Cross-day item embeddings improve click predictions and stability
Cross-Day Compatibility of Daily Item Embeddings: An Empirical Study of Alignment, Scale Management, and CTR Adaptation
Information Retrieval
Summary
Predicting what items people will click on over several days is tricky because the way items are represented can change day to day. The study compares two methods to create these item representations and shows that certain techniques can make them more consistent across days and improve click predictions. They found ways to keep item vectors aligned and manage scale changes to keep predictions reliable. Additionally, adapting the prediction model itself helps handle leftover differences between days.
What this means in practice
- •For ecommerce data scientists: Improve daily item recommendation models by stabilizing item embeddings and managing scale across multiple days to enhance click predictions.
- •For online advertising engineers: Increase accuracy of click-through rate prediction models by using alignment and rescaling techniques on item co-occurrence embeddings updated daily.
Authors
Yicheng Wang
Abstract
Daily item co-occurrence embeddings are paired with next-day behavior without future-behavior leakage. Multi-day click-through rate (CTR) training therefore mixes versions whose directions and norms may differ for the same item. We compare Word2Vec and an item-graph XSimGCL adaptation across initialization, inheritance, post-processing, training constraints, and downstream augmentation. The inventory contains 48 configurations, including 32 common-protocol CTR evaluations. Observed metrics are reported as relative percentage changes from the corresponding Word2Vec baseline (0%), with raw values and exact dataset sizes withheld. Three findings emerge. First, multiple upstream interventions improve compatibility and CTR. ID-init and orthogonal alignment increase cosine by 51.2% and 93.4%, and GAUC by 2.2% and 2.6%, relative to ordinary Word2Vec. Retrieval and geometry remain insufficient downstream surrogates. Second, warm starts perform strongly but accumulate scale. The day-matched norm increase over Word2Vec reaches 102.5% after eight days; daily global rescaling limits it to 8.2% and attains the highest observed GAUC increase, 4.0%. Third, CTR learning through a dimensionality-reducing encoder further adapts to residual drift. Against dimension-matched random-projection, SVD, and untrained-network controls, trained encoders attain the highest cross-day cosine for four of six upstream methods. The other two rank third and second in train-to-test cosine. These industrial results support complementary roles for upstream stabilization, inherited-scale management, and downstream task learning.