Can you predict the price of a house from its features? Identify customers who might leave? Discover groups of shoppers with similar habits?
In PITMA’s AI Lab, you can build these experiments from blocks. Pick a dataset, configure preprocessing, choose a model and evaluate the result.
You can build the workflow without writing code. AI Lab turns it into a Python notebook that you run in Google Colab.
In this tutorial, you’ll follow the entire process by predicting house prices.
Choose the problem you want to solve
The beta works with tabular data. Each row describes an item, and the columns contain its features. A housing table might have one house per row and columns for area, number of rooms and price.
Choose among three task types:
| Task | What the model does | Example |
|---|---|---|
| Classification | Predicts a category | Will a customer stay or leave? |
| Regression | Predicts a number | How much does a house cost? |
| Clustering | Finds similar groups | Which customer segments appear in the data? |
For a house price, choose regression because the outcome is a number.
JamsProjectsCommunityAI Lab BETANO-CODE EDITOR
FOR ML PROJECTS
Build ML projects without writing code. Prepare data, train a model, evaluate it and visualize the results.
Start with a preset
On the AI Lab home page, select Browse presets, choose Predict a number, then select Open preset.
A preset is a ready-to-use experiment. Its blocks are already connected, so you can focus on the dataset and settings.
You can change the dataset and the parameters available on each block. The task type and workflow structure are fixed, so you cannot accidentally change the regression preset into a classification task.
When you’re ready to choose and connect the blocks yourself, select Create experiment.
Ready-to-use workflows
Predict a category
Flower species or customer churn? Train a model to recognize classes.
Predict a number
Price, demand or another quantity? Compare predictions with actual values.
Find similar groups
Group similar items to discover patterns in unlabeled data.
Select the housing dataset
In the Data block, select Choose dataset to open a searchable catalog of datasets for your task. Find Housing prices and open it.
AI Lab shows a preview. If a dataset contains multiple CSV files, choose the table you want first.
- What are we predicting? is the target column. Here it’s
price, the value you want the model to predict. - Training features are the inputs used to make that prediction, such as floor area and number of bedrooms.
Keep the suggested features for the first run. Look at several rows to see what the model will receive, then select Use this dataset.
The notebook loads the table directly into Colab when it runs. The beta catalog includes a curated selection of datasets from Kaggle.
Choose data for your experiment
Curated tables with a verified target for this task.
Housing prices
Predict a house price from its features.
First 6 rows · 545 rows checked| priceNumber | areaNumber | bedroomsNumber | bathroomsNumber | storiesNumber | mainroadCategory | guestroomCategory |
|---|---|---|---|---|---|---|
| 13300000 | 7420 | 4 | 2 | 3 | yes | no |
| 12250000 | 8960 | 4 | 4 | 4 | yes | no |
| 12250000 | 9960 | 3 | 2 | 2 | yes | no |
| 12215000 | 7500 | 4 | 2 | 2 | yes | no |
| 11410000 | 7420 | 4 | 1 | 2 | yes | yes |
| 10850000 | 7500 | 3 | 3 | 1 | yes | no |
Read the workflow
Each block represents a step. The connections show how data moves between them.
| Block | Purpose |
|---|---|
| Data | Selects the table, target and features |
| Train / test | Separates training examples from evaluation examples |
| Missing values | Fills in missing entries |
| Categories → numbers | Encodes text categories as numbers |
| Scaler | Brings numeric features onto a common scale |
| Model | Selects the learning algorithm |
| Metrics | Measures prediction quality |
| Explore data | Examines the table and creates a visualization |
Explore data receives the original table through a separate branch, so you can inspect the raw data alongside training the model.
Change settings directly on each block. Zoom the canvas or choose Full screen to inspect the workflow.
Configure the first experiment
In Train / test, keep the test size at 25%. The model learns from 75% of the examples. The remaining 25% are held out to evaluate it on examples it did not see during training.
The random seed determines the split. Keep it unchanged so you can compare models against the same test set later.
In Model, keep Linear Regression as your baseline. In Metrics, select:
- MAE, mean absolute error in the dataset’s price units. Lower is better.
- RMSE, root mean squared error. It gives larger errors more weight. Lower is better.
- R², a measure relative to predicting the mean. Values closer to 1 are better, and a score can be negative.
Get your first result with these settings and then test one change at a time.
Hold out a separate set for evaluation
Preprocessing is fitted on the training set only.
Train on the training set, predict on the test set
Choose several metrics for the same predictions
View results in Colab and in the Results tab after import.
Ask a question about the data
In Explore data, choose what to examine. To see which prices are most common:
- Under What would you like to explore?, select How are the numbers distributed?
- Under What should the chart show?, select
price.
Running the notebook produces a histogram. Prices run along the horizontal axis. The vertical axis shows how many houses fall within each price range.
You can also ask How are two features related? Try area on the X axis and price on the Y axis. Each dot represents a house.
The charts appear once the notebook has run.
Choose a question to see its chart in the results
The chart appears after you run the notebook in Colab.
Run the experiment in Colab
Select Open in Colab. If any settings are missing, AI Lab points to the blocks that need attention. Fill them in and try again.
The first time, Google may ask you to connect your account and grant access to the Drive files created by PITMA. Return to the experiment after authorizing it. If the notebook hasn’t opened yet, select Open in Colab again.
PITMA creates the notebook from your workflow and settings and opens it directly in Colab.
In Colab, select Run all and wait for all cells to finish. They load and prepare the data, train the model and calculate the results.
Each block has a corresponding code cell. Read the cells in order to see how the workflow translates into Python.
Predict a numeric value
Each block becomes a cell. Run them from top to bottom. Preprocessing is fitted only on the training set.
# Libraries for tables, plots and saving results
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from IPython.display import display
import json
import io
import base64
reports = {}# Data
import kagglehub
from pathlib import Path
dataset_folder = kagglehub.dataset_download(
"yasserh/housing-prices-dataset/versions/1"
)
data = pd.read_csv(Path(dataset_folder) / "Housing.csv")Import the results into PITMA
Once all cells have finished:
- Download the executed notebook from Colab as an
.ipynbfile with its cell outputs saved. - Return to the experiment in PITMA.
- Open Results.
- Select Add result and choose the downloaded file.
Your metrics and configured charts appear in the tab. Import the executed notebook each time you want to bring results back from Colab. If a required cell failed, fix the error in Colab and run it again before importing.
How did the experiment perform?
Run the notebook in Colab, save the .ipynb file and add it here.
Your metrics and charts will appear here
Choose a dataset and run the experiment.
Continue in ColabCompare a second model
Now you can test a specific question: Can another model predict more accurately on the same data?
Go back to Model and change Linear Regression to Random Forest. Run the new notebook in Colab and import it under Results.
Keep the dataset, features, test size and seed unchanged. This makes it easier to attribute the difference in results to the model.
Switch between runs in Results. A lower MAE means smaller errors on average. Check RMSE as well, because it gives larger misses more weight.
A single run tells you how the model performed in this experiment. Its performance on other data still needs to be tested.
Train on the training set, predict on the test set
Build your own workflow
On the home page, select Create experiment. In Data, choose the task type first, then the dataset.
Add blocks from the library or use the suggested blocks on the canvas. To connect a pair, select the output of the first block and the input of the second. You can also drag between the connection points.
For classification and regression, follow this main path:
Data → Train / test → Preprocessing as needed → Model → Metrics.
For clustering, use Clustering sample in place of Train / test. The model looks for groups without a labeled target.
Connect Explore data to the table and Prediction quality to the model output.
Prepared datasets for your task
Hold out a separate set for evaluation
Preprocessing is fitted on the training set only.
Where your work is saved
Find your experiments in My experiments. In the current beta, drafts and imported results are stored in this browser. They do not automatically sync across devices. The notebook created by the Colab integration is saved in your Google Drive.
Start with one question
Choose a preset, run your first experiment and change a single parameter. Compare the results. As you see how the data and model settings affect predictions, you can move from presets to ML workflows of your own.