Uporabnost vložitev v večmodalnih modelih za napovedovanje verjetnosti klika na oglas

Accurate click-through rate (CTR) prediction under strict real-time bidding latency constraints is central to modern programmatic advertising. Models powering large-scale recommender systems rely on high-cardinality categorical IDs and neglect the rich semantic cues present in ad creatives and page context, limitations that become acute in cold-start setting. This thesis sys- tematically explores the integration of off-the-shelf image and text embed- dings into production-ready CTR models. We introduce three fusion strate- gies (early,intermediate,late), a block-wise multi-optimiser training scheme which reduces training memory by approximately threefold and computa- tional demand by approximately fivefold, and the Aligned Deep & Cross Network,which explicitly aligns categorical and external embedding spaces. Experiments on a large, real-world impression log show consistent lifts of up to 0.69% in relative information gain (RIG) over a highly optimised produc- tion model, with pronounced improvements for previously unseen creatives, all while respecting serving-time budgets.The proposed framework offers a practical path to multi-modal CTR prediction at production scale.