Group-Based Recommendation System Using Bi-Stage Adaptive Deep Learning Model

Recommender systems (RS) are utilized in various domains, including travel, movies, and music. The increase in social activity has led to an increase in the usage of RS in individual and group recommender systems (GRS). A GRS recommends perfect items to users according to their preferences. A bi-stage adaptive deep learning-based group recommendation system model is proposed to overcome these challenges. The aim of the proposed Bi-stage Adaptive Deep Learning-based GRS (BADLGRS) is to enhance the effectiveness of GRSs. At the GRL level, an undirected Tripartite Graph (TG) represents the interaction among groups, users, and items. Then, constructing a TG effectively represents the semantic features of both users and items within the group context. Then, a novel Deep Learning (DL) network, the Gated Recurrent Unit-based Attention Neural Network, is used to learn the semantic features of the group. Generate optimized semantic features to produce refined and optimized semantic feature representations for both users and items, which are fed into the next stage. A two-layer graph convolutional network (TGCN) is employed for user preference learning at the GPL level, enabling the accurate learning and capture of individual user preferences. After learning the group’s preferences, we employ the Pairwise Learning Method (PLM) to effectively learn and model the aggregated preferences of the group. Additionally, the Network model optimizes the parameters of the two-layered network within the GPL stage using the PLM. Additionally, the proposed model is validated using four different datasets and outperforms existing models in terms of HR, NDCG, MAP, accuracy, recall, and f1-score for group recommendation. The proposed model acquired enhanced outcomes in terms of various assessment metrics like accuracy of 0.893, which is 28.33%, 28.89%, 29.79%, 26.54%, 25.20%, 23.96%, 22.73%, and 17.02% superior to DFM-AVG, DFM-LM, DFM-MS, COM, DPMF-CNN, AGR, AGREE, and MAGRM methods.