<p>Session-based recommendation aims to provide the next recommendation for users based on anonymous sessions. Traditional methods typically assume that the order of interactions is crucial for generating recommendations. However, in many real-world scenarios, the order is often unreliable and may not accurately reflect user preferences. To address this issue, we reframe session-based recommendation as a set expansion task, which reduces reliance on interaction order and provides a more flexible approach to capturing user behavior. We propose self-supervised multi-stages set expansion for session-based recommendation (SMSE-SR), a novel framework that combines global preferences with dynamic shifts in user interests to make highly precise recommendations. This method utilizes a permutation-invariant network with attention mechanisms to capture global preference features from session data. Additionally, our method employs a multi-stage iterative session feature extraction process in a self-supervised learning manner, dividing sessions into sub-sessions to better reflect dynamic shifts in user interests. Extensive experiments on two real-world datasets demonstrate that SMSE-SR improves recommendation accuracy across various evaluation metrics. Furthermore, this method exhibits higher adaptability and lower volatility when handling sessions of varying lengths.</p>