Candidate pre-screening layer for large-scale product recommendations
A fast screening layer that narrows millions of candidate products down to the strongest few thousand, so a slower, more careful ranking model only has to work through the shortlist.
A lightweight ranking layer that runs in front of a heavier, more precise recommendation model (a multi-task deep-learning model that predicts each product's click-through rate and conversion rate, see the project above) in an e-commerce recommendation pipeline. It screens a huge pool of candidate products, far too many for that heavier model to score one by one, down into a smaller, high-quality shortlist.
A recommendation pipeline can only afford to carefully score a few hundred products before a page has to load, but the full catalog holds millions of candidates. This layer's job is to make sure the shortlist that reaches the slower, more careful model already contains the strongest matches, instead of that model wasting its limited budget scoring products that were never going to be shown anyway.
A lighter-weight ranking model that sits in front of the heavier ranking model, tuned toward the same business goals so the strongest matches bubble up first before the heavier model spends its budget on them. Drove +13% GMV, +9.24% orders, and +0.73% CTR.
Co-built as part of the Algorithm team at ByteDance.
Want the full technical depth, the tradeoffs, what broke, what I'd do differently? Ask the agent about this project.