Accurate bundle matching and generation via multitask learning with partially shared parameters.

Area of Science:

• E-commerce
• Recommender Systems
• Machine Learning

Background:

• Bundle recommendation, encompassing bundle matching and generation, is crucial for enhancing user and provider satisfaction in e-commerce.
• Existing models often struggle with heterogeneous data and personalized bundle generation, limiting their accuracy.

Purpose of the Study:

• To propose BundleMage, an accurate approach for both bundle matching and bundle generation.
• To address the limitations of current models in handling diverse data and creating customized bundles.

Main Methods:

• BundleMage employs an adaptive gate technique to integrate user preferences for items and bundles, enhancing bundle matching accuracy.
• A personalized bundle generation module leverages user preferences and incomplete bundle characteristics.
• Multi-task learning with partially shared parameters is utilized to further optimize performance.

Main Results:

• BundleMage demonstrated superior performance, achieving up to 6.6% higher normalized Discounted Cumulative Gain (nDCG) in bundle matching.
• The system achieved a 6.3× higher nDCG in bundle generation compared to state-of-the-art competitors.
• Qualitative analysis confirmed BundleMage's ability to generate bundles aligned with user tastes and bundle characteristics.

Conclusions:

• BundleMage offers a significant advancement in bundle recommendation accuracy for both matching and generation tasks.
• The proposed methods effectively address challenges posed by heterogeneous data and the need for personalized bundle creation.
• BundleMage provides a robust framework for improving e-commerce recommendation systems through accurate and tailored bundle suggestions.