A Weighted Hybrid Filtering Approach for Cold-Start and Diversity Evaluation in Culinary Recommender Systems

Abstract

Recommender systems play a crucial role in helping users discover relevant and diverse products tailored to their preferences. However, challenges such as the cold-start problem and low recommendation diversity still persist, particularly in single-approach models such as Content-Based Filtering (CBF) and Collaborative Filtering (CF). This study implements and evaluates a Weighted Hybrid Filtering approach that linearly combines prediction scores from CBF, which recommends items based on content similarity to a user's preference profile, and CF, which recommends items based on rating patterns from similar users, using a weighting parameter α. The model was evaluated on a real-world transaction dataset from Tantri Point of Sale (POS) consisting of more than 300,000 user–item interactions, with a user–item interaction sparsity of 99.93%, using Precision@10, Recall@10, F1@10, Intra-List Diversity (ILD@10), RMSE, and MAE. Cold-start conditions were assessed through simulation by masking user interaction histories. Results show that the hybrid configuration with α = 0.9 achieves the highest Recall@10 (0.8153) among all evaluated models, along with an F1@10 of 0.6634, an ILD@10 of 0.8030, an RMSE of 0.2984, and an MAE of 0.1924, although the standalone CF model attains higher F1@10 and ILD@10, and the standalone CBF model attains higher precision, individually. Compared to existing weighted hybrid recommender approaches, which have primarily been evaluated on non-culinary benchmark datasets, this study applies and evaluates the method on a large-scale, real-world culinary transaction dataset, providing an explicit assessment of the accuracy–diversity trade-off across multiple weighting configurations. These findings provide empirical insight into the trade-offs of weighted hybrid filtering for cold-start and diversity assessment in culinary recommender systems.