Enum MetricType

Namespace

AiDotNet.Enums

Assembly

AiDotNet.dll

Defines the types of metrics used to evaluate machine learning models.

public enum MetricType

Fields

AIC = 24

A measure of the relative quality of statistical models for a given set of data.

For Beginners: Akaike Information Criterion (AIC) helps you compare different models and choose the best one. It balances how well the model fits the data against how complex the model is. A lower AIC is better. It's like choosing a car - you want one that's fast (fits the data well) but also fuel-efficient (not too complex). AIC helps you find that balance.

AICAlt = 62

Alternative Akaike Information Criterion - A variant of AIC with a different penalty term.

For Beginners: AICAlt is another version of AIC that uses a slightly different approach to penalize model complexity. It's particularly useful when sample sizes are small. Like AIC and BIC, lower values are better.

AUC = 64

Alias of AUCROC. Area Under the Curve - Measures the area under the ROC curve for classification models.

For Beginners: AUC (Area Under the Curve) is another name for AUCROC. It measures how well your classification model can distinguish between classes. Values range from 0 to 1: - 0.5 means random guessing - 1.0 means perfect classification Higher AUC values indicate better model performance at separating classes.

AUCPR = 63

Area Under the Precision-Recall Curve - Measures classification accuracy focusing on positive cases.

For Beginners: AUCPR is especially useful for imbalanced classification problems (where one class is rare). It ranges from 0 to 1, with higher values indicating better performance.

Precision measures how many of your positive predictions were correct. Recall measures what fraction of actual positives your model identified. AUCPR considers how these trade off across different threshold settings.

AUCROC = 64

Area Under the Receiver Operating Characteristic Curve - Measures classification accuracy across thresholds.

For Beginners: AUCROC is a common metric for classification models. It ranges from 0 to 1: - 0.5 means the model is no better than random guessing - 1.0 means perfect classification - Values below 0.5 suggest the model is worse than random

It measures how well your model can distinguish between classes across different threshold settings.

AbsoluteRelativeError = 97

Accuracy = 5

The proportion of correct predictions among all predictions made.

For Beginners: Accuracy simply measures what percentage of your predictions were exactly right. For example, if your model made 100 predictions and got 85 correct, the accuracy is 85%. This metric is most useful when all types of errors are equally important and your data is balanced.

AdjustedR2 = 1

A modified version of R² that accounts for the number of predictors in the model.

For Beginners: Adjusted R² is similar to R², but it penalizes you for adding too many input variables that don't help much. This prevents "overfitting" - when your model becomes too complex and starts memorizing the training data rather than learning general patterns. Use this instead of regular R² when comparing models with different numbers of input variables.

AdjustedRandIndex = 86

Measures the similarity between two clusterings adjusted for chance.

For Beginners: Adjusted Rand Index (ARI) compares two different ways of grouping the same data and measures how similar they are, while accounting for random chance. It ranges from -1 to 1: - 1 means the two clusterings are identical - 0 means the similarity is what you'd expect from random clustering - Negative values mean the clusterings are worse than random

This is useful for:

• Comparing your clustering results to known "ground truth" labels
• Evaluating how stable your clustering algorithm is across different runs
• Comparing different clustering algorithms on the same data

For example, if you cluster customer data and want to see how well it matches predefined customer segments, ARI tells you how similar your automated clustering is to the manual segmentation, beyond what random grouping would achieve.

AverageEpisodeReward = 87

The average total reward per episode (reinforcement learning).

For Beginners: In reinforcement learning, an "episode" is one full run in the environment. This metric measures how much reward the agent earns on average per episode. Higher is better.

AveragePrecision = 35

Measures the average precision across different recall levels.

For Beginners: Average Precision summarizes the precision-recall curve as a single number. It's like getting an overall grade that takes into account how well your model performs at different levels of strictness. A higher average precision means your model is good at identifying positive instances without raising too many false alarms, even as you adjust how strict it is. It's particularly useful in tasks like information retrieval, where you want to know how well your model ranks relevant items.

BIC = 25

A criterion for model selection that penalizes model complexity more strongly than AIC.

For Beginners: Bayesian Information Criterion (BIC) is similar to AIC, but it's stricter about model complexity. It's useful when you have a lot of data and want to avoid overfitting. If AIC is like choosing a car based on speed and fuel efficiency, BIC is like also considering the car's price more heavily. It helps you find a model that's good enough without being unnecessarily complex.

BhattacharyyaDistance = 43

Measures the similarity between two probability distributions.

For Beginners: Bhattacharyya Distance measures the similarity of two probability distributions. It's used in classification, feature selection, and image processing. A distance of 0 means the distributions are identical, while larger values indicate more different distributions. It's like comparing two different brands of the same product - how similar are their characteristics?

BrierScore = 28

A proper score function that measures the accuracy of probabilistic predictions.

For Beginners: Brier Score is used for probabilistic predictions, especially in classification problems. It's like a "penalty" for how far off your probabilities are. If you predict a 70% chance of rain and it rains, you get a small penalty; if you predict a 70% chance and it doesn't rain, you get a larger penalty. The lower the Brier Score, the better your predictions.

CRPS = 98

Continuous Ranked Probability Score for probabilistic forecasts.

For Beginners: CRPS measures how well a probabilistic forecast matches actual observations. Unlike point forecast metrics (MAE, RMSE), CRPS evaluates the entire predicted probability distribution.

Think of it this way:

• A weather forecast that says "30% chance of rain" can't be evaluated with just MAE
• CRPS rewards forecasts that assign high probability to what actually happens
• Lower CRPS values indicate better probabilistic predictions

CRPS generalizes MAE to probabilistic forecasts. When the forecast is deterministic (a single point), CRPS equals MAE. This makes it ideal for evaluating models like DeepAR, TFT, and other probabilistic time series forecasters.

The metric is commonly used in:

• Weather forecasting evaluation
• Energy demand prediction
• Financial risk assessment
• Any scenario where uncertainty quantification matters

CalibrationError = 36

Measures how well the predicted probabilities match the actual outcomes.

For Beginners: Calibration Error measures how well your model's predicted probabilities match the actual outcomes. A well-calibrated model with a 70% confidence should be correct about 70% of the time. Lower calibration error is better. It's like checking if a weather forecast is reliable - if it says there's a 30% chance of rain, it should actually rain on about 30% of such days. This is important when you need to trust the confidence levels your model provides, not just its final decisions.

CalinskiHarabaszIndex = 81

A measure of how similar an object is to its own cluster compared to other clusters.

For Beginners: The Calinski-Harabasz Index measures how well-defined your clusters are. It compares the average between-cluster distance to the average within-cluster distance. Higher values indicate better-defined clusters, like well-separated groups in a crowd.

ChamferDistance = 88

Chamfer Distance for point cloud similarity measurement.

For Beginners: Chamfer Distance measures how similar two point clouds are by computing the average distance from each point in one cloud to its nearest neighbor in the other. Lower values indicate better similarity. Used for 3D reconstruction, point cloud completion, and generative 3D model evaluation.

CohenKappa = 38

Measures the agreement between two sets of labels.

For Beginners: Cohen's Kappa measures how much better your model is at classification compared to random guessing. It's especially useful when your classes are imbalanced. A score of 1 indicates perfect agreement, 0 indicates no better than random chance. It's like comparing a student's test answers to the correct answers, but also taking into account how many they might get right just by guessing.

ConditionNumber = 67

Measures the sensitivity of a system to small changes in input.

For Beginners: The Condition Number tells you how much the output of a system can change for a small change in the input. A high condition number means the system is very sensitive to small changes, which can lead to numerical instability. It's like measuring how wobbly a table is - a high condition number means even a small nudge could make things very unstable.

CosineSimilarity = 75

A measure of similarity between two non-zero vectors.

For Beginners: Cosine Similarity measures the cosine of the angle between two vectors. It's like comparing the directions two arrows are pointing, regardless of their length. A value of 1 means they point in the same direction, 0 means they're perpendicular, and -1 means opposite directions.

CrossEntropyLoss = 30

Measures the average log-loss across all classes in classification problems.

For Beginners: Cross-Entropy Loss is like a penalty system for your model's confidence in its predictions. If your model is very confident about a correct prediction, it gets a small penalty. But if it's very confident about a wrong prediction, it gets a big penalty. It's useful in classification problems, especially when you care about how certain your model is. Lower values are better. It's like grading a test where you not only mark answers right or wrong, but also consider how sure the student was about each answer.

DaviesBouldinIndex = 82

A metric for evaluating clustering algorithms.

For Beginners: The Davies-Bouldin Index measures how similar each cluster is to its most similar other cluster. It's like measuring how distinct each group in a party is from the group it's most likely to be confused with. Lower values indicate better clustering, with more distinct, well-separated clusters.

DurbinWatsonStatistic = 59

Durbin-Watson Statistic - Detects autocorrelation in prediction errors.

For Beginners: DurbinWatsonStatistic helps identify if there are patterns in your prediction errors over time. Values range from 0 to 4: - Values near 2 suggest no autocorrelation (good) - Values toward 0 suggest positive autocorrelation (errors tend to be followed by similar errors) - Values toward 4 suggest negative autocorrelation (errors tend to be followed by opposite errors)

Autocorrelation in errors suggests your model might be missing important patterns in the data.

EarthMoversDistance = 89

Earth Mover's Distance (Wasserstein) for point cloud comparison.

For Beginners: Earth Mover's Distance measures the minimum "work" needed to transform one point cloud into another, where work is defined as moving mass times distance. Lower values indicate better similarity. More computationally expensive than Chamfer Distance but provides a true metric on probability distributions.

EffectiveNumberOfParameters = 72

The effective number of parameters in a model, useful for model complexity assessment.

For Beginners: The Effective Number of Parameters measures how complex your model really is. It's like counting the number of knobs you can turn to adjust your model, but some knobs might be more important than others. This helps you understand if your model is overly complicated, which could lead to overfitting.

EuclideanDistance = 73

The straight-line distance between two points in Euclidean space.

For Beginners: Euclidean Distance is the "as the crow flies" distance between two points. If you plot your data points on a graph, it's the length of the straight line you'd draw between them. It's used to measure how different or similar data points are to each other.

ExplainedVarianceScore = 2

Measures the proportion of variance in the dependent variable explained by the model.

For Beginners: Explained Variance Score measures how much of the variation in your data is captured by your model. Like R², it ranges from 0 to 1, with higher values being better. The main difference is that this metric focuses purely on variance explained, while R² also considers how far predictions are from the actual values.

F1Score = 8

The harmonic mean of precision and recall, providing a balance between the two metrics.

For Beginners: F1 Score combines precision and recall into a single number. It's useful when you need to balance between not missing positives (recall) and not incorrectly flagging negatives (precision). The score ranges from 0 to 1, with higher values being better. It's especially useful when your data has an uneven distribution of classes.

F2Score = 32

Measures the harmonic mean of precision and recall, with more weight on recall.

For Beginners: F2 Score is similar to F1 Score, but it gives more importance to recall than precision. This is useful when the cost of false negatives is higher than false positives. For example, in medical testing, you might prefer to have some false alarms (false positives) rather than miss any actual cases of a disease (false negatives). F2 Score ranges from 0 to 1, with higher values being better. It's like a grading system that cares more about not missing any important information than about including some extra, unnecessary details.

FBetaScore = 34

FirstQuartile = 49

The first quartile (25th percentile) of the dataset.

For Beginners: The first quartile is the value where 25% of your data falls below it.

If you divide your sorted data into four equal parts, the first quartile is the value at the boundary of the first and second parts.

For example, in a class of 20 students, the first quartile test score is the score that 5 students scored below and 15 students scored above.

GMean = 33

Measures the geometric mean of precision and recall.

For Beginners: G-Mean (Geometric Mean) balances precision and recall by multiplying them together and then taking the square root. It's useful when you have imbalanced classes and want to give equal importance to the performance on both the majority and minority classes. A high G-Mean indicates that your model is performing well on both classes. It's like judging a decathlon athlete - you want someone who's good at all events, not just exceptional in one or two.

HammingDistance = 77

The number of positions at which the corresponding symbols are different.

For Beginners: Hamming Distance counts how many changes you need to make to turn one string into another. It's like comparing two versions of a document and counting how many characters are different. It's often used in error detection and correction in data transmission.

InterquartileRange = 19

The range of the middle 50% of the data.

For Beginners: Interquartile Range (IQR) is like range, but it focuses on the middle 50% of your data. It's the difference between the 75th percentile (Q3) and the 25th percentile (Q1). IQR is useful because it's not affected by outliers. It gives you an idea of where the "bulk" of your data lies. For example, if Q1 is 20 and Q3 is 30, the IQR is 10.

IoU3D = 90

Intersection over Union for 3D volumetric segmentation.

For Beginners: 3D IoU measures the overlap between predicted and ground truth volumes. It's the 3D extension of the 2D IoU metric used in image segmentation. Higher values (closer to 1) indicate better segmentation accuracy.

JaccardSimilarity = 76

A statistic used for comparing the similarity and diversity of sample sets.

For Beginners: Jaccard Similarity measures how similar two sets are by comparing what they have in common to what they have in total. It's like comparing two people's music libraries - how many songs do they both have compared to their total unique songs? A value of 1 means the sets are identical, and 0 means they have nothing in common.

KLDivergence = 41

Measures the difference between two probability distributions.

For Beginners: KL Divergence measures how one probability distribution differs from another. It's often used in machine learning to compare models or optimize algorithms. Lower values indicate more similar distributions. It's like comparing two recipes for the same dish - how different are the ingredients and their proportions?

KendallTau = 12

Measures the ordinal association between predicted and actual values.

For Beginners: Kendall Tau measures how well your model preserves the correct ordering of values. It compares every possible pair of data points and checks if your model predicts the same relationship (is A greater than B, less than B, or equal to B?). This is useful when you care more about getting the ranking right than the exact values. For example, in a recommendation system, you might care more about showing the most relevant items first, rather than predicting exact relevance scores.

Kurtosis = 45

LPIPS = 93

Learned Perceptual Image Patch Similarity for deep perceptual quality.

For Beginners: LPIPS uses deep neural networks to measure perceptual similarity between images. Lower values indicate more similar images. Often better at matching human judgment than traditional metrics like PSNR or SSIM.

LevenshteinDistance = 42

Measures the similarity between two sequences.

For Beginners: Levenshtein Distance counts the minimum number of single-character edits (insertions, deletions, or substitutions) required to change one word into another. It's useful in spell checking, DNA sequence analysis, and more. Lower values indicate more similar sequences. It's like counting how many changes you need to make to turn one word into another.

Likelihood = 26

A measure of how probable the observed data is under the model.

For Beginners: Likelihood measures how well your model explains the observed data. A higher likelihood means your model is more consistent with the data you've observed. It's like measuring how well a theory explains the evidence. However, likelihood values can be very small and hard to interpret directly, which is why we often use log-likelihood instead.

LogLikelihood = 27

The natural logarithm of the likelihood, used for numerical stability and easier interpretation.

For Beginners: Log-Likelihood is the natural logarithm of the Likelihood. We use it because likelihood values can be extremely small and hard to work with. Log-Likelihood turns these tiny numbers into more manageable negative numbers. Higher (closer to zero) is better. It's like using the Richter scale for earthquakes - it makes it easier to compare and work with a wide range of values.

LogPointwisePredictiveDensity = 68

A measure of the model's predictive accuracy on a point-by-point basis.

For Beginners: Log Pointwise Predictive Density (LPPD) measures how well your model predicts each individual data point. It's like grading a test where each question is scored separately, and then all scores are combined. A higher LPPD means your model is doing a better job at predicting individual data points.

MAD = 51

The median absolute deviation (MAD) of the dataset.

For Beginners: MAD is another way to measure spread that's less affected by outliers.

To calculate MAD:

1. Find the median of your data
2. Calculate how far each value is from the median (absolute deviation)
3. Find the median of those distances

MAD is useful when your data has outliers that might skew other measures like standard deviation. Think of it as measuring the "typical" distance from the center, ignoring extreme values.

MAE = 52

Mean Absolute Error - The average absolute difference between predicted and actual values.

For Beginners: MAE measures the average size of errors without considering their direction (positive or negative). Lower values indicate better accuracy. If MAE = 5, your predictions are off by 5 units on average.

MAPE = 55

Mean Absolute Percentage Error - The average percentage difference between predicted and actual values.

For Beginners: MAPE expresses error as a percentage, which helps you understand the relative size of errors. MAPE = 10 means that, on average, your predictions are off by 10% from the actual values. Note: MAPE can be problematic when actual values are close to zero.

MSE = 53

Mean Squared Error - The average of squared differences between predicted and actual values.

For Beginners: MSE squares the errors before averaging them, which penalizes large errors more heavily than small ones. Lower values indicate better accuracy. Because of squaring, the value is not in the same units as your data.

MahalanobisDistance = 78

A multi-dimensional generalization of measuring how many standard deviations away a point is from the mean of a distribution.

For Beginners: Mahalanobis Distance measures how many standard deviations a point is from the mean of a distribution. It's like measuring how unusual a data point is, taking into account the overall pattern of the data. It's useful for detecting outliers in multi-dimensional data.

ManhattanDistance = 74

The sum of the absolute differences of the coordinates.

For Beginners: Manhattan Distance is like measuring the distance a taxi would drive in a city with a grid layout. Instead of going directly ("as the crow flies"), it's the sum of the distances along each dimension. It's useful when you can't move diagonally through your data space.

MarginalLikelihood = 70

The probability of the observed data under all possible parameter values.

For Beginners: Marginal Likelihood is like asking "How likely is this data, considering all possible explanations?" It's used to compare different models, taking into account both how well they fit the data and how complex they are. A higher marginal likelihood suggests a better balance between fitting the data and keeping the model simple.

Max = 47

The largest value in the dataset.

For Beginners: The maximum is simply the largest number in your data.

For the numbers [5, 12, 3, 8, 9], the maximum is 12.

Together with the minimum, it tells you the full range of your data.

MaxError = 31

Measures the maximum error between predicted and actual values.

For Beginners: Max Error is simply the largest mistake your model makes. It tells you the worst-case scenario for your predictions. For example, if your model predicts house prices and the Max Error is $50,000, you know that your model's biggest mistake was being off by $50,000. This is useful when you really want to avoid large errors, even if they're rare. It's like focusing on the lowest grade in a class - it doesn't tell you about typical performance, but it shows you the biggest problem area.

Mean = 13

The arithmetic average of a set of values.

For Beginners: Mean is what most people call the "average". You calculate it by adding up all the numbers and dividing by how many there are. It gives you a sense of the typical value in your data. For example, if you have test scores of 80, 85, 90, 95, and 100, the mean is 90.

MeanAbsoluteError = 20

The average of the absolute differences between predicted and actual values.

For Beginners: Mean Absolute Error (MAE) measures how far off your predictions are on average, ignoring whether they're too high or too low. It's easy to understand: if your MAE is 5, it means your predictions are off by 5 units on average. MAE treats all errors equally, unlike some metrics that penalize large errors more heavily.

MeanAbsolutePercentageError = 23

The average of the absolute percentage differences between predicted and actual values.

For Beginners: Mean Absolute Percentage Error (MAPE) measures the average percentage difference between predicted and actual values. It's useful when you want to know the typical size of errors relative to the actual values. For example, a MAPE of 5% means that, on average, your predictions are off by 5% of the actual value. However, MAPE can be misleading when actual values are close to zero.

MeanAveragePrecision = 83

The mean of the average precision scores for each query.

For Beginners: Mean Average Precision is often used in information retrieval to evaluate search algorithms. It's like measuring how good a search engine is at putting the most relevant results at the top of the list. A higher MAP indicates that relevant items are generally ranked higher in the search results.

MeanBiasError = 56

Mean Bias Error - The average of prediction errors (predicted - actual).

For Beginners: MeanBiasError helps determine if your model tends to overestimate (positive value) or underestimate (negative value). Ideally, it should be close to zero, indicating no systematic bias. Unlike MAE, this doesn't take the absolute value, so positive and negative errors can cancel out.

MeanIoU = 94

Mean Intersection over Union across all classes.

For Beginners: mIoU averages the IoU score across all classes in a segmentation task. Used for semantic segmentation of point clouds, meshes, and volumes. Higher values indicate better overall segmentation quality.

MeanPredictionError = 3

The average difference between predicted values and actual values.

For Beginners: Mean Prediction Error simply calculates the average difference between what your model predicted and what the actual values were. A lower value is better. This metric helps you understand if your model tends to overestimate or underestimate the results, as positive and negative errors don't cancel each other out.

MeanReciprocalRank = 85

The average of the reciprocal ranks of the first relevant items.

For Beginners: Mean Reciprocal Rank measures how high the first correct answer appears in a list of results. It's like measuring how quickly a search engine gives you the answer you're looking for. A higher MRR means the correct answer tends to appear earlier in the list of results.

MeanSquaredError = 21

The average of the squared differences between predicted and actual values.

For Beginners: Mean Squared Error (MSE) is similar to MAE, but it squares the errors before averaging them. This means it penalizes large errors more than small ones. For example, being off by 2 is more than twice as bad as being off by 1. MSE is often used in optimization because it's mathematically convenient, but its units are squared, which can be hard to interpret.

MeanSquaredLogError = 66

Mean Squared Logarithmic Error - Penalizes underestimates more than overestimates.

For Beginners: MeanSquaredLogError is useful when you care more about relative errors than absolute ones. It's calculated by applying logarithms to actual and predicted values before computing MSE.

MSLE penalizes underestimation (predicting too low) more heavily than overestimation. This is useful in scenarios where underestimating would be more problematic, like inventory forecasting.

Median = 14

The middle value in a sorted list of numbers.

For Beginners: Median is the middle number when all your values are sorted. If you have an odd number of values, it's the middle one. If you have an even number, it's the average of the two middle numbers. It's useful when you have some extreme values that might skew the mean. For example, in the list 1, 3, 3, 6, 7, 8, 100, the median is 6.

MedianAbsoluteError = 57

Median Absolute Error - The middle value of all absolute differences between predicted and actual values.

For Beginners: Unlike MAE which uses the average, MedianAbsoluteError uses the middle value of all absolute errors. This makes it less sensitive to outliers (extreme errors) than MAE. For example, if you have errors of [1, 2, 100], the median is 2, while the mean would be 34.3.

MedianPredictionError = 4

The middle value of all differences between predicted values and actual values.

For Beginners: Median Prediction Error finds the middle value of all the prediction errors when sorted. Unlike the mean, the median isn't affected by extreme outliers, so it gives you a more robust measure of your model's typical error when some predictions are way off.

Min = 46

The smallest value in the dataset.

For Beginners: The minimum is simply the smallest number in your data.

For the numbers [5, 12, 3, 8, 9], the minimum is 3.

It's useful to know the extreme values in your data, especially when looking for outliers or setting valid ranges.

Mode = 15

The most frequently occurring value in a dataset.

For Beginners: Mode is the value that appears most often in your data. It's especially useful for categorical data (like favorite colors) or when you have distinct groups in your data. For example, in the list 2, 3, 3, 4, 5, 5, 5, 6, the mode is 5.

MutualInformation = 39

Measures the mutual dependence between two variables.

For Beginners: Mutual Information measures how much knowing one variable reduces uncertainty about the other. It's useful for feature selection and understanding relationships between variables. Higher values indicate stronger relationships. It's like measuring how much knowing a student's grade in one subject tells you about their grade in another subject.

N = 48

The number of values in the dataset.

For Beginners: N is just the count of how many numbers are in your data.

For the numbers [5, 12, 3, 8, 9], N is 5.

The sample size is important because larger samples generally give more reliable statistics.

NormalConsistency = 96

Normal consistency for surface reconstruction quality.

For Beginners: Normal Consistency measures how well the predicted surface normals match the ground truth. Higher values indicate better surface reconstruction quality. Important for mesh reconstruction and NeRF geometry evaluation.

NormalizedDiscountedCumulativeGain = 84

A measure of ranking quality used in information retrieval.

For Beginners: Normalized Discounted Cumulative Gain (NDCG) measures the quality of ranking in search results. It's like grading a search engine on how well it puts the most relevant results at the top. It takes into account that users are less likely to look at results further down the list.

NormalizedMutualInformation = 79

A normalization of the Mutual Information to scale the results between 0 and 1.

For Beginners: Normalized Mutual Information measures how much knowing one variable reduces uncertainty about another. It's like measuring how much knowing someone's job tells you about their income, but adjusted so it's always between 0 and 1. This makes it easier to compare across different datasets.

ObservedTestStatistic = 69

The value of the test statistic computed from the observed data.

For Beginners: The Observed Test Statistic is a number calculated from your data that you use to test a hypothesis. It's like measuring how far your results are from what you'd expect if your hypothesis were false. You compare this number to a threshold to decide if your results are statistically significant.

PSNR = 91

Peak Signal-to-Noise Ratio for rendered image quality.

For Beginners: PSNR measures the quality of reconstructed or rendered images compared to a reference. Higher values indicate better quality. Commonly used to evaluate NeRF and other neural rendering methods. Values above 30 dB are generally considered good quality.

PartIoU = 95

Part-averaged IoU for part segmentation tasks.

For Beginners: Part IoU is specifically designed for object part segmentation tasks, averaging IoU over different parts of objects (e.g., chair legs, back, seat).

PearsonCorrelation = 10

Measures the linear correlation between predicted and actual values.

For Beginners: Pearson Correlation measures how well the relationship between your predictions and actual values can be described with a straight line. It ranges from -1 to 1, where: - 1 means perfect positive correlation (when actual values increase, predictions increase) - 0 means no correlation - -1 means perfect negative correlation (when actual values increase, predictions decrease) A high positive value indicates your model is capturing the right patterns, even if the exact values differ.

Perplexity = 40

Measures the log-likelihood of held-out data under a topic model.

For Beginners: Perplexity is often used in natural language processing to evaluate how well a model predicts a sample. Lower perplexity is better. It's like measuring how surprised a model is by new text - a good model should find typical text unsurprising (low perplexity).

PopulationStandardError = 61

Population Standard Error - The standard deviation of prediction errors without adjustment for model complexity.

For Beginners: PopulationStandardError measures how much prediction errors typically vary, but unlike SampleStandardError, it doesn't adjust for model complexity. It gives you an idea of the typical size of the errors your model makes.

Precision = 6

The proportion of true positive predictions among all positive predictions.

For Beginners: Precision measures how many of the items your model identified as positive were actually positive. For example, if your spam filter marked 10 emails as spam, but only 8 were actually spam, your precision is 80%. High precision means few false positives - you're not incorrectly flagging things that are actually negative.

PredictionIntervalCoverage = 9

The percentage of actual values that fall within the model's prediction intervals.

For Beginners: Prediction Interval Coverage checks if your model's uncertainty estimates are reliable. Instead of just making a single prediction, some models provide a range (like "between 10-15 units"). This metric tells you what percentage of actual values fall within these predicted ranges. Ideally, a 95% prediction interval should contain the actual value 95% of the time.

R2 = 0

Coefficient of determination, measuring how well the model explains the variance in the data.

For Beginners: R² (R-squared) tells you how well your model fits the data, on a scale from 0 to 1. A value of 1 means your model perfectly predicts the data, while 0 means it's no better than just guessing the average value. For example, an R² of 0.75 means your model explains 75% of the variation in the data.

RMSE = 54

Root Mean Squared Error - The square root of the Mean Squared Error.

For Beginners: RMSE converts MSE back to the original units of your data by taking the square root. It's often preferred over MSE for interpretation because it's in the same units as your data. Like MAE, lower values indicate better accuracy.

ROCAUCScore = 29

Measures the trade-off between true positive rate and false positive rate.

For Beginners: ROC AUC Score measures how well your model can distinguish between classes. It's like measuring how good a bouncer is at a club. A perfect bouncer (ROC AUC = 1) lets in all the VIPs and keeps out all the non-VIPs. A bouncer who just guesses randomly (ROC AUC = 0.5) is no better than flipping a coin. The higher the ROC AUC, the better your model is at telling the classes apart.

RSquared = 0

Alias of R2. R-Squared (R²) - Coefficient of determination measuring model fit quality.

For Beginners: RSquared is another name for R². It measures how well your model explains the variance in the data on a scale from 0 to 1. A higher value means better fit. For example, RSquared = 0.80 means your model explains 80% of the variation in the target variable.

Range = 18

The difference between the largest and smallest values in a dataset.

For Beginners: Range is simply the difference between the highest and lowest values in your data. It gives you a quick idea of how spread out your data is, but it can be misleading if you have outliers (extreme values). For example, if your data is 1, 2, 3, 4, 100, the range is 99, even though most of the values are close together.

Recall = 7

The proportion of true positive predictions among all actual positives.

For Beginners: Recall measures how many of the actual positive items your model correctly identified. For example, if there were 20 spam emails, and your filter caught 15 of them, your recall is 75%. High recall means few false negatives - you're not missing things that should be flagged as positive.

ReferenceModelMarginalLikelihood = 71

The marginal likelihood of a reference or null model.

For Beginners: The Reference Model Marginal Likelihood is the marginal likelihood of a simpler, baseline model. It's used as a point of comparison for your main model. If your main model's marginal likelihood is much higher, it suggests your model is substantially better than the basic reference model.

RootMeanSquaredError = 22

The square root of the mean squared error.

For Beginners: Root Mean Squared Error (RMSE) is the square root of MSE. It brings the error metric back to the same units as your original data, making it easier to interpret than MSE. Like MSE, it penalizes large errors more than small ones. RMSE is often preferred in practice because it's both mathematically convenient and interpretable.

SMAPE = 65

Symmetric Mean Absolute Percentage Error - A variant of MAPE that handles zero or near-zero values better.

For Beginners: SMAPE is similar to MAPE but uses a different formula that handles cases where actual values are zero or very small. It's bounded between 0% and 200%, with lower values indicating better performance.

SMAPE treats positive and negative errors more symmetrically than MAPE, which can be important in some forecasting applications.

SSIM = 92

Structural Similarity Index for perceptual image quality.

For Beginners: SSIM measures perceived visual quality by comparing luminance, contrast, and structure between images. Values range from -1 to 1, with 1 indicating perfect similarity. Often considered more aligned with human perception than PSNR.

SampleStandardError = 60

Sample Standard Error - An estimate of the standard deviation of prediction errors, adjusted for model complexity.

For Beginners: SampleStandardError estimates how much prediction errors typically vary, taking into account how many parameters (features) your model uses. It's useful for constructing confidence intervals around predictions and is adjusted downward based on the number of parameters in your model.

SilhouetteScore = 37

Measures the quality of clustering algorithms.

For Beginners: Silhouette Score measures how well each item fits into its assigned cluster compared to other clusters. It ranges from -1 to 1, where a high value indicates that the object is well matched to its own cluster and poorly matched to neighboring clusters. It's like measuring how well students are grouped in classes - are students in each class similar to each other and different from students in other classes?

Skewness = 44

SpearmanCorrelation = 11

Measures the monotonic relationship between predicted and actual values.

For Beginners: Spearman Correlation is similar to Pearson, but it measures whether predictions and actual values increase or decrease together, without requiring a straight-line relationship. It works by ranking the values and then comparing the ranks. This makes it useful when your data has outliers or when the relationship isn't strictly linear but still follows a pattern. Like Pearson, it ranges from -1 to 1.

StandardDeviation = 17

The square root of the variance, measuring the average deviation from the mean.

For Beginners: Standard Deviation is like Variance, but easier to understand because it's in the same units as your original data. It tells you, on average, how far each value is from the mean. About 68% of your data falls within one standard deviation of the mean, 95% within two, and 99.7% within three. This helps you understand how spread out your data is in practical terms.

TheilUStatistic = 58

Theil's U Statistic - A measure of forecast accuracy relative to a naive forecasting method.

For Beginners: TheilUStatistic compares your model's accuracy to a simple "no-change" prediction. Values less than 1 mean your model is better than the naive approach. Values equal to 1 mean your model performs the same as the naive approach. Values greater than 1 mean your model performs worse than the naive approach. This is especially useful for time series forecasting evaluation.

ThirdQuartile = 50

The third quartile (75th percentile) of the dataset.

For Beginners: The third quartile is the value where 75% of your data falls below it.

If you divide your sorted data into four equal parts, the third quartile is the value at the boundary of the third and fourth parts.

For example, in a class of 20 students, the third quartile test score is the score that 15 students scored below and 5 students scored above.

Variance = 16

A measure of variability in the data, calculated as the average squared deviation from the mean.

For Beginners: Variance measures how spread out your data is. A low variance means the numbers are clustered close to the average, while a high variance means they're more spread out. It's calculated by finding the average of the squared differences from the Mean. Variance helps you understand the distribution of your data, but it's in squared units, which can be hard to interpret.

VariationOfInformation = 80

A measure of the amount of information lost when compressing two random variables.

For Beginners: Variation of Information measures how much information is lost when you group data. It's like measuring how much detail you lose when you summarize a long story. Lower values mean you've kept more of the original information in your grouping.