This article is part of the Academic Alibaba series and is taken from the WWW 2019 paper entitled “Personalized Bundle List Recommendation” by Jinze Bai, Chang Zhou, Junshuai Song, Xiaoru Qu, Weiting An, Zhao Li, Jun Gao. The full paper can be read here.
For online retailers, product bundling is a crucial marketing and sales tool. A product bundle groups together a variety of items that are frequently bought by the same groups of people. During big shopping events, retailers often promote specific bundles with significant discounts to drive sales. Even during regular shopping periods, bundles are used by retailers on the back-end. When you add a product to your cart or when you watch a show on Netflix, the recommendations you see for similar products or shows come from bundles.
For platforms such as Taobao and Amazon, creating high-quality bundles requires selecting products that are highly correlated and personalized for customers while also diverse. For example, when you buy a TV, high-quality bundling would ensure that it does not recommend you a second TV, and that it includes a variety of related items such as cables, TV stands, and sound systems. In the past, however, creating these bundles involved a difficult tradeoff: While manually applying human insight to the problem was time consuming, using automatic mining methods tended to sacrifice personalization.
Now Alibaba’s tech team, in collaboration with researchers from Peking University, is showing how it can automate personalized bundle generation by decomposing the problem into parts that focus on quality and diversity, respectively. Based on a bundle generation network (BGN) that uses a sequence generation approach and a determinantal point process (DPP), the proposed model greatly improves on previous techniques to provide better quality, diversity, and response times for bundling tasks.
Building on the Past
While work in the field of neural networks has not been applied directly to product bundling before, previous works have studied problems with similar challenges. With quality and diversity as the key requirements for product bundles, determinantal point processes (originally introduced to model fermion behavior) are perfect for modeling and analyzing their contents, as they are designed to effectively balance these two key criteria.
For the other half of the solution, researchers looked at sequence generation methods to generate the bundles. The standard encoder-decoder architecture for sequence generation provided the best foundation for the BGN. However, unlike the standard sequential recommendation model, the researchers’ bundle recommendation model uses both the encoder and decoder components. The encoder first models the user context, and the decoder then models the bundle distribution. In addition to having separate trainable parameters, this approach allowed the researchers to control diversity and know when to end bundle generation.
Finally, as part of this approach, the model applies a feature-aware softmax to alleviate the shortcomings of a traditional softmax. An integrated masked beam search maximizes sequences, and the DPPs are integrated for selection to produce high-quality and diversified bundle lists.
How Does It Stack Up?
The dataset was tested with three public datasets and one industrial dataset, with two of the datasets generated from co-purchase records and the other two extracted from real-world online bundle services. The BGN significantly outperformed the existing methods in terms of quality, diversity and response time for all datasets. The BGN beat the next best competitor’s precision by 16% on average while maintaining the highest diversity for all four datasets. It also provided a 3.85x improvement to response time in the bundle list recommendation problem compared to the next-best solution. As a neural network solution, it promises the best of both worlds for product bundling: It can effectively personalize bundles while also providing the speed of an automated solution.
The full paper can be read here.