According to a recent McKinsey survey, just 16 percent of companies have managed to take data science models involving deep learning techniques beyond the pilot stage. One reason for this could be the sheer complexity of these models. A single, large, complicated model can easily be overwhelming for data scientists to conceptualize and build. In addition, these models are often difficult for project managers to oversee effectively.

The above statistic is already a compelling argument for moving toward less complex data science models. But if that’s not enough, we’ve outlined five reasons below for why data science should embrace simpler models.

  1. Reliability

Large data science projects can be broken down into sub-models that can be conjoined or stacked. Companies can then perform a ceiling analysis to identify the strongest and weakest models and determine where model accuracy needs improvement. This approach is extremely reliable and ensures that the data science team can efficiently isolate and fix the weakest parts of the stack.

  1. Accountability

Simple, effective knowledge transfer is critical for the data scientists that build and maintain the models and for the stakeholders on the business side that utilize the model outputs. When working with simple models, it’s easy to designate and transfer ownership if there are changes in team members or new additions to the team.

  1. Interpretability

Communicating the insights derived from data science models to the technology, product, and business teams is a central part of the data scientist role. This job is exponentially easier when using a combination of simple models. 

In addition, it’s easier to determine where resources are needed to invest in data collection and refinement, create a clear roadmap for further investment in model building, and assign internal resources to act upon the results of the model output.

  1. Sustainability

Using multiple simple models makes monitoring and maintaining model performance significantly easier. Another benefit is that if a query breaks or part of the model needs retraining, the other models can continue producing business output.

  1. Executability

There is a direct correlation between the simplicity of a model and its transparency. The more transparent a model is, the more likely key stakeholders will be to act upon its results as they will have a well-defined understanding of its scope and output.

For more on these steps and why simple models can be so effective in solving complex problems, check out this recent Forbes article.