Harsh RB

Data Mining Blooms Taxonomy Questions

11 min read

Unit 3 Classification

Bloom’s Taxonomy Level 5: Evaluation Questions

Model Selection and Critique

  1. Critique and Justify: A health organization is building a system to predict a rare but fatal disease, where missing a positive case (False Negative) is significantly more costly than an incorrect positive prediction (False Positive). Critique the suitability of a Decision Tree (C4.5) versus a Support Vector Machine (SVM) with an RBF kernel for this task. Justify your final model recommendation by explicitly arguing which performance metric (Accuracy, Recall, or F1-Score) should be prioritized in the Cost Matrix setup, and how that priority would guide the selection and tuning of the recommended algorithm.

  2. Evaluate and Defend: You are tasked with classifying a massive dataset with millions of instances and a large number of irrelevant features (noise). -Nearest Neighbor (-NN) is rejected due to the Curse of Dimensionality and its lazy learning nature. Evaluate whether a Naïve Bayes Classifier or a Multilayer Feedforward Neural Network (NN) is the superior choice for high-speed, scalable classification under these conditions. Defend your choice by contrasting the models’ inherent assumptions about feature independence/dependence and their mechanisms for handling irrelevant features.

Ensemble Methods and Overfitting

  1. Assess and Optimize: A classification model is suffering from high variance (Overfitting) on the training set, exhibiting perfect accuracy but poor performance on the test set. Assess the use of Bagging (Random Forest) versus Boosting (AdaBoost) as an Ensemble Method to address this issue. Decide which ensemble technique is more appropriate for reducing variance in the base classifier, and justify your decision by explaining the fundamental difference in how they combine/weight the results of the base classifiers to manage the bias-variance trade-off.

Hyperplane and Feature Space Analysis

  1. Analyze and Defend: Consider a complex classification problem where the class labels are not linearly separable in the original feature space. An engineer proposes using a Decision Tree (CART) due to its interpretability, while another suggests an SVM utilizing a Polynomial Kernel. Evaluate the inherent limitations of the Decision Tree’s piece-wise linear approach against the SVM’s Kernel Trick ability to implicitly map data into a higher-dimensional space. Defend which approach is fundamentally more capable of finding the optimal decision boundary for highly non-linear data, and explain the key trade-off involved (interpretability vs. boundary complexity).

Rule-Based and Metrics Evaluation

  1. Design and Validate: You have developed a Rule-Based Classifier (e.g., using RIPPER) to categorize legal documents. The goal is high Precision, meaning when the system assigns a label, it must be correct. Design a validation strategy using the ROC Curve and AUC metrics that would allow you to definitively evaluate your Rule-Based model against a benchmark model (like a simpler k-NN) over all possible threshold settings. Justify why relying solely on Accuracy on the Test Set would be an insufficient metric for validating a system prioritizing high Precision.

Bloom’s Taxonomy Level 4: Analysis Questions

Classification Methods (Comparison and Contrast)

  1. Analyze Assumptions: Analyze the fundamental difference in how Naïve Bayes and Decision Tree Induction (C4.5/CART) handle feature interdependence. Explain how the core assumption of feature independence in Naïve Bayes can lead to superior speed but potentially lower accuracy compared to the localized, feature-splitting approach of a Decision Tree.

  2. Relate Mechanism to Challenge: Contrast the approach of Support Vector Machines (SVM) using the Kernel Trick with -Nearest Neighbor (-NN) in handling non-linear classification boundaries. Analyze why the -NN method is highly susceptible to the Curse of Dimensionality, while SVM, despite mapping to high dimensions, can often mitigate this effect through the use of the Maximum Margin Hyperplane concept.

  3. Differentiate Learning Styles: Differentiate between the “lazy learning” nature of -NN and the “eager learning” of a Multilayer Feedforward Neural Network (NN) using Backpropagation. Analyze the impact of each learning style on the training time versus the classification time for new instances.

  4. Component Analysis (Splitting Criteria): Analyze the difference in how Information Gain (used in ID3/C4.5) and the Gini Index (used in CART) determine the optimal split in a Decision Tree. Explain the primary computational advantage or disadvantage of one over the other in terms of minimizing impurity.

Model Evaluation and Enhancement

  1. Relate Metrics to Bias: Differentiate between the insights provided by Precision and Recall. Analyze a scenario where a classifier has a high Precision but low Recall, and explain how this pattern relates to either Overfitting or under-representation of the minority class in the Training Set.

  2. Deconstruct Ensemble Impact: Analyze the fundamental mechanisms by which Bagging (Random Forest) and Boosting (AdaBoost/Gradient Boosting) reduce classification error. Explain how Bagging primarily targets the reduction of variance, while Boosting primarily focuses on reducing bias, and how this leads to different training procedures.

  3. Interpret Confusion Matrix/ROC: Analyze the relationship between the values in a Confusion Matrix (True Positives, False Negatives, etc.) and the construction of the ROC Curve. Explain why the Area Under the Curve (AUC) is considered a more robust performance metric than simple Accuracy when dealing with imbalanced Class Labels and varying classification thresholds.

Integration and Rule Analysis

  1. Process Analysis (Rule-Based): Analyze the process of generating a set of IF-THEN Rules using a Sequential Covering algorithm like RIPPER. Explain how the process of Rule Pruning is essential to prevent the generated rules from resulting in Overfitting the specific structure of the training data.

Unit IV Cluster Analysis & Outlier Detection

Bloom’s Taxonomy Level 5: Evaluation Questions

Cluster Analysis

  1. Critique and Recommend: A data scientist is tasked with clustering a massive, high-dimensional dataset of customer transactions. The data contains dense clusters of arbitrary shapes and is known to have significant noise and outliers. Critique the suitability of -Means and AGNES for this specific task, citing the four fundamental Requirements of clustering. Propose and defend a robust two-step methodology (e.g., using a method like BIRCH or DBSCAN) that best balances scalability with the ability to handle arbitrary shapes and outliers, explicitly justifying your choice of method parameters (, MinPts, or CF-Tree parameters).

Outlier Detection and Integration

  1. Justify and Prioritize: You are working for a bank monitoring financial transactions. You need a system to detect Collective Outliers that signal coordinated fraud, while also identifying legitimate, high-value transactions that are Global Outliers. Evaluate the relative merits of a Density-based detection method (LOF) versus a Clustering-based method (e.g., using a -Medoids result) for this scenario. Justify which approach provides superior interpretability for regulatory reporting, and defend a specific method for handling the high dimensionality challenge inherent in financial data before applying your chosen detection algorithm.

Method Selection and Trade-offs

  1. Assess and Decide: A remote sensing team needs to identify anomalous geological formations (Contextual Outliers) within satellite imagery data, which is stored in a complex spatial format. The volume of data is large, but the formations of interest are small and sparse. Assess the trade-offs between using a Partitioning Method (-Medoids) and a Density-Based Method (DBSCAN) for simultaneously achieving cluster discovery and outlier detection in this context. Decide which single method is most appropriate for robustly identifying the specific contextual outliers, and defend your choice by explaining how its core mechanism (medoid distance or /MinPts) inherently addresses the challenge of noise masking or high dimensionality.

Bloom’s Taxonomy Level 4: Analysis Questions

Cluster Analysis Methods (Comparison and Contrast)

  1. Differentiate and Justify: Differentiate the operational mechanism and primary output of -Means (a Partitioning method) and AGNES (an Agglomerative Hierarchical method). Analyze how the reliance on centroids versus inter-cluster distances affects their ability to meet the requirement of handling arbitrary shapes in the data.

  2. Analyze Parameters and Scalability: Analyze how the introduction of the CF-Tree in BIRCH specifically addresses the scalability requirement of clustering, which poses a significant challenge for traditional methods like -Medoids (PAM) when dealing with massive datasets.

  3. Relate Mechanism to Challenge: Contrast -Medoids (PAM) with -Means in terms of robustness to outliers. Analyze the exact mechanism (centroid vs. medoid) that makes -Medoids more effective in dealing with data that violates the implicit assumption of spherical clusters, thereby improving its resistance to noise.

  4. Component Analysis (DBSCAN): Explain the roles of the (epsilon) and MinPts parameters in DBSCAN. Analyze how the definition of Core, Border, and Noise points allows DBSCAN to inherently handle clusters of arbitrary shapes and simultaneously identify outliers without relying on a pre-determined number of clusters ().

Outlier Detection and Clustering Integration

  1. Differentiate Outlier Types: Differentiate between a Global Outlier and a Contextual Outlier. Provide a scenario where identifying a Collective Outlier would require an analysis technique that first performs clustering (e.g., using DBSCAN) before the anomaly itself becomes apparent.

  2. Compare Detection Strategies: Contrast the fundamental principle of a Distance-based outlier detection method (-NN) with a Density-based method (LOF). Analyze why LOF is often more effective at tackling the challenge of noise masking in datasets where clusters have significantly different densities.

  3. Challenge Analysis: Analyze the challenges of High Dimensionality and Interpretability in the context of outlier detection. Explain how converting a high-dimensional dataset into a lower-dimensional representation before applying a Statistical method (like a Z-score) might alleviate one challenge but exacerbate the other.

UNIT V: Advanced Concepts in Data Mining

Bloom’s Taxonomy Level 5: Evaluation Questions

Web Mining

  1. Critique and Justify: A major e-commerce company wants to improve user engagement and site navigation. Design a comprehensive system that integrates Web Content Mining, Web Structure Mining (using a sophisticated link analysis technique beyond basic PageRank), and Web Usage Mining (clickstream analysis and session reconstruction). Justify which component provides the most critical insight for identifying the highest-value content pages, and defend your choice of algorithms for each component against alternatives, considering scalability and real-time performance.

Spatial Data Mining

  1. Evaluate and Prioritize: Imagine you are a city planner analyzing the spread of a localized infectious disease. You have access to geospatial data (patient locations, infrastructure, demographics). Evaluate and prioritize the use of either Spatial Classification Algorithms (like a Spatial Decision Tree with Moran’s I consideration) or Spatial Clustering Algorithms (like STING or GDBSCAN) to identify both the high-risk areas and the underlying spatial features contributing to the risk. Argue which approach is superior for generating actionable policy recommendations (e.g., resource allocation, infrastructure changes), and defend your choice of specific spatial data mining primitives (predicates, distance) necessary for a reliable analysis.

Temporal Data Mining

  1. Synthesize and Assess: A financial institution is analyzing high-frequency stock trading data. They need a system to predict short-term price movements and detect fraudulent activity. Propose an integrated methodology that utilizes Time Series Analysis (e.g., ARIMA or similar forecasting), Sequential Pattern Mining (GSP or PrefixSpan), and Temporal Association Rules. Assess the trade-offs between using a purely statistical model (Time Series) versus a pattern-based model (Sequences/Rules) for robust anomaly detection, and defend a criterion (e.g., false positive rate, predictive window accuracy) for evaluating the overall success of your proposed combined system.

Bloom’s Taxonomy Level 4: Analysis Questions

Web Mining

  1. Differentiate and Relate: Differentiate between the primary input data (e.g., text vs. links vs. logs) and the main objective of Web Content Mining and Web Structure Mining. Analyze why a researcher interested in the authority of information on a specific topic must necessarily integrate results from both to achieve a comprehensive view, and how HITS and Text Mining are related in this context.

  2. Analysis of Process: Analyze the key steps and inherent challenges in transforming raw web server log files into structured data suitable for Web Usage Mining. Specifically, explain the critical role of Session Reconstruction and how errors in this process (e.g., misinterpreting bot traffic or back-button usage) could lead to misleading results in User Profiling.

Spatial Data Mining

  1. Compare and Contrast: Contrast the fundamental goals and output of a Spatial Classification Algorithm (like a Spatial Decision Tree) versus a Spatial Clustering Algorithm (like GDBSCAN). Describe a scenario in urban planning where identifying Co-location Patterns (Spatial Rules) would be a necessary prerequisite for performing a meaningful spatial classification.

  2. Relate and Infer: Analyze the role of Spatial Autocorrelation (Moran’s I) within the spatial data mining process. Explain how a high Moran’s I value for a feature (e.g., crime rate) should influence the design of a Spatial Decision Tree or the parameter settings (e.g., epsilon and minimum points) for a spatial clustering algorithm like GDBSCAN.

Temporal Data Mining

  1. Distinguish and Apply: Differentiate between the nature of the output generated by Sequential Pattern Mining (e.g., PrefixSpan) and Temporal Association Rules. Provide a concrete example in a hospital patient dataset to illustrate how each technique yields a distinct, but equally valuable, clinical insight regarding event sequences and time-constrained relationships.

  2. Component Analysis: Analyze how the Autoregressive (AR) and Moving Average (MA) components of an ARIMA model specifically account for trend and seasonality in a time series of sales data. Explain the function of the Integrated (I) component in a non-stationary time series, and how its application fundamentally alters the data’s properties for successful forecasting.

Graph View

Backlinks