1 views collected around this technical thread.
The paper presents CAIG, a CTR‑driven advertising image generation pipeline that pre‑trains a multimodal LLM on e‑commerce data, trains a reward model on CTR‑labeled image pairs, and fine‑tunes generation via product‑centric preference optimization, achieving state‑of‑the‑art online and offline performance.
This article reviews the evolution of large‑model recommendation techniques, analyzes the specific challenges of health‑oriented e‑commerce recommendation, and details practical deployments such as LLM‑enhanced cold‑start recall, DeepI2I expansion, and scaling‑law‑driven CTR models within JD Health.
This article reviews the evolution of large‑model recommendation techniques, analyzes the specific demands and obstacles of health‑focused e‑commerce, and details JD Health's practical implementations—including LLM‑enhanced recall, deep item‑to‑item models, and scaling‑law‑driven CTR improvements—while discussing open research questions and future directions.
By fusing sparse‑feature operators, enabling multi‑stream execution, consolidating data copies, and merging inference batches, iQIYI reduced GPU CTR model latency to CPU‑level, boosted throughput over sixfold, and cut operational costs by more than 40%, overcoming launch‑overhead bottlenecks.
The article introduces STAN, a multi‑task recommendation framework that leverages user lifecycle segmentation to jointly optimize CTR, stay‑time, and CVR, detailing the business context, key challenges, solution architecture, offline and online evaluations, and future research directions.
This article reviews Shopee and Tsinghua University's latest work on multi‑task CVR optimization, introducing the Click‑aware Structure Transfer with Sample Weight Assignment (CSTWA) method, which tackles knowledge sharing and conflict between CTR and CVR through a three‑part architecture, and demonstrates its superior performance on industrial and public datasets.
This article presents Shopee's calibration‑based meta‑learning approach for personalized click‑through‑rate prediction in live streaming, detailing business context, modeling goals, model evolution from Calibration4CVR to CBMR, EmbCB and MlpCB optimizations, and multi‑task and multi‑scene extensions that achieve significant AUC and business metric improvements.
This article introduces STAN, a user‑lifecycle‑aware multi‑task recommendation model proposed by Shopee that refines CTR, CVR, and stay‑time predictions by identifying and tracking user states, demonstrates offline gains on Shopee and public datasets, and reports online improvements in click‑through, dwell‑time, and order metrics.
This article outlines the challenges of multi-task and multi-scenario modeling for large-scale C-end services, reviews key industry approaches such as Shared-Bottom, MMoE, PLE, ESMM, and LHUC, and details Zhaozhuan’s own EPNET-based solution that improved click-through and conversion rates.
AntM2C is a publicly released, billion‑sample click‑through‑rate (CTR) dataset covering five distinct Alipay business scenarios, providing both ID and rich multi‑modal (text and image) features to enable comprehensive evaluation of multi‑scenario, cold‑start, and multi‑modal CTR models at industrial scale.
This article describes the Hybrid Embedding solution proposed by Ant AI Infra to address storage, resource, and feature‑governance challenges of massive sparse CTR models, detailing its multi‑layer storage design, KV‑based parameter server, and performance gains in large‑scale recommendation systems.
Social4Rec introduces a social interest‑enhanced video recommendation framework that tackles user cold‑start by extracting coarse‑ and fine‑grained social interests via a self‑organizing neural network and meta‑path neighborhood aggregation, integrating these embeddings with a YouTube DNN model to improve CTR and AUC.
This article introduces a neuron‑level shared multi‑task learning framework that jointly estimates click‑through rate (CTR) and conversion rate (CVR), discusses the background and advantages of multi‑task learning, reviews classic shared‑bottom models, describes the proposed pruning‑based architecture, and presents experimental results demonstrating its effectiveness in large‑scale recommendation systems.
This article describes the architecture, data processing, city‑intent modeling, resource recall strategies, and multi‑task ranking models—including PLE‑CGC and ESMM—used to improve click‑through and conversion rates of the Tongcheng travel app's homepage waterfall‑flow recommendation, and outlines experimental results and future optimization directions.
This article explains how containerd works as a high‑level container runtime and demonstrates practical usage of its three command‑line clients—ctr, nerdctl, and crictl—for pulling images, managing containers, debugging Kubernetes pods, and performing low‑level runtime operations.
This guide explains what containerd is, how to manage images and containers with its command‑line clients ctr, nerdctl, and crictl, and provides practical examples of pulling, running, inspecting, and cleaning up containers in a Linux environment.
Bayesian hierarchical calibration applies a lightweight, interpretable Bayesian GLM with variational inference to correct pre‑ and post‑click scoring biases, using risk‑aware objectives that reduce calibration error by up to 66%, lift revenue by 5%, and cut conversion costs, while handling cold‑start and dimension‑wise sparsity in online advertising.
This article compares advertising and recommendation algorithms, highlighting distinct optimization goals, model design focuses, training methods, implementation principles, auxiliary strategies, and model characteristics, emphasizing how ads aim to increase revenue while recommendations prioritize user engagement and diversity.
The article introduces FiBiNET, a CTR prediction model that incorporates a SENet module for dynamic feature‑importance learning and a bilinear‑interaction layer for enhanced second‑order feature interactions, then details its improved variant FiBiNET++ which reduces parameters with Bi‑Linear+ and an enhanced SENet+.
The study reveals that industrial deep click‑through‑rate models often overfit dramatically after the first training epoch—a “one‑epoch phenomenon” caused by the embedding‑plus‑MLP architecture, fast optimizers, and highly sparse features, with performance dropping sharply unless sparsity is reduced or training is limited to a single pass.