Due to the dynamic nature of Wi-Fi channel state information (CSI), the traditional transfer learning (TL) approaches in CSI-based human activity recognition (HAR) are largely impacted by the issue of data mismatch. Data mismatch occurs when the training and testing data represent different distributions, causing pre-trained models to struggle to adapt to target CSI datasets. This issue is particularly evident when a single CSI dataset, often limited in size and diversity, is used to obtain a pre-trained model, which is subsequently utilized for TL on a CSI dataset representing another domain. Moreover, even when extended to multiple datasets, traditional TL methods are susceptible to the forgetting problem, where learning from the initial pre-trained dataset gradually diminishes. To address these challenges, a novel framework called Inter CSI-data Selective Sequential Transfer Learning (InSSeqTra) is proposed. This framework pre-trains a base model by leveraging multiple existing CSI datasets in a selective sequential pre-training scheme guided by divergence scores obtained from GAN-learned distributions of the pre-training datasets relative to the target dataset. InSSeqTra was evaluated by using four existing CSI-based HAR datasets, where three datasets (SignFi, WiAR, and UT-HAR) were used for the selective sequential pre-training, and the Wi-Gitation dataset served as a target dataset. The proposed framework shows promising results over traditional TL approaches, particularly achieving 11.3% improvement in person-wise and 6.5% in receiver-wise domain adaptation.