Logloss (or Logarithmic Loss) measures classification performance; specifically, uncertainty. This metric evaluates how closely a model’s predicted values are to the actual target value. For example, does a model tend to assign a high predicted value like .90 for the positive class, or does it show a poor ability to identify the positive class and assign a lower predicted value like .40? Logloss ranges between 0 and 1, with 0 meaning that the model correctly assigns a probability of 0% or 100%. Logloss is sensitive to low probabilities that are erroneous.
- Area Under the Curve or AUC
- Artificial Intelligence (AI)
- Artificial Neural Networks (ANNs)
- Big Data
- Black Box Algorithms
- Computer Vision
- Confusion Matrix
- Data Science
- Decision tree
- Deep Learning
- Embodied AI
- Ethical AI
- Explainable AI (XAI)
- Few-shot Learning
- Generative Adversarial Networks (GANs)
- Linear Algebra
- Machine learning
- Max F1
- Mean Absolute Error or MAE
- Mean Per Class Error
- Mean Square Error or MSE
- Natural Language Processing
- Pragmatic AI
- Predictive Analytics
- Reinforcement Learning
- Residual Deviance
- Root Mean Square Error or RMSE
- Root Mean Square Logarithmic Error or RMSLE
- Supervised Learning
- Transfer Learning
- Unsupervised Learning
- Variable Importance
- Weak AI
“A field of study that gives computers the ability to learn without being explicitly programmed.” (Arthur Samuel, 1959)
Max F1 is the point at which predictions will occur or not. When a row’s P1 value is at or above the Max F1, the predicted outcome will happen in the future. If a row’s PO value is below the Max F1, the predicted outcome will not occur. Thus the Max F1 is considered the cutoff point for probabilities to come true in the future or not. This cutoff point is not 50% as you might expect. If you see a prediction that has a probability of occurring (P1) at 90%, Seer may predict that outcome to not occur. This happens because the Max F1 rate is higher than the P1. The key concept to understand about the Max F1 is that Squark uses it to cut the probabilities. So in AI, a prediction may be more than 50% likely to occur and still not be predicted to occur. This AI metric is why you may not see 50/50 cutoff points for predicted probabilities.
MAE or the Mean Absolute Error is an average of the absolute errors. The smaller the MAE the better the model’s performance. The MAE units are the same units as your data’s dependent variable/target (so if that’s dollars, this is in dollars), which is useful for understanding whether the size of the error is meaningful or not. MAE is not sensitive to outliers. If your data has a lot of outliers, then examine the Root Mean Square Error (RMSE), which is sensitive to outliers.
Mean Per Class Error (in Multi-class Classification only) is the average of the errors of each class in your multi-class data set. This metric speaks toward misclassification of the data across the classes. The lower this metric, the better.
MSE is the Mean Square Error and is a model quality metric. Closer to zero is better. The MSE metric measures the average of the squares of the errors or deviations. MSE takes the distances from the points to the regression line (these distances are the “errors”) and then squares them to remove any negative signs. MSE incorporates both the variance and the bias of the predictor. MSE gives more weight to larger differences in errors than MAE.
The discipline within A.I. that deals with written and spoken language.
Pragmatic AI is designed to solve well-defined problems, as opposed to being allowed to seek its own purpose.
Statistical techniques gathered from predictive modeling, machine learning, and data mining that analyze current and historical facts to make predictions about future or otherwise unknown events.
Regression models are the mainstay of predictive analytics. The focus lies on establishing a mathematical equation as a model to represent the interactions between the different variables in consideration. Depending on the situation, there are a wide variety of models that can be applied while performing predictive analytics. Some of them are briefly discussed below.
Linear regression model
The linear regression model analyzes the relationship between the response or dependent variable and a set of independent or predictor variables. This relationship is expressed as an equation that predicts the response variable as a linear function of the parameters. These parameters are adjusted so that a measure of fit is optimized. Much of the effort in model fitting is focused on minimizing the size of the residual, as well as ensuring that it is randomly distributed with respect to the model predictions.
The goal of regression is to select the parameters of the model so as to minimize the sum of the squared residuals. This is referred to as ordinary least squares (OLS) estimation and results in best linear unbiased estimates (BLUE) of the parameters if and only if the Gauss-Markov assumptions are satisfied.
Once the model has been estimated we would be interested to know if the predictor variables belong in the model—i.e. is the estimate of each variable’s contribution reliable? To do this we can check the statistical significance of the model’s coefficients which can be measured using the t-statistic. This amounts to testing whether the coefficient is significantly different from zero. How well the model predicts the dependent variable based on the value of the independent variables can be assessed by using the R² statistic. It measures predictive power of the model i.e. the proportion of the total variation in the dependent variable that is “explained” (accounted for) by variation in the independent variables.
Discrete choice models
Multiple regression (above) is generally used when the response variable is continuous and has an unbounded range. Often the response variable may not be continuous but rather discrete. While mathematically it is feasible to apply multiple regression to discrete ordered dependent variables, some of the assumptions behind the theory of multiple linear regression no longer hold, and there are other techniques such as discrete choice models which are better suited for this type of analysis. If the dependent variable is discrete, some of those superior methods are logistic regression, multinomial logit and probit models. Logistic regression and probit models are used when the dependent variable is binary.
Main article: Logistic regression
In a classification setting, assigning outcome probabilities to observations can be achieved through the use of a logistic model, which is basically a method which transforms information about the binary dependent variable into an unbounded continuous variable and estimates a regular multivariate model (See Allison’s Logistic Regression for more information on the theory of logistic regression).
The Wald and likelihood-ratio test are used to test the statistical significance of each coefficient b in the model (analogous to the t tests used in OLS regression; see above). A test assessing the goodness-of-fit of a classification model is the “percentage correctly predicted”.
Multinomial logistic regression
An extension of the binary logit model to cases where the dependent variable has more than 2 categories is the multinomial logit model. In such cases collapsing the data into two categories might not make good sense or may lead to loss in the richness of the data. The multinomial logit model is the appropriate technique in these cases, especially when the dependent variable categories are not ordered (for examples colors like red, blue, green). Some authors have extended multinomial regression to include feature selection/importance methods such as random multinomial logit.
Probit models offer an alternative to logistic regression for modeling categorical dependent variables. Even though the outcomes tend to be similar, the underlying distributions are different. Probit models are popular in social sciences like economics.
A good way to understand the key difference between probit and logit models is to assume that the dependent variable is driven by a latent variable z, which is a sum of a linear combination of explanatory variables and a random noise term.
We do not observe z but instead observe y which takes the value 0 (when z < 0) or 1 (otherwise). In the logit model we assume that the random noise term follows a logistic distribution with mean zero. In the probit model we assume that it follows a normal distribution with mean zero. Note that in social sciences (e.g. economics), probit is often used to model situations where the observed variable y is continuous but takes values between 0 and 1. Logit versus probit The probit model has been around longer than the logit model. They behave similarly, except that the logistic distribution tends to be slightly flatter tailed. One of the reasons the logit model was formulated was that the probit model was computationally difficult due to the requirement of numerically calculating integrals. Modern computing however has made this computation fairly simple. The coefficients obtained from the logit and probit model are fairly close. However, the odds ratio is easier to interpret in the logit model. Practical reasons for choosing the probit model over the logistic model would be: There is a strong belief that the underlying distribution is normal The actual event is not a binary outcome (e.g., bankruptcy status) but a proportion (e.g., proportion of population at different debt levels). Time series models Time series models are used for predicting or forecasting the future behavior of variables. These models account for the fact that data points taken over time may have an internal structure (such as autocorrelation, trend or seasonal variation) that should be accounted for. As a result, standard regression techniques cannot be applied to time series data and methodology has been developed to decompose the trend, seasonal and cyclical component of the series. Modeling the dynamic path of a variable can improve forecasts since the predictable component of the series can be projected into the future. Time series models estimate difference equations containing stochastic components. Two commonly used forms of these models are autoregressive models (AR) and moving-average (MA) models. The Box–Jenkins methodology (1976) developed by George Box and G.M. Jenkins combines the AR and MA models to produce the ARMA (autoregressive moving average) model, which is the cornerstone of stationary time series analysis. ARIMA (autoregressive integrated moving average models), on the other hand, are used to describe non-stationary time series. Box and Jenkins suggest differencing a non-stationary time series to obtain a stationary series to which an ARMA model can be applied. Non-stationary time series have a pronounced trend and do not have a constant long-run mean or variance. Box and Jenkins proposed a three-stage methodology involving model identification, estimation and validation. The identification stage involves identifying if the series is stationary or not and the presence of seasonality by examining plots of the series, autocorrelation and partial autocorrelation functions. In the estimation stage, models are estimated using non-linear time series or maximum likelihood estimation procedures. Finally the validation stage involves diagnostic checking such as plotting the residuals to detect outliers and evidence of model fit. In recent years time series models have become more sophisticated and attempt to model conditional heteroskedasticity with models such as ARCH (autoregressive conditional heteroskedasticity) and GARCH (generalized autoregressive conditional heteroskedasticity) models frequently used for financial time series. In addition time series models are also used to understand inter-relationships among economic variables represented by systems of equations using VAR (vector autoregression) and structural VAR models. Survival or duration analysis Survival analysis is another name for time-to-event analysis. These techniques were primarily developed in the medical and biological sciences, but they are also widely used in the social sciences like economics, as well as in engineering (reliability and failure time analysis). Censoring and non-normality, which are characteristic of survival data, generate difficulty when trying to analyze the data using conventional statistical models such as multiple linear regression. The normal distribution, being a symmetric distribution, takes positive as well as negative values, but duration by its very nature cannot be negative and therefore normality cannot be assumed when dealing with duration/survival data. Hence the normality assumption of regression models is violated. The assumption is that if the data were not censored it would be representative of the population of interest. In survival analysis, censored observations arise whenever the dependent variable of interest represents the time to a terminal event, and the duration of the study is limited in time. An important concept in survival analysis is the hazard rate, defined as the probability that the event will occur at time t conditional on surviving until time t. Another concept related to the hazard rate is the survival function which can be defined as the probability of surviving to time t. Most models try to model the hazard rate by choosing the underlying distribution depending on the shape of the hazard function. A distribution whose hazard function slopes upward is said to have positive duration dependence, a decreasing hazard shows negative duration dependence whereas constant hazard is a process with no memory usually characterized by the exponential distribution. Some of the distributional choices in survival models are: F, gamma, Weibull, log normal, inverse normal, exponential etc. All these distributions are for a non-negative random variable. Duration models can be parametric, non-parametric or semi-parametric. Some of the models commonly used are Kaplan-Meier and Cox proportional hazard model (non parametric). Classification and regression trees (CART) Main article: Decision tree learning Globally-optimal classification tree analysis (GO-CTA) (also called hierarchical optimal discriminant analysis) is a generalization of optimal discriminant analysis that may be used to identify the statistical model that has maximum accuracy for predicting the value of a categorical dependent variable for a dataset consisting of categorical and continuous variables. The output of HODA is a non-orthogonal tree that combines categorical variables and cut points for continuous variables that yields maximum predictive accuracy, an assessment of the exact Type I error rate, and an evaluation of potential cross-generalizability of the statistical model. Hierarchical optimal discriminant analysis may be thought of as a generalization of Fisher's linear discriminant analysis. Optimal discriminant analysis is an alternative to ANOVA (analysis of variance) and regression analysis, which attempt to express one dependent variable as a linear combination of other features or measurements. However, ANOVA and regression analysis give a dependent variable that is a numerical variable, while hierarchical optimal discriminant analysis gives a dependent variable that is a class variable. Classification and regression trees (CART) are a non-parametric decision tree learning technique that produces either classification or regression trees, depending on whether the dependent variable is categorical or numeric, respectively. Decision trees are formed by a collection of rules based on variables in the modeling data set: Rules based on variables' values are selected to get the best split to differentiate observations based on the dependent variable Once a rule is selected and splits a node into two, the same process is applied to each "child" node (i.e. it is a recursive procedure) Splitting stops when CART detects no further gain can be made, or some pre-set stopping rules are met. (Alternatively, the data are split as much as possible and then the tree is later pruned.) Each branch of the tree ends in a terminal node. Each observation falls into one and exactly one terminal node, and each terminal node is uniquely defined by a set of rules. A very popular method for predictive analytics is Leo Breiman's random forests. Multivariate adaptive regression splines Multivariate adaptive regression splines (MARS) is a non-parametric technique that builds flexible models by fitting piecewise linear regressions. An important concept associated with regression splines is that of a knot. Knot is where one local regression model gives way to another and thus is the point of intersection between two splines. In multivariate and adaptive regression splines, basis functions are the tool used for generalizing the search for knots. Basis functions are a set of functions used to represent the information contained in one or more variables. Multivariate and Adaptive Regression Splines model almost always creates the basis functions in pairs. Multivariate and adaptive regression spline approach deliberately overfits the model and then prunes to get to the optimal model. The algorithm is computationally very intensive and in practice we are required to specify an upper limit on the number of basis functions.