Video summary

Classification in Orange (CS2401)

Main summary

Key takeaways

Technology

Summary of the Video: Technological Concepts + Orange Features for Classification

Orange overview (data mining / ML tool)

  • Orange is a free, component-based visual programming software for data mining, machine learning, and data analysis.
  • Workflows are created on a canvas by connecting predefined (or user-created) widgets—typically no coding required.
  • Orange includes many machine learning algorithms and supports tasks such as classification and clustering.

What “classification” means in this tutorial

  • Classification is presented as organizing data into categories / class labels.
  • Orange “learns from past data” by finding relationships between:
    • attributes (features) and
    • labels (classes)
  • The goal is to predict unknown instances.

How to start in Orange

  1. Download Orange from the provided link.
  2. Install it.
  3. Open it from Applications.

Tutorial workflow: Training → Cleaning (outliers) → Model selection → Evaluation → Prediction

1) Load and inspect a dataset (Vehicles)

  • The tutorial begins with a vehicles numerical dataset (and later shifts to categorical examples).
  • A Data Table widget is used to open the dataset.
  • Mentioned dataset stats:
    • 846 examples
    • 18 attributes

2) Detect and remove outliers (important before training)

  • Linear Projection is used to visually spot points far from the rest (outliers).
  • Outliers are detected via an outlier-focused widget by choosing a distance-based parameter (stated as 3).
  • Outliers are then excluded:
    • Route/export outliers for inspection.
    • Keep only inliers for the next ML step.

3) Switch to a categorical dataset (Zoo)

  • The categorical dataset used is Zoo.
  • Mentioned dataset stats:
    • 101 instances
    • 16 features

4) Choose and connect multiple classification learners

Learners compared/used in the workflow:

  • Naive Bayes
  • Neural Network
  • Classification Tree (tree)
  • CN2
  • Random Forest
  • K-Nearest Neighbors (KNN)

These learners are connected to the dataset for training and comparison.

5) Evaluate learners using “Test Learner” + classification accuracy (CA)

  • Evaluation is performed using Test Learner (under the Evaluate tab).
  • The tutorial emphasizes that Test Learner requires connections:
    • connect data + learners to obtain results.

Assessment settings discussed:

  • Random sampling: about 91.1% accuracy (as shown in the tutorial)
  • Cross-validation: about 94% accuracy (noted for larger datasets)
  • Leave-one-out:
    • longer runtime
    • more accurate
    • recommended for small datasets

Training/test split (handled directly in Orange)

  • Orange lets you set the percentage split directly inside the widget.
  • Example: 66% training / 34% testing
  • Result noted: KNN achieves the highest classification accuracy (closest to 1).

6) Confirm using confusion matrix

  • A Confusion Matrix widget is used to compare correct vs incorrect predictions.
  • The tutorial highlights that KNN still produces the most correct classifications, consistent with the accuracy results.

Prediction demonstration + probability outputs

  • Data Sampler is used to prepare data for prediction, with options such as:
    • random sampling
    • stratified sampling
  • Selected models (e.g., Naive Bayes and Classification Tree) are connected to Prediction.
  • Connection cleanup is mentioned (drag across to remove multiple connections).

Probability/score viewing

  • The tutorial shows how to view predicted probabilities using an additional widget (referred to as “Normal gram”).
  • It also mentions target encoding like “bird: yes/no”, where:
    • 1 = yes
    • 0 = no

Example described

  • Example outcome: “Bear is a mammal” (based on chosen attributes).
  • For a specific point:
    • Naive Bayes predicted insect
    • Classification Tree predicted mammal
    • the point is actually an insect
  • This is used to argue that Naive Bayes was more accurate for that point/model comparison.

Visualization section: 4 visualization widgets used for analysis and interpretation

The tutorial demonstrates the Visualize tab and focuses on four widgets:

  1. Distribution

    • Select variables to see frequency/probability plots.
    • Example mentioned: variable related to hair and the target “amphibian”.

  2. Attribute Statistics

    • Shows nominal values and their proportions.
    • Example described: hair categories mapped to 0/1, with percentages mentioned (wording is somewhat inconsistent), such as roughly 57.4% vs 42.6%.

  3. Scatter Plot

    • Choose X-axis and Y-axis attributes.
    • Points are colored by class type (e.g., amphibian/bird/fish).
    • Zooming helps identify anomalies/outliers.
    • Customization is mentioned for point labels/shapes/sizes.

  4. Linear Projection

    • Provides reduced-dimensional axes (e.g., feathers, hair, milk, eggs).
    • Classes are used to color/organize points by animal type.
    • Zooming supports close inspection.

Using visualization after prediction

  • Select Attributes is connected to the prediction flow to choose which attributes feed the prediction view.
  • Prediction results are connected to visualization (e.g., Scatter Plot) to compare:
    • where models predict the correct vs incorrect classes
  • Zoomed views are used to inspect success/failure regions.

Main speakers / sources

  • Source: A single instructor/narrator delivers the Orange classification tutorial.
  • No additional named speakers or sources are mentioned in the subtitles.

Original video