Podcast
Questions and Answers
Which type of neural network is best suited for processing image data?
Which type of neural network is best suited for processing image data?
What type of data is best processed using techniques like RNN and LSTM?
What type of data is best processed using techniques like RNN and LSTM?
For spatial problems with dense data, which deep learning method is most suitable?
For spatial problems with dense data, which deep learning method is most suitable?
Which of the following is NOT explicitly mentioned as a method for handling unstructured text data?
Which of the following is NOT explicitly mentioned as a method for handling unstructured text data?
Signup and view all the answers
What is a characteristic of the text data being processed by recurrent neural networks (RNNs)?
What is a characteristic of the text data being processed by recurrent neural networks (RNNs)?
Signup and view all the answers
What is the primary function of the system described in the 'Managerial decision support' diagram?
What is the primary function of the system described in the 'Managerial decision support' diagram?
Signup and view all the answers
According to the 'Data-driven risk model', what factors directly influence the probability of a crime, $P(crime_{it} = 1)$?
According to the 'Data-driven risk model', what factors directly influence the probability of a crime, $P(crime_{it} = 1)$?
Signup and view all the answers
What is the key difference between the 'Current practice' and 'Dynamic risk map' as shown in the 'Dynamic estimations of crime risk' diagram?
What is the key difference between the 'Current practice' and 'Dynamic risk map' as shown in the 'Dynamic estimations of crime risk' diagram?
Signup and view all the answers
In the 'Managerial decision support' system, where do officers on the ground fit into the process?
In the 'Managerial decision support' system, where do officers on the ground fit into the process?
Signup and view all the answers
What is the purpose of 'Theory' within the overall system process?
What is the purpose of 'Theory' within the overall system process?
Signup and view all the answers
In machine learning, what is the primary objective regarding task performance?
In machine learning, what is the primary objective regarding task performance?
Signup and view all the answers
Which of the following is a core component of defining a task in machine learning?
Which of the following is a core component of defining a task in machine learning?
Signup and view all the answers
What does the parameter w
represent in the mathematical function $y = f(x, w)$ within the context of machine learning?
What does the parameter w
represent in the mathematical function $y = f(x, w)$ within the context of machine learning?
Signup and view all the answers
The 'curse of dimensionality' in machine learning refers to:
The 'curse of dimensionality' in machine learning refers to:
Signup and view all the answers
What is the typical method used during the ‘learning’ phase in machine learning?
What is the typical method used during the ‘learning’ phase in machine learning?
Signup and view all the answers
What is the primary goal of the optimization process in machine learning?
What is the primary goal of the optimization process in machine learning?
Signup and view all the answers
Which of the following is an example of a regression task in machine learning?
Which of the following is an example of a regression task in machine learning?
Signup and view all the answers
Which of these examples is a classification task that results in a discrete output?
Which of these examples is a classification task that results in a discrete output?
Signup and view all the answers
In the context of machine learning, what does the term 'experience' primarily refer to?
In the context of machine learning, what does the term 'experience' primarily refer to?
Signup and view all the answers
Which of these is a key characteristic of supervised learning?
Which of these is a key characteristic of supervised learning?
Signup and view all the answers
What is the defining feature of unsupervised learning?
What is the defining feature of unsupervised learning?
Signup and view all the answers
A machine learning model achieves '99% correct classification.' What crucial information is needed to make this statement meaningful?
A machine learning model achieves '99% correct classification.' What crucial information is needed to make this statement meaningful?
Signup and view all the answers
In machine learning, what is the role of the 'actuating signal/error'?
In machine learning, what is the role of the 'actuating signal/error'?
Signup and view all the answers
If a machine learning model is described as 'speaker independent,' what does this primarily imply?
If a machine learning model is described as 'speaker independent,' what does this primarily imply?
Signup and view all the answers
Which of the following actions would best represent a task that uses unsupervised learning?
Which of the following actions would best represent a task that uses unsupervised learning?
Signup and view all the answers
What is the primary aim of measuring 'performance' in a machine learning task?
What is the primary aim of measuring 'performance' in a machine learning task?
Signup and view all the answers
In k-NN classification, what is the primary method used to assign a class label to a new data point?
In k-NN classification, what is the primary method used to assign a class label to a new data point?
Signup and view all the answers
What is a key characteristic of k-NN classification in terms of its training phase?
What is a key characteristic of k-NN classification in terms of its training phase?
Signup and view all the answers
A new data point is equidistant between 2 classes when using k-NN with k=5. Two neighbors are labeled 'A' and 3 are labeled 'B'. What label is assigned to the new data point?
A new data point is equidistant between 2 classes when using k-NN with k=5. Two neighbors are labeled 'A' and 3 are labeled 'B'. What label is assigned to the new data point?
Signup and view all the answers
Which of the following is NOT explicitly mentioned as a potential problem with k-NN classification?
Which of the following is NOT explicitly mentioned as a potential problem with k-NN classification?
Signup and view all the answers
What is the purpose of distance metrics in the context of k-NN classification?
What is the purpose of distance metrics in the context of k-NN classification?
Signup and view all the answers
Which of the following best describes the primary function of template-based wrappers in the context of ML.NET?
Which of the following best describes the primary function of template-based wrappers in the context of ML.NET?
Signup and view all the answers
According to the content, what is a significant limitation of some existing tools for business analytics?
According to the content, what is a significant limitation of some existing tools for business analytics?
Signup and view all the answers
Within the CRISP-DM framework, approximately what percentage of time is typically spent on the 'data understanding and preparation' phase?
Within the CRISP-DM framework, approximately what percentage of time is typically spent on the 'data understanding and preparation' phase?
Signup and view all the answers
Which of the following is NOT mentioned as a recommendation in the modeling process?
Which of the following is NOT mentioned as a recommendation in the modeling process?
Signup and view all the answers
Based on the content, what is considered more critical for improving model accuracy, investing in improving models or improving the data?
Based on the content, what is considered more critical for improving model accuracy, investing in improving models or improving the data?
Signup and view all the answers
According to the content, collecting data from which source is NOT part of creating a 360 customer view?
According to the content, collecting data from which source is NOT part of creating a 360 customer view?
Signup and view all the answers
What is recommended during the initial implementation phase of an AI project?
What is recommended during the initial implementation phase of an AI project?
Signup and view all the answers
According to this content, when should you consider re-training your AI model?
According to this content, when should you consider re-training your AI model?
Signup and view all the answers
What is the primary goal of the AI value chain as described in this content?
What is the primary goal of the AI value chain as described in this content?
Signup and view all the answers
According to the content, what does 'decision backwards' refer to within the AI implementation?
According to the content, what does 'decision backwards' refer to within the AI implementation?
Signup and view all the answers
Which of the following best describes the concept of 'test and learn' in the context of AI?
Which of the following best describes the concept of 'test and learn' in the context of AI?
Signup and view all the answers
What is the main objective of machine learning according to the content?
What is the main objective of machine learning according to the content?
Signup and view all the answers
Which of the following is NOT explicitly mentioned as a platform or package for machine learning?
Which of the following is NOT explicitly mentioned as a platform or package for machine learning?
Signup and view all the answers
What does the content emphasize about the relationship between the business decision and AI model development?
What does the content emphasize about the relationship between the business decision and AI model development?
Signup and view all the answers
According to the content, what is typically considered when defining a good AI case?
According to the content, what is typically considered when defining a good AI case?
Signup and view all the answers
Study Notes
Motivation
- Solving management problems with artificial intelligence (AI) is a key area of research.
- The research focuses on information, innovation, and impact.
- Information is gathered from data science to solve relevant management problems.
- Innovation involves developing AI algorithms like statistics and computer science.
- The impact of the tools is rigorously evaluated in terms of added value in management practice.
Example of Research with Impact
- Effective police patrolling
- Effective disease management
- Early warnings for fake news in social media
Impact During COVID-19 Epidemic
- Nationwide data on micro-level human mobility during the epidemic (~1.5 billion movements).
- Artificial intelligence algorithm linking mobility to case growth.
- Knowledge transfer through membership in COVID-19 Working Group of the WHO.
- Dissemination to the public through media appearances.
Combining AI Technologies and Real-World Applications
- Real-world AI demonstrations (field experiments, financial implications, organizational and behavioral implications, etc)
- AI for decision-making (causal ML, sequential settings, off-policy learning, dynamic treatment regimens)
- AI for good (healthcare applications, Sustainable Development Goals)
- AI & Web (social media data, clickstream data, mobility data)
In Joint Industry Collaborations
- Partnership for 3-year project. Full funding for a PhD position by the company.
- Companies (e.g., GEBERIT, SIEMENS, HITACHI, ABB).
- Challenges include insufficient data integration, missing policies for data use; absence of advanced analytics; inadequate analytics packages; lacking trust in data; insufficient capabilities in using or designing reporting and analytics; missing strategies for using big data, and advanced analytics.
Research to Publish in Leading Outlets
- Artificial intelligence research focuses on thought leadership in applied AI and ensures rigorous research and state-of-the-art performance.
- Application domains include contributing to relevant research and demonstrating practical impact.
- Example outlets like SIGKDD, WWW, EMNLP, ACL, CHI for AI. PNAS, Management Science, and Marketing Science are examples of application-specific journals/publications.
- Research collaborations with universities (e.g., Harvard, Berkeley, Carnegie Mellon, ETH, Texas).
Challenges for Bringing AI into Business Management
- Missing accountability of AI decisions.
- Incomplete human-in-the-loop analytics frameworks.
- Organizational inertia.
- Clarify boundaries of AI interventions, implement governance structures, and establish risk management frameworks.
- Develop human-centered approaches to frameworks for human learning and exploration, advance prescriptive algorithms, and promote a transformation workforce.
- Encourage managers to explore and experiment and incentivize adoption.
Course Overview
- 4 SWS / 6 ECTS.
- 4-day block course (online with optional Q&A).
- Exam with programming focusing on implementing analytics solutions.
- Constraints: BSc BWL.
- Required skills: programming and basic math (regression).
- Emphasis on practical application of machine learning rather than proofs, with a strong method focus.
Books
- James et al. (2013): Introduction to Statistical Learning.
- Wickham & Grolemund (2017): R for Data Science.
Mastering Business Analytics
- Understand business to derive KPIs.
- Prepare data from internal and external sources.
- Apply predictive analytics models and evaluate performance for decision-making.
Naïve Predictions in Decision-Making
- Forecasting sales volume as input to production quota, with calibration and go-live phases.
- Replicates observed patterns from past data, using neural networks and leveraging external predictors.
- Illustrative example only.
Predictive Analytics
- Predictive analytics anticipates outcomes from a given input.
- This includes descriptive (what happened?), diagnostic (why did it happen?), predictive (what will happen?), and prescriptive (what should happen?) steps.
- It learns from past decisions and replicates them, aiming to identify optimal decisions.
Prediction of Churn
- Based on multiple input variables in statistical propensity models.
- Identifies variables like customer behavior 6 months prior (e.g., bad debt) and inbound calls to customer care.
- Models are iteratively improved using "learning" from actual customer behavior.
- Examples: logistic regression and decision trees are used.
Banking Example Dataset
- Includes customer specifics (age, gender, marital status); product holdings (product, average balances, total assets, liabilities); product usage/transactions (total volumes); and contact history & other (last contact, campaigns/channels, etc).
- Desirable but challenging data - life changes, personal/professional events, income, and job type; transaction details (channel, purpose).
Example: Predicting Customers
- Categorizes customers as "persuadables," or "sure buyers" to target marketing campaigns effectively.
- Algorithm identifies subgroups based on customer response to treatment in the past. Instead of targeting all customers, it targets those likely to respond positively.
Targeting Customers to Prevent Churn
- Implementing churn early warning systems based on analytical propensity models.
- Yields concrete customer interactions in the call center, using AI to monitor real-time social distancing compliance, and provide knowledge transfer by the WHO.
- Customer-specific actions (e.g., targeted calls with individualized scripts) based on factors like the customer lifetime value (CLV) and segment.
Group Work: Pitch your idea of business analytics
- Form groups of 2-3 students and come-up with an innovative/impactful idea based on data analytics.
- Prepare short presentation (4 mins, max 4 slides) with name/study programs on front page, ready for Moodle submission.
- Questions answered include targeted benefits, value quantification metrics, necessary data sources, project complexity, and competitors.
Managerial Decision Support for Effective Police Management
- Using data to estimate crime risk and provide support for patrolling, risk reduction, and alert prioritization.
- Systems identify elevated risk areas/times for officers on the ground, assisting in proactive efforts.
Dynamic Estimation of Crime Risk
- Historic crime data and current practice (hotspot mapping) are compared to dynamic risk maps.
- This uses data from Kadar, Maculan & Feuerriegel (2019).
Data-Driven Risk Model
- Data collection, preprocessing, predictive modeling, performance measurement.
- Imbalance-aware machine learning (undersampling, cost-sensitive learning).
- Statistical measures like hit rate/PAI/AUC plot evaluate performance.
- Equation uses features of spatial/temporal data/crime to generate a predictive solution (e.g., probability of a crime at a given time/place).
Effectiveness
- Evaluation plot compares spatial/temporal/crime features with all features to determine the best approach toward an algorithm.
Primer
- A general overview.
Modeling Basics
- Supervised learning uses historical data to predict the future.
- Unsupervised learning finds patterns in historical data.
- Supervised learning involves a clear input (x) and output (y) relationship, while unsupervised learning works with unlabeled data.
- Key implications of applying unsupervised learning are unclear.
Today's Lecture - Objectives
- Learn common concepts of machine learning, able to evaluate predictive performance, and distinguish between predictive and explanatory power (e.g., in predicting customer behavior or crime occurrence).
Course Outline
- A detailed list of topics for course coverage, including predictive modeling, machine learning, and applications for use in business.
Machine Learning: Examples
- This section covers broad, real-world examples like speech recognition, handwriting recognition, fraud detection, text filtering, image processing, and robotics.
Machine Learning: Task
- Learning a mathematical function mapping inputs (x) to outputs (y) via parameters (w).
- Outputs can be continuous (y ∈ R) or discrete (e.g., y ∈ {0,1}).
- Complexities associated with high-dimensional datasets often arise.
Machine Learning: Performance
- Performance is typically presented as a single number, such as percentage of correctly classified letters, or frequency of success in a certain process (e.g., 99% correct classification; car driving on country roads without human intervention 99% of time).
- Further clarification on what is being categorized is essential.
- Type of data (e.g., characters, words, sentences, speaker/writer, and specific data set) is also essential.
- Example: learning how to drive based on labeled (data with labels) or unlabeled data (without labels).
Supervised vs. Unsupervised Learning
- Supervised learning relies on labeled data with input(x) and output/result(y) pairs.
- Unsupervised learning deals with unlabeled data, aiming to discover patterns and structure within it.
Statistical data types
- Data type defines variables' characteristics and suitable models.
- Categorical (discrete) variables include nominal (without inherent order) and ordinal (with implied order).
- Numerical (continuous) data represents values from an interval.
- Multivariate data includes multiple numerical variables.
Taxonomy of Machine Learning
- Broad categories for machine learning algorithms.
- Regression (predicting a continuous variable)
- Classification (predicting an outcome from pre-defined category labels)
- Clustering (grouping data points by similarity)
Multi-class prediction
- Method for handling more than two categories or outputs in algorithms.
k-Nearest Neighbor (k-NN) Classification
- Classifies new observations by majority vote among the k-nearest neighbors in the training set.
- No internal training.
- Distinctive features: no internal training parameters; classification via neighbors; high computational cost when dealing with many data points.
Distance Metrics
- Algorithms use metrics such as Euclidean and Manhattan distance to calculate distance between two points.
Choosing Number of Nearest Neighbors (k)
- Shows examples of decision boundaries produced for values of k in k-NN classification.
- Choices of k influence classification models for predicting different outcomes.
Prediction Performance
- Model Choice (e.g. linear, non-linear)
- Training and testing (splitting dataset)
- Performance metrics (evaluation)
- Statistical model comparison
- Overfitting (when training data adapts poorly to unseen data)
Assessment of Models
- Predictive performance (accuracy, recall, F1 Score, ROC, ...)
- Computation time (modeling and prediction)
- Robustness to noise in predictor values
- Interpretability (e.g., transparency, understanding)
Prediction Power vs. Interpretability
- Trade-off between model accuracy/predictive power, and the capability to easily understand model output/decisions.
Training and Test Set
- Splitting data into training and test sets to evaluate how well modeling generalizes to new, unseen data.
- 80/20 split (80% for training, 20% for testing) is a common rule-of-thumb.
Overfitting
- A phenomenon that arises when a model learns and fits characteristics of training data that aren't generalizable to unseen data.
- Overfitted models achieve high accuracy in training, but poor performance in evaluating unseen data.
- Common causes for overfitting arise from overly complex models (e.g., many predictors) when the training data is small.
Remedies to Overfitting
- Using a large training set
- Mathematical penalties to prefer simpler models (regularization)
- Tuning hyperparameters.
Overview: Approaches to Regularization
- Subset selection
- Dimension reduction
- Shrinkage methods (regularization)
Management Guidelines
- Identifying tasks suitable for applying machine learning.
- Guidelines for identifying suitable tasks for implementing machine learning applications
- Examples; learning a function for well-defined inputs and outputs; availability of large digital data; clear, definable task metrics; straightforward relationships (A→B); and no lengthy required logic or background knowledge.
Model Tuning
- Methods to control model parameters to optimize predictive performance using cross-validation approach.
- Validation sets (e.g. train/test/validation) is crucial for tuning models.
- Cross-validation technique improves estimation of generalization properties by using multiple parts of the dataset alternately for training/testing; common k-values are k = 5, and 10.
Validation Set
- Three-fold split into training, validation and test sets for evaluating different hyperparameter values.
- Fit multiple models to training data with different hyperparameters.
- Select the best hyperparameter set based on the performance on validation.
- Evaluate the final model on the test data to assess true generalizability to predict outcomes not in training dataset.
k-Fold Cross-Validation
- Dividing the data into multiple subsets for cross-validation purposes, which provides for robust estimates of the test error.
Leave-One-Out Cross-Validation (LOOCV)
- Special case of k-fold cross-validation; suitable for high computational costs.
Model Tuning in R
- Specific instructions using 'caret' package to perform train/test split, validation, and tuning.
- Tuning a model involves setting parameters and testing the models' performance across varying parameter values, and checking how well it performs on a variety of datasets.
Bringing Machine Learning into Practice
- Management process; involves aspects like data, insights, analytics, software, people, strategies, and decision-making.
- Crucial considerations like data integration policies, software suitability, team capabilities, and integration of analytics into the decision-making process.
Data vs. Prediction Model
- Data is the most significant aspect to consider first, as high-quality data leads to reliable predictive models.
- Improving data quality beforehand leads to better models overall than focusing exclusively on model enhancements.
The Big Data Fallacy
- Deep learning benefits from a large amount of data, but other methods may be suitable with lesser amounts of data.
Comparison: Traditional Machine Learning vs. Deep Learning
- Performance comparison of traditional and deep models given amounts of data.
- Deep learning is sometimes more advantageous for certain analysis, especially when lots of data is available.
Handling Imbalanced Data
- A common problem; common vs. less frequent classes occur in data.
- Bias towards a frequent category with standard metrics can be mitigated by employing alternative metrics such as balanced accuracy.
- Oversampling and undersampling strategies may provide improvements.
Interpretability
- Understanding of how a model functions and makes decisions.
- Choosing an effective graph/model to identify trends and relationships in data to aid in interpreting models, for example, partial dependency or local variable importance plots.
Studying That Suits You
Use AI to generate personalized quizzes and flashcards to suit your learning preferences.
Related Documents
Description
Test your knowledge on various deep learning methods and their applications in processing different types of data. This quiz covers neural networks like CNNs and RNNs, as well as their suitability for image and text data. Assess your understanding of the roles these models play in tasks such as crime risk assessment and managerial decision support.