How do you identify AI/ML problems and objectives?

When everyone around is talking about GenAI it is difficult to not solve all problems using AI. However, it is very important to understand the problem at hand. One of the important aspects to consider while solving a problem with AI is the right assessment of the underlying data. 3 parameters to asses the data - - Data Accessibility: Are there controls around the data which makes it difficult to use for our purposes? - Data Quantity: Is there enough data to derive useful insights and patterns? - Data Quality: Does the data have all the necessary values required to solve the problem? If any of the above parameters is not met, I would reassess the use of AI for sol inh the problem.

🎯 Prioritize Problem Identification! I'd stress the critical nature of laying a robust foundation for AI/ML projects. Before diving deep into the world of algorithms, you'll need to understand the essence of the problem. This isn't merely about technological prowess but about aligning technology with real-world needs. Emphasize the distinction between problems that genuinely demand AI/ML intervention and those that might be better addressed with more straightforward methods. Whichever direction organizations opt for, the initial step of problem definition remains the cornerstone of AI/ML success. 🤖🎓

With new technologies there is a tendency to look for a problem that fits the solution at hand, rather than the other way around. Starting with a strong problem identification process is critical. Success criteria need to be clearly defined. All too often I find that we are tempted to define soft goals which are not measurable.

Start by clearly articulating the problem you want to address. What are the pain points, challenges, or opportunities? Consider the context in which the problem exists and how it impacts your organization or stakeholders. Determine what you want to achieve with AI/ML. Your objectives should be specific, measurable, achievable, relevant, and time-bound (SMART). For example, if you're working on a recommendation system, your objective might be to increase user engagement by 20% within six months.

AI/ML is merely a tool, out of many tools, used to solve real problems. While it is immensely powerful and versatile, it may not always be the best tool to solve a problem and one shouldn't fall into the trap of "to someone who has only a hammer, everything looks like a nail". A common cause for AI projects not being valuable is that the wrong problems are solved. AI/ML people are frequently not experts in domains other than AI/ML. As such, they frequently do not recognise which problems are valuable. Similarly, domain experts frequently don't know which problems are best solved by AI, and therefore do not make the right requests of their AI experts. It is best to have people who are expert in AI and the domain to specify the problem.

Data plays a pivotal role in AI/ML projects. Based on my experience, leveraging interactive workshops that engage various teams within a company and stakeholders proves to be an effective tactic for overcoming data-related challenges and choosing the right approaches for AI initiatives. Such collaborative sessions are instrumental in pinpointing the so-called 'low-hanging fruits' – tasks or projects that can promise quick wins with relatively little effort, like potential Proofs of Concept. However, it also depends on the maturity level of the organization as someone mentioned here.

Depending on maturity level of the organization, consider launching several POCs to explore use cases and opportunities. These should run a small scale, agile fashion, to allow participants more of a chance to learn and explore, rather than be focused on succesful deliveries.

While the 'quality' of data is frequently mentioned, it encompasses multiple facets. Completeness, accuracy, timeliness, relevance, and consistency come under data quality. A common misconception is that larger datasets are inherently better. In truth, a smaller, high-quality dataset can sometimes be more valuable than a vast yet inconsistent one. A question often overlooked is, "Who or what does this data represent?" Unrepresentative data can lead to models that exhibit bias or lack generalizability. For instance, a facial recognition model trained primarily on data from one demographic may perform poorly on others, leading to fairness and equity concerns.

Before you even think about using AI or ML, you've got to know what kind of data you're working with. Is it relevant to the problem you're trying to solve? Is it accurate and up-to-date? If your data is a mess, you've got to clean it up first. Otherwise, you're setting yourself up for failure. If you find out you don't have enough good data, you've got a decision to make. You can either go out and get more data, which might cost you, or decide that AI/ML isn't the way to go this time. Bad data leads to bad decisions, so ensure you've got the good stuff before diving in. Remember, garbage in, garbage out.

Assess the availability and quality of data. Data is a critical component of AI/ML projects, so ensure you have access to the data required to solve the problem. Consider issues like data volume, variety, velocity, and veracity

First, match the problem you want to solve to the appropriate type of learning algorithm: * Consider supervised learning models like linear regression, random forests, or neural networks to predict outcomes based on labeled training data. * Try unsupervised learning techniques such as clustering algorithms or dimensionality reduction to find patterns and groupings within unlabeled data. * Reinforcement learning models that optimize actions over time may be suitable for real-time decision-making, as needed in gaming or robotics. Evaluate specific algorithms that fit your data type, volume, and resources. Experiment with different models; the data may be better suited to particular algorithms. Test, monitor, review and repeat.

A problem can be solved by many approaches, a regression can be solved by classification in logit space, can be solved using RL, or why not use LLMs, it really comes down to how you mathematically formulate the problem statement. But we can start simple by choosing the most easy way, and then improve iteratively. Many times i have found myself over complicating things, leading to endless complications.

Example: as a data scientist at an e-commerce company, your goal is to enhance customer retention by predicting customer churn. This involves defining the problem as a binary classification task, exploring available data, and engineering relevant features like RFM and customer lifetime value. Model selection should focus on recall and the F1-score due to the imbalanced nature of churn data. After splitting the data for training, validation, and testing, you should experiment with various algorithms, select the best-performing model, and deploy it. Continuous monitoring and retraining ensure long-term efficacy. Additionally, consider ethical implications, model explainability, and the business impact of your predictions.

It is also important that not just traditional AI/ML metrics (like RMSE, MAE, F1 etc)are used to assess model performance, but also real-world metrics, like profit, savings etc. The approach should be appropriate for the type of data used (and that will reliably be available going forward), the cost/benefit of using individual pieces of data, and the broader system into which the model will be integrated.

While accuracy or error rates are commonly used metrics, other metrics like precision, recall, F1-score, or AUC-ROC might be more relevant depending on the application. Additionally, consider real-world implications. A 99% accurate model might seem great, but it demands reconsideration if the 1% error can have severe consequences (like in medical applications). The theoretical alignment of a problem to an approach is a starting point. Iteratively experimenting with various algorithms and refining them based on performance is key. This iterative process, often called "the ML loop," is fundamental to achieving the best results.

Tactical view of the scope is important, and it is important to highlight the vision considering the strategic view. AI is a tool for transforming the business, and it is important to add mechanism of how you will calculate the ROI.

AI/ML is not a cure-all. Just like the internet and a cashless society, it can enhance, supplement, and increase the quality of life and business. On the flip side, it creates new avenues for exploitation and corruption. Strengths, weaknesses, opportunities, and threats must be evaluated for initiating a AI/ML campaign itself.

Defining the project scope is key. How many of us have been in never-ending projects where "just one more" feature is needed to make it perfect? Setting clear and measurable objectives is essential. It's also crucial to be aware of the scope and limitations of any project, not just in AI/ML. Remember the limitations of the technology and present them clearly to all stakeholders. Additionally, it's always vital to measure progress and have clear KPIs to evaluate the project's success. It's better to have a limited project that meets objectives and budget than to aim for a "perfect" project that never ends.

Defining the scope of an AI/ML project is crucial for its success. Key steps include identifying business objectives, assessing feasibility and value, clearly defining project scope, considering data needs, planning modeling and deployment, allocating team resources, and addressing ethical and legal considerations. By following these steps, you align your project with business goals, ensure technical feasibility, and uphold ethical standards, setting the project up for success.

Key considerations for AI/ML solutions: 1. Define features and functionalities. 2. Address dependencies for feasibility. 3. Set realistic goals for time and budget. 4. Build the right team with the necessary skills. 5. Manage pitfalls and uncertainties effectively for success.

Effectively communicating AI/ML project objectives, limitations, and challenges to stakeholders and users is crucial for managing expectations and building trust. Key steps include defining clear objectives and success metrics, tailoring communication to different audiences, being transparent about limitations and challenges, using visuals and storytelling, and creating a comprehensive communication plan. This approach fosters understanding, emphasizes value, and ensures project success.

This is especially the case with AI/ML projects, which are unfamiliar to many stakeholders. It is also frequently the case that certain stakeholders may feel threatened by the AI/ML in that the project may intrude into areas where they have previously worked or led. Without proper communication and consultation around this, it is very easy for "corporate antibodies" to hamper, or even entirely frustrate, an AI/ML project.

Key considerations for successful AI/ML projects include securing support and resources by defining project goals and viability, gathering input to mitigate risks, building trust through transparency and stakeholder involvement, and emphasizing effective communication and collaboration. These steps ensure technical requirements are met and value is delivered to stakeholders and users.

Problem identification is not a one-time activity. Iterate and refine your problem statement and objectives as you gather more insights and information

In many cases, it may be possible to build a "quick and dirty" model with data that one has immediately on hand. Although this is not intended to the final deployed model, it is frequently the case that such models are already capable of generating value, and can generate value while improved models are developed. Additionally, such models allow one to learn about, and deal, with any integration and operationalisation issues.

To continuously improve an AI/ML project, follow these steps: evaluate the model's performance using thorough testing and relevant metrics; gather and analyze feedback to gain insights; iteratively refine through feature engineering, hyperparameter tuning, and data augmentation; monitor and maintain the model's performance over time; and foster collaboration and communication among team members. This iterative process ensures the model remains valuable and effective while adapting to user and business needs.

Key aspects to enhance your AI/ML project: 1. Testing in various scenarios 2. Optimization and fine-tuning 3. Scaling and deployment 4. Maintenance and updating 5. Learning and innovation Improve your project, adapt to new data, and deliver effective solutions.

AI/ML projects are complex with many "moving parts". The more moving parts, the more likely a project is to fail. As such, projects that require the involvement of the fewest critical decision makers are usually the most successful and quickest to complete. Avoid projects that span multiple departments / decision makers, at least at first until there is wider buy-in to the AI/ML process.

Driving AI is a change management process. Make sure you have a champion for change management, or else you will find tons of resistance and slow progress.

Examples showcasing AI/ML applications: 1. Customer Churn Prediction: Using ML to analyze data for proactive retention. 2. Fraud Detection: Training AI on past transactions for efficient fraud detection. 3. Image Classification for Manufacturing: Enhancing defect detection using computer vision. Key insights: Focus on data quality, start iteratively, measure success, and consider ethics from the beginning.