CHRONOSYNC SPARSE BUNDLE DRIVER

But identifying the correct driver item leading to the purchase of relevant go-together items is challenging. Image-based approaches have also been used for bundles recommendation, but extending them to frequently bought together items is not feasible due to sparsely available labeled data.Īnother popularly used method for bundles recommendation is by using behavioral signals such as customer co-browse or co-purchases . We need to build a robust system that works even in the cold start situation. They cannot produce satisfactory results for i) users with no historical data and ii) products with less or no purchase data. But they suffer from the cold start problem. Many popular state-of-the-art models for bundles recommendations include personalized product bundles. Bottom - Product bundle of baby food comprising similar products. Figure 2: Top - Product bundle of camping gear comprising complementary products. Thus, using just one relation for bundle formation limits the potential of bundling.

For example, Figure 2 shows a bundle for a camping tent with other camping gear such as an air mattress, lantern, and burner (top), while a bundle for baby snacks can have similar snacks with different flavors (bottom). However, in practice, product bundles can have both similar and complementary products. Product recommendation is a widely studied research problem, and many works focus on similar or complementary product recommendations forming bundles using either of these relations. Recommending bundles is different from recommending purely similar or complementary products. Figure 1: Sample example of a product bundle presented to a customer intending to purchase a car seat. In the rest of the paper, we will call the item clicked by the user as the driver item and all the recommended items in the bundle as the go-together items. Depending on the retailer's understanding of user intent, such recommendations can surface on the product display pages when users add items to their cart or even when they initially enter the site. When a user clicks on the car seat, a bundle with a back seat mirror, a support cushion, and a car seat cover is displayed. For example, in Figure 1, we show an example of the product bundle formed for the car seat. These recommendations are often labeled as ‘Frequently Bought Together’ items in the e-stores. Digitally, creating this experience requires a careful understanding of the product assortment and user behaviors. Appropriately created bundles can lead to serendipitous, and more satisfying purchases for the customer, often called the ‘Target effect’ in the retail jargon. The product bundles are highlighted in mannequins or other displays across the stores. Retailers often bundle products to offer an improved shopping experience to the customers and increase their sales by expanding the order size. The model was also validated over the Amazon product metadata dataset. BundlesSEAL based link prediction leads to amelioration of the above-mentioned cold start problem by increasing the coverage of product bundles recommendations in various categories by 50% while achieving a 35% increase in revenue over behavioral baseline. Further, we also apply BundlesSEAL for predicting the edge weights instead of just link existence. We also propose a heuristic to identify relevant pairs of products for efficient inference. First, we build a graph using add-to-cart data and then use BundlesSEAL to predict the link representing bundles relation between products represented as nodes. A graph neural network (GNN) based product bundles recommendation system, BundlesSEAL is presented. This work alleviates this problem by modeling product bundles recommendation as a supervised graph link prediction problem.

These scenarios are frequent for retailers like Target, where much of the stock is seasonal, and new brands are launched throughout the year. Hence, no bundles recommendations will be provided to users for such product categories. When a new product is launched or for products with limited purchase history, the lack of user-product interactions will render these algorithms inaccessible. Moreover, these approaches require robust user-product interactions and cannot be applied to cold start scenarios. In these methods, user-product interactions provide a more personalized experience.

Existing work on modeling bundle recommendations consists of graph-based approaches. In e-commerce, a group of similar or complementary products is recommended as a bundle based on the product category.

Chronosync sparse bundle