The accelerating evolution of online advertising necessitates improved prediction accuracy and optimized placement strategies as critical research priorities. Conventional predictive approaches struggle to process complex, heterogeneous advertising data—particularly in capturing temporal shifts and periodic fluctuations. Addressing these limitations, this paper proposes an integrated methodology employing iTransformer with a Periodicity Decoupling Framework (PDF) for enhanced advertising performance forecasting. iTransformer preserves Transformer’s architecture while innovatively redefining attention mechanisms and feedforward networks to model distinct variables as independent tokens. This paradigm enables superior capture of cross-variable dependencies and multi-scale temporal relationships, significantly enhancing adaptability to intricate datasets. Concurrently, PDF examines periodicity patterns through spectral analysis, precisely isolating regular fluctuations to fortify long-sequence forecasting robustness. Further leveraging self-supervised learning minimizes labeled data dependency, maintaining high generalizability under data scarcity. Empirical validation demonstrates substantial performance gains over state-of-the-art methods, particularly in managing periodic complexities within real-world advertising datasets.
Zhang et al. (Wed,) studied this question.