Your team is divided on machine learning models for data mining. How will you navigate conflicting opinions?

I hold a strong opinion that "Data hides more than it reveals". Most data science projects fail due to improper cleaning and lack of business logic before creating data. Additionally, 90% of models will need to be retrained after a certain duration because real data changes its distribution at the production level. Hence, it is crucial to be very cautious in data mining for any project, ensuring it is driven by business understanding. Navigating conflicting opinions on ML models in DS team is like trying to decide on a pizza topping. Some swear by the classic simplicity of linear regression while others advocate for the NN. personally i usually prefer classical ML model over NN due its robust nature which and can be correlated with bussines

When navigating conflicting opinions on machine learning models for data mining, foster open dialogue, evaluate evidence with performance metrics, conduct pilot tests, and build consensus based on data and strategic goals. This approach ensures a data-driven decision and promotes team collaboration.

To navigate conflicting opinions on machine learning models for data mining, focus on aligning discussions with clearly defined goals. Start by revisiting the project's objectives and key performance indicators (KPIs) with the team. Facilitate a data-driven comparison of the proposed models, evaluating how each aligns with the goals and performs against the KPIs. Encourage open dialogue, allowing each team member to present their model's strengths and potential drawbacks. Use objective metrics and empirical results to guide the decision-making process. By centering the conversation on the project's defined goals and evidence-based evaluation, you can foster consensus and a collaborative resolution.

I like taking a technical and pragmatic perspective. MLOps have matured. Once you have your business requirements, a well-defined problem, goals, and appropriate data and you have done your intense processing and analysis, you want to decide on the model, note that the best model is the one that will give you the best performance. I don’t think you can really have it by just looking at the data. My strategy has always been to create multiple pipelines for multiple models’ trainings and optimisation. And I have another pipeline that compare the outputs of the primary models. The best model, evaluated based on predefined metrics - depending on the type of task, is selected for the final task - it always works for me.

Begin by setting up a criteria list that includes key performance metrics, scalability, ease of implementation, and alignment with project goals. Encourage team members to present their preferred models, backed by empirical data and real-world examples. For instance, one model may excel in terms of accuracy but require extensive computational resources, while another might be less accurate but more efficient. Discuss trade-offs openly and highlight each model’s strengths and weaknesses in the context of the project’s specific requirements. Use comparative analysis tools and visualization techniques, such as performance graphs and confusion matrices, to provide a clear, objective basis for the discussion.

From my viewpoint, the importance of setting clear goals cannot be overstated. In my experience, having well-defined objectives not only guides the technical aspects of the project but also fosters better communication and collaboration among team members. It helps in prioritizing tasks and allocating resources efficiently, ensuring that everyone is on the same page. Moreover, clear goals provide a framework for measuring success and identifying areas for improvement. For instance, if the aim is to enhance customer segmentation, my focus would be on understanding customer behavior deeply and using insights to tailor marketing strategies effectively. By aligning our efforts with these goals, we can maximize the impact of our data mining .

?? Without data, you’re just another person with an opinion. ?? - W. Edwards Deming. According to a recent study, 85% of data science projects fail due to poor data quality (Gartner). In my experience, assessing data quality is crucial. For example, preprocessing steps like data cleaning can expand model options. One thing I’ve found helpful is that diverse data sources improve model accuracy. A common mistake is ignoring bias. You should leverage ethical AI frameworks to ensure fairness and transparency.

Our data is the foundation of our machine learning efforts. Advanced models need clean, abundant data like high-performance cars need premium fuel. With dirty, gappy data, we might need a more forgiving model. Let's assess our data quality first. If improvable, focus on: Cleaning: Handle missing data and outliers. Optimizing: Encode categories, engineer features. Balancing: Address imbalances, reduce noise, augment if needed. This process helps choose an appropriate model - both theoretically sound and practically effective. Enhancing data quality can open up options for sophisticated models or improve simpler ones' performance. Our goal is to match our model choice to our data reality for best results.

In my experience, to navigate conflicting opinions on machine learning models for data mining within your team, I would choose option 4, "Test Rigorously." By implementing thorough testing of each proposed model, you can objectively assess their performance based on predefined metrics and criteria. This approach minimizes bias and personal preferences, ensuring that the selection process is grounded in empirical evidence. Rigorous testing will help identify the model that best meets the project's goals, effectively addressing concerns related to model merits, data quality, and future-proofing, while also providing a feedback loop based on concrete results.

Bring in External Expertise: If the team remains divided, consider consulting an external expert or someone with more experience to provide an impartial perspective.

Incorporating a feedback loop is crucial for ensuring that models remain accurate and relevant over time. By continuously monitoring model performance and integrating real-world feedback, data scientists can identify and address issues that weren't evident during initial testing. This iterative process allows for fine-tuning the model, ensuring it adapts to changing data patterns or business needs. Additionally, it deepens the understanding of model behavior in production environments, which is essential for improving data-driven decision-making.

To navigate conflicting opinions, focus on two key criteria: Define Desired Outcomes: Set clear goals with its KPIs to evaluate success and align on them. In a recent identity graph project, goal ambiguity caused friction. We refined our goals iteratively, achieving clarity while remaining flexible on details. Gather Opinions: Understand team reasoning. In the same project, we debated using actual data versus dummy data for validations. Actual data provided better insights, avoiding analysis paralysis and the pitfall of making decisions based on "this is how it’s always been done." Ultimately, the most valuable models in my experience have involved close collaboration between all teams, consistently adding value year over year.

Reflect and Document: After a decision is made, take time to reflect on the process with the team to understand what worked well and what could be improved for future decisions. Document the rationale behind the final choice to help align future discussions.