Product Similarity for New Product Demand Forecasting in Apparel

The use of machine learning for time series techniques in demand forecasting has been growing in recent years in Retail, CPG and manufacturing industries making demand forecasts more granular and accurate. A new wave of innovation in Demand Forecasting and Supply Chain optimisation is coming from a wider adoption of the latest advancements in AI ML technologies- i.e. Computer Vision and Natural Language Processing (NLP).

Data Reply Visual Product Similarity solution is using a combination of Computer Vision and Natural Language Processing (NLP) on AWS to identify products that are similar so that the historic demand and sales data on existing products can be used to forecast the demand for new similar products.

This is a scalable solution incorporating training and inference on images for new products over time uses a bespoke deep learning model fine-tuned by training on a dataset of images provided by the client, to create image embeddings of their products. In addition, NLP techniques are used to generate text embeddings (‘line descriptions’) using product attributes to augment the image embeddings of each product, to further capture their characteristics. In cases where no product image is available, text embeddings are used as substitutes. Hard text-based constraints with 'human in the loop' are used to ensure that outcomes meet rational business expectations.

AWS services and products used: Amazon S3 and Amazon SageMaker.

Solution approach The solution implementation starts with a time boxed 'Discovery' /'Proof of Value' engagement lasting 4-5 weeks (elapsed) followed by the productionisation phase of 5-6 weeks (elapsed)

At the start of the engagement the customer needs to provide a number of data sets:

• product images : new products and current product catalogue;
• product ‘line descriptions’ or product attributes for image labelling
• historic sales data