# A network-based transfer learning approach to improve sales forecasting of new products

**Authors:** Tristan Karb, Niklas Kühl, Robin Hirt, Varvara Glivici-Cotruta  
**arXiv:** [2005.06978](https://arxiv.org/abs/2005.06978)  
**Date:** 2020-05-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2005.06978v1/

## Key points

- Design Science Research (DSR) approach used in three cycles of investigation: rigor, relevance, and design
- Research questions focus on assessing performance of transferred models for new products compared to models without transfer
- Study aims to identify characteristics of product groups indicating success in transferring models
- Knowledge contribution defined as providing new solutions for known problems or extending known solutions to new problems
- Focus on Transfer Learning for sales forecasting in the food retail domain
- Exploration of transferring knowledge between listed stock products and new products using deep neural networks
- Application of Transfer Learning techniques allows leveraging domain knowledge and historical sales data for accurate sales forecasting
- Ensemble strategies recommended for implementing a robust Transfer Learning approach with better results than models without transfer
- Real-world representational characteristics analyzed to enhance understanding of transfers and improve forecasting accuracy

## Summary

This study adopts a Design Science Research (DSR) approach to evaluate the developed artifact in three cycles of investigation: rigor, relevance, and design. The research questions aim to assess the performance of transferred models compared to models without transfer for newly introduced products. Additionally, the study aims to draw conclusions on the similarity between source and target domains after experimenting with different transfer options and identify characteristics of product groups that could indicate success in transferring models. The developed artifact is considered a knowledge contribution if it provides new solutions for known problems or extends known solutions to new problems. The focus of this research is on Transfer Learning for sales forecasting in the food retail domain. Specifically, it explores how knowledge can be transferred between listed stock products and new products to improve prediction accuracy using deep neural networks. By applying Transfer Learning techniques, human experts can leverage their domain knowledge and historical sales data of similar products to forecast sales for newly introduced items. The study delves into understanding the effectiveness of transferring models with different degrees of freedom (TL_models) compared to models without transfer (models w/o_TL) for newly introduced products. Through experimentation with different transfer options, the study aims to draw conclusions on the similarities between source and target domains and identify characteristics of product groups that may influence the success of transferring models. The findings suggest that ensemble strategies can be used to implement a robust Transfer Learning approach that consistently delivers better results than models without transfer. By analyzing real-world representational characteristics of data in a network-based transfer scenario, the study aims to provide insights that could enhance understanding of transfers and improve forecasting accuracy in food retail sales.