Chapter 5: Classification: Cytochrome P450 Inhibition
Introduces drug metabolism as a drug-discovery concern and explains why changes in cytochrome P450 activity can affect drug exposure and interactions.
After this chapter you will be able to
- Drug metabolism and clearance. Explain what drug metabolism does and why it matters to a drug’s effects and safety.
- Altered CYP activity and interactions. Explain how inhibiting or accelerating CYP3A4 can change exposure to another drug, using the section’s examples.
- Logistic Regression Classification Role. Explain why logistic regression is introduced for this binary task despite its name: it transforms inputs into values from 0 to 1, inclusive, to make predictions for binary outcomes.
- Linear Signal and Sigmoid. Explain how logistic regression differs from linear regression and linear classification: it passes a weighted feature signal through the sigmoid function to produce a value between 0 and 1.
- Inhibition Decision Threshold. Describe how the section uses a 0.5 threshold on the sigmoid output to classify and handle a compound.
- Binary Cross Entropy Optimization. Explain what binary cross-entropy measures and why the model uses an optimizer such as gradient descent to minimize its aggregated cost.
- Classifier performance in context. Explain why the logistic regressor’s validation accuracy of 0.78 is not sufficient on its own: an always-negative model would reach 0.55, while precision of 0.76 and recall of 0.74 give a more nuanced view.
- Meaning of Calibrated Probabilities. Explain what a calibrated predicted probability means for groups of molecules and why a sigmoid output is not automatically calibrated. Describe why unreliable probabilities can mislead drug-discovery decisions.
- Calibration Diagrams and Scores. Describe how a reliability diagram compares binned predicted probabilities with observed frequencies, including what points above or below the diagonal mean. Explain how RMSCE and Brier score summarize calibration and why one score alone may not tell the whole story.
- Post Hoc Calibration Methods. Distinguish Platt scaling from isotonic regression as ways to calibrate an already trained model. Explain why calibration is fitted on validation data and assessed on separate held-out test data, and describe the trade-off illustrated by their results here.
- Threshold Dependent Classifier Evaluation. Explain how changing the classification threshold affects precision and recall, and distinguish what PR and ROC curves plot across thresholds. Describe their random baselines, when the section favors PR evaluation, and how a use-dependent threshold can be chosen with Youden’s Index.
- Decision tree prediction path. Trace how a decision tree makes a classification by starting at the root, following branches based on feature-specific decisions, and reaching a leaf that gives the final prediction.
- Decision Tree Node Information. Describe what a visualized decision-tree node tells you about its split, impurity, samples, class distribution, and predicted class.
- Molecule Classification Path. Trace how a molecule receives a class prediction and probability estimate from the trained tree.
- Recursive Tree Construction. Explain how training grows a decision tree from labeled examples and when splitting stops.
- CART Split Criterion. Explain how CART evaluates candidate binary splits using Gini impurity and child-node sizes.
- Minority Inhibitor Classification Risk. Explain why even the modest imbalance in this CYP3A4 data set can cause a classifier to miss inhibitors.
- Resampling Methods And Tradeoffs. Distinguish random undersampling, random oversampling, and SMOTE by how each changes the training data and the drawback or benefit described here.
- Cost Sensitive Class Weighting. Distinguish a cost matrix from inverse-frequency class weighting and explain how weighting changes the loss used in training.
- Decision tree bias variance trade-off. Explain why a sufficiently deep decision tree can fit a target function with low bias yet change substantially when its training data changes slightly.
- Split limiting hyperparameters. Describe how the listed hyperparameters regularize a decision tree by restricting its growth or splits, and state the resulting bias-variance trade-off.
- Random forest construction and benefit. Explain how bagging and random feature selection build a random forest, and why combining its trees can improve predictions over a single decision tree.
- Consensus-based applicability domain. Explain how agreement among a forest's predictions can help assess whether to trust a prediction for a molecule.
- Adversarial validation of chemical space. Trace how adversarial validation tests whether new molecules resemble the training set, and interpret a low versus high classifier score.
- Bag-Level Learning Task. Explain how multi-instance learning differs from using a single input instance: a bag contains multiple instances, receives the label, and is classified as a whole.
- Bag and Instance Wrappers. Distinguish bag-wrapper from instance-wrapper MIL by identifying whether conformer descriptors or conformer predictions are aggregated to make one molecule-level prediction.
- Binary and multiclass classification. Distinguish binary classification from multiclass classification, and describe what a multiclass model predicts for a new input.
- Softmax multiclass predictions. Explain how softmax regression extends logistic regression to predict across more than two classes, including its class-specific weights, probability outputs, and training objective.
- CYP inhibition class mismatch. Explain why treating CYP inhibition as a multiclass choice gives the wrong question for a molecule.
- CYP Classification Task Types. Distinguish multiclass, multilabel, and multioutput classification, and explain why predicting which of five CYPs a molecule inhibits is multilabel.
- Multilabel Model Choices. Explain how independent logistic predictions and Random Forests are used for this multilabel task, including when independently trained per-label models are reasonable.
- Observed CYP Label Coverage. Justify restricting the merged CYP data to molecules reported in all five datasets, and describe the alternative assumption behind negative sampling.
- Per Label And Ranking Evaluation. Describe the section's per-label evaluation and distinguish ranking metrics from metrics for final class predictions, including the calibration caution.
- Micro Versus Macro Averaging. Distinguish micro from macro averaging across classes and explain when the section favors each interpretation.
What it builds on
- Linear Models (chapter 4) · Introduces linear models, the model family revisited here for classification.
- Gradient Descent (chapter 4) · Gradient descent as an iterative optimization method.
- Evaluating our Model (chapter 3) · Assumes prior experience evaluating a trained model.
- Bias-Variance Decomposition (chapter 4) · Assumes the distinction between bias and variance.
- Hyperparameter Tuning (chapter 3) · Assumes familiarity with hyperparameter tuning.
- Hashed Fingerprints (chapter 2) · Assumes familiarity with fingerprint features representing molecular structure.
Section by section
Introduction
Introduces drug metabolism as a drug-discovery concern and explains why changes in cytochrome P450 activity can affect drug exposure and interactions.
- Drug metabolism and clearance: Explain what drug metabolism does and why it matters to a drug’s effects and safety.
Why it matters: Metabolism helps determine both how long a drug acts and whether its breakdown products pose risks. - Cytochrome P450 enzyme role: Describe the role of cytochrome P450 enzymes in drug metabolism and why they receive particular attention.
Why it matters: CYPs are a useful focus because they handle a substantial share of drug metabolism, with effects that are not always benign. - Altered CYP activity and interactions: Explain how inhibiting or accelerating CYP3A4 can change exposure to another drug, using the section’s examples.
Why it matters: A compound’s effect on CYP activity can change the exposure and effectiveness of a drug taken alongside it.
Key terms: metabolism, xenobiotics, clearance, drug metabolites, Cytochrome P450, CYP, CYP-mediated metabolism, reactive metabolites, CYP inhibition, CYP3A4
5.1 Binary Classification of CYP3A4 Inhibition
The section frames CYP3A4 inhibition as a binary classification task and introduces logistic regression as a model for it.
- CYP3A4 Inhibition Classification Task: Describe the prediction task and the dataset used for it: determining whether a molecule inhibits CYP3A4 using data on 12,326 drugs, including 5,110 inhibitors.
Why it matters: The prediction target gives the classification model a concrete drug-discovery purpose. - Logistic Regression Classification Role: Explain why logistic regression is introduced for this binary task despite its name: it transforms inputs into values from 0 to 1, inclusive, to make predictions for binary outcomes.
Why it matters: It explains why a model called regression is relevant to an inhibitor-or-not decision.
Key terms: CYP 3A4, inhibitor, binary classification, logistic regression, binary outcomes, classification
5.1.1 Logistic Regression in Theory
The section explains how logistic regression turns a linear signal into a binary CYP3A4 inhibition prediction and trains that model with binary cross-entropy loss.
- Linear Signal and Sigmoid: Explain how logistic regression differs from linear regression and linear classification: it passes a weighted feature signal through the sigmoid function to produce a value between 0 and 1.
Why it matters: This transformation is what makes the model’s output usable for the inhibition decision described here. - Inhibition Decision Threshold: Describe how the section uses a 0.5 threshold on the sigmoid output to classify and handle a compound.
Why it matters: The threshold connects a model output to the practical decision to keep or filter a compound. - Binary Cross Entropy Optimization: Explain what binary cross-entropy measures and why the model uses an optimizer such as gradient descent to minimize its aggregated cost.
Why it matters: This explains both what training rewards and why it proceeds through iterative parameter updates.
Key terms: signal, weights, sigmoid function, logistic function, decision boundary, noninhibitor, inhibitor, binary cross-entropy (BCE) loss, true class, predicted probability
5.1.2 Logistic Regression in Practice
This section applies logistic regression to CYP3A4 inhibition data and shows why its validation accuracy should be considered alongside a baseline, precision, recall, and probability estimates.
- Classifier performance in context: Explain why the logistic regressor’s validation accuracy of 0.78 is not sufficient on its own: an always-negative model would reach 0.55, while precision of 0.76 and recall of 0.74 give a more nuanced view.
Why it matters: Comparing several measures keeps one attractive accuracy score from standing in for the whole performance story. - Predictions and probability estimates: Distinguish the class predictions obtained with predict() from the per-class probability estimates obtained with predict_proba(), and state that listing 5.1 records both.
Why it matters: The two outputs serve different purposes in the classification discussion that follows.
Key terms: accuracy, validation set, class imbalance, dummy model, precision, recall, class predictions, probability estimates, predict(), predict_proba()
5.1.3 (Mis)calibration: Questioning Probabilistic Output Assumptions
This section explains how to assess and improve logistic regression probabilities, evaluate classification across decision thresholds, and interpret logistic regression weights for CYP3A4 inhibition.
- Meaning of Calibrated Probabilities: Explain what a calibrated predicted probability means for groups of molecules and why a sigmoid output is not automatically calibrated. Describe why unreliable probabilities can mislead drug-discovery decisions.
Why it matters: A probability-based decision threshold is useful only if the scores reliably reflect outcomes. - Calibration Diagrams and Scores: Describe how a reliability diagram compares binned predicted probabilities with observed frequencies, including what points above or below the diagonal mean. Explain how RMSCE and Brier score summarize calibration and why one score alone may not tell the whole story.
Why it matters: The diagram and scores help you spot unreliable confidence estimates before using them to make decisions. - Post Hoc Calibration Methods: Distinguish Platt scaling from isotonic regression as ways to calibrate an already trained model. Explain why calibration is fitted on validation data and assessed on separate held-out test data, and describe the trade-off illustrated by their results here.
Why it matters: Recalibration can make model scores more usable, but its flexibility and evaluation procedure affect whether the improvement is trustworthy. - Threshold Dependent Classifier Evaluation: Explain how changing the classification threshold affects precision and recall, and distinguish what PR and ROC curves plot across thresholds. Describe their random baselines, when the section favors PR evaluation, and how a use-dependent threshold can be chosen with Youden’s Index.
Why it matters: Evaluating multiple thresholds lets you choose a trade-off that fits the cost of missed inhibitors and false positives. - Logistic Weight Interpretation: Explain why logistic regression weights should be interpreted through log-odds and odds ratios rather than as direct changes in predicted probability. Describe the effects of positive and negative weights while holding other features fixed.
Why it matters: Reading weights as effects on odds avoids mistaking them for constant changes in inhibition probability.
Key terms: predicted probabilities, well calibrated, miscalibration, reliability plot, bins, RMSCE, Brier score, underconfident, overconfident, Platt scaling
5.2 Tree-based Models
Introduces decision trees as classifiers that reach predictions through a sequence of feature-based decisions.
- Decision tree prediction path: Trace how a decision tree makes a classification by starting at the root, following branches based on feature-specific decisions, and reaching a leaf that gives the final prediction.
Why it matters: This gives you the basic picture needed to follow how a decision tree classifies an example.
Key terms: decision tree, feature, node, root node, branch, leaf node
5.2.1 Decision Trees
This section shows how a decision tree classifies molecules, how CART chooses its splits, and what makes trees useful and limited.
- Decision Tree Node Information: Describe what a visualized decision-tree node tells you about its split, impurity, samples, class distribution, and predicted class.
Why it matters: Reading the nodes lets you explain a tree’s decisions rather than treating its predictions as a black box. - Molecule Classification Path: Trace how a molecule receives a class prediction and probability estimate from the trained tree.
Why it matters: This explains both where a tree’s prediction comes from and what its reported probability represents. - Recursive Tree Construction: Explain how training grows a decision tree from labeled examples and when splitting stops.
Why it matters: The recursive training process explains how the readable sequence of decisions is created. - CART Split Criterion: Explain how CART evaluates candidate binary splits using Gini impurity and child-node sizes.
Why it matters: The split criterion explains why CART favors some molecular-descriptor thresholds over others. - Tree Strengths and Limits: Distinguish the practical advantages of decision trees from the limitations described here.
Why it matters: These trade-offs help you decide when a tree is convenient and when its predictions need caution.
Key terms: decision criteria, Gini, samples, value, class, root node, leaf node, class predictions, probability estimates, top-down
5.2.2 Dealing with Data Set Imbalance
Explains why class imbalance can harm CYP3A4 inhibitor classification and compares data-level and algorithm-level ways to address it.
- Minority Inhibitor Classification Risk: Explain why even the modest imbalance in this CYP3A4 data set can cause a classifier to miss inhibitors.
Why it matters: Missing inhibitors is particularly problematic when the task is to identify them. - Resampling Methods And Tradeoffs: Distinguish random undersampling, random oversampling, and SMOTE by how each changes the training data and the drawback or benefit described here.
Why it matters: Balancing class counts is not enough; how the balance is achieved affects what the model can learn. - Augmentation For Robustness: Explain how data augmentation differs from simply duplicating examples and why it can improve generalization, using a molecular example from the section.
Why it matters: Variation can help a model handle unseen inputs without relying on identical copies of minority examples. - Cost Sensitive Class Weighting: Distinguish a cost matrix from inverse-frequency class weighting and explain how weighting changes the loss used in training.
Why it matters: Weighting can address imbalance during training without changing the data set's class composition. - Validation Versus Training Performance: Describe what the section's model comparison shows about class weighting and decision-tree overfitting.
Why it matters: A mitigation strategy should be judged on validation results, not a perfect score on its training examples.
Key terms: class imbalance, minority class, majority class, CYP3A4 inhibition, Random Undersampling, Random Oversampling, SMOTE, downsampling, upsampling, data set augmentation
5.3.1 Decision Tree Bias-Variance Trade Offs
This section explains why decision trees can have low bias but high variance, how regularization and pruning address that trade-off, and how to compare tree models more reliably.
- Decision tree bias variance trade-off: Explain why a sufficiently deep decision tree can fit a target function with low bias yet change substantially when its training data changes slightly.
Why it matters: This trade-off explains why fitting a tree well is not enough to make it stable. - Split limiting hyperparameters: Describe how the listed hyperparameters regularize a decision tree by restricting its growth or splits, and state the resulting bias-variance trade-off.
Why it matters: These controls offer a direct way to make an unstable tree less sensitive to its training data. - Validation guided tree pruning: Describe pruning as removing splits from an already constructed tree when validation evidence does not justify keeping them.
Why it matters: Pruning provides another way to control tree instability after a tree has been built. - Reliable decision tree comparisons: Explain the two shortcuts noted in the previous evaluation and how tuning and repeated validation would make model comparisons more robust.
Why it matters: A single untuned comparison can make one model appear better without establishing a reliable performance difference.
Key terms: hypothesis space, low bias, high variance, training set, regularization, max_depth, max_features, max_leaf_nodes, min_samples_split, min_samples_leaf
5.3.2 Random Forests: A Strong Collective of Weak Decision Trees
This section explains how random forests combine decision trees, how to evaluate and interpret their predictions, and how to assess whether predictions are trustworthy for new molecules.
- Random forest construction and benefit: Explain how bagging and random feature selection build a random forest, and why combining its trees can improve predictions over a single decision tree.
Why it matters: This explains why a collection of individually sensitive trees can make more stable predictions. - Out-of-bag performance estimation: Describe where out-of-bag samples come from and how they provide a performance estimate without a separate validation set.
Why it matters: Out-of-bag evaluation offers a performance check while retaining scarce data for training. - Gini feature importance: Explain what random-forest Gini importance measures and how importance is combined across trees.
Why it matters: It gives a model-level indication of which input features influence the forest's splits. - Molecule-level similarity maps: Describe how a similarity map identifies parts of a molecule that influence its predicted CYP 3A4 inhibition.
Why it matters: A similarity map connects a prediction to the molecular regions affecting that particular prediction. - Consensus-based applicability domain: Explain how agreement among a forest's predictions can help assess whether to trust a prediction for a molecule.
Why it matters: Agreement helps distinguish predictions the model handles well from those that warrant more caution. - Adversarial validation of chemical space: Trace how adversarial validation tests whether new molecules resemble the training set, and interpret a low versus high classifier score.
Why it matters: This checks for a shift toward unfamiliar chemical space even when activity labels for new molecules are unavailable.
Key terms: ensemble learning, bagging, bootstrap sampling, Random Forests, feature subsampling, out-of-bag (OOB) evaluation, OOB samples, OOB PR AUC, feature importance, impurity reduction
5.3.3 Multi-Instance Learning: Model-agnostic Ensemble Learning
This section explains multi-instance learning through molecular conformers and compares two ways to combine conformer information into a molecule-level prediction.
- Bag-Level Learning Task: Explain how multi-instance learning differs from using a single input instance: a bag contains multiple instances, receives the label, and is classified as a whole.
Why it matters: The bag-level label explains why the model must combine information from multiple inputs. - Molecular Conformer Bags: Describe how different 3D conformations of one molecule can serve as instances in a bag and why considering multiple conformers can matter for a property such as CYP 3A4 inhibition.
Why it matters: The conformer example shows why one molecule can usefully be represented by several instances. - Bag and Instance Wrappers: Distinguish bag-wrapper from instance-wrapper MIL by identifying whether conformer descriptors or conformer predictions are aggregated to make one molecule-level prediction.
Why it matters: The distinction is whether conformer information is combined before or after individual predictions.
Key terms: multi-instance learning, MIL, bag, instance, conformational sampling, conformers, CYP 3A4 inhibition, bag-wrapper MIL, instance-wrapper MIL, descriptors
5.4 Multiclass & Multilabel Classification
This section distinguishes binary classification from multiclass classification by the number of possible classes and explains what a multiclass model predicts.
- Binary and multiclass classification: Distinguish binary classification from multiclass classification, and describe what a multiclass model predicts for a new input.
Why it matters: The number of possible outcomes determines whether a classification task is binary or multiclass.
Key terms: binary classification, multiclass classification, classes
5.4.1 Multiclass Classification
This section explains how softmax regression and random forests handle multiclass prediction, then shows why mutually exclusive classes do not fit CYP inhibition.
- Softmax multiclass predictions: Explain how softmax regression extends logistic regression to predict across more than two classes, including its class-specific weights, probability outputs, and training objective.
Why it matters: It shows what changes when logistic regression is extended from two classes to several. - Random forest multiclass support: State how the section says a random forest handles multiclass classification.
Why it matters: It distinguishes a model that already supports multiclass prediction from the logistic-regression extension described here. - CYP inhibition class mismatch: Explain why treating CYP inhibition as a multiclass choice gives the wrong question for a molecule.
Why it matters: The choice of classification setup must match the possible inhibition outcomes.
Key terms: softmax regression, softmax function, weight matrix, cross-entropy loss, Random Forests, multiclass classification, probability estimate, multiclass scenario, CYPs, one and exactly one CYP
5.4.2 Multilabel Classification
This section frames CYP inhibition as a multilabel problem and explains how to prepare data, train models, and evaluate predictions across five CYP labels.
- CYP Classification Task Types: Distinguish multiclass, multilabel, and multioutput classification, and explain why predicting which of five CYPs a molecule inhibits is multilabel.
Why it matters: The possible combinations of CYP inhibition determine what kind of prediction task and model output are appropriate. - Multilabel Model Choices: Explain how independent logistic predictions and Random Forests are used for this multilabel task, including when independently trained per-label models are reasonable.
Why it matters: These choices determine whether CYP labels are predicted separately or together when their relationships may matter. - Observed CYP Label Coverage: Justify restricting the merged CYP data to molecules reported in all five datasets, and describe the alternative assumption behind negative sampling.
Why it matters: Treating missing CYP results as confirmed non-inhibition would change the labels the models learn from. - Representative Multilabel Data Splits: Explain why the section uses stratified sampling for training and test sets in a multilabel dataset.
Why it matters: A representative split makes evaluation more informative about performance on the different CYP labels and their combinations. - Per Label And Ranking Evaluation: Describe the section's per-label evaluation and distinguish ranking metrics from metrics for final class predictions, including the calibration caution.
Why it matters: A multilabel model can be assessed both on its calls for each CYP and on how it orders likely labels. - Micro Versus Macro Averaging: Distinguish micro from macro averaging across classes and explain when the section favors each interpretation.
Why it matters: The averaging choice changes whether a reported score emphasizes common activities or treats rare, potentially important activities equally.
Key terms: multiclass classification, multilabel classification, multioutput classification, CYP inhibition, softmax regression, logistic function, MultiOutputClassifier, Random Forests, correlated labels, CYP inhibitor data sets
5.5 Summary
Where this is used later
- Downstream Applications: Antibody Classification (chapter 12) · Assumes basic interpretation of binary classifier outputs and evaluation.
- The XGBoost Magic Trick (chapter 6) · The earlier introduction to decision trees.
Concepts explained in this chapter
- random forestA random forest is an ensemble model that combines predictions from multiple decision trees trained with randomized data samples or feature selections.
- model calibrationCalibration is the agreement between a model's predicted probabilities and the observed frequencies of outcomes among examples receiving those probabilities.
- Brier scoreBrier score is the mean squared error between predicted probabilities and binary observed outcomes.
- Platt scalingPlatt scaling is a parametric probability-calibration method that fits a logistic regression mapping from model scores to outcome probabilities.
- ROC curveA receiver operating characteristic curve plots true positive rate against false positive rate as the classification threshold varies.
- Youden IndexYouden’s Index is the difference between a classifier's true positive rate and false positive rate, equivalently sensitivity plus specificity minus one.
- isotonic regressionIsotonic regression is a non-parametric regression method that fits a monotonic function to map scores to calibrated outcomes.
- miscalibrationMiscalibration is the discrepancy between a model's predicted probabilities and the observed frequencies of the corresponding outcomes.
- precision-recall curveA precision-recall curve plots classifier precision against recall as the classification threshold varies.
- reliability diagramA reliability diagram, or reliability plot, compares predicted probabilities with observed outcome frequencies across probability bins to assess calibration.
- root mean squared calibration errorRoot mean squared calibration error is the square root of the mean squared difference between predicted probabilities and observed outcome frequencies.
- ensembleEnsemble learning combines multiple individual predictive models to form a composite model with improved robustness or accuracy.
- recallTrue positive rate is the proportion of actual positive instances correctly predicted as positive.
- AUC-PRThe area under the precision-recall curve is a summary metric that aggregates precision and recall across classification decision thresholds.
- area under the ROC curveAUC is the area under the receiver operating characteristic curve and measures ranking-based discrimination between classes.
- logitA logit is an unnormalized real-valued score for a possible output token before conversion to probabilities.
- oddsOdds are the ratio of the probability of an event to the probability of its complement.
- odds ratioAn odds ratio is the ratio of two odds, often comparing the odds under a changed feature value with the odds under a reference value.
- decision treeA decision tree is a hierarchical predictive model that applies feature-based decision rules to reach an output.
- baggingBagging is an ensemble technique that trains multiple instances of a learning algorithm on bootstrap-sampled subsets of data and aggregates their predictions.
- random feature selectionFeature subsampling is the random selection of a subset of input features for constructing an individual model component or tree split.
- coverage errorCoverage error is a multilabel ranking metric measuring how many top-ranked labels must be included to cover all true labels for an instance.
- micro averagingMicro-averaging pools classification outcomes across classes before calculating an aggregate metric, giving more influence to frequent classes or labels.
- multilabel classificationMultilabel classification is a supervised learning task in which each instance can be assigned multiple class labels simultaneously.
- pairwise ranking lossRanking loss is a multilabel metric measuring the fraction of incorrectly ordered positive-negative label pairs.
- sigmoidThe logistic function, or sigmoid, is an S-shaped function that maps real-valued inputs to values between zero and one.
- adversarial validationAdversarial validation trains a classifier to distinguish examples from two data partitions in order to detect distribution differences between them.
- bootstrap samplingBootstrap sampling creates a sample by repeatedly drawing observations from a dataset with replacement.
- ensemble consensusEnsemble consensus is the degree to which the individual models in an ensemble produce mutually agreeing predictions for an input.
- feature importance scoreFeature importance is a quantitative estimate of how much an input feature contributes to a predictive model's decisions.
- gini importanceGini importance is a tree-ensemble feature-importance measure based on the weighted reduction in node impurity attributable to each feature.
- out-of-bag errorOut-of-bag evaluation estimates an ensemble's performance using the bootstrap samples excluded from training each individual model.
- similarity mapA similarity map is a molecular visualization that assigns atom-level weights according to how much each atom affects a model prediction when its contribution is ablated.
- class imbalanceClass imbalance is a dataset condition in which some outcome values or classes occur much more frequently than others.
- SMOTESMOTE is an oversampling technique that creates synthetic minority-class instances along line segments connecting existing minority-class instances.
- class-balanced lossA class-balanced loss is a training loss whose contributions are weighted to compensate for unequal class frequencies.
- cost matrixA cost matrix specifies the cost assigned to each combination of predicted class and actual class.
- cost-sensitive learningCost-sensitive learning incorporates unequal costs for different prediction errors into model training.
- data augmentationData augmentation is the creation or transformation of training examples to increase the effective diversity of a dataset.
- undersamplingRandom undersampling is a sampling technique that randomly removes instances from the majority class to reduce class imbalance.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.
