ThingWorx Analytics Training: Module 6 Part 2
This video continues Module 6: Predictive Models & Model Validation of the ThingWorx Analytics Training videos. It describes the remaining machine learning algorithms used by ThingWorx Analytics to build predictive models that weren't covered in Part 1. In addition, this video describes the different kinds of ensembles you can build that utilize multiple algorithms.
This video contains a hands-on exercise, and it is recommended to complete it before moving on to subsequent sections. This exercise involves training a predictive model using ThingWorx Analytics dataset created in Module 4. NOTE: If you did not yet create a dataset, the process is reviewed in this video.
This video has chapters — use the menu in the player to jump to any section
Chapter Summaries:
0:00 – Neural Network Models: Hidden Layers, Adaptive Weights & Binary Classification
Explains how neural networks use interconnected input, hidden, and output nodes with adaptive weights to approximate functions, handle numerical and categorical data, and how hidden layers are configured inside ThingWorx Analytics for binary/Boolean classification.
1:49 – Random Forests: Bagging, Random Decision Trees & Mean Prediction
Walks through the four-step bagging technique where random forests sample records into bags, build decision trees on randomly chosen features, construct hundreds of trees, and average their outputs into an unbiased final prediction.
3:10 – Gradient Boosting Machines (GBM): Sequential Trees & Residuals
Details the six-step GBM algorithm that grows shallow decision trees sequentially on residuals, and weighs its higher predictive power against overfitting on noisy data, slower training times, and tuning the number of trees.
4:33 – Support Vector Machines (SVM): Hyperplanes, Margin & Linear Kernel
Covers how SVMs classify Boolean goal variables by finding the hyperplane with the largest margin between classes, illustrated with good vs. bad margin examples and the linear kernel implemented in ThingWorx Analytics.
6:07 – Ensemble Techniques: Average, Best, Elite Average, Majority Vote & Soloist
Compares ensemble combination methods — average, best, elite average, majority vote, and soloist — explaining how grouping learners often outperforms a single learner and how elite average selects the top-N learners within a score threshold.
8:15 – Ensemble Trade-offs: 80/20 Split, 64% Training Data & Model Iteration
Explains how best and elite average ensembles run an extra 80/20 split (training on just 64% of the data) and why iterative model building, validation, and feature engineering are needed to find the best-performing learner.
9:29 – Model Development Process: Training Data to Production Scoring
Gives a high-level overview of the workflow: passing training data into one or more machine learning algorithms to produce a trained model that can then be deployed into production to score records.
9:52 – Uploading the BeanPro Dataset into ThingWorx Analytics
Demonstrates creating a new dataset named beanpro_data from the Analytics > Data menu, choosing the BeanPro file, unchecking Upload Metadata and Time Series Data, and submitting so ThingWorx can infer the metadata.
10:56 – Reviewing Inferred Metadata: Opt Types, Boolean & Informational Fields
Reviews how ThingWorx infers opt types and data types (continuous/double, categorical, Boolean, informational) for each field, and changes Record Purpose from categorical to informational since it's only used later for filtering.
12:31 – Creating the Dataset: JSON Export, Dataset URI & Job Results
Shows the optional Download as JSON option for replicating a dataset in another environment, clicking Create Data Set, and reviewing the completed job — roughly 53,000 rows, 34 fields, the copied dataset URI, and non-time-series status.
13:32 – Building a Neural Net Model: BeanPro_NN & Grinder Error Occurrence Goal
Covers creating a model under Models > New named BeanPro_NN with a neural net learner, adding an optional description, selecting the dataset, and choosing Grinder Error Occurrence as the goal variable.
14:54 – Creating Filters & Excluding Fields: Record Purpose, Training vs Scoring
Walks through building a Training_Validating_Data filter on Record Purpose (52,000 training vs. 1,000 scoring records) and using Exclude Fields to remove Identifier and Record Purpose before the model is trained.