Untitled Quiz

Machine Learning for Business Analytics

• The presentation is about machine learning for business analytics, focusing on different aspects of data analysis.
• The instructors are Dr. Marc Hilbert and Dr. Andrii Kleshchonok.
• The presentation language is English.
• The date of the presentation is October 20, 2024.

Part 1: Introduction to Business Analytics

• This section introduces the core concepts of business analytics.
• Define the task (e.g., prediction, clustering, classification, anomaly detection).
• Define objectives, error metrics, and performance standards.
• Data collection: set up data streams, storage, input, parallelisation, and Hadoop.
• Preprocessing: noise and outlier filtering, completing missing data using histograms and interpolation, normalization to scale data.

Dimensionality Reduction/Feature Selection

• Choose features to use and extract data from.
• Explore methods such as PCA, LDA, LLE, GDA.
• Consider goals, questions related to tractability.
• Design experiments, including train/validate/test data sets and cross-validation.
• Perform deployment.

Classification vs. Clustering

• Classification: uses labeled data, requires training phases, and is domain sensitive. Easy to measure performance. Includes methods like Naive Bayes, KNN, SVM, Decision Trees, Random Forests.
• Clustering: uses unlabeled data, organizes patterns based on similarity, difficult to evaluate, and includes methods like K-means, Fuzzy C-means, Hierarchical Clustering, DBScan.

Examples of ML Problems

• Predict how much customers spend in online retail.
• Explore different types of online retail customers.
• Find categories for items in an online store.
• Suggest items users might want to buy online.

Part 2: Elements of Statistics

• Discusses fundamental statistical concepts.
• Random variable descriptions, discrete and continuous.
• Probability function mapping.
• Probability function area always equals 1.

Description of Random Variables

• A random variable takes on a range of values with specific probabilities.
• The probability is how often we expect different outcomes in repeated experiments.

Discrete vs. Continuous Random Variables

• Discrete: countable number of outcomes. Examples: dead/alive, treatment/placebo, dice rolls.
• Continuous: infinite continuum of values. Examples: blood pressure, weight, speed of a car, real numbers from 1 to 6.

Probability Function

• A probability function maps possible values of a variable against the probability of their occurrence. This value is between 0 and 1.
• The area under the probability function is equal to 1.0.

Continuous Case

• For continuous variables, the probability function is a continuous mathematical function that integrates to 1.
• Example: the negative exponential function (exponential distribution) integrates to 1.

Continuous Case (cont.)

• The probability function for continuous random variables is called the probability density function (PDF).
• Probabilities of continuous variables are associated to ranges, not single values.

All Probability Distributions

• All probability distributions are characterized by an expected value (mean) and a variance (standard deviation squared).

Mean or Expectation Value

• Discrete case mean (expected value): E(X) = Σxᵢp(xᵢ) for all x.
• Continuous case mean (expected value): E(X) = ∫xᵢp(xᵢ)dx for all x.

Variance

• σ² = Var(X) = E(x-μ)²
• Variance is the expected squared distance from the mean.

Variance (cont.)

• Discrete case: Var(X) = Σ(xᵢ - μ)² p(xᵢ) for all x.
• Continuous case: Var(x) = ∫(xᵢ - μ)² p(xᵢ)dx for all x.

Normal Distribution

• A bell-shaped curve.
• Defined by the mean (μ) and standard deviation (σ).
• Changing μ shifts the distribution left or right. Changing σ increases or decreases the distribution spread.

The Normal Distribution: Mathematical Function

• f(x) = 1/(σ√2π) * e^(-(x-μ)²/(2σ²)).

The Normal PDF

• The area under the PDF curve always integrates to 1.

Normal Distribution Definition

• Mean = E(X) = μ
• Standard deviation (Std Dev) = √Var(X) = σ

Central Limit Theorem

• The mean of many random samples will be normally distributed around the true mean of the population, as the sample size increases.
• Standard deviation of the sampling distribution decreases as the sample size increases.

68-95-99.7 Rule

• 68% of the data falls within one standard deviation of the mean.
• 95% within two standard deviations.
• 99.7% within three standard deviations.

Confidence Interval

• μ = x ± t * (s/√n).
• Uses t-Students value, dependent on sample size and confidence level.

Testing Hypothesis

• Comparison between population distribution vs. sampling distribution.
• Test on the sample mean to either reject or accept the null hypothesis.

Deterministic vs. Statistical Testing

• Deterministic: observe the event and decide (reject/don't reject null).
• Statistical: observe the event and decide, with a chance of error (reject/don't reject with chance p%).

Types of Errors in Hypothesis Testing

• Type I error (α): Rejecting the null hypothesis when it's true (false positive).
• Type II error (β): Failing to reject the null hypothesis when it's false (false negative).

p-value

• Probability of an observed event to occur by pure chance.
• Informal significance levels help us interpret the results.

Anomaly Detection

• Techniques used to isolate and identify data points or values that are considered unusual or don't align with the rest of the data.

Examples of ML problems (cont.)

• Identify potential scams in online retail outlets.

A/B Testing

• A method for testing two different design options by comparing their success rate.
• Used to quantitatively asses if there's a statistical difference between the two.

Underlining Links

• Assess the effect of underlining links on click-through rate.

Correlation

• Describes a linear relationship between two variables.
• Positive correlation (increasing X → increasing Y).
• Negative correlation (increasing X → decreasing Y).
• No correlation.

Correlation (cont.)

• cov(X,Y) = Σ((xᵢ - X̄)(yᵢ - Ȳ))/(n - 1).

Linear Correlation

• Linear relationships are visualized and evaluated on scatterplots.
• Assess the strength of the relationship between variables.
• Visual assessment of the relationship: Strong, weak, no relationship.

Linear Regression Model

• Assumes a linear relationship between variables.
• Defines the relationship using an equation (Y = β₀ + β₁X₁ + εᵢ).
• The dependent variable is Y, and the independent variable is X.
• Random error (ε) accounts for the fact the linear relation is an approximation.

Estimating Parameters: Least Squares Method

• The best fit is when the differences between prediction values and the actual values are minimal.
• Least squares minimizes the sum of the squared differences.
• The method is used for parameter estimation in linear regression, and can also be applied to other models.

Least Squares Graphically

• Visual representation of the error minimisation using a line graph and the points.

Residual Analysis for Linearity, Homoscedasticity, and Independence

• Residual analysis assess the validity of the assumptions for the linear model.
• Linearity: The errors should be randomly distributed around the line.
• Homoscedasticity: The variance of errors should be constant across the independents variable.
• Independence: The residual values at one point should not be correlated with the residual values at a different point.

Estimating Parameters: Classification

• This section covers estimation techniques specific to classification problems.

Comparing LP and Logit Models

• Comparison of linear predictive models vs. logistic predictive models focusing on the shape differences.

Confusion Matrix/Crosstabs

• Calculates the performance of a classification model.
• True positives (TP), True negatives (TN), False Positives (FP), False Negatives (FN).

Confusion Matrix

• A table that records the counts of the classifications.

Underfitting and Overfitting

• Underfitting: The model is too simple to capture the true relationship (e.g., a flat line through data with a curve).
• Overfitting: The model is too complex, fitting the training set too closely and losing generalisability (e.g., following the noise in data, which does not reflect the underlying pattern)

Overfitting

• The model fits the training data very well.
• The model does not generalise well for the test data.
• There is a gap between training and test error.

Overfitting (cont.)

• Overfitting is a problem in machine learning.
• It occurs when a model is too complex.
• Good fit on training data but poor on unseen (test) data.

Overfitting of ANNs

• Parameters (e.g., number of neurons, initial weights).
• Activation functions (sigmoid, etc.).
• Learning rate, momentum (increase in flexibility due to increase in neurons).

Training and test data set

• Training data is used to train or learn a model.
• Test data is used to evaluate the performance.

Goodness-of-fit of ANN

• Measure of how well the ANN models the data.
• Similar to R-squared for linear regression, but applied to ANNs.
• Close to 1.0 = better fit.

MSE related with training over time

• Plotting MSE vs. Epochs helps choose the optimal model, as the training and test MSE.

Advantages of ANNs

• Efficient for massively parallel processing.
• Robust, tolerant to missing or noisy data.
• User-friendly programming.

Disadvantages of ANNs

• Difficult to design models for arbitrary applications.
• Difficult to assess internal operation of the ANN.
• Not easy to know which variable is influential (black box).

Part 5: Python Implementation

• This section outlines components for designing and training dense artificial neural networks using the Keras Python library.
• This includes Data (housing dataset), Problem statement (regression or classification), Preprocessing (standard scaling, one-hot encoding), and Architecture (number of layers/neurons, activation functions, dropout layers), Training parameters (optimizer ADAM, batch size, epochs), Evaluation metrics and learning curves, Analysis of errors and residuals.

Dropout Layers

• Randomly remove nodes within the NN during a forward path to train an ensemble of subnetworks.
• Effectively improves generalisation ability.
• Leads to improved uncertainty estimation of predictions.

Classification Implementation in Keras

• Steps to use Keras for classification models include defining a sequential model, defining the layers (input, hidden, output), specifying activation functions, compiling the model, and training it.

High-level Language Model Overview

• Large language models are described and their parameters vs. the year are displayed.

Intuition behind LLM trainings

• Autoregressive models predict future tokens given past history. Autoencoders predict tokens, given the rest of the context.

LLM Capabilities

• LLMs cover various tasks such as text classification, entity recognition, summarization, paraphrase, translation, and data generation.

GPT (Generative Pre-trained Transformer)

• Generative Large Language Model (LLM). Zero-shot and few-shot learning on diverse tasks. Includes a Chat Functionality and Human Feedback Loop.

Part 6: Exam Information

• This section contains exam-related details (e.g., dates, topics).