Building Powerful Machine Learning Models Without Writing Code Using Orange
Machine learning (ML) is one of the hottest areas of data science and artificial intelligence. ML models power everything from product recommendations to fraud detection to self-driving cars. But the conventional wisdom has been that to do machine learning, you need strong programming skills in languages like Python or R.
However, a new generation of visual, drag-and-drop machine learning tools is challenging this assumption. Open-source tools like Orange aim to democratize ML and empower non-programmers to build powerful predictive models and uncover insights from their data.
In this post, we‘ll dive deep into using Orange for a real-world supervised learning project. By following along, you‘ll learn essential machine learning concepts and see how you can apply them, all without writing a single line of code!
What Makes Orange Unique?
Orange is a data science and machine learning platform developed by an academic team at the University of Ljubljana in Slovenia. Orange‘s lead developers Blaž Zupan and Janez Demšar first conceived of the tool as a way to interactively teach machine learning concepts to their students.
As an academic project, Orange‘s philosophy emphasizes interactivity, experimentation, and education. The tool is designed to allow users to easily explore data, try different models, and see the results visually.
Some key tenets of Orange‘s design:
- Visual, drag-and-drop interface using "widgets"
- Open-source and free to use and modify
- Strong emphasis on interactive visualizations
- Extensible with add-ons and custom widgets
- Suitable for both beginners and advanced users
While there are several other visual machine learning tools like KNIME and RapidMiner, Orange stands out for its educational focus, interactive visualizations, and fully open-source nature. Let‘s see what it can do!
Machine Learning Project Walkthrough: Predicting Bank Term Deposit Subscriptions
To illustrate Orange‘s capabilities, we‘ll walk through a supervised classification project using a popular dataset from the UCI Machine Learning Repository. The dataset contains information on marketing campaigns for term bank deposit subscriptions, with the goal of predicting which customers will subscribe.
Step 1: Import and Explore Data
After launching Orange, we start by importing our dataset using the File widget. Upon connecting the File to a Data Table widget, we can inspect the raw data:
Already, we can see some important characteristics of our dataset:
- Our target variable is
y– whether the customer subscribed to a term deposit - We have a mix of numerical (age, balance, day), categorical (job, marital), and binary (default, housing, loan) predictor variables
- The dataset contains over 45,000 records
Next, let‘s do some exploratory data analysis and visualization. With Orange‘s Scatter Plot widget, we can examine the pairwise relationships between variables:
This plot reveals some interesting insights:
- There appears to be a positive relationship between age and account balance
- Customers with higher balances were more likely to subscribe to a term deposit (blue dots higher on y-axis)
The Distributions widget lets us see the univariate distributions of variables. For example, we can see that our dataset has an imbalanced class problem – there are far more non-subscriptions (red) than subscriptions (blue):
Class imbalance is a common issue in ML where the target class we‘re interested in predicting is underrepresented. This can lead to models that have high overall accuracy but perform poorly at predicting the minority class. We‘ll discuss ways to handle imbalance later.
Step 2: Data Preprocessing
Although Orange handles a lot of data prep tasks automatically, it‘s still important to think carefully about how to treat different variable types:
- Numerical variables like age and balance can be used as-is
- Categorical variables like job and education should be one-hot encoded into binary dummy variables
- Binary variables like default and housing can be used as-is
We also may want to standardize the numerical variables so they‘re on the same scale.
To handle these preprocessing steps in Orange, we can use a combination of widgets:
- Categorical to Numeric for one-hot encoding
- Impute for filling in missing values
- Normalize for standardization
Here‘s what the preprocessed data looks like:
Step 3: Train and Evaluate Models
Now for the fun part – building ML models! Orange has a variety of classification algorithms we can try, including:
- Logistic Regression
- Decision Tree
- Random Forest
- Neural Network
- Support Vector Machine
For this problem, we‘ll compare three models: Logistic Regression as a linear baseline, Random Forest as a powerful ensemble method, and Neural Network as a black-box nonlinear model.
With Orange, training a model is as easy as connecting the preprocessed data to the model widget of your choice. We can view a trained model‘s predictions by connecting it to a Predictions widget:
Evaluating a model is important to gauge its performance and see how well it generalizes to unseen data. With the Test and Score widget, we can easily get evaluation metrics based on a held-out test set or cross-validation:
Our logistic regression achieves 90.6% accuracy and 0.67 AUC (area under the ROC curve). Not bad for a simple linear model!
Let‘s compare that to a random forest:
The random forest performs even better with 92.6% accuracy and 0.77 AUC. This shows the power of ensemble models that combine predictions from many individual decision trees.
Finally, let‘s try a neural network:
The neural network performs comparably to the random forest at 92.4% accuracy and 0.76 AUC.
So which model should we choose? It depends on our priorities. If interpretability is important, logistic regression or a single decision tree may be preferable since we can easily see the learned coefficients or decision rules. If performance is the main concern, the random forest or neural network will likely do better on unseen data.
The confusion matrices give an idea of each model‘s error patterns:
We can see that the models are better at predicting the majority non-subscription class than the minority subscription class. This goes back to the class imbalance issue we identified earlier.
To address imbalance, we could try oversampling the minority class, undersampling the majority class, or using a cost-sensitive learner that weights the minority class more heavily. In Orange, the Preprocess widget has options for over/undersampling:
After oversampling, we see that the models‘ precision and recall on the minority class improve:
Step 4: Inspect and Interpret Models
While it‘s important to have accurate models, we also want to understand what they‘ve learned and how they make predictions. ML interpretability is a hot research area and Orange provides some nice tools for it.
For example, Orange has a unique interactive visualization for inspecting decision trees:
We can see how the tree splits on the most informative features and what decision paths lead to subscription predictions.
The Nomogram widget shows a more concise visual representation of a logistic regression model‘s coefficients:
The top features contributing to a subscription prediction are duration of the last contact and number of contacts. This gives valuable insight into what factors drive the model‘s decisions.
Examining permutation feature importances is another way to interpret ML models and see which variables have the biggest impact on predictions:
For the random forest, we see that duration is by far the most important feature, followed by month of last contact and age. Less important features could potentially be removed to simplify the model.
Advanced Techniques in Orange
Beyond the core supervised learning workflow we‘ve focused on so far, Orange supports many advanced ML techniques including:
- Unsupervised learning for clustering, dimensionality reduction, and anomaly detection
- Text mining and natural language processing
- Image analytics and computer vision
- Time series forecasting
- Recommender systems
- Network analysis
Orange also has robust tools for data integration (merging multiple datasets), automated data preprocessing (one-hot encoding, imputation, normalization), and automated machine learning (AutoML).
AutoML has become a hot area that aims to automate the process of algorithm selection and hyperparameter optimization to arrive at high-performing models faster. With Orange‘s AutoML widget, you can quickly try many algorithms and preprocessing steps to get a strong baseline model:
After running AutoML, we see that it found a boosted tree model that slightly outperforms our random forest, achieving 93.1% accuracy and 0.79 AUC. The AutoML widget is a great way to get a quick baseline and see if simple models can perform well with minimal tuning.
The No-Code ML Revolution
We‘ve seen how visual tools like Orange are making machine learning more accessible than ever before. Data scientists are using Orange to accelerate their workflows, while domain experts and "citizen data scientists" are using it to derive insights without having to learn programming.
But Orange is just one part of a larger trend toward no-code AI and ML. Other visual tools like KNIME and RapidMiner as well as automated machine learning (AutoML) services from major cloud providers are putting the power of ML into more people‘s hands.
I spoke with Orange‘s lead developer Dr. Blaž Zupan about why he sees no-code tools as the future of data science:
"Machine learning is too important to be left only to highly-trained data scientists and ML engineers. Visual tools like Orange allow anyone to ask questions of their data and quickly test their ideas. It‘s really about democratizing ML and empowering more people to solve problems with data."
Of course, no-code ML is not a panacea. It‘s still important for serious practitioners to understand the underlying algorithms and techniques. And coding is still necessary for developing custom algorithms and deploying models to production.
But for many common business use cases, no-code ML tools can provide a huge productivity boost and allow data science to scale beyond a small team of experts. As the tools continue to advance and add features like AutoML, the skill barrier to entry will keep declining.
Getting Started with Orange
If you‘re new to data science and ML or don‘t have a coding background, Orange is a great way to start learning essential concepts and doing real projects. Here are some tips for getting started:
- Download and install Orange from the official site
- Work through the built-in interactive tutorials to learn the basics
- Explore the sample workflows for examples and inspiration
- Try replicating a project you find interesting on Kaggle or elsewhere
- Join the Orange community on Discord to get help and connect with other users
You can find detailed Orange documentation and tutorials at:
Wrapping Up
Machine learning doesn‘t have to be rocket science or require years of coding experience. With user-friendly tools like Orange, anyone can start asking interesting questions of data, building powerful predictive models, and extracting valuable insights.
In this walkthrough, we covered essential ML concepts like exploratory data analysis, data preprocessing, model training and evaluation, and model interpretation, all without writing a line of code. But we‘ve only scratched the surface of what Orange can do.
Whether you‘re an aspiring data scientist looking to build your skillset or a domain expert wanting to solve business problems with ML, Orange is a fun and powerful platform for learning and doing data science. Try it out, join the community, and see what you can discover!
