Document Type : Research Paper
Authors
1 Department of Computer Engineering, Abadan Branch, Islamic Azad University, Abadan, Iran
2 Department of Computer Engineering, Na.C., Islamic Azad University, Najafabad, Iran,
3 Department of Computer Engineering, Arv. C., Islamic Azad University, Abadan, Iran
Abstract
Early-stage startups face the problem of cold start, as they have limited real-world data to train AI models. This lack of data, combined with the incompatibility of generic data with specific business needs, reduces the accuracy of predictions and recommendations. Rapid changes in data and concepts (such as data and concept drift), the risk of forgetting prior knowledge in transfer learning, and the heterogeneous quality of user feedback are the main challenges in this area. The proposed framework is an integrated and scalable architecture that combines transfer learning and crowd intelligence. The framework consists of four parts: collection and preprocessing of (limited), generic, and user feedback real-world data; transfer learning with a pretrained model and efficient optimization to prevent forgetting prior knowledge; model enhancement with filtered and weighted user feedback; and continuous prediction by monitoring data and concept changes with mathematical criteria. The training data is composed of a combination of real, generic, and user feedback data, and optimization is performed by minimizing error and controlling complexity. Evaluation on three real datasets. Other metrics such as prediction accuracy, positive sample detection, balance between the two, error reduction, and data stability were also improved in all three datasets, especially in investment data that is more scattered. This framework increases the efficiency of limited data and ensures the stability of the model.
Keywords
Main Subjects
)