Metrics Glossary#
This section contains guides for different metrics used to measure model performance.
Each ML use case requires different metrics. Using the right metrics is critical for understanding and meaningfully comparing model performance. In each metrics guide, you can learn about the metric with examples, its limitations and biases, and its intended uses.
-
Accuracy measures how well a model predicts correctly. It's a good metric for assessing model performance in simple cases with balanced data.
-
Average precision summarizes a precision-recall (PR) curve into a single threshold-independent value representing model's performance across all thresholds.
-
Averaging Methods: Macro, Micro, Weighted
Different averaging methods for aggregating metrics for multiclass workflows, such as classification and object detection.
-
Coefficient of Determination (R²)
R² provides a scale that is generally from 0 to 1, where higher values indicate a better fit and imply that the model can better explain the variation of the model inferences.
-
Cohen's Kappa measures the agreement between model inferences and ground truths. It's a good performance metric that considers class imbalance.
-
Confusion matrix is a structured plot describing classification model performance as a table that highlights counts of objects with predicted classes (columns) against the actual classes (rows), indicating how confused a model is.
-
F1-score is a metric that combines two competing metrics, precision and recall with an equal weight. It symmetrically represents both precision and recall as one metric.
-
False positive rate (FPR) measures the proportion of negative ground truths that a model incorrectly predicts as positive, ranging from 0 to 1. It is useful when the objective is to measure and reduce false positive inferences.
-
MAE measures the average magnitude of errors in predictions, without considering their direction. It represents the mean of the absolute differences between predicted and actual values across a dataset, treating each discrepancy equally.
-
MSE measures the average squared difference between the predicted values and the actual values, emphasizing larger errors more significantly than smaller ones. It is mainly different from MAE in that larger errors are penalized more heavily.
-
Precision measures the proportion of positive inferences from a model that are correct. It is useful when the objective is to measure and reduce false positive inferences.
-
Precision-recall curve is a plot that gauges machine learning model performance by using precision and recall. It is built with precision on the y-axis and recall on the x-axis computed across many thresholds.
-
Recall, also known as true positive rate (TPR) and sensitivity, measures the proportion of all positive ground truths that a model correctly predicts. It is useful when the objective is to measure and reduce false negative ground truths, i.e. model misses.
-
Receiver Operating Characteristic (ROC) Curve
A receiver operating characteristic (ROC) curve is a plot that is used to evaluate the performance of binary classification models by using the true positive rate (TPR) and the false positive rate (FPR).
-
Root Mean Squared Error (RMSE)
RMSE measures the square root of the average squared difference between the predicted values and the actual values, emphasizing larger errors more significantly than smaller ones while maintaing the same unit as the predicted and actual values which is what differentiates it from MSE.
-
Specificity, also known as true negative rate (TNR), measures the proportion of negative ground truths that a model correctly predicts, ranging from 0 to 1. It is useful when the objective is to measure the model's ability to correctly identify the negative class instances.
-
The counts of TP, FP, FN and TN ground truths and inferences are essential for summarizing model performance. They are the building blocks of many other metrics, including accuracy, precision, and recall.
Computer Vision#
-
Geometry matching is the process of matching inferences to ground truths for computer vision workflows with a localization component. It is a core building block for metrics such as TP, FP, and FN, and any metrics built on top of these, like precision, recall, and F1-score.
-
IoU measures overlap between two geometries, segmentation masks, sets of labels, or time-series snippets. Also known as Jaccard index in classification workflow.
Large Language Models#
-
The consistency score is a sampling-based approach that computes a score by aggregating across
nsample responses obtained by using the same promptntimes. The model is considered hallucinating if there exists sample responses that contradict one another or are not consistent. -
NLI classification is an insightful tool to measure if candidate texts and reference texts are contradictory or consistent, providing extra detail towards hallucination detection.
-
Hughes Hallucination Evaluation Model (HHEM) is an open-source model that can be used to compute scores for hallucination detection.
-
Prompt-based Hallucination Metric
Large Language Model (LLM) Prompt-based Metric involves using an LLM and various prompt engineering techniques to perform evaluations. These methods include chain-of-thought, self-consistency, and others to determine whether or not an inference text contains a hallucination.
Natural Language Processing#
-
BERTScore is a metric used in NLP workflows to measure textual similarity between candidate texts and reference texts.
-
BLEU is a metric commonly used in a variety of NLP workflows to evaluate the quality of candidate texts. BLEU can be thought of as an analog to precision for text comparisons.
-
Diarization Error Rate (DER) is an important metric used to evaluate speaker diarization systems. It quantifies the overall performance of a speaker diarization system by measuring the ratio of the duration of errors to the total ground truth speech duration.
-
METEOR is a widely recognized and vital metric used in NLP. It is used to measure the quality of candidate texts against reference texts. Though it is an n-gram based metric, it goes beyond traditional methods by factoring in elements such as precision, recall, and order to provide a comprehensive measure of text quality.
-
Perplexity is a metric commonly used in natural language processing to evaluate the quality of language models, particularly in the context of text generation. Unlike metrics such as BLEU or BERT, perplexity doesn't directly measure the quality of generated text by comparing it with reference texts. Instead, perplexity assesses the "confidence" or "surprise" of a language model in predicting the next word in a sequence of words.
-
ROUGE-N, a metric within the broader ROUGE metric collection, is a vital metric in the field of NLP. It assesses the quality of a candidate text by measuring the overlap of n-grams between the candidate text and reference texts. ROUGE-N can be thought of as an analog to recall for text comparisons.
-
Word, Character, and Match Error Rate
Word Error Rate (WER), Character Error Rate (CER), and Match Error Rate (MER) are essential metrics used in the evaluation of speech recognition and natural language processing systems. From a high level, they each quantify the similarity between reference and candidate texts, with zero being a perfect score.
Kolena Insights#
-
Statistical significance represents the amount of tolerance in your evaluation results based on a chosen confidence level and the size of your test case. Significance is derived by estimating the margin of error, which ranges from 0% to 100%. The larger the margin of error, the less confidence one should have that a result would reflect the result of a fully representative test set.