What machine learning is and when it makes sense
Machine learning is the branch of artificial intelligence that lets systems learn patterns from historical data and use them to make decisions or predictions on new data. Unlike traditional programming, where a developer explicitly codes the rules, in machine learning the system infers the rules from examples.
That has important practical implications for businesses. Machine learning is powerful when there’s a recurring problem, enough historical data about it, and rules too complex or too changeable to code by hand. It isn’t the right tool for every problem, and part of our implementation work is helping you identify when ML adds real value and when a simpler solution is more efficient.
Types of machine learning and their business applications
Supervised learning: the model learns from labeled examples — historical data where we know the right answer. It’s the most common type in business applications:
- Classification: Will this customer churn? Is this transaction fraudulent? Will this component fail?
- Regression: How many units will we sell next month? What’s the optimal price for this product?
Unsupervised learning: the model finds structure in unlabeled data. Useful for customer segmentation, anomaly detection, and dimensionality reduction.
Reinforcement learning: the model learns by trial and error in a simulated environment. It applies to complex process optimization, though it’s less common in midsize companies.
The machine learning model lifecycle
An ML project doesn’t end when the model posts good metrics in development. The full cycle includes phases many projects underestimate:
Problem definition: translating the business problem into a well-defined ML problem. What exactly are we predicting? What’s the target variable? What time period matters? This phase is critical and requires close collaboration between the technical team and the business.
Data preparation: collecting, integrating, cleaning, and transforming the required data. In real-world implementations, this phase takes up most of the project’s time.
Model selection and training: experimenting with different algorithms and configurations, cross-validation to estimate real performance, and hyperparameter tuning.
Evaluation: measuring performance with business-relevant metrics, not just technical ones. A model with 95% accuracy can be useless if the 5% it gets wrong are precisely the most expensive cases.
Production deployment: integrating the model into the company’s operational systems and processes — APIs, batch jobs, integration with ERPs or CRMs.
Continuous monitoring: real-world data changes over time. A model trained two years ago may have lost accuracy because the patterns it learned no longer hold. Monitoring performance in production is essential.
When to use ML and when to use something else
Machine learning isn’t always the answer. For companies weighing whether ML suits their problem, the key questions are: Do we have enough historical data (generally thousands of examples at a minimum)? Is the problem too complex for manual rules? Does the value of better accuracy justify the cost of development and maintenance?
If the answer to any of these is no, a well-designed rules engine, classic statistical analysis, or a standard analytics solution may be more appropriate and less expensive.
Frequently asked questions
How much data do we need to train a machine learning model?
It depends on the type of problem and the model’s complexity. As a rough guide, simple binary classification models can perform acceptably with a few thousand examples. More complex problems need tens or hundreds of thousands. For highly industry-specific problems, transfer learning lets you start from models pretrained on large volumes of general data and adapt them with a smaller set of your own.
Can we use machine learning models without our own data?
You can start with pretrained or industry models fine-tuned on limited proprietary data. Still, the best results always come from models trained on your specific business data, because they capture the particulars of each company and market.
What happens when the model starts performing worse in production?
That’s known as model drift or data drift. A production model’s performance has to be monitored systematically. When degradation shows up, the options are retraining on more recent data, reviewing the input variables, or — if the business has changed structurally — redesigning the model from its original assumptions.