Multitask Model with an Attention Mechanism for Sequentially Dependent Online User Behaviors to Enhance Audience Targeting

This paper proposes a multitask learning approach with an attention mechanism to predict audience behavior as sequential actions. The goal is to improve click-through and conversion rates by effectively targeting audience behavior. The proposed model introduces specific task sets designed to address the challenges specific to each prediction task. In particular, the first task, click prediction, suffers from data sparsity and a lack of prior knowledge, limiting its predictive power. To address this, a one-dimensional convolutional network (1D CNN) tower is used in the first task to learn local dependencies and temporal patterns of user activity. This design choice allows the model to better detect potential clicks, even without rich historical data. The task of conversion prediction is tackled by a fully connected convolution tower that selectively combines the corresponding features extracted from the first task using an Attention Mechanism, as well as the original shared embedding input data, enabling richer context for performing more accurate prediction. Experimental results show that the proposed multitask architecture significantly outperforms existing state-of-the-art models that do not consider tower architecture design to predict sequential online audience behavior.

@article{El-Sherief2025,
  title     = {Multitask Model with an Attention Mechanism for Sequentially Dependent Online User Behaviors to Enhance Audience Targeting},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {16},
  number    = {4},
  year      = {2025},
  publisher = {The Science and Information Organization},
  author    = {Marwa Hamdi El-Sherief and Mohamed Helmy Khafagy and Asmaa Hashem Sweidan},
  doi       = {10.14569/IJACSA.2025.01604112},
  url       = {https://doi.org/10.14569/IJACSA.2025.01604112}
}