Untitled Quiz

Final Online Test Details

• Date of test: During Week 12 Tutorial Sessions
• Group 1: Tuesday (November 26th) at A312, 8 am - 10 am
• Group 2: Thursday (November 28th) at B219, 1 pm - 3 pm
• Group 3: Friday (November 29th) at A312, 10 am - 12 pm
• Group 4: Wednesday (November 27th) at A312, 8 am - 10 am
• Group 5: Friday (November 29th) at B219, 2 pm - 4 pm
• Mobile phones and ChatGPT prohibited during the test
• Test must be taken in person on campus

Assessment Details

• Close-book test
• 50 questions
• Total points: 120
• 40 questions x 2 points = 80 points
• 10 questions x 4 points = 40 points
• Question formats:
  • Multiple choice (one correct answer)
  • Multiple choice (up to two correct answers)
  • Matching questions

Big Data Ecosystem Components

• Data Devices: Cell phone, GPS, MP3, eBook reader, video player, cable box, ATM, credit card reader, RFID
• Data Collectors: Law enforcement, government, insurance companies, individual medical information brokers, advertising, marketers, employers
• Data Users/Buyers: Media archives, credit bureaus, financial institutions, banks, delivery services, websites, private investigators
• Data Aggregators: Websites, data aggregators, etc

Data Devices

• Gather data from multiple locations
• Continuously generate new data about subject data
• For each gigabyte of data created, a petabyte of data is also generated about the subject data

Data Collectors

• Entities that collect data from devices and users
• Example: Cable TV provider tracks:
  • Shows watched
  • Channels subscribed to/not willing to pay for
  • Prices for premium TV content

Data Aggregators

• Entities that compile and make sense of data collected by collectors
• Companies that transform and package data to sell

Data Users and Buyers

• Direct beneficiaries of the data collected and aggregated
• Example: Corporate customers, analytical services, media archives, advertising companies, information brokers, credit bureaus, catalog co-ops

Four V's of Big Data

• Scale (volume)
• Distribution
• Diversity (variety)
• Timeliness (velocity)
• Accuracy (veracity)

Data Science vs Enterprise Data Warehouse

• Data Warehouse (DW) is a relational database designed for querying and analysis rather than for transaction processing.
• Data warehouse contains cleaned, selective historical data.
• Includes ETL (Extraction, Transformation, and Loading), OLAP (Online Analytical Processing) processes
• Data Science processes deal with diverse data sets (4 Vs of big data) and often need different architectures and analytics.

Analytic Sandbox (Workspaces)

• Resolves conflicts between analysts' needs and traditional enterprise data warehouses.
• Stores data from various sources and technologies
• Enables flexible, high-performance analysis in non-production environments
• Reduces costs and risks of data replication to "shadow" file systems
• "Analyst-owned" rather than "DBA-owned"

Data Science vs Business Intelligence

• Data Science is exploratory and predictive, focusing on past and future trends and scenarios using various types of data.
• Business Intelligence is focused on historical and current data to present trends, performance, and issues via reports.

Big Data Data Structures

• Unstructured: Data with no predefined format (e.g. text documents, images)
• Quasi-structured: Data with inconsistent formats that can be structured (e.g., clickstream data)
• Semi-structured: Data with a defined pattern or format that can be parsed (e.g., spreadsheets, XML)
• Structured: Data with defined formats, models, and structures (e.g., databases)

Data Scientist Definition (Academic)

• A scientist trained in diverse fields from social science to biology.
• Works with large amounts of data
• Addresses computational issues of data structure, size, and messiness
• Solves real-world problems simultaneously.

Data Scientist Definition (Industry)

• Someone who has the capacity to extract meaning and interpret data via using tools and methods in statistics and machine learning as well as being human.

Data Types

• Text data: Limited symbols, usually encoded with ASCII, Unicode, or other standards.
• Audio data: Amplitude corresponds to volume, frequency corresponds to pitch
• Image data: Pixels are basic representation unit.
• 3-D Image data: Voxels instead of pixels to indicate points in a 3D space
• Video/Streaming data: Image frames displayed in a timeline of events
• Trajectory data: Collected by GPS, including geo-location and timestamp

Data Analytics Lifecycle

• Discovery: Evaluating resources, framing the analytics problem, identifying stakeholders, and determining initial hypotheses.
• Data Prep: Preparing the analytic sandbox by extracting and cleaning the relevant data
• Model Planning: Determining the best methods, techniques, and workflow for the next modeling phase
• Model Building: Creating the model, using data sets for training and testing
• Communicating Results: Presenting findings and determining if the project achieved intended goals.
• Operationalizing Results: Implementing models in a production environment.

Key Activities in Phase 1 (Discovery).

• Learns the business domain and assesses the resources needed for the project (people, technology, time, and data)
• Formulating initial hypotheses that are testable against the data.
• Determining the key stakeholders (those who benefit or are affected by the project).
• Articulating the key stakeholders' pain points.
• Interviewing the analytics sponsor

Key Activities in Phase 2 (Data Preparation).

• Preparing the analytics sandbox
• Performing ETL/ETLT on large datasets
• Gathering insights about the data's characteristics
• Building a dataset inventory
• Performing data conditioning (cleaning, normalizing, transforming data)

Data Discrepancies

• Poorly designed forms, human errors, deliberate errors (e.g. not providing information), data decay (outdated info), system errors, data integration issues causing attribute name inconsistencies.
• Detection strategies: examining metadata, using rules regarding uniqueness, consecutiveness or null values and employing commercial data scrubbing tools.

Data Reduction Strategies

• Data cube aggregation
• Attribute subset selection (removing irrelevant, weakly relevant, or redundant attributes)
• Dimensionality reduction (reducing data set size using encoding schemes)
• Numerosity reduction (replacing the data or estimating it with smaller representations)
• Discretization and concept hierarchy generation (replacing raw attribute values with ranges or high-level concepts)

Data Transformation Strategies

• Data smoothing (removing noise)
• Attribute/feature construction (creating new attributes)
• Aggregation (building data cubes)
• Normalization (scaling attributes to fall within a specified range)
• Discretization (replacing values with numerical intervals or conceptual labels)
• Concept hierarchy generation (generalizing attributes into higher-level categories)

Data Normalization Methods

• Min-Max: Transforming data to a specific range (e.g., 0 to 1)
• Z-score: Normalizing data using standard deviation measure from the mean. It can adjust data to be within a -1,1 range
• Decimal scaling: Scaling data by dividing by 10j where j is the integer that would place the absolute maximum value within a -1,1 range

Data Discretization Methods

• Binning, Histogram analysis, Cluster analysis, Decision tree analysis, correlation analysis
• Concept Hierarchy Approach: Method of transforming data into various levels of granularity (e.g., age, zipcode, country)

K-Means Clustering

• Exploratory model, unsupervised.
• Groups data based on attributes into clusters (using centroid values for cluster center).

DBSCAN Clustering

• Density-based
• Locates areas of high density, clusters are regions where data density exceeds some threshold
• Sensitive to parameters (ε, MinPts)

Hypothesis Testing

• Assessing the difference in means of two data samples, or the significance of the difference
• Two types of hypotheses:
  • Null Hypothesis(HO): No difference between the two data samples
  • Alternative Hypothesis(HA): A difference exists between the two data samples
• Outcome can lead to rejection or non rejection of HO

Predictive Models

• Identifying attributes of a data object in advance (e.g., guessing whether a customer will subscribe or not)

Regression vs Classification

• Classification deals with making decisions based on categorical results.
• Linear regressions give numerical values (as opposed to classes)

Training and Test Sets

• Training set: Used to train the model
• Test set: Used to evaluate the model
• Both sets are independent of each other and non-overlapping

Validation Set

• Part of the data set that's not used in training or testing and is separated to tune the model's parameters

Naïve Bayes Model

• Classifier based on probabilities and the Bayes' theorem.
• Simplifying assumption of attribute values being independently dependent

Naïve Bayes Classifier Metrics

• Accuracy, TPR, FPR, FNR, Precision, AUC

Cross-Validation

• Holdout (percentage split): Divides data into training and test sets based on pre-determined percentages. The performance is dependent on how the data is split
• K-fold cross validation: Data divided into K subsets, K trials performed. For each trial, one subset is used for testing, and the remaining K-1 subsets are for training

Model Deployment Best Practices

• Specify performance requirements (accuracy, TPR, precision, etc.)
• Separate model coefficients from the program
• Develop automated tests of the model (testing on a smaller portion of the data outside the training/testing data sets)
• Develop a back-test and now-test infrastructure. (testing models on historical data for updates, and ensuring that the model still works as expected when used on new data points).
• Evaluate each model update (testing each update to verify if performance requirements are still met).

Decision Tree & Ensemble Learning

• Prediction/classification by creating a tree-like structure based on attributes and criteria
• Splitting attributes, pruning, information gain, and Gini index

Ensemble learning

• A learning model that combines multiple learners to make predictions that are stronger and more accurate than individual models.
• Approaches like bagging (Bootstrap Aggregation) create multiple models from sampled training data and find a 'majority' vote for predictions
• Boosting method builds successive models and builds predictions based on weaker models, and giving weights to those outputs.