What is no-code AI?
No-code AI means building AI models with a visual tool instead of typing code. You drag, drop and connect blocks. The tool does the programming for you.
- Good points: quick to try ideas, easy for beginners, you see every step.
- Limits: fewer options than code, large data can be slow, and you still must understand the data and the results.
Examples of no-code tools: Orange (free data-mining tool with widgets), browser tools that train an image or sound classifier from your webcam, and spreadsheet add-ons that make predictions.
Orange basics: canvas, widgets and links
Orange has a canvas (work area) and a widget panel. A widget is a block that does one job. You join widgets with links; data flows from the output of one to the input of the next. A chain of widgets is a workflow.
| Group | Useful widgets |
|---|---|
| Data | File, Data Table, Select Columns, Data Sampler |
| Visualise | Scatter Plot, Distributions, Box Plot |
| Model | kNN, Tree, Logistic Regression, Naive Bayes, Random Forest, Linear Regression |
| Evaluate | Test and Score, Confusion Matrix, Predictions |
| Unsupervised | k-Means, Hierarchical Clustering |
Building a classification model step by step
- File: load a dataset (for example a flower or student-marks table). Set the answer column as target; the others are features.
- Data Table: check rows, columns and missing values.
- Scatter Plot: colour points by target to see if classes separate.
- Model widgets: add kNN and Tree, linked from File.
- Test and Score: link the data and the models. Choose cross-validation (e.g. 10 folds) or random sampling (e.g. 80% train). Read CA (classification accuracy), precision, recall and F1.
- Confusion Matrix: link from Test and Score to see mistakes per class.
- Predictions: link a model and a new data file to label new rows.
How kNN decides
k-nearest neighbours finds the k closest examples to the new one and picks the most common class. Small k follows the data closely; large k gives smoother decisions.
Reading the results
Accuracy (CA) = correct ÷ total. Precision = of the items the model called "yes", how many were really "yes". Recall = of the real "yes" items, how many the model found. F1 balances the two.
In the confusion matrix, rows are actual classes and columns are predicted classes. Numbers on the diagonal are correct. Off-diagonal numbers show which classes the model confuses.
Compare models side by side in Test and Score and choose the best one for your problem. Also check for bias: does the data fairly represent everyone the model will be used on?
Try it: the 3D and at home
In free play, put the new flower right between type B and type C. Change k from 1 to 15. Does the answer change? Why is k usually an odd number?
At home or in the lab: make a small table of 20 fruits with weight (g), length (cm) and type. Load it into Orange (or a free browser classifier), draw a scatter plot, train kNN and Tree, and compare their accuracy in Test and Score.
Key formulas and definitions
- Accuracy (CA) = correct predictions ÷ all predictions
- Precision = true positives ÷ (true positives + false positives)
- Recall = true positives ÷ (true positives + false negatives)
- F1 = 2 × precision × recall ÷ (precision + recall)
- Workflow: File → Data Table / Scatter Plot → Model → Test and Score → Confusion Matrix → Predictions
Worked examples
1. You want to see if three flower types form separate groups. Which widget helps?
Scatter Plot, with points coloured by the target (flower type).
2. Test and Score shows kNN CA = 0.94 and Tree CA = 0.90. Which model is better on this data, and by how much?
kNN, by 0.04 (4 percentage points) in accuracy.
3. A confusion matrix for spam has: actual spam predicted spam 40, actual spam predicted not-spam 10, actual not-spam predicted spam 5, actual not-spam predicted not-spam 45. Find accuracy, precision and recall for "spam".
Accuracy = (40 + 45) ÷ 100 = 85%. Precision = 40 ÷ (40 + 5) ≈ 0.89. Recall = 40 ÷ (40 + 10) = 0.80.
4. kNN with k = 3: the nearest neighbours of a new fruit are apple, orange, apple. What is the prediction?
Apple (2 votes out of 3).
Common mistakes
- Forgetting to set the target column in the File widget, so models have nothing to learn.
- Judging a model on the training data only. Use Test and Score with cross-validation or a hold-out set.
- Thinking "no code" means "no thinking". You still need a clear problem, clean data and careful reading of results.
- Looking only at accuracy when classes are unbalanced. Check the confusion matrix, precision and recall too.