How do you start a new machine learning project?

Every machine learning project consists of six key steps once the problem is defined: 1. Data Collection: Gather relevant data from various sources. 2. Analysis: Examine and understand the data's characteristics and patterns. 3. Visualization: Create visual representations to gain insights from the data. 4. Cleaning: Preprocess and clean the data to remove inconsistencies. 5. Building a model: Develop and train machine learning models using the data. 6. Deployment: Implement the model in a real-world environment. 7. Bonus Step - Impact: Evaluate the practical impact and effectiveness of the machine learning solution.

An integral aspect of defining the problem that's often underestimated is the continuous dialogue with domain experts. While machine learning can uncover patterns and insights from data, a domain expert can provide invaluable context and nuance to the problem definition. This collaborative approach ensures that the developed solution is not only technically sound but also genuinely addresses the real-world challenges faced by the organization.

To start with a ML project, we need to follow the DIKW pyramid. It is: D for Data, I for Insights, K for Knowledge, and W for Wisdom. From Zero to Hero. Above all these, one crucial aspect is U, which is Understanding the problem and getting the domain expertise.

In my experience data cleaning is one of the most important steps in building a model. Each model has different data requirements, so the same data cannot be used for a variety of models. Additionally, data will most likely contain missing values, so being straightforward with assumptions on missing values is a must. Whether you are using any matrix completion algorithm or not, if you are dropping cases or not.

An essential aspect of building a model is parsimony. Starting from the simplest model possible, make it the benchmark, and then improve. Another important recommendation would be to work with a sample of the data when exploring models. It can make the process easier and faster.

Maintaining a sense of skepticism is essential when building your model. Overconfidence in initial results, especially if they seem too good to be true, can be a pitfall. Always remember to re-evaluate and challenge your assumptions. On multiple occasions, I've found that models which initially seemed near-perfect were, in fact, overfitted to the training data or had overlooked certain biases. Regularly cross-checking with out-of-sample data or using techniques like cross-validation can be instrumental in ensuring your model's robustness in varied scenarios.

There are many metrics to evaluate your models. This variety is an advantage than a one-size fits all deal. Ensure you understand what each metric measures and how you can capitalize on multiple evaluation metrics.

Post-deployment monitoring is as critical as the initial deployment phase. Over time, data distributions can shift and user behaviors can evolve, potentially affecting the model's performance. Implementing a robust monitoring system to track the model's accuracy, latency, and other key metrics, and setting up alerts for any drastic changes, ensures that the model remains reliable and continues to meet users' expectations throughout its lifecycle.

To kickstart your machine learning project, begin by thoroughly researching existing literature and white papers relevant to your project domain. Explore academic research and publications to gain insights into established methodologies, techniques, and best practices. This foundational understanding will not only inform your project's design but also ensure that you build upon and contribute to the existing body of knowledge. Leveraging insights from prior work enhances the robustness and innovation of your machine learning approach.