Graph Method for Improving Food Recommendation

Tourampfood has been working on Food Intelligence (FI) in order to better understand our inventory and make it more relevant to our users. We give a series of two articles on using FI to propose meals to consumers, as a follow-up to an earlier blog. The problem is addressed in this first installment, and random walk-based graph embedding solutions are demonstrated.

Food Recommendation That Is Customized

The high intent and immediacy of the customer demand — her meal — makes food discovery, ordering, and delivery a complicated environment. Today’s business models are defined by how machine learning improves the consumer experience. For a user with individual taste preferences, associating food with its specific qualities aids personalization.

When compared to more common situations such as movie or book recommendations, item suggestions in the food realm are intrinsically different. Customers who have seen and reviewed a film or book are unlikely to want to see the same item recommended again. When it comes to food, though, many people prefer to order and enjoy the same products they did previously. At the same time, they’re open to proposals on related topics. A user profile can be established based on previous ordering history to learn what things the user orders

As an obvious method, the most favored items can be offered (our baseline) (our baseline). When things similar to these can be suggested, this will be an enhancement. Following this line of reasoning, how many orders must be placed before a user preference emerges? What is a user’s degree of confidence when they are new to the platform and just have a few orders to learn from?

The following is one method for analyzing customer behavior that has been offered. Consider a three-month timeframe for ordering. We can calculate the number of times a client ordered an item based on their orders (say, customer A ordered paneer biryani 25 times). A percentile score can be provided to each item based on this support as a proxy for the customer’s choice for that item. A higher score indicates that the buyer has a strong preference for this item. Higher confidence in utilizing this score (the greater percentile rated item) as a proxy for preference if there is a substantial split (higher deviation).