Modeling Algorithmic Influence On User Behavior In Digital Environments

Abstract

Algorithms have a considerable impact on users' decision making process on online e-commerce websites as recommendation systems affect their buying decisions beyond personal preferences. Traditional methods of Deep Learning (DL) and Machine Learning (ML) are mainly concerned with modeling click-through rate prediction and purchase prediction. However, these methods cannot take into account the behavioral changes due to continuous recommendation exposure. To overcome this problem, an algorithm named the Influence-Aware Proximal Policy Self-Attention Memory Network (IA-PP-SM Net) has been developed for understanding the dynamic relationship between users and recommendation algorithm in online digital commerce context. The IA-PP-SM Net comprises of Proximal Policy Optimization (PPO) with attention-enhanced Long Short-Term Memory (LSTM). The first one is used for learning the sequence of user behavior and the latter one learns recommendation policies. In addition, the algorithm uses an influence-aware reward system to estimate the behavioral change after exposure to recommendations. Moreover, self-attention is introduced to recognize influential past interactions that can affect future user behavior. Min-max normalization was used for data preprocessing for the Kaggle dataset, and Principal Component Analysis (PCA) was utilized to perform characteristic retrieval combined with simplification of data complexity. The proposed approach was modeled through the use of Python programming language utilizing the help of TensorFlow and Keras packages. Experimental outcomes revealed that IA-PP-SM Net model achieves higher performance than other models as it is able to attain a higher accuracy rate of 0.962. This model is able to efficiently identify the effect of dynamic algorithm on recommendation systems for online commerce.