<p>Modern recommender systems must effectively capture and model multiple aspects of users to provide accurate recommendations. Social network services offer valuable insights into user behavior and preferences, traditionally studied through explicit relationships or even the analysis of communities. On the other hand, deep learning-based models have achieved significant success in providing more scalable and intelligent recommendations, but they are typically applied to user–item interactions and often focused solely on features or structural information. As a result, the generated user embeddings do not simultaneously capture the diverse aspects such as users’ multiple interests, features, and the characteristics of the items they find appealing. Graph neural network (GNN) models provide a solution for hybrid recommendation systems without being specifically focused on the underlying community-based insights. In this paper, we propose a stacked and hybrid recommendation framework. This framework contrasts several aspects, including community-based hybrid recommendations versus community-based collaborative filtering, explicit user–item interactions versus implicit user–user ties, item embeddings versus user embeddings, and communities of interest in movie recommendations versus book recommendations. The framework is first based on the recognition and modeling of implicit interest ties derived from similar/shared interests. Unlike conventional user–item interactions, the proposed approach focuses on underlying interest networks, which encode deeper connections and allow us to study communities of interest. Several prediction scenarios, grouped into two main phases, are proposed and compared. In the first phase, community-based recommendation scenarios, using different ranking techniques, are studied: community-driven, item-based, and user-based predictions, addressing the challenges of scalability and the integration of clustered embedded items to predict the best scores. The most accurate scenario in this communitarian approach will be identified. In the second phase, insights from interest communities will be used to propose an IntereSage-powered prediction. Unlike conventional GNN approaches that focus on link prediction in user–item bipartite graphs, IntereSage will be trained on the underlying interest network using a community-driven training strategy. It merges both the underlying structural information and user attributes and propagates users’ diverse interests (memberships) through implicit interest ties on a large scale. Finally, the user embedding outcomes are used for an embedding-based prediction. The findings show that IntereSage-powered recommendations not only preserve core preferences highly localized within a community of interest but also capture cross-community relationships, leading to more diverse preferences that are more accurate and better aligned with larger top-k recommended item lists. Findings also highlight disparities due to the density of the interest network and the extent to which interests overlap, which vary between recommendation domains (movies or books).</p>