ThingWorx Analytics Training: Module 6 Part 1
This video begins Module 6: Predictive Models & Model Validation of the ThingWorx Analytics Training videos. It gives examples of different types of goal variables. It also discusses data considerations in predictive modeling, It begins describing the machine learning algorithms used by ThingWorx Analytics to build predictive models.
This video does not contain any hands-on exercises, and you don't need access to a running ThingWorx Analytics environment.
This video has chapters — use the chapter menu in the player to jump to any section.
Chapter Summaries
0:00 – Module 6 Overview: Predictive Models & Model Validation in ThingWorx Analytics
This introduction outlines what the module covers: predictive analytics fundamentals, predictive modeling techniques, model training, setting options and parameters in ThingWorx Analytics, and model accuracy, performance, and advanced modeling techniques.
0:32 – Data Foundation: Source Data, Contextual Repository & Feature Engineering
Explains how the ThingWorx Analytics process begins with raw data from social, client, text, spreadsheet, or database sources, is structured in a contextual repository into a dimensional model of facts and dimensions, then feature-engineered into new static and time-series columns.
1:43 – Descriptive & Predictive Analytics to Action: Signals, Clusters, Profiles & Prescriptive Plans
Covers loading structured data into ThingWorx Analytics to derive signals, clusters, and profiles, then turning results into action through predictive and prescriptive analytics, tactical action plans, and optimized service, quality, and part-performance decisions.
2:40 – ThingWorx Analytics Data Flow: Explore, Scope Use Case, Metadata & Champion Model
Walks through the end-to-end data flow — exploring raw data, scoping a use case, structuring data, defining metadata with automatic detection, building the analytics-ready view, and the iterative build-validate loop that exposes a champion model fed into a smart connected mashup.
4:29 – Predictive Analytics Defined: Use Cases & When It Works Well
Defines predictive analytics as statistical and machine learning techniques applied to historical data, with use cases like elevator service prediction, manufacturing rework, and demand forecasting, plus the success factors: good data collection, clearly defined outcomes, accessible features, and strong generalizing algorithms.
5:59 – Model Development Steps: Data Prep, Model Training & Champion Selection
Breaks down the high-level, agile and iterative workflow for developing a predictive model: data preparation (collection, cleansing, feature engineering), model training, assessing performance across models and parameters, and choosing a champion model.
6:39 – Three Goal Types: Continuous, Boolean & Categorical Predictions
Describes the three prediction goal types in ThingWorx Analytics — continuous variables (electrical output, rejects per hour, sales dollars), Boolean yes/no outcomes (asset failure, purchase intent), and categorical multi-value goals — with class probabilities available through the APIs.
8:15 – Data Considerations: Random 80/20 Train-Validation Split
Explains the best practice of a random 80/20 split, using 80% of historical data for training and withholding 20% for validation to give an unbiased estimate of model performance, better generalization, and scoring against live test data streaming from connected assets.
9:27 – Predictive Modeling Techniques Overview: Seven Learning Techniques Plus Ensemble
Introduces the seven distinct learning techniques plus an eighth ensemble technique, and the characteristics compared for each — supported goal types, model complexity, training speed, and the amount of parameter tuning required.
10:29 – Boolean Classifier: Model Output & Threshold Optimization
Shows how a ThingWorx Analytics classifier converts a numeric model output between zero and one into a Boolean prediction by testing thresholds from 0.1 to 1.0 against a validation set, selecting the threshold with the best accuracy (0.3 yielding 84.6%) for scoring new records.
12:35 – Multi-Class Predictions: One-vs-Rest Boolean Models & Max Probability
Explains how multi-class problems (e.g., red, green, or blue car) are handled by training an individual Boolean model per class, scoring a new record through each, and returning the class with the maximum probability as the final prediction.
13:23 – Linear Regression: Continuous Goals & Residual Error Minimization
Details linear regression for predicting continuous goals by learning feature weights that minimize residual errors, illustrated with a drill-accuracy example (operating hours, pump pressure, chamber temperature, drill-bit changes), the beta coefficients, Y-intercept, and epsilon error term — fast, simple, and best without nonlinear relationships.
15:27 – Logistic Regression: Log-Odds Ratio & Classification
Covers logistic regression as a classification algorithm that learns feature weights to minimize misclassifications, using a pump-failure example (operating hours, max pump pressure, air temperature, pump type) and the log-odds ratio that indicates the probability of a record belonging to one class versus another.
17:15 – Decision Trees: Tree Splitting, Categorical Handling & Failure-Rate Prediction
Introduces decision trees that partition data from input features to improve accuracy, highlighting their interpretability, handling of numerical and categorical data, automatic feature-interaction learning, and fast training — shown via a pump inline/pressure/temperature branching example producing failure rates of 45% and 92%.