Fundamentals & Problem Framing
What constitutes 'structure' in an unlabeled dataset?
What are the primary task categories under unsupervised learning?
Why are unsupervised learning problems inherently harder to specify than supervised problems?
How do you determine whether a business problem should be framed as an unsupervised learning task?
When would you choose an unsupervised method over waiting to collect labeled data?
Why is overfitting possible in an unsupervised model when there are no labels to overfit to?
Why does learning the data distribution P(X) qualify as an unsupervised generative objective?
When would you frame a problem as unsupervised learning rather than weakly supervised learning?
What are the primary shortcomings of using histograms for density estimation?
What is density estimation in unsupervised learning?
What is the effect of choosing a bandwidth that is too small or too large in kernel density estimation?
What is the difference between parametric and non-parametric density estimation?
How can density estimation be used to score anomalies?
How can a Gaussian Mixture Model serve as a density estimator?
How do you select the bandwidth for kernel density estimation in practice?
When would you use a Gaussian mixture model rather than kernel density estimation as a density estimator?
How does the curse of dimensionality degrade density estimation?
How can you compare two density estimators when you do not know the true underlying distribution?
How do you decide whether a business problem should be framed as density estimation rather than clustering?
What is the cluster assumption in semi-supervised learning?
How does pseudo-labelling work?
How does consistency regularization work in semi-supervised learning?
How does label propagation spread labels across a graph of data points?
How does confidence thresholding mitigate error propagation during self-training?
What is confirmation bias in the context of pseudo-labeling?
Under what conditions can incorporating unlabeled data hurt model performance?
What is the conceptual difference between label propagation and label spreading?
How would you detect that pseudo-labels are degrading your model during self-training?
Association Rules & Market Basket Analysis
What is support in market basket analysis?
What does confidence measure in an association rule?
What does lift indicate about the relationship between two itemsets?
What is the difference between an itemset and an association rule?
What is the main conceptual advantage of the FP-Growth algorithm over Apriori?
Why is confidence alone insufficient to prove a genuine association between items?
What causes spurious association rules in transactional data?
What makes an association rule redundant?
How does the Apriori property reduce the search space when mining frequent itemsets?
How do you choose minimum support and confidence thresholds when mining association rules for a business?
How do you deal with combinatorial explosion when mining association rules from large, sparse catalogues?
When would you use association rules for product recommendations instead of embedding-based or collaborative filtering approaches?
How do you evaluate an unsupervised learning model when you have no ground-truth labels?
What is downstream task evaluation for an unsupervised model?
How do you evaluate the stability of an unsupervised model across random seeds or resampling?
Why is evaluating an unsupervised model harder than evaluating a supervised model?
How do you ensure that unsupervised clusters or patterns are genuine rather than artifacts of preprocessing or noise?
What are the limitations of relying purely on proxy intrinsic metrics to evaluate unsupervised algorithms?
How do you verify whether an unsupervised discovery is actionable versus an artifact of data distribution drift?
How do you design a structured human review process to validate unsupervised groupings?
Representation & Self Supervised Learning
What defines positive versus negative pairs in contrastive representation learning?
What is representation learning?
What is a pretext task in self-supervised learning?
What is masked prediction in self-supervised learning?
Why is data augmentation critical to contrastive self-supervised frameworks?
Why has self-supervised learning become the dominant approach for large-scale pre-training?
Why is hard negative mining important for contrastive learning objectives?
Why can pre-training on unlabelled data act as a regulariser for downstream tasks?
Why are negative pairs critical to prevent representation collapse in contrastive methods?
When would you choose a masked-prediction objective over a contrastive objective for self-supervised pre-training?
How can a self-supervised model learn shortcuts that make its pretext task trivial without learning useful representations?
What does the bottleneck layer of an undercomplete autoencoder learn?
What is the reconstruction loss in an autoencoder measuring conceptually?
What is the primary conceptual difference between a standard autoencoder and a variational autoencoder?
What prevents an autoencoder from simply learning the identity function?
How does a denoising autoencoder prevent overfitting to the input data?
What is the conceptual difference between reconstruction loss and KL divergence in a variational autoencoder?
What is posterior collapse in variational autoencoders?
Applications & Production
What is the difference between customer segmentation and clustering?
How can unsupervised models assist human annotators in data labelling?
How do you use the outputs of an unsupervised model as features in a downstream supervised model?
How do you decide whether to freeze or fine-tune unsupervised pre-trained representations for a downstream task?
How do you monitor for drift in an unsupervised model running in production?
When would you use customer segmentation over supervised propensity scoring for personalizing content?
What are the tradeoffs between using pre-trained embeddings out of the box versus training a domain-specific unsupervised representation?
How do you determine whether a dataset should be addressed via unsupervised representation learning or standard feature engineering?