Digital twins are not inherently generative in nature, but the evolution of generative AI is poised to revolutionize them with unprecedented intelligence and predictive capabilities, thus transforming them into "intelligent twins”.   Digital twin is a virtual representation of a physical entity – an object, behavior, process, system, or product. It is essentially an animate ecosystem that is trained to mirror and sync with the physical entity. A digital twin is created by collecting real-time data from sensors embedded in physical entities and using that data to create a digital replica. This digital replica can then be used for various purposes such as monitoring, analysis, simulation, and optimization.   Select any image to see a larger version. Mobile users: To view the images, select the "Full" version at the bottom of the page.   A digital twin ecosystem is an inherently dynamic, interconnected network comprised of software, generative and non-generative algorithms, batch and streaming data, and of course, business logic. Digital twins often converge AI/ML, simulation, forecasting, optimization, and visual and streaming analytics to stress test a physical entity in the digital world to prescribe actions that optimize the entity in the physical world – to improve the lives of individuals, the lives of populations, the environment, cities, organizations, systems, products, and more.   A simple example of a Factory Digital Twin is illustrated in a SAS post authored by Biller, B. (2022) titled Digital twin development: Why simulations are critical.   Generative AI as a Digital Twin   Generative AI is set to revolutionize digital twins by transforming them into “intelligent twins”.   Envision a friendly looking doctor wearing a white lab coat with a stethoscope around her neck. She has a warm and approachable smile, exuding confidence and professionalism. Her eyes convey empathy and understanding, and she appears focused yet compassionate. This is an AI-generated version of your doctor.     AI-generated version of doctor is created using https://app.simplified.com/graphic-design/tools/text-to-image/results    Now, imagine a scenario where you can post a question to this virtual persona who embodies expertise, care, and a dedication to helping others and could serve as a virtual advisor. This AI-powered digital twin would respond to your queries based on the knowledge it has acquired from the imperfections and behavioral traits of your doctor. It goes beyond simply analyzing real data. Instead, it taps into the essence of your doctor's persona and assisting in decision-making, providing expert insights, and even offering personalized treatment recommendations.   This concept of generative AI digital twins is still in its early stages, but it has already captured the attention of researchers and industry experts.   Generative Ecosystem of Digital Twins   Digital twins can be generative in nature. The simulation aspect of a digital twin is often emphasized, but the role of generative AI is severely undervalued.   The interconnected networks in a digital twin are comprised of numerous deep learning algorithms that are inherently generative. For example, generative adversarial networks (GANs) are increasingly being adopted in the field.   A digital twin uses real-time data to generate or create simulations that mirror the behavior and characteristics of real-world entities. Besides, a digital twin uses generative outputs as automated inputs to the ecosystem. For example, it generates synthetic data as an output from the ecosystem that is, in turn, used as a feedback loop and contributes back as an input into the ecosystem to optimize the ecosystem.   A digital twin generates new data and insights about an entity’s behavior in different scenarios and conditions to proactively guide real-world strategy and to improve real-world outcomes – not as a static replica, but as an animate virtual representation of an entity.   Similarly, it generates new observations to continuously learn and adapt to changes in the real-world entity, leading to more accurate predictions and better decision making about the real-world entity.   What is more, a digital twin generates large amounts of synthetic data in a controlled ecosystem when real-world data is limited or too expensive to collect. This real-time synthetic data creates simulations to mirror and emulate real-world objects, systems, or processes.   Methodological Perspective of Digital Twin Ecosystem   Frank Emmert-Streib presented a methodological perspective of a digital twin system (shown below in a light blue rectangle) with an interface to a physical entity and external data. Its structural architecture highlights the various contributions that AI can offer for optimizing the outcomes of a digital twin. These contributions will be represented from 1 to 6.   A digital twin is created corresponding to a dynamic system that is implemented via a computer simulation capturing essential features of the physical entity. The physical entity to be modeled can come from a wide range of applications including engineering, manufacturing, healthcare, urban development, and more. Usually, this involves external data.     Source: Emmert-Streib, F. What Is the Role of AI for Digital Twins? AI 2023, 4, 721-728.   The additional external data is used to estimate the parameters of the digital twin. This is an optimization process of model creation, the first AI contribution. Importantly, there is another optimization step involving AI that is different from the first one.   The second optimization step ensures the synchronization of the digital twin and its physical counterpart during its operation. This leads to the optimization of the updating mechanism of the digital twin called model updating.   The third AI involvement is generative modeling. For example, a GAN can be used to generate data with characteristics learned from large-scale data to replace or complement a traditional simulation model. That means AI does not only help optimize a simulation model, but it can constitute the simulation model itself.   The fourth and fifth AI associations are principally the data analysis steps. However, the sources of the data for such an analysis are entirely different.   External data is thoroughly analyzed to identify patterns and relationships. This data analysis step is called descriptive analytics. Descriptive analytics seeks to describe an event, phenomenon, or outcome. It helps understand what has happened in the past and provides businesses the perfect base to track trends and can also be used to compare the performance with others within the same industry.   A lot of data is generated by the digital twin. The data generated by the digital twin and the data that is coming from descriptive analytics using the external data are combinedly used to predict future outcomes like component failure of a particular component in a manufacturing unit. This process is called predictive analytics. Evidently, the source of data has a crucial influence on the interpretation of an analysis. Hence, descriptive analytics and predictive analytics provide complementary means, and the technique we discuss next aims at their integration.   The sixth AI application, called decision making, is used for summarizing all individual results achieved up to this point to enable decision makers to base their choices on factual data and insights rather than intuition or assumptions without testing the physical entity. This step integrates everything together and produces a quantitative or qualitative summary that can be seen as the ultimate output of a digital twin system.   How Digital Twin Is Different from Simulation   Data simulation is not a new idea, and it has been a fundamental tool for statistical modeling. Digital twin technology is similar to simulations as both utilize digital models to replicate various processes of a system. However, there are some key differences between the two.     The most notable is that a simulation model is a brute-force computational technique that relies on repeating a computation on many different static random samples to estimate a statistical quantity where modelers must set all parameters manually, whereas digital twin technology uses real-time data to mimic processes by integrating all relevant data into a digital system, effectively mirroring the entirety of the product, service, or process.   It is important to understand that a simulation model is not equipped with an updating mechanism, whereas a digital twin is a mathematical model with an updating mechanism. Digital twins are designed around a two-way flow of information that first occurs when object sensors provide relevant data to the system processor and then happens again when insights created by the processor are shared back with the original source object.   A simulation replicates what could happen to a product, service, or process. Any changes to a simulation are limited to the imagination of the modeler. But a digital twin replicates what is happening to an actual outcome in the real world as it offers active interaction. Thus, the modeler can see if it is working as intended and then determine any improvements based on actual use.   The difference between digital twin and simulation is largely a matter of scale. A simulation studies one process, but a digital twin can itself run any number of useful simulations to study multiple processes. A digital twin is far more flexible, allowing modelers to test the entity through all stages in many different settings. As long as you have the data, the digital twin will replicate the entire process. Digital twin simulations are more versatile and offer a deeper understanding of processes when compared to simulations.   Concluding Remarks   Generative AI represents a paradigm shift in the realm of digital twins, unlocking new possibilities for intelligent decision-making, predictive analysis, and scenario forecasting. By harnessing the power of generative models, digital twins are evolving into intelligent twins that not only replicate physical entities but also simulate, predict, and optimize their behavior in real-time. As this transformative technology continues to mature, it holds the promise of revolutionizing diverse industries and shaping the future of smart, interconnected systems.     Find more articles from SAS Global Enablement and Learning here. ... View more

The Decision Tree node in Model Studio pipeline can build decision tree models autonomously or interactively. To build a model autonomously, simply run the Decision Tree node. Building models interactively, however, gives you more control and can be more informative and is helpful when you first learn about creating decision trees, and even when you are more expert at predictive modeling.   After you run a Decision Tree node in Model Studio pipeline, you can modify the splitting and pruning logic of nodes when you open the decision tree. The splitting logic that you can modify varies, depending on whether you want to modify a category predictor (class input) or a measure predictor (continuous input).   Interactively Editing a Decision Tree   To edit split for a node, right-click the Decision Tree node and select Open. Switch to the Stack view from the top right corner of the window. Stack view displays the objects as if they are in a slide deck. Only one object is displayed at a time, however, a control bar lets you move between objects. By default, Decision tree object is open.     Select any image to see a larger version. Mobile users: To view the images, select the "Full" version at the bottom of the page.   Right-click the node in the decision tree that you want to split and select Edit split for node. The Split Node window appears.   Decision tree models involve recursive partitioning of the training data in an attempt to isolate concentrations of cases with identical target values. Regardless of the type of predictor that you select, you can edit rules when splitting a node.   Under Variable, select the variable that you want to use for splitting. Variables are sorted by -Log(p) or logworth. The logworth is used to determine how well a variable divides the data. As the logworth increases, the partition better isolates cases with identical target values. You can choose to split the data on any of the listed inputs and also edit the splitting point given additional information.     Under Branches, specify the splitting criteria for each branch that stems from a node. You can directly replace the values that you want to use for your splitting criteria in the Values field.     You can also add branches and specify the splitting criteria for each of those branches. If you want to add branches to your splitting criteria, click Add new branch. You can clear all of the branches or reset the branches to reflect the best split by clicking the Actions (the three vertical dots icon) and selecting either Set to best split or Clear all branches. Any changes that are made at this step override the value that is defined for Maximum number of branches in the options pane for the Decision Tree node.   Under Missing values, specify how to treat missing values when splitting the node.     Once done with editing the splitting rules, click OK to apply your changes to the decision tree. You can undo or redo edits that were made to your decision tree by clicking Undo or Redo icons in the upper right corner of the window.   Interactively Pruning and Training a Decision Tree   In order to prune your decision tree to a node, right-click the node that you want to prune to and select Prune to node. The Prune to node option removes all nodes beneath the selected node and turns that node into a leaf node, resulting in a little simpler tree.     If you want to undo pruning, you have two options – you can either split the node one level or you can train more than one level beyond the leaf node.   For the former one, select Split from node and the Split Node window appears to specify the splitting rules. You can select the variable that will be used to split the node one level and edit the splitting point.   For the latter one, select the Train from node option and the Train Node window appears to select the variables to use in splitting. Variables are still sorted in descending order by their logworths at that node.     The Train node window appears with all the variables selected by default. However, you can deselect variables in the list based on business requirements or based on any other reasons. You can also change the Maximum depth of subtree.   Interactive Editor   When you close the interactive editor, you are prompted to save your changes. To save your changes, click Save and Run. If you do not want to save your changes, click Don’t Save. To continue editing your decision tree, click Cancel.   You cannot copy a Decision Tree node when it is in interactive mode. To copy a Decision Tree node, right-click the node and select Enable Properties. When the Enable Properties window appears, select Yes. You can now copy the Decision Tree node. When you enable properties after interactively editing a decision tree, the changes that you made to the decision tree are lost. During an interactive editing session, the Decision Tree node is locked so that other users cannot simultaneously make edits in interactive mode. The node is also locked when a second user who has access to the project attempts to open the editor while it is in use by another user or the same user attempts to open the editor in a different tab or browser. If you encounter an error when you try to save your work, you are returned to the interactive editor. The interactive editing session lasts 30 minutes by default.   Concluding Remarks   In a Model Studio pipeline, you can interactively edit splitting variable or split point or both in one or more than one decision tree nodes, of course, one at a time. You can interactively control the missing values allocation when splitting the node. Additionally, you can create entire tree interactively by first pruning it up to the root node and then use the Split from node or the Train from node options. The tree will then be created by splitting the source data, constituting the root node of the tree, into subsets, which constitute the child nodes. The splitting is based on a set of splitting rules based on classification features (Shai and Shai 2014). This process is repeated on each derived subset in a recursive manner. The recursion is completed when the subset at a node has all the same values of the target variable, or when splitting no longer adds value to the predictions. This process of top-down induction of decision trees (Quinlan 1986) is an example of a greedy algorithm, which is based on the concept of heuristic problem solving by making an optimal local choice at each node. By making these local optimal choices, you reached the approximate optimal solution globally.   If you're just getting started with Model Studio, check out Build Models with SAS Model Studio | SAS Viya Quick Start Tutorial. It guides you through the process of training and deploying machine learning models using Model Studio pipelines.     Find more articles from SAS Global Enablement and Learning here. ... View more

Generative Artificial Intelligence, abbreviated as GenAI or Generative AI, refers to a type of artificial intelligence that is designed to generate new content, data, or information that is similar to, or indistinguishable from, what a human might produce. One of the prominent models in generative AI is the Generative Pre-trained Transformer (GPT), which is capable of generating coherent and contextually relevant text based on the input it receives.   Generative AI systems learn real-world data to generate seemingly new content, such as text, images, audio, video, tabular data, synthetic data, and code, with similar probabilistic distributions and characteristics of the real-world data. Generative AI can produce highly realistic and complex content that mimics human creativity, making it a valuable tool for many industries such as art, writing, software development, product design, health care, finance, marketing, fashion, and more. Generative AI is an exploding area within the field of artificial intelligence. Therefore, it is important to explore the AI backdrop and see where generative AI fits in.   Artificial Intelligence (AI)   SAS describes Artificial Intelligence (or AI) as “the science of designing ethical and transparent systems to support and accelerate human decisions and actions.”   AI is a branch of computer science that deals with the creation of intelligence agents, which are systems that can reason, learn, and act autonomously. Essentially, AI has to do with the theory and methods to build machines that think and act like humans. Advancements in computer processing and data storage made it possible to ingest and analyze more data than ever before. Around the same time, we started producing more and more data by connecting more devices and machines to the internet and streaming large amounts of data from those devices. All these advancements brought artificial intelligence closer to its original goal of creating intelligent machines, which we're starting to see more and more in our everyday lives. From recommendations on our favorite retail sites to auto-generated photo tags on social media, many common online conveniences are powered by artificial intelligence. Most artificial intelligence examples that you hear about today, from chess-playing computers to self-driving cars, rely heavily on machine learning that gives the computer the ability to learn without explicit programming.   Machine Learning (ML), Deep Learning (DL) and Beyond   Artificial intelligence is the broad science of mimicking human abilities, and machine learning is a specific subset of artificial intelligence that trains a machine how to learn. Machine learning automates analytical model building. Early work with neural networks stirred excitement for "thinking machines." A neural network is a kind of machine learning inspired by the workings of the human brain. It's a computing system made up of interconnected units (like neurons) that processes information by responding to external inputs, relaying information between each unit. The process requires multiple passes at the data to find connections and derive meaning from undefined data.   Machine learning and deep learning are subfields of artificial intelligence, and deep learning is a specific field of machine learning. Whereas machine learning mostly dealt with huge tabular data, deep learning generally dealt with images, audio, and text as input data that results in even larger tables. Recent deep learning breakthroughs are driving an AI boom. Deep learning uses huge neural networks with many layers of processing units, enabling them to process more complex patterns than machine learning in large amounts of data. Deep neural networks leverage artificial neurons, backpropagation, weights, and biases to identify patterns based on the inputs.   Artificial Intelligence applies machine learning, deep learning, and other techniques to solve problems in areas such as computer vision (CV), natural language processing (NLP) and internet of things (IoT). In addition, several technologies enable and support AI including GPUs, advanced algorithms and APIs.   Graphical processing units, or GPUs, are key to AI because they provide the heavy compute power that's required for iterative processing. Training neural networks requires big data plus compute power. Advanced algorithms are being developed and combined in new ways to analyze more data faster and at multiple levels. This intelligent processing is key to identifying and predicting rare events, understanding complex systems, and optimizing unique scenarios. Application programming interfaces, or APIs, are portable packages of code that make it possible to add AI functionality to existing products and software packages. They can add image-recognition capabilities to home security systems and Q&A capabilities that describe data, create captions and headlines, or call out interesting patterns and insights in data.   In summary, the goal of AI is to provide software that can reason on input and explain on output. AI will provide human-like interactions with software and offer decision support for specific tasks, but it's not a replacement for humans. And it won't be any time soon.   Generative AI: Subset of All Select any image to see a larger version. Mobile users: To view the images, select the "Full" version at the bottom of the page.   Now we finally get to where generative AI fits into this AI landscape. Generative AI is a specific field of artificial intelligence that uses several techniques that continue to evolve. Foremost are deep learning algorithms such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs), which consist of multiple layers of interconnected nodes that accept large quantities of data as inputs and identify patterns or structures within the data set. CNN is mostly used for image recognition and computer vision tasks. It works by pooling important features from an input image. RNN is used in advanced fields like natural language processing, handwriting recognition, time series analysis, and so on.   Generative AI can process both labeled and unlabeled data using many machine learning algorithms like neural network, k-nearest neighbor, and more. Generative AI also uses some of the reinforcement learning algorithms that focuses on sequential decision-making problems. The goal for such algorithms is to train an agent to perform actions in order to optimize an objective based on a reward signal.     A recent breakthrough of transformer models has revolutionized how we use machine learning to analyze unstructured data. The transformer in NLP is an innovative architecture that aims to solve sequence-to-sequence tasks while handling long-range dependencies with ease. One of the key innovations in transformer models is the use of the self-attention mechanism, which enables the model to weigh the importance of different parts of the input and output without using sequence-aligned RNNs or CNNs. Large language models (LLMs) are also a subset of deep learning. The use of deep learning models which will employ a Variational autoencoder or Generative Adversarial Network model uses methods for generating synthetic data.     AI systems can be used to do different things for us, such as forecast the demand for a product or the weather, recommend customized product suggestions for e-commerce sites, identify objects from images and video, detect fraudulent bank transactions, or optimize large scale production systems. On the other hand, generative AI is a type of AI whose job is simply to generate content.   Arguably, because machine learning and deep learning are inherently focused on generative processes, they can be considered types of generative AI, too. Broadly speaking, whereas traditional AI/ML systems recognize patterns and make predictions, GAI systems learn real-world data to generate data – like text, images, audio, video, synthetic data, code – with similar probabilistic distributions and characteristics of the real-world data.     Find more articles from SAS Global Enablement and Learning here. ... View more

Supervised learning techniques focus on using the data that have labels to train classifiers and then use the trained classifiers to do predictive modeling. However, in real-world applications such as fraudulent insurance claims, the labeled data are usually limited and very expensive to obtain.   Moreover, the key to supervised learning algorithms, such as k-NN and SVM, is that, in general, they depend on the assumption that nearby points are likely to have the same label, regardless of the underlying data structure. This is more an assumption of local consistency and thus might result in inferior classifiers. However, points on the same structure (typically referred to as a cluster) are likely to have the same label. This assumption is often called the cluster assumption and ensures global consistency. Semi-supervised learning ponders both of these assumptions.   Select any image to see a larger version. Mobile users: To view the images, select the "Full" version at the bottom of the page.   Semi-supervised learning techniques combine labeled data and unlabeled data to address challenges posed by supervised learning and to improve classification performance. Semi-supervised learning is used for the same applications as supervised learning. But it uses both labeled and unlabeled data for training—typically, a small amount of labeled data with a large amount of unlabeled data (because unlabeled data are less expensive and take less effort to acquire). The target variable specifies the variable that contains the label. Input variables are used for distance calculation. Only interval variables are supported as input variables.   This type of learning can be used with methods such as classification, regression, and prediction. The goal is to predict the labels of the unlabeled points. The performance of an algorithm is measured by the error rate on these unlabeled points only. Semi-supervised learning is useful when the cost associated with labeling is too high to allow for a fully labeled training process. Early examples of this include identifying a person's face on a webcam. The semi-supervised learning algorithm has numerous applications, including fraud detection, web page classification, image recognition, medical imaging, natural language processing, and action recognition.   Graph-Based Semi-supervised Learning   A principled approach to semi-supervised learning is to design a classifying function that is sufficiently smooth with respect to the intrinsic structure collectively revealed by labeled and unlabeled observations.   The SEMISUPLEARN procedure in SAS Viya implements the graph-based semi-supervised learning algorithm, which is well known for its good performance and scalability for big data. Here the assumption is that the data (both labeled and unlabeled) are embedded within a low-dimensional cluster or manifold that can be reasonably expressed by a graph.   In the graph-based methods, label information of each sample is propagated to its neighboring sample until a global stable state is reached on the complete data set. For each observation in the query data table, PROC SEMISUPLEARN returns the predicted labels for the observations in both the unlabeled data table and the labeled data table.   Graph-based semi-supervised learning includes a 3-step process:   Similarity Measure: Similarities (distances) between the pairs of the observations in the unlabeled and labeled data tables are computed to determine similar data points. The SEMISUPLEARN procedure offers both a dense kernel (the radial basis function, or RBF, kernel) and a sparse kernel (the k-nearest-neighbors, or KNN, kernel) for computing the similarity measure or distance. Graphical Network: A network graph is constructed to represent the intrinsic structure of the data. This graphical network consists of vertices and edges, where the vertices are specified by unlabeled and labeled observations, whereas the edges specify the similarities among the labeled as well as unlabeled observations. Thus, each data sample is represented by a vertex in a weighted graph with the weights providing a measure of similarity between vertices. Label Propagation: Label of each data sample is progressed to its neighboring points to infer labels on unlabeled nodes in the graph. Here the assumption is that the points within the same structure (such as a cluster or a manifold) are likely to have the same label. Given this assumption, the aim is to use unlabeled data to uncover this structure.   Model Studio Demo: Detecting Fraud   The semi-supervised learning algorithm to the PaySim data from Dr. Lopez-Rojas et. al. is applied. A graph-based semi-supervised learning model is created using a SAS Code node in Model Studio pipeline. The PaySimLabeled data set consists of 430 observations, 130 of which are labeled as fraud. The PaySimUnlabeled data set has 5000 observations, and none are labeled. The objective is to detect fraud in the unlabeled data.   In the Code Editor window, first, a CAS session is started, and all the default libraries are assigned.   options cashost='&dm_cashost' casport=&dm_casport; cas; caslib _all_ assign;   The next part of the program loads labeled and unlabeled SAS data files from a client location into the specified caslib and saves it with specified table name.       proc casutil; load file="/home/PaySimLabeled.sas7bdat" outcaslib="casuser" casout="PaySimLabeled"; load file="/home/PaySimUnlabeled.sas7bdat" outcaslib="casuser" casout="PaySimUnlabeled"; run;   The following piece of code invokes the SEMISUPLEARN procedure to perform graph-based semi-supervised learning on the unlabeled data table casuser.PaySimUnlabeled and the labeled data table casuser.PaySimLabeled and output the results to ODS tables. The INPUT statement specifies that the interval variables are to be used as inputs. The OUTPUT statement requests that the predicted labels for the unlabeled and labeled tables be written to the data table casuser.PaySimOut.   proc SEMISUPLEARN data= casuser.PaySimUnlabeled label = casuser.PaySimLabeled gamma = 1000; input %dm_interval_input; output out = casuser.PaySimOut copyvar=(_all_); target %dm_dec_target; run;   By default, the radial basis function is used as a kernel to calculate the similarity in distance computation using the Gaussian kernel. The GAMMA parameter specifies 1000 as the inverse of the variance for Gaussian kernels when you compute the pairwise distance between data observations. This is essentially the RBF kernel width.   The following statements sort the output of PROC SEMISUPLEARN by id and show the observations from 100 to 200:   data PaySimOut; set casuser.PaySimOut; run; proc sort data=PaySimOut; by %dm_id; run; proc print data=PaySimOut(firstobs=100 obs=200); run;   The successful run of the SAS Code node results in the following:     The Model Information table shows the number of unlabeled observations, number of labeled observations, number of levels for the target variable, gamma value, maximum number of iterations, the kernel used in the computation, number of nearest neighbors, and the loss.   You can try experimenting with GAMMA values to minimize the loss.   The PAYSIMOUT table is printed.     The table shows the values of the input and ID variables, the predicted labels for the 100 observations, and the indicators for the labeled or unlabeled data. For each row in the output data table, the _WARN_ column indicates whether the row of data is coming from labeled data or unlabeled data. The _WARN_ value is 1 for labeled data and 0 for unlabeled data. I_isFraud is the predicted target variable, which is generated for labeled as well as unlabeled samples.   PaySim first paper of the simulator: Lopez-Rojas, E. A., A. Elmir, and S. Axelsson. 2016. PaySim: A financial mobile money simulator for fraud detection." In the 28th European Modeling and Simulation Symposium-EMSS, Larnaca, Cyprus.     Find more articles from SAS Global Enablement and Learning here. ... View more

  The purpose of supervised machine learning is to generate accurate predictions for the future. Good models strive to achieve low training error, but it is just as important to achieve low generalization error. The training process needs to account for this compromise and make an honest assessment of the accuracy of the model. Assessing a candidate model on the data that is used to train the model would direct the algorithm to overfit to that training data. Instead of doing this, we normally take the historical data with known target, and split it into three partitions: training, validation, and test.   A major portion is used for fitting the model - the training data set.   Validation data is used to assess the model during training for the purpose of selecting variables and adjusting parameters. Validation data sets are instrumental in preventing overfitting. In lieu of a separate validation set (which might not be feasible for smaller data sets), SAS Viya Machine Learning offers k-fold cross validation through the Autotuning capability. Whether through a validation set or through k-fold cross validation, ensure that the training process assesses the error on data that is not used to train the model.   Test data is used at the end of model fitting to obtain a final assessment of how the model generalizes to new data. The reason for using test data (instead of validation data) is that validation data plays a role in the model training process. Hence, using validation data might lead to the same biased assessments as using training data. For this reason, a test data set should be used only at the end of the analysis and should not play a role in the model training process. For example, you make predictions for new data (say 2024) using a model trained, validated and tested on 2019–2023 data. This classic setup is illustrated below.   Select any image to see a larger version. Mobile users: If you do not see this image, scroll to the bottom of the page and select the "Full" version of this post   This is essentially an in-time testing approach that could be easily implemented in Model Studio while creating a Data Mining and Machine Learning project. You can either, choose to Create partition variable checkbox under the Advanced project settings and then specify the training, validation and/or test percentages, or, set a pre-existed partition variable within the data by specifying its Role as Partition.   By default, Model Studio uses the test data to select the champion model, if a test data set exists. Otherwise, Model Studio uses the validation data to select the champion model. If a validation data set does not exist, Model Studio uses the training data to select a champion.   Simple! Nonetheless.   “My team build a churn model for a telecommunication company where the churn rate was 10%. The model’s sensitivity (% of actual churners correctly predicted by the model) on test data was 72% but when they rolled out the model and tracked the result after 5 months, the sensitivity was dropped to 45%.”   Have you encountered such a problem? There could be many reasons for such a fallout. Perhaps doing an out-of-time testing would have helped identify the issue with the model before it was rolled out.   Out-of-time testing is the process of validating the model on the latest unseen data and capturing its performance to see whether there is any performance dip happening on the model prediction output. For example, you hold-out 2023, train the model using 2019–2022 data, and out-of-time test it on 2023 hold-out data. This out-of-time testing is illustrated below.     Holdout data is a capability that is new to Model Studio. Since test data was used in model comparison, using it again in comparing different pipelines might introduce bias. By setting aside data for the holdout partition, you allow for a further safeguard against generalization error. However, one of the biggest drawbacks of this approach is that you are not using the most recent data to make predictions.   Note that the holdout data set generally comes from a different time period than the training, validation and/or test sets and thus cannot be partitioned while creating project in Model Studio. The proof of a model's stability is in its ability to perform well year after year. A data set from a different time period, often called an out-of-time holdout data, is a good way to verify model stability, although such a holdout data set is not always available.   Once you have a holdout data you can score one or more models on it in Model Studio. To score holdout data, complete the following steps:   Select the three vertical dots icon in the upper right corner of the Pipeline Comparison tab and then select Score holdout data.       The Browse Data window appears. Select the data set that contains the holdout data that you want scored. Click OK. Model Studio will score the holdout data.     Once all the models are scored, click Close. To see the results of this process, use the Data menu below Pipeline Comparison to select Holdout.     This enables you to see the performance of the champion model (which was selected using the test data) on the holdout data to get an idea of expected performance (e.g., KS Youden) when this model is implemented in the field.     Close the Results. Switch your attention to the Insight tab by clicking on it.     The Insight tab provides model assessment results on holdout data along with other partitions.   Some analysis complications might introduce while interpreting these model assessment results. For example, if you have selected Enable event-based sampling, you must use model performance measures (like KS Youden and ROC curve) that are not drawn from the outcome proportions. Note that Model Studio uses the score code while scoring holdout data. The score code contains a section titled Adjust Posterior Probabilities. This code block modifies the posterior probability by multiplying it by the ratio of the actual probability to the event-based sampling values specified previously. Most model fit statistics, especially those related to prediction decisions (like misclassification rate), and some of the assessment plots (like cumulative lift) are closely tied to the outcome proportions in the training data. If the outcome proportions in the training, validation and test samples do not match the outcome proportions in the scoring holdout sample, model performance can be greatly misestimated.   So, what is an effective out-of-time testing approach in Model Studio? You might want to follow the following process for implementing out-of-time testing in Model Studio.   Find the best model based on Test data using Model Comparison node and/or Pipeline Comparison tab functionalities in Model Studio. You can also use only two partitions: train and validate instead of using three: train, validate and test. Use holdout data to find the best model based on the out-of-time sample. If the model based on holdout is different and if you feel it to be compelling enough to be chosen as the best model you can choose the best model from the comparison screen manually and pass it over to SAS Model Manager for scoring.   Find more articles from SAS Global Enablement and Learning here. ... View more

  It’s here!   Statistics is a core component of data analytics and machine learning. Despite the "bigness" of the data, statistics still has a lot of application. The role of statistics remains what it has always been and is even more important now. Perhaps the core statistical task in (traditional) statistics is inductive inference from data to models and scientific conclusions. This core task is still very relevant in the advent of massive data sets.   Replicability, stability, heterogeneity, causality, and uncertainty are the five basic principles of statistics, and they all hold equally well with big data.     Mobile users: To view the images, select the "Full" version at the bottom of the page.   Ideally, in big data scenario too, the conclusions and findings are replicable and generalizable. If you imagine running the analysis again, now on a new data set, would the outcome be similar, meaning that the model is stable? How would you find out what similarity in outcomes means and how to evaluate accuracy, to quantify uncertainty. Understanding heterogeneity in large-scale data sets is more crucial and comprehending causality and its connection to robust prediction is still interesting.   Are you interested in machine learning and want to grow your career in it? The key to machine learning is using the right data preprocessing techniques, understanding the algorithm, cutting through the equations and Greek letters, and making sense out of complex results.   Developing an accurate understanding of statistics will help you build robust machine learning models that are optimized for a given business problem. SAS launched a new course that provides a comprehensive overview of the fundamentals of statistics that you'll need to start your data science journey. This course is also a prerequisite to many courses in the SAS data science curriculum.   In this course, you learn how to: explain the relevance of statistics in the big data and machine learning world relate statistical and data science terminology generate descriptive statistics, explore data with graphs and plots, and perform testing of hypotheses detect associations among variables and perform linear regression compare explanatory modeling with predictive modeling describe trade-offs between bias and variance fit a logistic regression model and score new data explain the statistical foundations of machine learning prepare your data for machine learning modeling using transformations, imputation, standardization, and variable reduction discuss data difficulties and modeling issues, and their statistical solutions.   It also gives you opportunity of hands-on using SAS Studio tasks to perform your data analysis. This course is available in three modes: face-to-face classroom training, remotely connected live web, and self-paced e-learning mode.   More details on this brand-new SAS course can be checked out at: SAS Training in the United States -- Statistics You Need to Know for Machine Learning   See this course on the Data Science Learning Path: SAS Training - Data Scientist learning path   Also check out On-Demand Webinar: The Data Scientist Learning Journey: Statistics You Need to Know for Data Science | SAS     Find more articles from SAS Global Enablement and Learning here. ... View more

Today it is imperative to deal with the high dimensional data. The dimension of your data refers to the number of input variables. The analysis of high-dimensional data offers a great challenge because the human intuition about geometry of high dimensions fails. One of the biggest challenges in data visualization is to find general representations of data that can display the multivariate structure of more than two variables. Appropriate visuals are especially helpful when you're trying to find relationships among hundreds or thousands of variables to determine their relative importance – or if they are important at all.   The t-distributed stochastic neighbor embedding (t-SNE) is a method in SAS Viya for visualizing high-dimensional data. The t-SNE method computes a low-dimensional representation, also called an embedding, of high-dimensional data into two or three dimensions.   Unlike other dimension reduction methods, such as principal component analysis (PCA), t-SNE is appropriate for nonlinear data and emphasizes existing groupings in the data. The method is named t-SNE because it models the pairwise distances in low dimensions according to Student’s t-distribution. The t-distribution with one degree of freedom has heavier tails than the Gaussian distribution, which means that it assigns higher probability values to large distances. This enables t-SNE to relax pairwise distances for non-neighboring observations, whereas distances between closely neighboring observations are more exactly preserved. This behavior is desirable because it mitigates the crowding problem in high-dimensional data representation and makes existing groups in the data visually evident.   Select any image to see a larger version. Mobile users: To view the images, select the "Full" version at the bottom of the page.   For example, in the above illustration, the point-of-interest observation is represented in green, and the other two observations, 1 and 2, are represented in orange. Observation 1 has a higher similarity score than observation 2. Thus, it is closer to the point-of-interest observation.   The t-SNE plot is available using the Data Exploration node in Model Studio, which might help identify naturally occurring clusters in the data. You can also use the TSNE procedure in a SAS Code node.   The t-SNE method operates on an input data table. For each observation in the input data table, the procedure returns either two or three computed columns that contain the embedding coordinates. It computes the embedding by minimizing the Kullback-Leibler divergence between the joint probabilities in high dimensions and the joint probabilities in low dimensions, and it stores the embedding in the output data table. To speed up computations, the t-SNE method uses the Barnes-Hut approximation to the perplexity loss.   In the Data Exploration node, the t-SNE projection property specifies whether to perform t-SNE projections of input variables. This option is deselected by default. The t-SNE perplexity property specifies the t-SNE perplexity that controls the separation of points in the projected space. Possible values range from 1 to 100. The default value is 30.   Following example shows how to use the TSNE procedure to obtain an embedding from observations in a data table. The example uses the Iris data from Fisher (1936), which contain morphological measurements of 50 specimens from each of three different species of iris flowers: Iris setosa, I. versicolor, and I. virginica.         proc tsne data = mycas.iris nDimensions = 2 perplexity = 5 learningRate = 100 maxIters = 500; input SepalLength SepalWidth PetalLength PetalWidth; output out = mycas.tsne_out copyvars=(id species); run;         The PROC TSNE statement and at least one INPUT statement are required. You can specify multiple INPUT statements. The NDIMENSIONS=2 option requests that the model return two embedding dimensions; the PERPLEXITY=5 option specifies the perplexity value; the LEARNINGRATE=100 option specifies the learning rate for the optimization; and the INPUT statement specifies that the SepalLength, SepalWidth, PetalLength, and PetalWidth variables be used as inputs.   PROC TSNE returns a two-dimensional representation of each observation. The OUTPUT statement requests that the embedding be written to the data table mycas.tsne_out, and the COPYVARS= option requests that the ID and Species variables be copied to the output.   The following PROC SGPLOT statements plot embedding dimension _DIM_2_ against embedding dimension _DIM_1_:         proc sgplot data=tsne_out; title "Iris embedding"; title1 "Scatter plot of iris embedding"; scatter x=_DIM_1_ y=_DIM_2_ / group=species markerattrs=(symbol=CircleFilled); run;         The visualization below shows the resulting scatter plot.     The colors in the scatter plot indicate three distinct clusters, which correspond to the three iris species. The two axes here in a low dimensional space do not have a particular meaning.   There are two methods available for the algorithm to use for gradient calculation: METHOD=BARNES_HUT | EXACT. By default, METHOD=EXACT. For larger data sets, Barnes-Hut method is recommended over Exact method as it gives good visualisation accuracy when plotted.   The t-SNE method doesn't construct explicit mappings relating the high dimensional and low dimensional spaces. In fact, t-SNE captures structure in the sense that neighboring points in the high-dimensional input space will tend to be neighbors in the resulting low-dimensional space. Care should be exercised while interpreting the t-SNE plot: Unlike PCA or MDS, larger distances can't necessarily be interpreted. If points are separated in input space, t-SNE separates them in the low dimensional space, however, it doesn't care how far they are. At times, t-SNE breaks overlapping or continuous segments of data into contiguous pieces and artificially separates them, particularly at low perplexity settings. Note that the t-SNE method is well suited for visualization of high-dimensional data, and less so for dimension reduction, feature engineering and preprocessing for subsequent clustering and modeling.   Related Links: For more information on The TSNE Procedure see the Data Mining and Machine Learning Procedures documentation: (https://go.documentation.sas.com/doc/en/pgmsascdc/default/casml/casml_tsne_overview01.htm) StatQuest: t-SNE, Clearly Explained (https://www.youtube.com/watch?v=NEaUSP4YerM) is a good resource to understand the concept behind the t-SNE method. Find more articles from SAS Global Enablement and Learning here. ... View more