scenarios/retail/README.md
Recommender systems have become a key growth and revenue driver for modern retail. For example, recommendation was estimated to account for 35% of customer purchases on Amazon, Alibaba reported a revenue increase of 19% YoY due to their recommendation solutions in their domestic retail platform Tmall. Other examples in ecommerce are BestBuy with 24% revenue increase, eBay with 15.9% click-through rate increase, Sephora with 15% increase in average order value.
Next we will describe several most common retail scenarios and main considerations when applying recommendations in retail.
A major task in applying recommendations in retail is to predict which products or set of products a user is most likely to engage with or purchase, based on the shopping or viewing history of that user. This scenario is commonly shown on the personalized home page, feed or newsletter. Most models in this repo such as ALS, BPR, LightGBM and NCF can be used for personalization. Vowpal Wabbit, which uses reinforcement learning in real-time, is a good solution to rerank the outputs from the personalization model.
In this scenario, the user is already viewing a product page, and the task is to make recommendations that are relevant to it. Personalized recommendation techniques are still applicable here, but relevance to the product being viewed is of special importance. As such, item similarity can be useful here, especially for cold items and cold users that do not have much interaction data.
In this task, the retailer tries to predict product(s) complementary to or bought together with a product that a user already put in to shopping cart. This feature is great for cross-selling and is normally displayed just before checkout. In many cases, a machine learning solution is not required for this task.
This scenario covers down-selling or out of stock alternatives to avoid losing a sale. Similar alternatives predict other products with similar features, like price, type, brand or visual appearance.
Datasets used in retail recommendations usually include user information, item information and interaction data, among others.
To measure the performance of the recommender, it is common to use ranking metrics. In production, the business metrics used are CTR, AOV and revenue per order. To evaluate a model's performance in production in an online manner, A/B testing is often applied.
Retailers use recommendation to achieve a broad range of business objectives, such as attracting new customers through promotions, or clearing products that are at the end of their season. These objectives are often achieved by re-ranking the outputs from recommenders in scenarios above.