Enhancing Predictive Accuracy in Online Advertising Services: A Data-Driven Approach Using iTransformer and Periodicity Decoupling

As the online advertising industry rapidly evolves, enhancing the accuracy of service predictions and optimizing ad placement strategies has emerged as a crucial area of research. Traditional prediction methods often encounter difficulties when processing complex and diverse advertising data, especially when addressing temporal dynamics and periodic fluctuations. To tackle these challenges, this paper introduces a data analysis approach based on iTransformer combined with a Periodic Decoupling Framework (PDF) for improving online advertising predictions. Rather than modifying the Transformer architecture, iTransformer redefines the attention mechanism and feedforward network to treat distinct variables as individual tokens. This design allows the model to effectively capture inter-variable relationships and temporal dependencies, boosting its adaptability to intricate datasets. Additionally, the Periodic Decoupling Framework delves into periodic patterns within sales data, accurately isolating regular variations and providing more robust support for long-term sequence forecasting. Moreover, the adoption of self-supervised learning minimizes dependence on labeled data, enabling the model to retain high generalization and performance even in low-data scenarios. Experimental results confirm that this method significantly outperforms in advertising predictions, particularly in managing complex datasets with periodic variations.