Scribd
Upload
17 views11 pages

Predictive Analytics: Key Concepts & Tools

Uploaded by

Dikshith Ml
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
17 views11 pages

Predictive Analytics: Key Concepts & Tools

Uploaded by

Dikshith Ml
Copyright
© All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

Predictive Analytics

Module-1 Questions

1. What is Business Analytics? Explain its types.


2. Explain different types of models.
3. List and explain the Predictive Analytics applications.

4. Explain Analytical Techniques? (Explain any one Analytical Techniques with its
types)
-Explain Regression Analytical Techniques.
-Explain Machine learning Analytical Techniques.
5. List and explain open source predictive analytic tools.
6. List and explain commercial predictive analytic tools.
7. Describe Propensity Model with its different types of models.
8. Brief the Application of Propensity Models.
9. List and explain the notable failures and limitations of Predictive Models.
10. Write and Explain the basic syntax of R Tool.

Module-2
Chapter 3
[Link] is Statistical Modeling ? Explain with the formal definition and an example.
[Link] is Machine Learning ? Explain the different ML tasks.
[Link] the desired output of machine-learned system, categorize the ML tasks.
[Link] the different ML approaches. List some applications of ML.

Chapter 4
1. What is the key difference between the Standard Bayesian approach and the
Empirical Bayes approach and how does this impact the estimation of prior
distributions?
2. How does the Hierarchical Bayes model structure uncertainty across multiple
levels and what role do hyperparameters play in the Empirical Bayes
estimation?
3. What is the EM Algorithm used for in the context of Empirical Bayes and how
does it contribute to point estimation?
4. Compare the Parametric and Non-Parametric Empirical Bayes methods. In
which scenarios would Robbins' NPEB method be preferred over parametric
approaches?
5. How does the Poisson–Gamma model function within the Parametric Empirical
Bayes framework, and what type of data is it best suited for?
6. In Bayesian linear regression, how are prior beliefs incorporated into the model
and what is the role of the posterior distribution?

Module-3
Chapter 3
1. Explain the importance of data transformation in machine learning. Discuss at
least three specific transformation techniques and how they help improve model
performance.
2. What is skewness in data distribution? Describe how to detect skewness and
explain at least two transformation methods used to address it, including the
Box-Cox transformation.
3. Discuss the role of Principal Component Analysis (PCA) in data transformation
for multiple predictors. Explain the process, goals and how scree plots are used
to decide the number of components to retain.
4. Describe different strategies for dealing with missing values in a dataset.
Compare removal, model-based handling, and imputation techniques and
discuss the challenges associated with each method.

Chapter 4
1. Explain the impact of removing predictors in a predictive model.
Discuss the concepts of zero and near-zero variance predictors, and
how their presence affects different types of models.
2. What is multicollinearity in the context of predictive modeling?
Describe methods to detect and handle collinear predictors with
examples.
3. Describe the process of adding predictors in machine learning. How
are categorical and nonlinear predictors handled, and why is dummy
encoding important in this context?
4. Discuss the concept of binning predictors. What are the perceived
advantages and disadvantages of binning, particularly in the context
of medical or survey data?
5. Define over-fitting in predictive analytics. Explain with examples
how over-fitting impacts model performance and describe strategies
to avoid it.
6. What is model tuning? Explain the importance of tuning parameters
with examples such as K in KNN and the cost parameter in SVM.
7. Compare different model evaluation methods such as holdout
method, cross-validation, and bootstrap sampling. When should
each be used and why?

Module-4
Chapter 5

1. What is Performance of the model? How performance is measured?


2. Explain the impact of Bias and Variance to performance(R2 score
Value) of the model.
3. Write the commands for computing the performance using R tool
and explain it.

4. Explain the key assumptions behind Linear Regression. Why are these
assumptions important, and what problems can arise if they are violated?
(Hint: Linearity, Independence, Homoscedasticity, Normality)

5. Compare and contrast Ridge Regression, Lasso Regression, and Elastic


Net. In which situations would you prefer each method? Provide examples
to support your answer.
(Hint: Think about overfitting, feature selection, multicollinearity, etc.)

6. In the Solubility case study, why were Box–Cox transformations applied?


What other preprocessing techniques could you use if predictors are
highly skewed? Explain how skewness affects linear regression.

Q: How does multicollinearity affect the coefficients in linear regression,


and what are two methods to address it?
● Expected Answer:
o Multicollinearity inflates the variance of coefficients, making
them unstable.
o Solutions: Remove highly correlated predictors or use
dimension reduction (e.g., PCA/PLS).
2. Model Comparison
Q: Why might PLS outperform PCR in predictive modeling, even though
both reduce dimensions?
● Expected Answer:
o PLS is supervised (maximizes covariance with the response),
while PCR is unsupervised (maximizes predictor variance).
PLS often needs fewer components to achieve similar/better
performance.

3. Practical Application
Q: In the solubility case study, why were Box–Cox transformations applied
to the predictors?
● Expected Answer:
o To correct right-skewness in continuous predictors (e.g.,
molecular weight), ensuring they meet linear regression’s
normality assumptions.

4. Penalized Regression
Q: When would you choose Lasso over Ridge regression, and vice versa?
● Expected Answer:
o Lasso: When feature selection is needed (sparse models).
o Ridge: When all predictors are potentially relevant but
correlated.
o Elastic Net: Hybrid for highly correlated predictors with
irrelevant features.

5. Interpretation
Q: How do you interpret a coefficient in Ridge regression if it’s shrunk
toward zero but not eliminated?
● Expected Answer:
o The predictor still contributes to the model, but its impact is
dampened to reduce overfitting. The direction
(positive/negative) of the relationship remains meaningful.

6. Critical Thinking
Q: Can linear regression handle nonlinear relationships? If not, how
might you modify the model?
● Expected Answer:
o No—it assumes linearity. Solutions:
▪ Add polynomial terms (e.g., x2x2).
▪ Use splines or nonlinear models (e.g., MARS, neural
networks).

Bonus: Interactive Question


Q: Looking at the PLS variable importance plot (Fig. 6.14), which
predictor seems most influential for solubility, and why might this make
sense chemically?
● Expected Answer:
o MolWeight (highest VIP score). Heavier molecules often
have lower solubility due to reduced interaction with solvents.

Chapter-7
1. Neural Networks
Q1:
Explain the structure and functioning of a single-layer feed-forward neural
network used in nonlinear regression. How does weight decay help in
preventing overfitting in such models?

2. Multivariate Adaptive Regression Splines (MARS)


Q2:
Describe how MARS models handle nonlinearities and interactions in data.
Discuss the process of feature creation and pruning in MARS, and
highlight its advantages and limitations.

3. Support Vector Machines (SVM) for Regression


Q3:
Discuss the concept of ε-insensitive loss in Support Vector Regression
(SVR). How do kernel functions enable SVMs to model nonlinear
relationships? Provide an example to support your explanation.

4. K-Nearest Neighbors (KNN) for Regression


Q4:
Explain how the KNN regression algorithm works. What are the key factors
that affect its performance, and how can they be optimized through
preprocessing and tuning?

Write the commands for computing the performance of


NN,MARS,KNN,SVM using R tool and explain it.

Chapter-8
1. Basic Regression Trees
Q1:
Describe the process of building a regression tree using the CART
methodology. How does cost-complexity pruning help in improving model
performance and interpretability?

2. Regression Model Trees


Q2:
Explain the structure and working of model trees such as M5. How do
pruning and smoothing contribute to enhancing prediction accuracy and
reducing overfitting in model trees?
3. Rule-Based Models
Q3:
What are rule-based models, and how are they derived from regression
trees or model trees? Compare the advantages and limitations of rule-
based models with those of full tree models.

4. Bagged Trees
Q4:
Discuss how bagging improves the performance of regression trees. What
are the roles of out-of-bag estimates and the bias-variance trade-off in
evaluating bagged models?

Module -5
Chapter-11
1. Class Predictions and Probability Calibration
Q1:
Discuss the importance of class probability calibration in classification
models. How do techniques like Platt scaling and Bayesian calibration
help produce well-calibrated probabilities, and why are they crucial for
real-world decision-making?

2. Evaluating Predicted Classes


Q2:
Explain the construction and use of the confusion matrix in classification
models. What are sensitivity, specificity, Kappa statistic, and how do they
provide a deeper understanding of model performance beyond overall
accuracy?

3. Cost-Sensitive Evaluation and Profit-Based Metrics


Q3:
How can classification model performance be evaluated using profit-based
metrics instead of accuracy? Illustrate with an example how unequal costs
for false positives and false negatives can influence model assessment in
applications like direct marketing or fraud detection.

4. Visual and Statistical Tools: ROC and Lift Charts


Q4:
Compare and contrast ROC curves and Lift charts as tools for evaluating
classification models. How can the area under the ROC curve (AUC) and
lift values help in choosing optimal thresholds for decision-making in
imbalanced classification problems?

Chapter-12
What are the key factors that influence the compressive strength of
concrete mixtures, and how can they be optimized?
How do machine learning techniques improve the accuracy of
predicting the compressive strength of concrete compared to traditional
statistical methods?
What role do model evaluation metrics such as RMSE, MAE, and R-
squared play in validating predictive models for concrete strength?
Explain the key differences between Linear Discriminant Analysis
(LDA), Logistic Regression, and Partial Least Squares Discriminant
Analysis (PLS-DA). In what scenarios would one method be preferred over
the others?
Discuss the role of feature engineering and predictor reduction in
handling high-dimensional data for classification models. Use examples
from the University of Melbourne case study to support your answer.
What is Elastic Net regularization, and how does it improve the
performance of logistic regression models in high-dimensional settings?
Include an explanation of how the GLMNET model performed in the grant
application dataset.
Evaluate the challenges and solutions in validating classification models
using time-based data splits. Why is this approach significant for real-
world deployment, and how was it implemented in the study?
Chapter-13
1. Nonlinear Discriminant Analysis (QDA, RDA, MDA)
Q1:
Compare and contrast Quadratic Discriminant Analysis (QDA),
Regularized Discriminant Analysis (RDA), and Mixture Discriminant
Analysis (MDA). How do these models address the limitations of Linear
Discriminant Analysis (LDA) in nonlinear classification problems?

2. Neural Networks for Classification


Q2:
Explain how neural networks are adapted for classification tasks. Discuss
the role of hidden units, weight decay, model averaging, and softmax
transformation in producing well-generalized class predictions.

3. Flexible Discriminant Analysis (FDA)


Q3:
Describe the concept of Flexible Discriminant Analysis (FDA) and explain
how it extends LDA using regression-based techniques. How does the use
of MARS hinge functions contribute to nonlinearity and model flexibility?

4. Support Vector Machines (SVM)


Q4:
Discuss the principle of maximum margin in Support Vector Machines and
explain how kernel functions enable SVMs to handle nonlinear
classification. Include the effect of cost and kernel parameters on model
performance.

5. K-Nearest Neighbors (KNN) for Classification


Q5:
Explain how the K-Nearest Neighbors (KNN) algorithm is applied to
classification tasks. What are the effects of different values of K on model
bias and variance, and how does it compare to boundary-based
classification methods?
6. Naïve Bayes Classifier
Q6:
Outline the working of the Naïve Bayes classifier. What assumptions does
it make about predictor independence, and how does it perform probability
estimation for classification? Discuss the role of Laplace correction in
handling sparse data.

Chapter-14
1. Basic Classification Trees (CART)
Q1:
Explain the process of constructing a classification tree using the CART
algorithm. Include a discussion on split selection using the Gini index, tree
pruning through cost-complexity, and the model’s advantages and
limitations.

2. Rule-Based Models (C4.5Rules, PART)


Q2:
Differentiate between the C4.5Rules and PART rule-based modeling
approaches. How are rules generated and pruned in each method, and
what are the benefits of using rule-based models over traditional
classification trees?

3. Bagged Trees
Q3:
Describe how bagging enhances the performance of classification trees.
Discuss how ensemble predictions are formed, and the impact of bagging
on model variance, bias, and overall classification accuracy.

4. Random Forests
Q4:
Explain the working of Random Forests for classification tasks. What is the
role of the 'mtry' parameter, and how does random subspace selection
contribute to improved model generalization?
5. Boosting (AdaBoost, Gradient Boosting)
Q5:
Compare AdaBoost and Stochastic Gradient Boosting in the context of
classification. How does each algorithm handle sample weighting and
iteration, and what makes boosting particularly effective for high-variance
base learners?

6. C5.0 Algorithm
Q6:
Discuss the key features of the C5.0 classification algorithm. How does it
improve upon C4.5, and what roles do pruning, boosting, and winnowing
play in refining model complexity and performance?

Common questions

Powered by AI

Predictive models can face limitations such as overfitting, where the model fits the training data too closely and performs poorly on unseen data. Another limitation is bias, which can arise from training data that is not representative of future cases, causing inaccurate predictions. To manage these, techniques such as cross-validation can be employed to assess model performance on different data subsets, and simpler models or regularization methods like Lasso or Ridge might be used to reduce model complexity. Understanding and addressing these limitations is crucial for developing more robust predictive models .

In predictive analytics, bias refers to the error due to overly simple assumptions in the learning algorithm, while variance is the error due to the model's complexity and sensitivity to fluctuations in the training data. A high-bias model may underfit, failing to capture the underlying trend, while a high-variance model may overfit, failing to generalize to new data. The trade-off involves finding a balance where the total error, combining both bias and variance, is minimized for optimal model performance. Techniques like cross-validation or regularization are often used to find the right balance .

Skewness can significantly impact linear regression models by violating the normality assumption of the error terms, which can lead to biased estimates and reduced model performance. Techniques such as applying a Box-Cox transformation can help in mitigating the skewness by stabilizing the variance and making the data more normally distributed, leading to better linear regression model performance. Other techniques include log transformations or assessing and removing outliers that contribute to skewness .

Model tuning is crucial in optimizing the performance of machine learning algorithms, as it involves selecting the best parameters that minimize error and maximize accuracy. For example, adjusting the value of K in K-Nearest Neighbors (KNN) can affect the model's sensitivity to noise, and tuning the cost parameter in Support Vector Machines (SVM) can control the trade-off between model complexity and misclassification. These adjustments help in achieving an optimal balance between bias and variance, leading to models that generalize well to new data .

Cross-validation offers the advantage of providing a more reliable estimate of model performance by using multiple subsets of the data for training and validation, which helps in mitigating overfitting. Unlike the holdout method that uses a single train-test split, cross-validation uses all data points several times for both training and testing, giving a comprehensive assessment of model robustness. However, it is more computationally intensive, especially with large datasets, and may require extensive computational resources, which is a notable challenge compared to simpler methods like the holdout approach or bootstrap sampling .

Lasso regression is preferred over Ridge regression when feature selection is necessary, as it can shrink some coefficients to zero, resulting in a sparser model with fewer predictors. This is particularly useful when it's suspected that many predictors are not relevant. Lasso is beneficial in scenarios with high-dimensional data where multicollinearity might also be present. Conversely, Ridge regression would be used when all predictors are likely relevant, even if correlated, as it cannot eliminate predictors but only reduce coefficient magnitudes .

Principal Component Analysis (PCA) is a dimensionality reduction technique that transforms multiple predictors into a new set of uncorrelated variables called principal components. The goal of PCA is to reduce the number of variables while retaining as much variance in the data as possible, thereby enhancing the performance of predictive models by improving computational efficiency and reducing overfitting. PCA can also help in visualizing complex data and handling multicollinearity by orthogonalizing the predictor variables. The selection of the number of components is often aided by scree plots, which help visualize the explained variance .

Neural networks adapt for classification tasks by using layers of neurons to learn complex, non-linear relationships between input features and class labels. Hidden units in these layers capture feature interactions and transformations needed for accurate predictions. The softmax transformation is applied in the output layer to convert raw output values into probabilities, facilitating multiclass classification by ensuring that predicted class probabilities sum to one. Weight decay is often employed to avoid overfitting by reducing the magnitude of weights in the network, promoting more generalized models. These components work together to produce reliable and interpretable class predictions .

The key difference between Standard Bayesian and Empirical Bayes approaches lies in how the prior distributions are estimated. The Standard Bayesian approach relies on subjective priors, often determined through expert knowledge or previous studies, while the Empirical Bayes approach determines priors empirically from the data itself. This empirical estimation often makes the model more objective and data-driven. The difference in prior estimation impacts statistical modeling by potentially reducing the influence of subjective bias in prior selection for the Empirical Bayes approach, potentially leading to more robust estimates in contexts where subjective priors are hard to define .

In Bayesian linear regression, prior beliefs are incorporated by defining prior probability distributions for the model parameters, reflecting any existing knowledge or assumptions about them. These priors are then combined with the likelihood of the observed data to form the posterior distribution via Bayes’ theorem. The posterior distribution provides a probabilistic representation of the parameter estimates given the data, including uncertainty, allowing for more informed decision-making. It helps in updating beliefs as new data is observed, thereby integrating evidence into the decision-making process .

You might also like