Predicting Actions of Users Using Heterogeneous Online Signals

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Advertising platforms require enhanced prediction quality for optimal performance.
• Existing deep sequence-learning models face challenges with data noise and temporal information in user activity trails.

Purpose of the Study:

• To propose novel extensions for modeling online user activity trails.
• To address the challenges of temporal information and noise in sequence data for predictive tasks.

Main Methods:

• Developed dedicated layers for deep sequence-learning approaches.
• Integrated methods to model the temporal aspects of user activities.
• Incorporated techniques to mitigate noise from diverse data sources.

Main Results:

• Achieved up to 3% improvement in area under the receiver operating characteristic curve (AUC) compared to production models.
• Demonstrated a 1.75% AUC improvement over the best baseline approaches.
• Validated the method across two major advertiser datasets and multiple predictive tasks.

Conclusions:

• The proposed extensions effectively enhance deep sequence-learning models for advertising prediction.
• The approach successfully incorporates temporal information and noise reduction, leading to significant performance gains.