Making Pour Decisions with AI
The multimodal WineSensed dataset includes fine-grained flavor annotations from human tasters to improve wine recommendation algorithms.
Recommendation algorithms have become a fundamental component of our online experiences, providing tailored recommendations for a broad array of products and services. These algorithms employ data analytics and machine learning techniques to study user preferences and behaviors, with the goal of predicting and suggesting items that individuals are likely to appreciate. This technology is prevalent on platforms like streaming services, e-commerce sites, and social media.
These algorithms have a significant advantage in that they can help users discover new and relevant content, such as movies or electronics, that is tailored to their tastes and preferences. By examining patterns in a user's past interactions, these algorithms can identify similarities with other users who share similar interests. As a result, users receive tailored recommendations, which enhances their overall experience and may expose them to products or content they would have otherwise overlooked.
However, it is important to recognize that recommendation algorithms are not without their limitations. In the case of food and drink, the subjective nature of taste presents a major obstacle. Unlike movies or electronics, where user preferences can be more readily quantified, individual tastes in food and beverages are highly nuanced and difficult to capture accurately. The sensory experience of consuming food and drink is influenced by personal preferences that are often shaped by cultural, regional, and even emotional factors. As a result, recommendation algorithms in this area may be less effective, as they struggle to account for the intricacies of individual taste preferences.
Taking advantage of recent advances in machine learning and the growing interest in multimodal models among researchers in the field, a group led by a team at the Technical University of Denmark has proposed a new path forward for food and drink recommendation algorithms. Initially, they focused their attention on wine recommendations, however, similar techniques could in principle be used for other types of food and beverages. The team’s primary contribution is the development of what they call WineSensed, a large multimodal wine dataset.
Existing wine recommendation services tend to focus on textual reviews written by people and images of the labels on the bottles. The WineSensed dataset includes this type of information, but also includes a crucial component that has been missing — characterization of the flavor of each wine. Paired with 897,000 label images, 824,000 reviews, and other metadata about the wine, are fine-grained flavor annotations collected from an experiment involving 256 tasters.
The tasters were given small cups of wine, and after taking a drink they were asked to place them closest to the other cups that they tasted the most similar to. This resulted in the creation of a sort of graph that expressed similarity relationships between the wines. The researchers took pictures of these cup arrangements and digitized them such that the relationships could be represented in more convenient ways for use in a recommendation algorithm.
A machine learning algorithm called Flavor Embeddings from Annotated Similarity & Text-Image (FEAST) was developed and trained using the WineSensed dataset. It was noted that by including the additional flavor similarity data, the model was able to make more accurate predictions of people’s wine preferences. Looking ahead, the team hopes to explore new ways that human sensory experiences can be incorporated into machine learning algorithms to produce better results for users. They hope others will build on their dataset in the future, and suggest beer and coffee as the next targets for new recommendation algorithms.