Merits of an Ensemble Deep Neural Net Vs. Machine Learning

Ankur Verma (https://www.linkedin.com/in/ankur-verma-350b7844/) wrote a thought-provoking research paper few years ago.

The paper talks about benefits of an Ensemble Deep Neural Network architecture (Deep learning) over Machine Learning models. Industrial application to use cases such as Demand Forecasting, Inventory, Promotion, and Assortment optimization.

The research was performed on a large data set. Fast forward today, this approach has been successfully executed with relatively small data set, for med-enterprise companies in e-commerce, CPG and Retail sector. In reality, this approach is proving to be very effective for med-enterprise considering the limited data history and growth /evolving business conditions.

I have tried to summarize the learnings for business and supply chain planning community. Complete paper is available here — https://arxiv.org/abs/2010.06952

Three big takeaways

Ensembling: Proved ensemble modelling is superior. Ensembling, whether in Deep Learning or Machine Learning (ML), involves merging multiple forecasting models to enhance the accuracy and reliability of predictions. This method capitalizes on the “wisdom of the crowd” principle, recognizing that aggregating predictions from various models typically yields superior outcomes compared to relying solely on any single model.

Non-linear relationships: Deep Neural Networks have demonstrated superior effectiveness in capturing non-linear patterns within demand data and producing precise forecasts. Non-linear relationships in statistical demand forecasting occur when the connection between predictor variables (e.g., time, price, promotions) and product or service demand doesn’t follow a linear pattern. Therefore, managing non-linear relationships in demand forecasting demands advanced modelling techniques capable of comprehending intricate data interactions and patterns. By considering and accounting for non-linear relationships in demand forecasting, businesses can develop more accurate and robust forecasting models that better reflect the complexities of consumer behavior and market dynamics.

Cross-learning: Deep Neural Networks have demonstrated superior effectiveness cross-learning. Cross learning in demand forecasting with neural networks means using what we’ve learned from one set of data to do better forecasting on another set that’s similar but not exactly the same. It’s useful when we’re predicting demand for different products, areas, or times where the data might be limited or different. Therefore, it proves to be very useful for use cases such as new product forecasting.

TrueGradient AI was founded with a mission to address time-to-value gaps in the Supply Chain Planning domain. We are committed to democratize AI and make advanced techniques like Deep Learning, Auto-ML and Stochastic Optimization accessible/ easy to adopt for all supply chain professionals and organizations.

If you’d like to learn more, please visit our website and sign up for demo: https://truegradient.ai/