Visual-Aware Deep Learning-Based Model for Predicting the Popularity of Online Finnish Recipes

Abstract

The internet has emerged as a prominent platform for culinary inspiration, dining experiences, and social gatherings that revolves around food. Consequently, significant number of individuals now depend on online recipes to a greater extent than conventional cookbooks. However, worries are increasing over the healthiness of these internet recipes. This thesis explores the determinants influencing the popularity of online recipes by analyzing a dataset of over 5,000 dishes from Valio, one of Finland’s largest firms. Valio's website showcases a diverse array of culinary tastes and preferences among Finnish users. Through the analysis of visual characteristics obtained from food images (such as sharpness, contrast, RGB contrast, entropy, saturation, naturalness, and brightness), as well as other characteristics obtained by deep learning techniques and recipe attributes such as nutritional content (energy, fat, salt, etc.), cooking complexity (preparation time, number of steps, required ingredients, etc.), and user engagement (number of comments, ratings, comment sentiment, etc.), our objective is to determine the prominent factors that impact the popularity of online recipes. Our best predictor, AdaBoost, exhibits considerable accuracy (with an accuracy of 95%), outperforming other models, demonstrating that unique visual aspects of food images play a major role in their appeal. Our results show that visual characteristics and deep learning-based feature extraction and prediction can greatly boost final prediction accuracy. By providing useful insights into contemporary cooking tastes in Finland and guiding possible future dietary policy changes, this study enhances our understanding of the variables leading to the popularity of online recipes.