Study reveals methods to enhance daily item embedding compatibility for CTR prediction

A new study on arXiv explores methods to improve the compatibility of daily item embeddings, which are crucial for predicting user behavior like click-through rates (CTR). The research compares Word2Vec with an adaptation of XSimGCL, evaluating various techniques such as initialization, alignment, and scale management. Key findings indicate that interventions like ID-init and orthogonal alignment significantly boost embedding compatibility and CTR, while daily global rescaling effectively manages accumulated scale over time. The study also demonstrates that dimensionality-reducing encoders can further adapt embeddings to residual drift, suggesting a combination of upstream stabilization, scale management, and downstream task learning is beneficial for industrial applications. AI

IMPACT Enhances prediction accuracy for user behavior in recommender systems and online platforms.

RANK_REASON The cluster contains an academic paper detailing empirical study results on information retrieval techniques. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.IR (Information Retrieval) →

• alphaXiv
• arXivLabs
• CatalyzeX
• DagsHub
• Geometry-Aware Uncertainty Coresets
• Gotit.pub
• Hugging Face
• ID-init
• orthogonal alignment
• ScienceCast
• singular value decomposition
• Word2Vec
• XSimGCL

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →