<p>Group recommendation system (GRS) is designed to deliver diverse and accurate recommendations for a groups of users rather than individual recommendation system. One of the major challenge in this domain is group formation to identify group of similar users based on their preferences, since user groups are not predefined. Traditional clustering techniques, such as K-Means and Spectral Clustering, often struggle with the dynamic nature of user preferences and fail to capture complex relationships in high-dimensional data. In this paper, we propose a novel group formation technique, adaptive similarity-driven deep embedded clustering (AS-DEC), designed specifically for GRS. AS-DEC adapts to user preference patterns more effectively by embedding users in a lower-dimensional space, capturing both linear and nonlinear relationships within the data. The AS-DEC framework is evaluated against traditional clustering techniques, including K-Means Clustering, Spectral Clustering, Fuzzy C-Means, Bisecting K-Means and the standard deep embedded clustering (DEC). To assess clustering performance, we have utilized three evaluation metrics Silhouette Score, Davies–Bouldin Index and Calinski–Harabasz Index. Additionally, to evaluate the overall performance of GRS model we have employed the standard evaluation metrics including accuracy, precision, recall and <i>F</i>1-score on different group sizes. The empirical results demonstrate that the proposed AS-DEC framework achieves superior results with a 92.8% accuracy and 34.28% <i>F</i>1-score for group sizes of 20 and 5. Furthermore, AS-DEC achieved the highest Silhouette Score of 0.614 and outperformed other traditional techniques across all evaluated metrics. This work highlights the potential of deep learning-driven clustering to enhance group recommendation systems by effectively capturing complex user preferences and behaviors.</p>