Content-Aware Few-Shot Meta-Learning for Cold-Start Recommendation on Portable Sensing Devices

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• The cold-start problem is a significant challenge in sequence recommendation systems, particularly for portable sensing devices.
• Current content-aware methods lack generalizability and struggle to prioritize content features effectively.
• New data processing often leads to performance degradation in existing models.

Purpose of the Study:

• To propose a novel content-aware few-shot meta-learning (CFSM) model to enhance cold-start sequence recommendation accuracy.
• To address the limitations of existing approaches in distinguishing feature importance and handling new data.
• To improve the robustness and generalizability of recommendation systems in data-scarce scenarios.

Main Methods:

• Developed a content-aware few-shot meta-learning (CFSM) model.
• Incorporated a double-tower network (DT-Net) for learning user and item representations.
• Utilized a meta-encoder and a mutual attention encoder to mitigate noisy auxiliary information.
• Employed a model-agnostic meta-optimization strategy for training across diverse tasks.

Main Results:

• CFSM demonstrated superior performance in cold-start recommendation scenarios across three real-world datasets (ShortVideos, MovieLens, Book-Crossing).
• The model achieved AUC improvements of 1.55%, 1.34%, and 2.42% over the second-best approach, MetaCs-DNN.
• The proposed DT-Net effectively mitigated the impact of noisy data on auxiliary information.

Conclusions:

• The CFSM model offers a significant advancement in addressing the cold-start problem for sequence recommendations.
• The meta-learning approach enhances model adaptability and performance with limited data.
• CFSM provides a robust and accurate solution for personalized recommendations in data-constrained environments.