An adaptive hybrid XdeepFM based deep Interest network model for click-through rate prediction system.

Area of Science:

• E-commerce technology
• Machine learning for advertising
• Data mining and analysis

Background:

• E-commerce platforms generate vast user data, necessitating accurate Click-Through Rate (CTR) prediction for effective advertising.
• Existing CTR prediction models, from Logistic Regression to deep neural networks, have limitations in capturing complex user interests and feature interactions.
• Deep Interest Network (DIN) and eXtreme Deep Factorization Machine (xDeepFM) are advanced models that address specific aspects of CTR prediction.

Purpose of the Study:

• To propose and evaluate a novel hybrid model for robust CTR prediction by integrating the strengths of DIN and xDeepFM.
• To leverage the attention mechanism of DIN for adaptive user interest learning and the Compressed Interactions Network (CIN) of xDeepFM for implicit feature interactions.
• To demonstrate the superior performance of the proposed hybrid model over existing methods in an e-commerce context.

Main Methods:

• Developed an end-to-end hybrid model as a parallel ensemble, combining DIN and xDeepFM components via a multilayer perceptron.
• Utilized the attention mechanism in DIN to adaptively learn user interests from historical behavior data.
• Incorporated the Compressed Interactions Network (CIN) within xDeepFM to generate vector-wise feature interactions implicitly.
• Employed the Alibaba e-commerce dataset and focal loss with online hard example mining (OHEM) for training and evaluation.

Main Results:

• The proposed hybrid model demonstrated superior performance compared to other benchmark models in CTR prediction tasks.
• The integration of DIN's attention mechanism and xDeepFM's feature interaction capabilities proved effective.
• The parallel ensemble architecture facilitated robust and accurate prediction by combining complementary model strengths.

Conclusions:

• The hybrid DIN and xDeepFM model offers a significant advancement in CTR prediction accuracy for e-commerce advertising systems.
• The model's ability to capture both user interest and complex feature interactions is key to its enhanced performance.
• This research provides a robust framework for future developments in personalized advertising and e-commerce optimization.