Automated recommenders have become a mainstay of modern information retrieval and e-commerce systems. You’ve probably used one before. For example, when you visit Netflix or Amazon and see recommended items for you to watch/buy. By suggesting related items (e.g., news articles, music, movies), recommender systems filter away irrelevant data, enabling users to efficiently locate items of interest. However, current work has focussed on providing recommendations of existing items to consumers (either personalized to individuals or groups). In this work, we take on the problem of recommending new items that could be designed and created to satisfy users.
New item recommendation would be beneficial for item makers—artists, designers, producers, and manufacturers—who generate novel items for the marketplace. For example, mobile phone makers often produce a set of phone models because different users have different preferences; one group of users may prefer a phone with a large touch-screen, whilst another group may prefer a smaller, lighter, device. For item makers, important questions include: Is it possible to craft a set of new phone models that appeal to the widest possible number of customers? Which users would each phone design attract? And how large is each of these user groups?
Generation Meets Recommendation: Movies and Art!
In this recent paper, we take a first step towards a new class of recommender systems: item recommendation for item makers. We propose a principled approach for addressing the aforementioned questions: first, we formalize the problem and show it to be a variant of maximum coverage, which is NP-hard in general. We then present an approximate scheme for generating plausible K items that are predicted to be appealing to different groups of users.
Our key insight is to leverage upon low-dimensional latent spaces, which can be efficiently searched in a greedy manner (with an approximation guarantee). To learn this low-dimensional space, we use a deep generative model, specifically, a collaborative variant of the Variational Autoencoder that is trained both to reconstruct ratings as well as observed item features. This model enables projection of existing items into the latent space, as well as the generation of item features from “imaginary” item representations. An example of the top-3 generated movie “genomes” using our model trained with the Movielens dataset is shown below:
We see a diverse set of generated movies genomes: (GM-1) a sentimental/touching drama with a mentor and friendship; (GM-2) a narrated social commentary with elements of dark-humor and loneliness; (GM-3) a big-budget franchise action film with special effects and violence (possibly starring Arnold Schwarzenegger).
We also use the method to generate new “positive” artwork using the MART abstract artwork dataset:
Generation meets recommendation: proposing novel items for groups of users Thanh Vinh Vo and Harold Soh, Proceedings of the 12th ACM Conference on Recommender Systems, 2018. (Best Paper Award Runner-up) [ PDF ]
Deep Language-based Critiquing for Recommender Systems, Ga Wu, Kai Luo, Scott Sanner and Harold Soh. ACM Recommender Systems Conference (RecSys), 2019 [ PDF ]