This article details the design of a low‑latency video recommendation system, covering requirements, overall architecture, data pipelines, consistency handling, write smoothing, and performance optimizations such as multi‑level caching and Elasticsearch tuning.
This article shares Baidu's experience and thinking on cross‑domain multi‑objective estimation and fusion for video recommendation, covering background, system overview, multi‑objective design and modeling, long‑term value attribution, cross‑domain network architecture, and the evolution‑strategy based fusion approach.
This article presents Baidu's technical experience on designing, estimating, and fusing cross-domain multi-objective models for its immersive video recommendation system, covering business background, system architecture, target design, long‑term value modeling, and evolution strategies.
This article introduces Social4Rec, a video recommendation algorithm that tackles user cold‑start problems by extracting and integrating social interest information through coarse‑ and fine‑grained interest extractors, attention‑based fusion, and extensive offline and online experiments demonstrating significant CTR improvements.
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.
The paper proposes a Bloom‑filter‑based deduplication service for short‑video recommendation that moves three‑month playback histories to disk‑backed Bloom filters while keeping the latest 100 served IDs in Redis, employing write batching, sharding, expiration policies, and an incremental migration strategy to replace memory‑intensive Redis ZSets and dramatically reduce storage costs.
iQIYI’s short‑video team solves the new‑video cold‑start problem by using a GAN that generates latent user features from video attributes and a discriminator to validate them, then matches these vectors to real users via cosine similarity, achieving double‑digit gains in exposure, CTR, and watch time.
An ecosystem‑wide intelligent promotion system applies a budget‑pacing algorithm with probabilistic throttling and fine‑ranking score adjustments in 5‑minute slots, guaranteeing uniform video exposure while minimizing impact on overall consumption, boosting daily exposure completion from under 5 % to up to 70 % and reducing watch‑time loss.
This article analyzes iQIYI’s evolution of multi‑interest recall techniques—from clustering‑based PinnerSage to MOE and single‑activation models—showing how extracting multiple user interests improves recall diversity, mitigates filter bubbles, and boosts key performance metrics in short‑video recommendation.
The article describes a real‑time video recommendation indexing system that replaces 30‑minute batch builds with an Elasticsearch‑based service, integrates prior and posterior data pipelines, ensures consistency via locking and version checks, enables zero‑downtime upgrades, smooths write spikes, and boosts recall performance through multi‑level caching and ES tuning, delivering sub‑40 ms latency and significant business growth.
The paper proposes a two‑stage framework that first fits a differential‑equation‑based exposure‑click (P2C) model for each new video and then uses a genetic‑algorithm multi‑objective optimization to allocate scarce scene‑level exposure slots, simultaneously maximizing total views and halving CTR variance while outperforming manual baselines.
This article presents a data‑driven approach for allocating limited video exposure resources by building a PV‑click‑CTR (P2C) sensitivity model and a multi‑objective optimization framework that balances overall view volume and fairness across scenes, validated through offline metrics and online bucket tests.
The KDD 2020 paper from Alibaba Entertainment presents a differential‑equation‑based hot‑content exposure sensitivity model and a multi‑objective optimization framework that, under exposure‑resource constraints, guarantees video delivery by accounting for nonlinear content exposure, timing, strategies, and user click habits, now deployed on Youku.
When you browse video apps by tags, receive personalized recommendations, or search with keywords, the seamless experience is powered by Natural Language Processing, which analyzes and interprets textual data to connect users with relevant content, and the article invites you to scan a QR code for further exploration.
The paper proposes a Siamese‑based PredictionNet that combines EfficientB3 image and VGGish audio features with user metrics to predict a HotValue score for newly uploaded short videos, using a margin loss with view‑value‑aware pair selection, enabling tiered cold‑start exposure that boosts overall platform efficiency.
This article explains the rapid growth of short‑video on Taobao, describes the video feature framework, details the RankI2V and RankV2V recall methods, outlines coarse and fine ranking models, and presents real‑time interest and business strategies that significantly improved click‑through rates and viewing time.
The paper proposes a video recommendation framework that combines recall and ranking modules, using a multi‑modal topic clustering approach—integrating audio, visual, and textual features via NeXtVLAD, PCA, and K‑Means—to generate unified video representations, improve candidate selection, and boost click‑through and viewing time, while addressing cold‑start and semantic relevance challenges.
This article details iQIYI's practical experience in building and iterating its video search system, covering basic relevance, semantic matching via translation and click models, deep‑learning approaches, and ranking model evolution from heuristic rules to learning‑to‑rank, highlighting challenges, solutions, and performance gains.
The article reviews Hulu’s 2018 iQIYI keynote on AI video applications, detailing how AI drives personalized recommendations, content analysis through computer vision and NLP, ad targeting across visual, linguistic, and semantic layers, and outlines the platform’s machine‑learning architecture and future directions.
