.g., screen width in pixels, completion duration in seconds, character count).

Google Sheets Operations and Data Tidying

• Crowdsourcing Data Cleansing:

  • Datasets containing raw text inputs require manual crowdsourced auditing to systematically convert free-form comments into standardized categorical data.

  • A random number generator selects specific row indices between row $2$ and the maximum dataset row (e.g., row $457$) to distribute auditing across reviewers.

• Google Sheets Functions and Formulas:

  • Column Insertion: Right-click the header of an adjacent column (e.g., column H) and select "Insert 1 column to the left".

  • Character Length Calculation:

  • Function syntax: $=\text{LEN}(\text{cell\_reference})$

  • Example: $=\text{LEN}(G127)$ calculates the total number of characters contained in cell $G127$.

  • Duplicate Removal:

  • Path: Data -> Data cleanup -> Remove duplicates.

  • Select target unique identifier columns (e.g., submission ID or comment row) to eliminate double submissions.

• Publishing Spreadsheets for Insight Lite Integration:

  • Path: File -> Share -> Publish to web.

  • Sheet Selection: Select the specific named worksheet tab.

  • File Format: Set output drop-down to "Comma-separated values (.csv)".

  • Link Verification: Ensure the generated URL ends explicitly with the .csv extension to guarantee readability by Insight Lite.

Exploratory Data Analysis Using Insight Lite

• Data Import Procedure in Insight Lite:

  • Open Insight Lite via course book tools.

  • Path: File -> Import dataset -> Paste published .csv URL -> Click Import -> Click Confirm.

  • Select Visualize tab to initialize graphical exploration.

• Variable Classifications in Bivariate Analysis:

  • Response Variable: The primary outcome variable being analyzed or predicted. Selected as Variable $1$ in Insight Lite and plotted on the primary vertical axis.

  • Explanatory Variable: The secondary variable used to explain variations or structural differences in the response variable. Selected as Variable $2$ in Insight Lite.

  • Terminology Standard: Replaces legacy scientific terms such as "dependent variable" (response variable) and "independent variable" (explanatory variable).

• Analysis of Student Course Sentiment:

  • Overall Sentiment Distribution:

  • Positive Sentiment: Represented the statistical majority at $52.5\%$.

  • Neutral and Negative Sentiment: Accounted for the remaining $47.5\%$ combined.

  • Sentiment Segmented by Prior Statistics Background:

  • Year $13$ Statistics Background: $65.55\%$ of students with Year $13$ high school statistics experience expressed a positive sentiment toward the course.

  • Less Prior Statistics Background: Approximately $33\%$ to $35\%$ of students with less prior statistics experience expressed positive sentiment.

  • Neutral Sentiment: Exhibited minimal variation between students with Year $13$ background and those with less background.

  • Evaluative Standards: Statistical evaluations must rely strictly on quantitative empirical data (e.g., percentages, proportions, counts) rather than anecdotal or contextual assumptions.

Classification Models, Baseline Models, and Spam Detection

• Terminology in Text Classification:

  • Spam: Unsolicited, unwanted automated messages or promotional communications.

  • Ham: Legitimate, essential text messages from personal, academic, or professional contacts.

• The Baseline Model Concept:

  • Definition: A baseline model is a naive classification benchmark that operates without incorporating any explanatory variables, predicting the overall majority class for every instance.

  • Baseline Performance: In a dataset where $52\%$ of total messages are ham, predicting "ham" for all incoming messages yields a baseline accuracy of $52\%$.

  • Purpose: Serves as the baseline threshold that any feature-driven predictive model must exceed to demonstrate predictive utility.

• Feature Engineering for Text Classification:

  • Categorical variables engineered from message text must feature clear, descriptive names and precise, unambiguous level definitions:

  • contains_numbers: Levels = contains numbers vs. no numbers.

  • contains_exclamation: Levels = contains exclamation vs. no exclamation.

  • contains_free_or_urgent: Levels = contains free/urgent vs. does not contain.

  • character_length_group: Levels = less than 100 characters vs. at least 100 characters.

  • contains_url: Levels = contains URL vs. no URL.

• Graphical Evaluation of Explanatory Power in Mosaic Plots:

  • Bar Width: Represents the sample size / proportion of observations belonging to that category level of the explanatory variable.

  • Bar Height: Displays the relative proportion of response variable categories (spam vs. ham) within that specific explanatory level.

  • Criterion for Superior Predictors: Optimal explanatory variables produce maximum vertical contrast across levels (e.g., one level showing a high percentage of spam and another showing a high percentage of ham).

  • Comparative Evaluation of Candidates:

  • contains_url: Ineffective predictor; very few text messages contain URLs, yielding minimal overall predictive separation.

  • contains_free_or_urgent: Strong positive identifier ($95\%$ spam when present), but weak negative identifier when absent ($65\%$ ham, $35\%$ spam).

  • contains_numbers: Highly effective dual-level predictor. $77\%$ (up to $82.4\%$ in sub-samples) of messages containing numbers were spam, while $83\%$ (up to $57.4\%$ ham in contrast tests) of messages lacking numbers were ham.

Confusion Matrices and Model Evaluation Metrics

• Structure of a Confusion Matrix:

  • A confusion matrix is a tabular grid displaying cross-tabulated counts of predicted classifications versus actual ground-truth classes.

  • Grid Components:

  • True Spam (Predicted Spam, Actual Spam)

  • False Spam / False Positive (Predicted Spam, Actual Ham)

  • False Ham / False Negative (Predicted Ham, Actual Spam)

  • True Ham (Predicted Ham, Actual Ham)

• Formula for Percentage Classified Correctly (PCC):

  • PCC measures total classification accuracy of a model as a percentage:     $\text{PCC} = \left( \frac{\text{Number of Correct Predictions}}{\text{Total Number of Observations}} \right) \times 100\%$

• Sample Calculation from Class Exercise:

  • Decision Rule: Predict "Spam" if the message contains numbers; predict "Ham" if the message contains no numbers.

  • Evaluated Sample Size: $20$ text messages.

  • Correct Predictions: $16$ out of $20$ messages correctly identified ($9$ True Ham + $7$ True Spam).

  • Computation:     $\text{PCC} = \left( \frac{16}{20} \right) \times 100\% = 80\%$

  • Model Performance Assessment: The feature-driven decision rule model ($\text{PCC} = 80\%$) significantly outperforms the naive baseline model ($\text{PCC} = 52\%$), establishing strong predictive value.