Given the vast array of food options available, food recommender systems play a crucial role in assisting users in discovering desired foods by analyzing their past preferences, represented as interactions between users and food items. To mitigate challenges stemming from data sparsity and cold start problems, these systems incorporate supplementary data resources in conjunction with rating data to enhance the accuracy of their recommendations. Furthermore, deep learning models offer a means to extract latent features from input data resources, thereby enhancing the accuracy of the recommendation process. Nevertheless, the performance of food recommender systems is intricately linked to the judicious selection and fusion of data resources. Previous research has overlooked this critical aspect in the design of food recommender systems. To address this challenge, this paper undertakes a comparative analysis of various data resources applicable to food recommender systems. We delve into the impact of leveraging deep neural networks to obtain deep representations of these data resources, as well as their fusion techniques, on the efficacy of food recommendations. Moreover, we comprehensively investigate the efficacy of various deep data representation fusion models in mitigating the challenges posed by data sparsity and cold start problems. To this end, we design extensive experiments utilizing two widely recognized food datasets. The findings indicate that the integration of additional data resources does not consistently improve recommendation accuracy. Also, varying combinations of these resources exert differential impacts on the performance of food recommender systems.
