Part 10 - Why Iterative Problem-Solving is important to improve ML models

10-part series about building your first Data stack from 0 to 1, and be ready for AI implementation.

Hello Readers!

I can’t believe we are already at last part of the series. This time we will reflect on what we learned so far, and most importantly, how to make our ML models get better by itself by implementing feedback loops.

Remember, how you learned to ride a bike, did you get it the first time ? No, right ? It takes practice!

Similarly, ML models will get better overtime as it learns from the patterns and data. This also means we as business owners need to make sure we feed the beast with feedback loop.

ML models will improve by implementing 5 level iteration process in your workflow :

• Human level- improving business strategy and KPI around ML use case

• Data level - Improve your Data and features needed to make better predictions

• Model level - Explore different models that may suit better with changing demands

• Parameter level - Tune different hyperparameter for type of model to get best results

• Culture level - Continued learning and understanding ML output at org level

Let’s dive into each point to understand better:

---

AI has become the buzzword that's been dominating tech conversations for years now. But it's not as complicated as most people make it out to be.

Sure, there's complex math involved under the hood. But for you, as a decision-maker, AI is more about iterative problem-solving than crunching numbers.

1. Your business strategy is the real driver of ML success

Think about it this way: the best algorithm in the world won't help if it's solving the wrong problem.

So, before you dive into the technical aspects of ML, start with your business strategy. Ask yourself:

• What are our key business objectives?

• What problems are we trying to solve?

• How could better predictions or insights help us achieve our goals?

These questions form the foundation of your ML journey. They guide every decision you'll make along the way.

For example, in our e-commerce example. Your main objective might be to increase customer lifetime value. With this in mind, you might use ML to:

• Predict which customers are likely to churn

• Recommend products that customers are most likely to buy

• Optimize pricing strategies for maximum revenue

But here's where the iteration comes in. As you implement ML solutions, you'll learn more about your business. You might discover new opportunities or challenges you hadn't considered before.

Maybe you find out that reducing churn isn't as impactful as you thought. Or perhaps you discover that optimizing your supply chain could have a bigger impact on your bottom line.

This iterative process - implementing, learning, and adjusting - is what truly drives ML success. It's not about getting it perfect the first time. It's about continuous improvement.

So, as a business leader, your job isn't to understand the intricacies of neural networks or gradient descent. Your job is to:

1. Set clear business objectives

2. Identify potential ML applications that align with these objectives

3. Implement solutions

4. Learn from the results

5. Adjust your strategy

6. Repeat

This iterative approach ensures that your ML initiatives always align with your business goals. And that's how you drive real value.

2. Turn your business insights into powerful ML features

Data is the fuel that powers machine learning, we learned that in this series.

As a business leader, you have invaluable insights into your industry, your customers, and your operations. These insights can be turned into powerful features for your ML models.

What's a feature, you ask? In ML terms, a feature is simply an input variable used in making predictions. It could be anything from a customer's age to the time of day they usually make purchases.

The key is to identify features that are likely to be predictive of the outcome you're interested in. And that's where your business knowledge comes in.

Let's go back to our e-commerce example. We did pick the following features in ML model we built :

• Time since last purchase

• Total spend in the last 3 months

But don't stop there. The iterative nature of ML means you should constantly be thinking about new features to add.

Maybe you notice that customers who purchase a particular product are more likely to become repeat buyers. That's a new feature to add to your model.

Or perhaps you realize that weather has a significant impact on your sales. Time to incorporate weather data into your predictions.

The process looks something like this:

1. Start with features based on your current business insights

2. Build and run your model

3. Analyze the results

4. Identify new potential features based on these results

5. Add these new features to your model

6. Repeat

This iterative approach to feature engineering ensures that your ML models are always improving, always incorporating the latest business insights.

As your models improve, they'll start generating new insights. These insights can then inform your business strategy, creating a virtuous cycle of improvement.

So don't get caught up in thinking you need a PhD in statistics to come up with good features. Your business knowledge is your superpower. Use it.

3. Choose the right ML approach for your specific needs

Choosing an ML approach is less about understanding complex algorithms and more about understanding your business needs.

We learned these common types of business problems and the ML approaches that might be useful:

1. Prediction problems (e.g., predicting customer churn):

   • Logistic Regression

   • Random Forests

   • Gradient Boosting Machines

2. Clustering problems (e.g., customer segmentation):

   • K-Means Clustering

   • Hierarchical Clustering

3. Recommendation problems (e.g., product recommendations):

   • Collaborative Filtering

   • Content-Based Filtering

4. Anomaly detection (e.g., fraud detection):

   • Isolation Forests

   • Autoencoders

Don't worry if these terms sound unfamiliar. The important thing is to understand the type of problem you're trying to solve.

But remember, choosing an ML approach isn't a one-time decision. It's part of the iterative process.

You might start with a simple approach, like logistic regression for a prediction problem. As you learn more about your data and your problem, you might decide to try more complex approaches, like gradient boosting machines.

The key is to start simple and iterate. Here's a process you can follow:

1. Start with a simple, interpretable model

2. Evaluate the results in business terms (not just statistical metrics)

3. If the results aren't satisfactory, try a more complex approach

4. Compare the results of different approaches

5. Choose the approach that gives the best business results

6. Repeat this process regularly as your data and business needs evolve

This iterative approach allows you to balance complexity with business value. You're not choosing an ML approach based on what's trendy or what sounds most impressive. You're choosing based on what works best for your specific needs.

And as your needs change, your approach can change too. That's the beauty of thinking about ML as an iterative process rather than a one-time implementation.

4. Optimize your models to maximize business ROI

Now we're getting to the heart of why you, as a business leader, care about machine learning: ROI.

Optimizing ML models isn't just about improving accuracy. It's about maximizing business value.

This is where many organizations go wrong. They focus on technical metrics like accuracy or F1 score, without considering the actual business impact.

So how do you optimize for business ROI? It's all about iteration and alignment with business goals.

Here's a process you can follow:

1. Define clear business metrics: What KPIs will you use to measure success? It could be increased revenue, reduced costs, improved customer satisfaction, or any other relevant business metric.

2. Establish a baseline: What are your current KPI values without the ML model?

3. Implement your initial model: Use the approach you chose in the previous step.

4. Measure the business impact: How did the KPIs change after implementing the model?

5. Identify areas for improvement: Where could the model be doing better from a business perspective?

6. Make incremental changes: This could involve adding new features, trying a different ML approach, or adjusting how you use the model's outputs.

7. Measure again: Did the changes improve your business metrics?

8. Repeat steps 5-7: Keep iterating until you're satisfied with the results.

So you've implemented a model to predict customer churn for your subscription service. Your KPI is the number of customers retained each month.

You implement your initial model and see a 5% increase in customer retention. That's good, but you think you can do better.

You notice that the model is particularly bad at predicting churn for new customers. So, you add some features specific to new customers, like their onboarding experience and initial usage patterns.

You run the model again and now see a 8% increase in customer retention. Better, but still room for improvement.

Next, you realize that you're not using the model's outputs effectively. Instead of just flagging customers at risk of churn, you start using the model to tailor retention offers based on the predicted risk level.

Now you're seeing a 12% increase in customer retention. That's a significant improvement in your KPI, driven by iterative optimization of your ML model.

This process never really ends. You should be constantly monitoring your model's business impact and looking for ways to improve.

Remember, the goal isn't to build a perfect model. The goal is to continuously improve your business outcomes. And that comes through iteration, not perfection.

5. Boost your bottom line through continuous ML improvement

The business world doesn't stand still, and neither should your ML initiatives.

Continuous improvement is also about creating a culture of learning and adaptation throughout your organization.

Here's how you can foster this culture:

1. Regular review cycles: Set up quarterly or bi-annual reviews of your ML initiatives. Are they still aligned with your business goals? Are they delivering the expected value?

2. Cross-functional collaboration: Encourage ongoing dialogue between your data science team and other departments. Business insights can inform model improvements, and model insights can inform business strategy.

3. Stay curious: Always be on the lookout for new data sources or business changes that could impact your models. Could social media sentiment improve your customer churn predictions? Could macroeconomic indicators enhance your sales forecasts?

4. Embrace failure: Not every ML initiative will be a home run. That's okay. Treat failures as learning opportunities. What went wrong? What can you do differently next time?

5. Invest in infrastructure: As your ML initiatives grow, you'll need robust systems for data management, model deployment, and monitoring. This infrastructure enables faster iteration and improvement.

6. Educate your team: The more your entire organization understands ML (at a high level), the more they can contribute to its success. Consider offering basic ML training to all employees.

7. Track long-term trends: Look at how your ML models perform over time. Are they getting better? If not, why?

Let's see how this might play out in practice.

Imagine you're using ML for inventory management. Your initial model does a decent job of predicting demand, reducing stockouts by 20%.

In your quarterly review, you notice that while overall performance is good, the model struggles during holiday seasons. You dig deeper and realize it's not accounting for seasonal trends effectively.

You work with your data science team to incorporate seasonal features into the model. Now, stockouts are reduced by 30%.

A few months later, your marketing team mentions they're planning a major promotional campaign. You realize this could significantly impact demand, so you quickly update your model to include marketing spend as a feature.

As the year progresses, you notice that your model's performance is plateauing. You invest in a new data pipeline that allows you to incorporate real-time sales data. This lets you update your predictions daily instead of weekly, pushing stockout reduction to 40%.

Each of these improvements came from continuous attention and iteration, not from getting it perfect the first time.

This approach – constant learning, adapting, and improving – is what separates successful ML initiatives from those that fizzle out after initial implementation.

And here's the best part: as you get better at this iterative process, you'll see compounding returns. Each improvement builds on the last, creating a virtuous cycle of increasing business value.

Conclusion: Embrace the iterative mindset

We've covered a lot in this series. We've seen how AI is far from being just about complex math, is really about iterative problem-solving.

This iterative mindset is your secret weapon in the world of ML. It allows you to:

• Start ML initiatives without needing to get everything perfect from the beginning

• Adapt quickly to changing business needs

• Continuously increase the value you get from your ML investments

• Foster a culture of learning and improvement throughout your organization

Embrace the simplicity combined with iterative mindset, and you'll be well on your way to AI success, because you've mastered the art of learning and adapting in a data-driven world.

And that, more than any algorithm, is what will drive your business forward, even when the hype cycle is over.

Happy learning and see you in brand new series next time!