Machine Learning Glossary/C

Verfasst von / Written by Sebastian F. Genter

---

• Machine Learning Glossary

calibration layer

Post-prediction adjustment aligning model outputs with observed distributions. Corrects systematic biases through:

• Platt scaling (logistic calibration)
• Isotonic regression

Critical for reliability in probabilistic forecasting.

candidate generation

Initial recommendation phase filtering large catalogs to manageable options. Strategies:

• Collaborative filtering
• Content-based filtering
• Embedding similarity search

Reduces 100,000 items → 500 candidates for detailed ranking.

candidate sampling

Training optimization for large output spaces:

• Evaluates all positive labels
• Samples subset of negative labels

Reduces computation in scenarios like extreme classification (millions of classes).

categorical data

Discrete features with finite possible values:

• Nominal: Colors {red, blue, green}
• Ordinal: Ratings {poor, fair, good}

Encoded via one-hot, embeddings, or target encoding.

causal language model

Unidirectional models predicting next token using left context:

• GPT architecture
• Masked future tokens during training

Contrasts with bidirectional models like BERT.

centroid

Cluster center in partitioning methods:

• k-means: Mean of cluster points
• k-median: Median reduces outlier sensitivity

Updated iteratively during clustering.

centroid-based clustering

Partitioning approach grouping data around central points:

• k-means (most common)
• k-medoids
• BIRCH (hierarchical variant)

Requires predefined cluster count (k).

chain-of-thought prompting

LLM technique eliciting step-by-step reasoning: "Calculate gravitational force between Earth and Moon. Show equations." Forces explicit calculation rather than direct answers.

chat

Conversational interface preserving dialog history:

• Context window management
• Multi-turn interaction tracking

Applications: Customer service bots, AI companions.

checkpoint

Model state preservation:

• Training: Resume from interruption
• Deployment: Version control

Contains weights, optimizer state, and metadata.

class

Discrete prediction category:

• Binary: {spam, not_spam}
• Multiclass: {cat, dog, horse}
• Multilabel: Multiple simultaneous classes

classification model

Predictor outputting discrete labels:

• Logistic regression (probabilistic)
• SVM (maximum margin)
• Decision trees (rule-based)

Contrasts with regression models.

classification threshold

Probability cutoff converting scores to classes:

• Default 0.5 for binary
• Tuning affects precision/recall tradeoff

Example: Cancer diagnosis vs spam filtering thresholds.

classifier

(See classification model)

class-imbalanced dataset

Skewed class distribution challenges:

• 1,000,000:1 negative:positive ratio
• Mitigation: Resampling, class weights, anomaly detection

clipping

Outlier handling techniques:

• Feature clipping: Cap extreme values (e.g., ages >100 →100)
• Gradient clipping: Prevent exploding gradients

Stabilizes training and numerical computations.

Cloud TPU

Google's tensor processing units:

• Matrix multiplication optimization
• Pod configurations for large models
• Integrated with TensorFlow/JAX

clustering

Unsupervised grouping methods:

• Centroid-based (k-means)
• Density-based (DBSCAN)
• Hierarchical (agglomerative)

Applications: Customer segmentation, anomaly detection.

co-adaptation

Neural network pathology where:

• Neurons over-specialize to specific patterns
• Reduces generalization

Mitigated via dropout regularization.

collaborative filtering

Recommendation technique using user-item interactions:

• User-based: "Customers like you bought..."
• Item-based: "People who bought X also bought Y"

Matrix factorization approaches (SVD, ALS).

concept drift

Data distribution shifts over time:

• Feature meaning changes (e.g., "fuel efficiency" standards)
• Requires model retraining/monitoring

Detection methods: Statistical process control.

condition

Decision tree splitting rule:

• Axis-aligned: single feature threshold
• Oblique: multiple feature combinations

Determines data partitioning path.

confabulation

(LLM Hallucination) Plausible but incorrect generations:

• Factual errors in summaries
• Fictional citations

Mitigation: Retrieval augmentation, grounding.

configuration

Model setup parameters:

• Hyperparameters (learning rate, layers)
• Architectural choices (optimizer type)

Managed via config files (YAML/JSON) or libraries (Gin).

confirmation bias

Human tendency favoring information confirming existing beliefs:

• Dataset collection bias
• Labeling subjectivity
• Metric selection skew

confusion matrix

Classification performance visualization:

Confusion Matrix

Actual\Predicted	Positive	Negative
Positive	TP	FN
Negative	FP	TN

Derives metrics: Accuracy, Precision, Recall, F1.

constituency parsing

Sentence structure analysis breaking text into nested phrases:

• Noun phrases
• Verb phrases
• Prepositional phrases

Used in grammar checking, information extraction.

contextualized language embedding

Word representations adapting to context:

• "Bank" → financial vs river meanings
• BERT-style dynamic embeddings

Superior to static embeddings (Word2Vec).

context window

LLM input token capacity:

• Early models: 512 tokens
• Modern: 8k-128k tokens

Determines document processing capabilities.

continuous feature

Numerical variables with infinite possible values:

• Temperature measurements
• Sensor readings

Requires normalization/scaling for model stability.

convenience sampling

Non-probabilistic data collection:

• Quick experiments
• Potential sampling bias

Should transition to stratified/random sampling.

convergence

Training stability state when:

• Loss plateaus
• Parameter updates become negligible

Indicates model readiness (or local optimum).

convex function

Mathematical property enabling optimization:

• Bowl-shaped curve
• Single global minimum

Examples: L2 loss, logistic loss.

convex optimization

Minimizing convex functions:

• Guaranteed convergence
• Gradient descent variants

Foundation of linear models.

convex set

Geometric property where line between any two points remains within set:

• Spheres
• Cubes

Non-convex example: Star shape.

convolution

Matrix operation extracting spatial/temporal patterns:

• Kernel sliding across input
• Element-wise multiplication + summation

Basis for CNNs in image processing.

convolutional filter

Feature detector kernels:

• Edge detection: [[-1,0,1], [-1,0,1], [-1,0,1]]
• Learned during training

Depthwise separable variants reduce parameters.

convolutional layer

CNN component applying filters:

• Stride controls overlap
• Padding preserves dimensions
• Channels manage feature depth

convolutional neural network

Architecture for grid-like data:

• Convolution → Pooling → Dense
• Local connectivity → Translation invariance

Dominates computer vision tasks.

convolutional operation

Feature map calculation process: $(I*K)[i,j]=\sum _m\sum _{n}I[i-m,j-n]K[m,n]$ Where I=input, K=kernel.

cost

(See loss)

co-training

Semi-supervised method using:

• Multiple views of data
• Complementary feature sets

Example: Web page classification using text + link structure.

counterfactual fairness

Fairness criterion requiring:

• Same prediction for individuals differing only in protected attributes
• "What if?" scenario analysis

Mathematically formalized through causal models.

coverage bias

Dataset incompleteness issues:

• Missing population segments
• Unrepresentative sampling

Leads to model underspecification.

crash blossom

Ambiguous phrasing challenging NLU: "Stolen painting found by tree"

• Found near tree?
• Found using tree method?

Requires world knowledge for disambiguation.

critic

Reinforcement learning component:

• Estimates value functions
• DQN: Q-value approximator

Guides policy improvement through evaluation.

cross-entropy

Multiclass loss function: $H(p,q)=-\sum _{x}p(x)\log q(x)$ Minimized during classification training.

cross-validation

Robust evaluation protocol:

• k-fold data partitioning
• Rotation of train/validation splits

Prevents overfitting to single split.

cumulative distribution function (CDF)

Probability analysis tool: $F(x)=P(X\le x)$ Used for:

• Statistical testing
• Data distribution analysis
• Quantile calculations