Advanced AI for Ads Ranking With Reddit Inc

Ai4

AI-assisted summary of Advanced AI for Ads Ranking With Reddit Inc — original source. Check the original for full context.

Source published: 2025-08-26

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 26, 2025, this talk features Vishal Gupta, Engineering Manager of Machine Learning at Reddit Inc. The session description highlights how modern advertising platforms enhance ads ranking and recommender systems by adopting advancements from natural language processing, computer vision, graph neural networks, and large language models.

## TL;DR Vishal Gupta outlines the end-to-end advertising funnel, explaining how candidate retrieval, heavy ranking models, and auction dynamics work together under strict calibration requirements. He traces the historical evolution of engagement prediction across four generations, transitioning from pre-deep learning linear models with decision trees to Wide & Deep paradigms, specialized feature-crossing architectures like DCN-v2 and DLRM, and generative sequence models governed by LLM scaling laws. Gupta also details how Reddit leverages multimodal representations and Graph Neural Networks across posts, subreddits, and ads. He concludes with practical lessons for applied ML practitioners, highlighting the realities of single-digit millisecond latency constraints, the primacy of training infrastructure, and the hidden metric gains achieved through codebase refactoring.

## Atmosphere & style The presentation carries the brisk, focused energy of an experienced engineering leader packing dense technical insights into a compact briefing. Gupta balances architectural depth with candid pragmatism and occasional dry humor, drawing upon career lessons spanning Google, DeepMind, X, and Reddit. The session feels refreshingly grounded, demystifying cutting-edge AI breakthroughs by viewing them through the practical lens of operational latency budgets, rigorous error analysis, and clean system maintenance.

## Key takeaways - [01:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=117s) The ad ranking pipeline filters candidates across retrieval, heavy ranking, and auction stages, distinguishing itself from organic recommender systems because paying advertisers directly influence outcomes through financial bids and budgets. - [04:06](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=246s) Heavy ranker models compute conditional engagement probabilities across upper-funnel and lower-funnel events to calculate estimated cost per mille (eCPM) for ad auction utility. - [04:06](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=246s) Precise model calibration is essential in ads ranking because uncalibrated probabilities directly distort monetary auction pricing and lead to systemic auction instability. - [07:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=420s) First-generation ad systems paired offline Gradient Boosted Decision Trees for non-linear feature extraction with online Logistic Regression, creating an engineering bottleneck around manual feature crossing. - [09:46](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=586s) Second-generation Wide & Deep architectures balanced the memorization of historical co-occurrences with the generalization capabilities of dense continuous embeddings. - [11:33](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=693s) Third-generation specialized architectures like DCN-v2 eliminate manual feature engineering by learning explicit, bounded-degree feature interactions automatically within dedicated cross layers. - [13:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=780s) Meta's DLRM architecture maps sparse categorical features into unified embedding dimensions and calculates explicit dot-product interactions across all feature pairs. - [14:22](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=862s) Platforms like Reddit build multimodal entity representations by fine-tuning foundation models (such as CLIP and Transformers) on ad copy, images, and chronological user activity sequences. - [14:22](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=862s) Graph Neural Networks (GNNs) capture structural community topology across users, subreddits, and posts to generate expressive entity embeddings for downstream retrieval and ranking. - [16:40](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1000s) Fourth-generation systems adopt generative LLM scaling paradigms by representing user engagements as discrete action tokens within sequential Transformer architectures. - [17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) Breakthroughs in AI research migrate to organic recommender systems before ads ranking due to the strict single-digit millisecond latency budgets and auction-level compute constraints of ad platforms. - [17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) Targeted cohort error analysis and periodic refactoring of legacy feature and serving infrastructure often yield larger online revenue gains than deploying complex architectures without optimization.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=4s) Introduction and Speaker Background

Vishal Gupta opens the presentation noting that while the session is brief, it is packed with dense material on advanced AI for ads ranking, meaning he may need to skim through certain technical nuances depending on how timing unfolds. Emphasizing a standard disclaimer, Gupta explains that the perspectives shared reflect his personal trajectory across the applied AI industry rather than the official stance of Reddit. His career path began at Google, transitioning in 2018 to DeepMind to tackle applied machine learning problems. In 2022, he moved to Twitter (now X), spending roughly a year there—including three months under Elon Musk's ownership, which he lightheartedly describes with a snort as "kind of fun." In February 2023, he joined Reddit initially as a machine learning tech lead before transitioning into engineering management.

The session outline is structured around three core pillars: an introduction to how ads ranking operates, an industry-wide historical evolution of engagement prediction models categorized into four distinct generations (from pre-deep learning to modern approaches), and key concluding takeaways for practitioners.

---

[01:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=117s) The Ads Funnel and Ads Auction vs. Recommendation Systems

To understand ads ranking, one must first grasp the broader end-to-end advertising funnel, a system refined across the tech industry over more than two decades. Given an incoming user request and an ad slot to be filled, the primary objective of the ad funnel is to discover and serve the single best candidate ad. The pipeline begins at the retrieval stage, where candidate spaces comprising millions—or on massive platforms, hundreds of millions—of candidate ads are narrowed down to a few hundred or a few thousand candidates. These candidates are then passed to the ranking stage, which computes fine-grained probabilities of user engagement. Finally, top-ranked candidates enter an auction where the winning ad is selected and rendered to the user.

```mermaid flowchart LR accTitle: High-Level Ads Funnel Architecture subgraph S1["Retrieval Stage"] A["Millions / 100M+ Ads"] --> B["Candidate Retrieval"] end subgraph S2["Ranking Stage"] B --> C["A Few Hundred / Thousand Ads"] C --> D["Heavy Ranker Model\nEngagement Probabilities"] end subgraph S3["Auction Stage"] D --> E["Ad Auction\n(eCPM + Relevance)"] E --> F["Winning Ad Served"] end ```

Gupta highlights a fundamental architectural distinction between standard recommender systems and ad systems. Having previously worked on recommendation systems across multiple Google products, he notes that while candidate retrieval and scoring look structurally similar, the auction mechanism introduces a completely different dynamic: in advertising platforms, paying advertisers directly influence which items are displayed through their bids and budgets.

---

[04:06](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=246s) Heavy Ranker Models, Contextual Nuances, and Calibration

The heavy ranker operates as an engagement prediction engine whose core task is calculating the conditional probability that a user will interact with a given creative:

$$\Pr(\text{Engagement} \mid \text{Ad}, \text{User}, \text{Context})$$

Engagement spans a wide taxonomy of event types. Upper-funnel (first-party) events encompass actions occurring directly within the platform's interface, such as clicking on an ad or viewing a video creative. Lower-funnel (third-party) events track post-click user actions occurring on external advertiser properties, including web conversions, mobile conversions, and application installs. These predicted probabilities integrate directly into the ad auction by computing the estimated cost per mille (eCPM)—multiplying the advertiser's bid by the model's predicted engagement probabilities—which is then combined with relevance scores to establish the final auction utility:

$$\text{eCPM} \propto \text{Bid} \times \Pr(\text{Engagement})$$

Gupta stresses that context is critical to prediction accuracy. Context factors include the time of day and the specific UI surface being browsed. On Reddit, this means accounting for distinct surface behaviors across the main home feed, individual subreddit feeds, and deep conversation/comment pages under specific posts. Furthermore, model calibration is paramount: because predicted probabilities are directly multiplied by financial bids to determine auction outcomes across heterogeneous engagement types, uncalibrated probabilities will distort pricing, degrade ad utility, and cause severe systemic auction instability.

---

[07:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=420s) Why AI in Ads Ranking and Generation 1: Pre-Deep Learning Systems

Relying solely on rigid mathematical formulas or basic statistical models fails because modern ad ranking requires delivering "the right ad to the right user at the right time." Historically, legacy systems relied heavily on historical user activity to forecast future behavior. However, if a user initiates a real-time search or enters a context entirely unrelated to their historical profile, past behavioral signals fall flat, underscoring the absolute necessity of real-time contextual modeling. Powerful AI models bridge this gap, optimizing both short-term performance and long-term ecosystem value for users and advertisers alike.

```mermaid flowchart TD accTitle: Generation 1 Pre-Deep Learning Architecture subgraph Offline["Offline Batch Training (Periodic)"] A["Historical Data"] --> B["Decision Trees / GBDT\n(Non-linear Feature Transforms)"] end subgraph Online["Online Real-Time Training"] B --> C["Leaf Node Index Features"] D["Real-Time Stream"] --> E["Logistic Regression\n(Online Linear Model)"] C --> E end E --> F["Click / Engagement Prediction"] ```

In Gupta's classification, First Generation (Gen 1) systems represent the pre-deep learning era: - **Core Architecture:** Centered on Logistic Regression (LR) computing linear combinations across extensive sets of sparse categorical and dense numerical features. - **Non-Linearity via GBDTs:** To capture non-linear feature interactions, platforms adopted architectures popularized by Facebook's seminal 2014 paper, pairing Gradient Boosted Decision Trees (GBDT) with an online linear model. - **Training Frequency:** GBDT components were retrained offline every few days or weeks to extract structural feature crosses, whereas the downstream logistic regression layer was updated online in real time on streaming data to catch shifting trends. - **Scaling Bottlenecks:** Maintaining these systems required hundreds to thousands of engineers manually engineering handcrafted feature crosses, which eventually became a severe organizational and computational scaling bottleneck.

---

[09:46](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=586s) Generation 2: Deep Learning and Wide & Deep Architectures

Second Generation (Gen 2) models emerged as ad platforms adopted deep learning, with industry pioneers like Google and Meta transitioning around 2016–2017 and other platforms following suit. The primary limitation of pure logistic regression was its tendency to merely "memorize" past co-occurrences without the capacity to generalize to unseen feature combinations. Deep Neural Networks (DNNs) resolved this by learning dense, low-dimensional continuous embeddings capable of generalizing across sparse feature spaces.

```mermaid flowchart TD accTitle: Wide and Deep Architectural Paradigm subgraph Wide["Wide Component (Memorization)"] W1["Cross-Product Transformations\n(Historical Co-occurrences)"] --> W2["Linear Model"] end subgraph Deep["Deep Component (Generalization)"] D1["Dense Embeddings\n(Continuous Representations)"] --> D2["Hidden Deep Layers"] end W2 --> Out["Final Engagement Probability Prediction"] D2 --> Out ```

Inspired by recommender system research, platforms widely embraced Google's *Wide & Deep* paradigm: - **Wide Component:** Preserved the memorization strengths of linear models over historical feature interactions. - **Deep Component:** Provided non-linear generalization through deep neural network stacks. - **Training Paradigm:** Gen 2 systems operated primarily via offline batch training over massive datasets. - **Impact:** Although extensive manual feature engineering remained necessary, deep learning unlocked substantial performance gains by extracting signal from massive data volumes.

---

[11:33](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=693s) Generation 3: Specialized Architectures and Automated Feature Crosses (DCN-v2)

Generation 3 (Gen 3) evolved away from massive, monolithic feed-forward DNNs toward specialized, domain-tailored neural architectures. While a single-layer feed-forward network with sufficient width is theoretically a universal function approximator, in practice, designing explicit architectural components for specific ranking subtasks yields far superior empirical performance and training efficiency.

Gupta highlights Deep & Cross Network V2 (DCN-v2) as an especially impactful milestone in this evolution: - **Eliminating Manual Crosses:** Rather than requiring ML engineers to manually hypothesize and craft high-degree feature crosses, DCN-v2 delegates interaction learning directly to dedicated model layers. - **Cross Layers:** The cross-layer architecture automatically computes explicit, bounded-degree feature interactions at each subsequent layer without combinatorial explosion. - **Controllable Interaction Degree:** Stacking $N$ cross layers explicitly models interaction degrees up to order $N+1$, allowing the network to systematically learn complex multi-way feature relationships.

[13:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=780s) Feature Crossing Architectures and DLRM

Deep & Cross Network (DCN) architectures make it possible to automatically learn explicit feature interactions up to multiple degrees. While practitioners can technically stack many cross layers, practical experience shows that two to three cross layers are more than sufficient. Furthermore, parallel cross networks—where the cross network runs alongside a standard deep network—consistently outperform sequentially stacked cross networks.

Building directly on the necessity of capturing sparse pairwise interactions, Meta pioneered the Deep Learning Recommendation Model (DLRM), with related architectural variations also developed at Google. In prior generations of recommendation systems, combining sparse signals was proven to yield the largest lift—such as understanding that a user watched a movie in a specific genre and subsequently bought a related product. DLRM formalized this by projecting all categorical sparse features into a unified embedding dimension, followed by computing explicit dot-product interactions between every feature pair. This design significantly expanded what was achievable at the intersection of deep learning and recommender systems.

- **Cross Layer Depth:** Two to three cross layers typically capture optimal non-linear interactions without unnecessary overhead. - **DLRM Core Innovation:** Maps distinct sparse categorical features to embedding vectors of identical dimensionality and computes dense dot-product crosses across all pairs.

---

[14:22](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=862s) Multimodal Entity Representations and Graph Neural Networks at Reddit

High-capacity entity representations form the backbone of both the retrieval and ranking stages. At Reddit, primary entities include posts, subreddits, and ads. As computer vision and natural language processing advanced, their foundational models were adopted to generate rich semantic embeddings. Image models extract visual features from creative assets, while LSTMs, GRUs, and Transformers encode post titles, descriptions, and comments. Multimodal vision-language models, such as OpenAI's CLIP, are frequently utilized to generate joint text-and-image embeddings. The standard operational recipe involves taking off-the-shelf pre-trained foundation models, deploying them directly or fine-tuning them on in-house platform data, and feeding the resulting embeddings downstream into core ranking models.

Beyond unstructured media, sequence encoders and graph representation learning unlocked relational modeling. Instead of tokenized natural language, sequence models ingest chronologically ordered sequences of user engagements. Simultaneously, Graph Neural Networks (GNNs) capture the structural topology of Reddit's ecosystem, where users subscribe to specific subreddits while independently interacting with posts outside their subscribed communities. Passing this heterogeneous graph into powerful graph encoders with supervised objectives yields expressive embeddings across all platform entities.

```mermaid flowchart TD accTitle: Reddit Multimodal and Graph Entity Representation Pipeline accDescr: Diagram illustrating how text, images, sequences, and graph topology are processed into unified entity representations for ranking.

subgraph Inputs ["Platform Signals & Entities"] IMG["Ad & Post Images"] TXT["Post & Ad Copy"] SEQ["User Action Sequences"] GRAPH["User-Subreddit-Post Graph"] end

subgraph Encoders ["Specialized Encoders"] CLIP["Vision & Multimodal Models (CLIP)"] TRANS["Sequence Models / Transformers"] GNN["Graph Encoders (GNNs)"] end

subgraph Output ["Downstream Ranking"] EMB["Unified Entity Representations\n(Posts, Subreddits, Ads)"] RANK["Ads Retrieval & Ranking Models"] end

IMG --> CLIP TXT --> CLIP SEQ --> TRANS GRAPH --> GNN

CLIP --> EMB TRANS --> EMB GNN --> EMB EMB --> RANK ```

---

[16:40](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1000s) Generative RecSys: Scaling Laws and Sequence Modeling for Ads

Modern fourth-generation recommendation systems increasingly adopt the generative scaling paradigms of Large Language Models (LLMs). Meta's landmark research paper, *"Actions Speak Louder Than Words,"* established that the fundamental scaling laws governing autoregressive LLMs apply directly to sequential recommendation architectures.

Instead of passing natural language tokens into a Transformer decoder, these architectures represent discrete user engagements as an ordered sequence of action tokens. Scaling compute, dataset volume, and parameter counts on these action sequences unlocks substantial performance gains across both organic content recommendation and real-time ads ranking pipelines.

- **Token Analogy:** User actions (clicks, views, community visits) replace textual words as the fundamental vocabulary of the sequence model. - **Power of Scale:** Model performance scales predictably with increased sequence length, compute, and parameter capacity.

---

[17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) The Applied ML Lifecycle: Latency Constraints, Infrastructure, and Clean Codebases

Machine learning breakthroughs follow a defined progression: breakthroughs originating in NLP, computer vision, and graph learning are first absorbed by organic recommendation systems, and only later migrate into ads ranking. Direct, immediate adoption in ads ranking is blocked by extreme operational constraints. Ads systems operate under strict, single-digit millisecond latency budgets and evaluate massive candidate pools. While organic feed ranking can compute recommendations once per session, ads ranking must run complex inference passes for every distinct ad slot auction across millions of concurrent requests, making compute efficiency paramount.

```mermaid flowchart LR accTitle: Research Innovation Flow to Ads Ranking accDescr: Flow diagram showing research migrating from foundational AI fields to organic recommender systems and finally to ads ranking.

AI["Core AI Research\n(NLP, Vision, Graphs, LLMs)"] REC["Organic Recommender Systems\n(Looser Latency Budgets)"] ADS["Ads Ranking Systems\n(Strict SLA & Per-Slot Cost)"]

AI --> REC REC --> ADS ```

Applied ML engineers must also avoid the common pitfall of treating advanced model architectures as a "hammer" where "every problem starts looking like nails." Squeezing maximum real-world value requires granular error analysis to discover exactly which user or advertiser segments are underperforming, followed by targeted modeling. Furthermore, machine learning training and serving infrastructure quickly becomes the ultimate ceiling on innovation. Long-term practitioners discover that periodically refactoring and rewriting legacy feature engineering and serving pipelines often produces dramatic, unexpected online revenue wins by eliminating silent technical debt, metric drift, and unintentional implementation bugs accumulated during rapid prototyping.

- **Inference Constraints:** Ads ranking faces tight real-time latency SLAs and multiplicative compute costs per ad impression auction. - **Targeted Improvements:** Error analysis on underperforming cohorts delivers greater incremental ROI than blindly deploying complex architectures. - **Infrastructure Investment:** Robust training and serving platforms prevent infrastructure from throttling applied modeling velocity. - **Code Refactoring Gains:** Rewriting legacy serving codebases frequently uncovers silent bugs, yielding substantial online metric improvements.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=4s) kind of fun — Vishal Gupta

> [07:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=420s) the right ad to the right user at the right time — Vishal Gupta

> [09:46](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=586s) memorize — Vishal Gupta

> [16:40](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1000s) Actions Speak Louder Than Words — Vishal Gupta

> [17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) hammer — Vishal Gupta

> [17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) every problem starts looking like nails — Vishal Gupta

## Mentioned Vishal Gupta, Reddit, Google, DeepMind, Twitter, X, Elon Musk, Meta, Facebook, OpenAI, CLIP, Wide & Deep, Deep & Cross Network V2 (DCN-v2), Deep Learning Recommendation Model (DLRM), Actions Speak Louder Than Words

## Video description Vishal Gupta, Engineering Manager, Machine Learning, Reddit Inc

Abstracts

How advanced AI can improve Ads Ranking?

AI models have been helping Ads platforms to show relevant Ads to their users since last couple of decades. The advancement in Ads ranking models has accelerated with the advancement in fields like NLP, Computer Vision, Recommender Systems.

In this talk you will learn how Ads ranking models have improved by drawing inspiration from Transformers, LLMs, Graph neural networks etc. Lets dive deep into concepts that can improve Ads ranking and other related domains like Recommender Systems, e-commerce etc.

[Source](https://www.youtube.com/watch?v=CbTWx11S9wQ)

中文简报

Reddit 高级广告精排 AI 技术 Vishal Gupta 概述了端到端广告漏斗机制，阐释了在严格的校准要求下，候选检索、重排序模型与竞价动态如何协同运作。他梳理了互动预测技术经历的四个演进世代：从深度学习时代前结合决策树的线性模型，过渡到 Wide & Deep 范式、DCN-v2 和 DLRM 等专用特征交叉架构，再到受 LLM 扩展律驱动的生成式序列模型。Gupta 还详细介绍了 Reddit 如何跨帖子、subreddit 与广告利用多模态表示和图神经网络。最后，他为应用机器学习从业人员总结了实用经验，强调了个位数毫秒级延迟限制的现实挑战、训练基础设施的首要地位，以及通过代码库重构所获得的隐性指标提升。广告排序漏斗包含召回、精排（Heavy Ranking）和竞价拍卖三个阶段；与自然内容推荐系统不同，广告系统由付费广告主通过出价和预算直接干预曝光结果。 (1:57) 精排模型负责计算漏斗上层与下层转化行为的条件交互概率以推导 eCPM，因此精准的模型校准（Calibration）至关重要，未校准的概率会直接扭曲拍卖计费并导致竞价体系失衡。 (4:06) 排序架构经历了四代演进：从早期的 GBDT 结合逻辑回归（LR），到平衡记忆与泛化的 Wide & Deep，再到通过 DCN-v2 与 Meta DLRM 实现显式、多阶特征自动交叉的专用架构。 (7:00) Reddit 等平台通过微调 CLIP 和 Transformer 等基础模型，结合时序用户行为与异构图神经网络（GNN），构建出覆盖帖子、Subreddit 与广告的多模态实体表征。 (14:22) 前沿推荐系统正借鉴大语言模型的自回归 Scaling Law，将用户行为序列 Token 化；同时在严格的毫秒级时延限制下，持续重构特征与推理底层基建往往能带来巨大的线上收益。 (16:40)

Reading summary

Modern ads ranking systems rely on multi-stage funnels where predictive machine learning directly determines auction pricing and ad selection. Drawing on industry experience across Google, Meta, and Reddit, Vishal Gupta traces the field's evolution from handcrafted linear pipelines to automated deep architectures and generative sequence models. Auction Dynamics and Probability Calibration (1:57) The advertising funnel winnows millions of candidate creatives down to hundreds or thousands in retrieval before a heavy ranker estimates the conditional probability of engagement given the ad, user, and context. Unlike organic recommender systems, advertising requires balancing relevance against commercial bids. Estimated probabilities for upper-funnel events (clicks, video views) and lower-funnel events (installs, conversions) are multiplied by advertiser bids to calculate estimated cost per mille (eCPM), which is combined with relevance metrics to determine auction winners. Because model outputs are directly multiplied by financial bids, Gupta emphasizes that precise probability calibration is critical. Uncalibrated predictions distort pricing across disparate engagement types, impair advertiser utility, and destabilize auction equilibria. Accuracy also depends heavily on real-time context, including browsing time and UI placement across Reddit's home feed, specific subreddits, or comment conversations. Architectural Evolution Across Four Model Generations (7:00) Ad ranking models have advanced through four generations. First-generation systems relied on online Logistic Regression paired with offline Gradient Boosted Decision Trees (GBDTs), following Facebook's 2014 design. GBDTs mapped structural feature crosses offline, while downstream linear models trained on real-time streaming data; however, managing handcrafted feature crosses created severe scaling bottlenecks. Second-generation platforms adopted Google's Wide & Deep paradigm around 2016–2017, combining linear memorization of historical co-occurrences with deep neural networks learning continuous dense embeddings for generalization. Third-generation systems replaced monolithic deep networks with specialized feature-crossing layers. Google's Deep & Cross Network V2 (DCN-v2) enabled automated, bounded-degree interaction modeling, with empirical implementations favoring two to three parallel cross layers. Meta's Deep Learning Recommendation Model (DLRM) similarly projected sparse categorical features into uniform dimensions to execute pairwise dot-product crosses. Fourth-generation architectures now adopt generative scaling paradigms. Citing Meta's "Actions Speak Louder Than Words," Gupta highlights how treating chronological user actions as discrete tokens allows autoregressive Transformer scaling laws to apply directly to recommendation and ranking. Multimodal Embeddings and Graph-Based Entity Modeling (14:22) Entity representation at Reddit relies on multimodal encoders to capture semantic context across ads, posts, and subreddits. Standard industry practice involves taking off-the-shelf foundational models, deploying them directly or fine-tuning them on proprietary data, and passing the embeddings into downstream rankers. Vision models process image creatives, sequential encoders (Transformers, LSTMs, GRUs) process text from titles and comments, and multimodal frameworks like OpenAI's CLIP generate joint text-and-image embeddings. Beyond unstructured media, Reddit models relational platform topology using Graph Neural Networks (GNNs). By ingesting the heterogeneous graph of user subscriptions and interactions with posts outside their subscribed subreddits, supervised GNN encoders generate unified structural representations that enrich both retrieval and ranking candidate scoring. Operational Latency and Infrastructure Hygiene (17:09) Ad ranking systems operate under distinct latency and cost boundaries compared to organic recommendation engines. While organic feeds often score content once per user session, ads ranking must compute multi-candidate inference within single-digit millisecond latency budgets for every auction slot across millions of concurrent requests. Consequently, foundational research from computer vision and natural language processing is typically validated in organic recommenders before adapting to production ads pipelines. Gupta advises practitioners against defaulting to architectural complexity, noting that targeted error analysis on underperforming cohorts yields higher returns than treating advanced models as universal solutions. Furthermore, machine learning infrastructure serves as the ultimate ceiling on modeling velocity. Gupta observes that periodically refactoring legacy feature-engineering and serving pipelines consistently delivers significant online revenue improvements by eliminating metric drift and silent technical debt. While ads ranking architectures increasingly adopt foundation models and generative sequence paradigms, real-world deployment remains tightly constrained by strict millisecond latency budgets, auction probability calibration requirements, and the operational integrity of underlying serving infrastructure.

精炼阅读摘要

Vishal Gupta结合其在Google、DeepMind、Twitter和Reddit的应用机器学习经验指出，广告排序是受制于竞价机制与毫秒级延迟预算的复杂系统工程；模型架构虽已从浅层非线性组合演进至专用交叉网络、图多模态表征及生成式序列模型，但实际落地成效依然高度取决于概率校准、定向误差归因与基础设施的工程治理。漏斗机制、竞价拍卖与概率校准 (1:57) 现代广告投放链路遵循由工业界二十余年沉淀形成的“检索召回—重排—拍卖”三层漏斗。在召回阶段，系统需从数百万乃至上亿级候选池中筛选出数百至数千个候选广告；随后的重排阶段（Heavy Ranker）则计算用户在特定上下文与创意下的多类型交互条件概率，包括应用内的点击、播放等浅层事件，以及应用外的转化、安装等深层事件。与自然推荐系统不同，广告系统的核心分流机制由广告主的出价与预算主导，最终排名取决于预估千次展示收益（eCPM，即出价与预测点击转化率的乘积）以及相关性得分的综合判定。时间、设备以及Reddit独有的主信息流、子版块（Subreddit）和深层评论串等UI上下文对预测精度具有决定性影响。Gupta强调，模型输出概率的校准（Calibration）是系统稳定的基石：由于预估概率会直接与广告主的真实财务出价相乘，若预估值发生漂移或未校准，将扭曲多目标异构事件的竞价公平性，破坏广告主效用，并最终导致整体拍卖生态失稳。架构演进：从手工特征到深度交叉网络 (7:00) 由于用户的实时检索与即时意图往往与其历史行为画像脱节，广告排序无法依赖粗糙的统计规则，必须借助强上下文AI模型。第一代（Gen 1）系统主要依托以Facebook在2014年论文为代表的GBDT与在线逻辑回归（LR）组合：由离线决策树每隔数天或数周更新以提取非线性交叉特征，下游LR则在流式数据上实时更新。然而，依赖成百上千名工程师手工构造特征交叉的做法，最终成为了组织与算力的扩展瓶颈。第二代（Gen 2）系统以2016至2017年前后Google提出的Wide & Deep架构为分水岭，引入连续低维稠密向量嵌入，兼顾了线性模型的记忆能力与深度神经网络的泛化能力。第三代（Gen 3）系统逐步摒弃了单体前馈全连接网络，转向领域专用的特征交互架构。Gupta特别指出DCN-v2（深度交叉网络V2）与Meta提出的DLRM是该阶段的关键突破。DCN-v2利用显式交叉层自动建模高达N+1阶的特征组合，有效遏制了组合爆炸，且两到三层并行结构的实测表现普遍优于深度堆叠；DLRM则将稀疏类别特征映射至相同维度空间，通过全特征对的点积交互高效挖掘高价值稀疏信号，彻底终结了对手工特征交叉的依赖。多模态融合、图拓扑与生成式序列 (14:22) 在Reddit的工程实践中，高质量的实体表征构成了召回与排序的核心基石。针对帖子、子版块与广告创意，系统利用主流基础模型提取语义：通过图像模型解析创意视觉素材，利用Transformer、LSTM或GRU编码文本标题与评论，并借助CLIP等跨模态模型对图文进行联合表征。对于结构化关联，序列模型负责捕捉用户随时间变化的交互行为；图神经网络（GNN）则用于建模用户、帖子与版块构成的异构拓扑网络——GNN能够有效学习用户在订阅特定社区之外与站内其他内容的跨圈层交互，为下游排序模型提供高表达力的实体嵌入。第四代（Gen 4）系统开始全面引入大语言模型的生成式扩展范式。Meta在论文《Actions Speak Louder Than Words》中验证，自回归大语言模型的扩展律（Scaling Laws）完全适用于序列推荐系统。广告排序模型将用户的点击、浏览、访问等离散动作抽象为“动作标记（Action Tokens）”，伴随参数量、训练计算量以及动作序列长度的成倍扩展，模型在自然推荐与广告实时排序中均展现出了高度可预测的性能增益。落地约束：延迟红线与工程管线治理 (17:09) 机器学习前沿创新的落地存在明确的梯次迁移规律：NLP、计算机视觉与图学习的突破通常先由自然推荐系统吸收，随后才逐步渗透到广告排序。阻碍技术直接迁移的根本障碍在于极度严苛的工程SLA——广告竞价必须在个位数毫秒级的时间窗口内完成推理打分，且每次并发曝光拍卖都需要对大量候选执行重算，计算成本远高于一次会话仅计算一次的自然流推荐。 Gupta告诫应用算法团队切忌盲目将前沿模型当作“四处找钉子的锤子”；相比于盲目堆叠网络复杂度，针对表现较差的细分受众群体与广告主进行详尽的误差分析，往往能带来更高的边际投资回报率。此外，训练与推理基础设施是算法迭代的物理天花板。在成熟的工业实践中，定期重构并重写积攒了技术债的历史特征工程管线与在线服务代码，通常能够消除潜藏的指标漂移与静默缺陷，带来意料之外的显著线上变现增益。 Gupta总结认为，广告排序算法的突破不仅在于模型参数与网络深度的堆叠，更在于工程实现能否严密契合毫秒级延迟预算与拍卖博弈逻辑。在专用交叉架构与生成式序列模型重构预估能力的同时，高精度的概率校准、以误差归因为导向的局部精细化建模，以及对陈旧工程管线的持续重构，仍是保障平台长期变现效能不可替代的底层支柱。

Detailed notes — zh

## Background 该视频由频道 Ai4 于 2025 年 8 月 26 日发布在 YouTube 平台。视频简介显示，本次演讲由 Reddit 机器学习工程经理 Vishal Gupta 带来，主要探讨如何借鉴 Transformer、大语言模型（LLM）和图神经网络（GNN）等前沿 AI 技术来优化工业级广告排序系统及推荐系统。 ## TL;DR 演讲系统梳理了广告排序漏斗的核心机制与排序模型（Heavy Ranker）的运作原理，强调了广告系统因竞价拍卖与单槽位成本约束而区别于常规推荐系统的本质特征。讲者将广告排序技术划分为四个演化代际：从第一代 GBDT+LR 依赖人工特征工程，到第二代 Wide & Deep 结合记忆与泛化，再到第三代以 DCN-v2 和 DLRM 为代表的自动特征交叉，直至第四代将用户行为作为 Token 的生成式扩展定律（Scaling Laws）架构。此外，分享还结合 Reddit 的实践阐述了多模态与图神经网络的应用路径，并强调了深耕基础设施与警惕“唯模型论”的实战落地经验。 ## Atmosphere & style 整场演讲呈现出干练、严谨且极具工业界一线实战色彩的分享风格。讲者 Vishal Gupta 叙述沉稳、条理清晰，开场穿插个人大厂经历的幽默自嘲，随后迅速切入高密度的系统架构解析与数学本质拆解。整体氛围既有技术演进脉络的宏观视野，又充满了对工程延迟、算力瓶颈和代码重构红利的诚恳洞察，给人一种极具实操指导价值的资深工程师对谈感。 ## Key takeaways - [01:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=117s) 广告漏斗包含召回（Retrieval）、排序（Ranking）与竞价拍卖（Auction）三大核心阶段，其中竞价机制是广告系统区别于常规推荐系统的本质特征。 - [03:59](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=239s) 核心排序模型（Heavy Ranker）本质上是条件概率预估模型，预估在给定广告、用户及特定上下文（Context）下的各类上漏斗与下漏斗互动概率。 - [03:59](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=239s) 在广告排序中概率校准（Calibration）至关重要，因为预估概率会直接与广告主出价相乘计算 eCPM，若概率失准将导致整个竞价链路失衡。 - [06:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=417s) 现代广告排序必须重度依赖 AI 捕捉实时意图与瞬时上下文，仅依赖用户历史行为画像在用户切换新话题时会全面失效。 - [08:15](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=495s) 广告排序演进第一代以 GBDT+LR 为代表（依赖繁重人工特征工程），第二代以 Google Wide & Deep 为代表，首次结合了 LR 的记忆能力与 DNN 的语义泛化能力。 - [11:33](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=693s) 第三代排序模型通过特化网络结构实现自动化特征交叉，例如 DCN-v2 通过跨层显式控制交互阶数，DLRM 则通过统一稀疏特征嵌入与两两点乘捕获交互。 - [13:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=780s) 在特征交叉工程实践中，通常堆叠 2 到 3 层交叉层即可捕捉绝大多数高阶特征组合，且并行交叉网络的综合表现普遍优于层叠式结构。 - [14:22](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=862s) 通过引入预训练多模态大模型（如 CLIP）提取特征或基于业务数据微调，可生成高质量的帖子、广告和社区表征并注入排序模型中。 - [15:32](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=932s) 利用 Transformer 对用户行为序列建模，以及利用图神经网络（GNN）挖掘 Reddit 独特的订阅和跨版块互动拓扑，能有效提升深层协同表征质量。 - [16:40](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1000s) 第四代广告架构探索将用户历史动作（Actions）作为 Token 进行序列化生成式建模，证明了大模型的 Scaling Law 同样适用于广告与推荐系统。 - [17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) 基础 AI 技术的工业界落地通常遵循“NLP/CV/GNN $\rightarrow$ 推荐系统 $\rightarrow$ 广告排序”的传导链，广告排序落地滞后主要受几十毫秒内完成打分的极严苛低延迟与高算力成本限制。 - [18:07](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1087s) 算法落地应避免“手里拿着先进架构看什么都像钉子”，针对薄弱数据切片定向优化、重投训练推理基础设施，并定期重构代码清理隐蔽 Bug，往往能带来扎实的线上收益。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=4s) 讲者背景与广告排序概览演讲开篇，Reddit 机器学习工程经理 Vishal Gupta 明确给出了免责声明：本次分享的内容纯属个人在各家公司从事应用 AI（Applied AI）实践的经验总结，并不代表 Reddit 的官方立场。回顾自己的职业路径，Vishal 在应用 AI 领域积淀深厚：最早在 Google 接触工业级机器学习系统；2018 年加入 DeepMind 负责一系列实际落地业务的机器学习项目；到了 2022 年，他跳槽至当时的 Twitter（现已更名为 X），并在埃隆·马斯克（Elon Musk）执掌期间工作了大约一年零三个月，他笑着打趣道“那段经历挺有意思的”；2023 年 2 月，他正式加入 Reddit，起初担任技术主管（Tech Lead），随后逐步转型走向工程管理岗位。整场分享的结构分为三个主要板块：首先解析广告排序（Ads Ranking）与广告漏斗的基础运作机理；其次是他个人对工业界广告排序模型演进历程的归纳与划分，将其清晰拆解为从“第一代”到“第四代”的发展路径；最后则总结落地高阶 AI 技术的关键洞察。 --- ### [01:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=117s) 广告漏斗架构与推荐系统的本质差异要理解复杂的广告排序，首先必须透彻认识经历了二十余年工业级优化的“广告漏斗”（Ads Funnel）。对于给定的用户访问请求以及待填充的广告位（Ad Slot），广告漏斗的核心任务是在严苛的时间延迟约束内，挑出最合适展示给该用户的广告。 ```mermaid flowchart LR accTitle: 广告漏斗基本流转过程 subgraph S1["召回阶段 (Retrieval)"] A["候选广告库 (数百万至数亿广告)"] --> B["粗筛缩小 (降至数百至数千广告)"] end subgraph S2["排序阶段 (Ranking)"] B --> C["Heavy Ranker 预估各类交互概率"] end subgraph S3["竞价与展现 (Auction)"] C --> D["竞价拍卖竞逐 (结合 bid 与 eCPM)"] --> E["胜出广告曝光展示"] end ``` 整个流程通常始于检索召回阶段（Retrieval Stage），从广告底池中的数百万甚至数亿候选广告中快速筛出几百到几千个相关候选；随后进入排序阶段（Ranking Stage），由复杂的模型预估各类用户行为概率；最终进入竞价拍卖（Auction），胜出的广告得以真正渲染展示。 Vishal 强调了一个许多从推荐系统（Recommender Systems）转行做广告的算法工程师容易忽视的关键差异：在 Google 多个核心产品做推荐系统时，整个流程同样是经典的“召回 + 排序”，但广告系统最大的区别在于**拍卖机制（Auction）**——在这里，广告主真金白银地花钱竞价，并拥有直接影响哪个广告能够展现在用户眼前的商业杠杆。 --- ### [03:59](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=239s) 核心排序模型（Heavy Ranker）、出价与校准机制排序阶段的核心引擎通常被称为“重排序模型”（Heavy Ranker）。从本质上讲，这类模型全部都是行为交互概率预估模型（Engagement Prediction Models），其数学形式可以精确表达为：在给定广告（Ad）、用户（User）及上下文环境（Context）的条件下，预估用户与该广告发生特定互动的概率： $$P(\text{Engagement} \mid \text{Ad}, \text{User}, \text{Context})$$ 平台需要预估的互动事件多种多样： - **上漏斗事件 / 第一方事件（Upper Funnel / First-party events）**：主要发生在媒体平台站内，如点击（Click）、视频播放（Video View）等。 - **下漏斗事件 / 第三方事件（Lower Funnel / Third-party events）**：通常来源于外部广告主的监测回传，包括网页转化、移动端转化（Web/Mobile Conversions）以及 App 安装（App Installs）等。在接入拍卖机制时，系统通常会将广告主的出价（Bid）与 Heavy Ranker 预估出的概率相乘，计算出千次展示期望收入（eCPM，Estimated Cost Per Mille）。拍卖引擎再通过综合衡量 eCPM 与用户相关度（Relevance），计算出最终效用函数（Utility）以确定赢家。 Vishal 特别提醒与会者，概率公式中的**上下文（Context）至关重要**。所谓上下文，涵盖了时间（Time of Day）以及广告展示的具体页面表面（Surface）。以 Reddit 为例，站内包含主信息流（Main Home Feed）、特定板块信息流（Subreddit Feed）以及单个帖子下方充满讨论的评论详情页（Comments / Conversations Page）。不同的产品表面存在极其微妙的用户心智与场景差异，排序模型必须全面建模这些细微区别。此外，**校准（Calibration）**是广告排序的生命线——因为预估概率将直接与真金白银的出价相乘，一旦跨多种互动类型的概率输出失准，下游整条竞价链路就会彻底失衡。 --- ### [06:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=417s) 为什么广告排序必须拥抱 AI 与上下文感知如果广告可以通过统计学规则或传统数学公式来打分，为什么平台必须重度依赖高阶 AI？Vishal 指出，AI 赋能的核心价值在于让平台能够“在正确的时间，将正确的广告展示给正确的用户”（Right ad to the right user at the right time）。 - **打破过去行为的束缚**：回看多年前的传统广告系统，排序严重依赖用户的历史行为特征来预测未来。但如果用户当下正在主动搜索或浏览一个与以往历史毫无关联的新话题，仅靠历史行为画像的预估就会全面失效。 - **瞬时意图捕获**：这就是为什么“当下时机”与“实时上下文”不可或缺。高阶 AI 能够深度捕捉用户当下的实时意图，平衡短期点击收益与长期用户体验，从而在生态中实现广告主收益与用户价值的最大化。 --- ### [08:15](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=495s) 广告排序演进第一代与第二代：从 GBDT+LR 到 Wide & Deep Vishal 将工业界广告排序模型的演变划分为不同代际，系统梳理了从统计学习到深度学习的跨越： | 世代划分 | 代表架构与代表工作 | 训练方式与数据流 | 核心特点与局限 | |:--- |:--- |:--- |:--- | | **第一代：深度学习前时代 (Pre-Deep Learning)** | GBDT + 逻辑回归 (LR) *(例如 Facebook 2014 年经典论文)* | GBDT 周期性离线批训练（数天或数周）；LR 线上流式实时训练（Online Training） | GBDT 挖掘非线性特征组合，LR 线性拟合捕捉最新趋势；极其依赖手工特征工程，无法应对数据与用户规模暴涨。 | | **第二代：早期深度学习时代 (Early Deep Learning)** | 基础 DNN、Google Wide & Deep (2016-2017) | 全量离线批处理数据训练（Offline Batch Training） | Wide 部分负责记忆历史高频模式（Memorization），Deep 部分利用 Embedding 和高层网络进行语义泛化（Generalization）。 | 在第一代模型中，成百上千名算法工程师耗费海量精力在特征交叉（Feature Crosses）与特征挖掘上。Facebook 当年提出的经典方案是用 GBDT 挖掘非线性交叉特征并输出叶子节点索引，再输入给在线学习的 LR 模型。虽然 LR 在线更新保证了对最新热点的敏锐度，但随着广告平台用户量和数据量的指数级爆发，人工构造特征的模型架构不可避免地遭遇了天花板。大约在 2016 至 2017 年间，以 Google 和 Meta 为代表的巨头率先引入深度学习，开启了第二代广告排序体系。推荐系统领域的突破迅速反哺广告领域，Google 提出的 Wide & Deep 架构极具里程碑意义：传统的 LR 擅长“记忆”（Memorization），即记住历史出现过的强相关组合；而多层感知机（DNN）擅长“泛化”（Generalization），能够推断从未出现过的组合。Wide & Deep 将二者融合，使得模型在处理超大规模数据时释放出惊人的性能提升，但这一阶段对高质量手工特征工程的依赖依然较高。 --- ### [11:33](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=693s) 第三代演进：网络结构特化与自动特征交叉（DCN-v2）进入第三代演进后，学术界与工业界不再满足于仅堆叠一个巨大的单体多层全连接网络（Large Monolithic DNN）。尽管从万能近似定理（Universal Approximation Theorem）的角度来看，理论上拥有足够神经元的单层或多层 DNN 可以逼近任何连续函数，但在工程实操中，为特定任务专门设计带有特定归纳偏置（Inductive Bias）的网络拓扑结构，能够带来巨大收益。 Vishal 重点拆解了业内非常经典的 **DCN-v2（Deep & Cross Network V2）** 架构。该网络的核心愿景非常明确：**用模型内部的自动化机制彻底取代工程师人工设计特征交叉**。 - **Cross Layer 的核心机制**：DCN-v2 的精妙之处完全体现在其跨层（Cross Layer）设计上。 - **阶数可控的显式特征交互**：网络通过级联 Cross Layer 的深度，精确定义了特征显式交叉的交互阶数（Degree of Feature Interactions）。例如，堆叠两层 Cross Layer 即可显式且低成本地学习到特征之间的二阶交叉组合，免去了人工繁琐罗列特征组合的沉重负担。 ### [13:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=780s) 特征交叉的实践经验与 DLRM 架构在特征交叉网络的实际工程落地中，堆叠过多的交叉层往往并不能带来预期的收益。根据工程经验，构建 2 到 3 层的特征交叉层通常就已经绰绰有余；同时，并行交叉网络（Parallel Cross Networks）在实际效果和性能表现上普遍优于层叠式交叉网络（Stacked Cross Networks）。谈到特征交叉，最具里程碑意义的模型是 Meta（前 Facebook）提出的 DLRM（Deep Learning Recommendation Model，Google 等公司内部也存在类似变体）。DLRM 的核心出发点在于：早期推荐系统的实践早已证明，稀疏特征之间的两两交互（Pair-wise Interactions）对预测转化至关重要——例如“用户观看了某特定流派的电影”与“用户购买了某款特定商品”之间的组合关联能够提供极强的信号。DLRM 提出的关键架构创新是将各类稀疏特征映射到统一维度的嵌入空间（Embedding），随后对所有特征向量两两计算点积（Dot Product）或特征交叉，这一设计不仅极大地推动了推荐系统的发展，也拓展了深度学习在稀疏数据场景下的能力边界。 - **交叉层数权衡**：实践中 2~3 层交叉层即可捕捉绝大多数有效的高阶特征组合，更深的网络层往往边际效应递减。 - **DLRM 核心机制**：将所有离散稀疏特征映射至相同维度，通过显式的向量点乘捕获两两特征交互，兼顾了表达能力与计算效率。 --- ### [14:22](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=862s) 实体表征的演进：从多模态到预训练模型迁移强大的实体表征能力是同时赋能召回（Retrieval）与排序（Ranking）阶段的关键底座。在 Reddit 的具体业务场景下，核心实体包括帖子（Post）、版块（Subreddit）以及广告（Ad）等。随着计算机视觉与自然语言处理领域的突破，推荐与广告系统逐步引入了这些先进模型来生成高阶表征：从早期使用卷积与视觉模型提取图像特征，到使用 LSTM、GRU 及 Transformer 编码文本，再到如今广泛采用多模态统一模型。例如 OpenAI 的 CLIP 模型能够直接对图像和文本进行多模态联合表征学习。在广告排序系统中使用这些技术的通用路径非常清晰：首先引入业界成熟的开箱即用（Off-the-shelf）预训练大模型，直接提取特征或者基于业务自有数据进行微调（Fine-tuning），随后将生成的实体表征向量注入到底层排序模型中，从而显著提升最终的预估准度。 - **Reddit 核心实体**：Post（帖子）、Subreddit（社区版块）、Ad（广告物料）。 - **表征升级路径**：CNN/LSTM/GRU $\rightarrow$ Transformer 文本编码 $\rightarrow$ 多模态对齐（如 CLIP）。 - **落地通用范式**：借助现成预训练大模型 $\rightarrow$ 业务数据微调或直接推断 $\rightarrow$ 导出高质量向量注入排序模型。 --- ### [15:32](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=932s) 用户行为序列建模与 Reddit 的图神经网络（GNN）除了针对单个静态实体的表征，NLP 与图表征学习（Graph Representation Learning）技术也迅速渗透到了广告排序领域。一方面，Transformer 和 LSTM 不再仅仅用于处理自然语言文本，而是被直接用来编码用户在平台上产生的一连串离散行为序列；另一方面，图表示学习能够利用不同实体类型之间复杂的拓扑关联，挖掘出更深层的协同信息。 Reddit 拥有极其丰富且独特的图结构数据：用户不仅会主动订阅特定的 Subreddit，还经常对他们并未订阅的版块中的帖子进行浏览、点赞或互动。通过将这些异构交互关系构建为图谱并输入给强大的图神经网络编码器（Graph Encoders），系统在统一的监督学习任务下为全图中的每一个实体生成高质量的表征向量，进而全方位赋能下游推荐和广告任务。 - **行为序列化建模**：将用户行为序列视作“句子”，使用 Transformer 捕捉长短期的行为转移模式。 - **异构图结构拓扑**：利用 User-Subreddit 订阅关系以及跨版块的 User-Post 互动行为构建全局图网络。 --- ### [16:40](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1000s) 第四代架构与 Scaling Law：以行为代替文本广告与推荐系统当前正在步入第四代架构演进。Meta 发布的重磅研究论文《Actions Speak Louder than Words》（行动胜于言语）揭示了一个核心趋势：在大语言模型（LLM）中被充分验证的扩展定律（Scaling Laws）同样适用于推荐与广告系统。在该架构范式下，模型的输入不再是传统的文本词元，而是将用户的历史动作（Actions）直接作为输入 Token。Meta 的研究表明，基于生成式/序列式大模型框架训练的架构，不仅在传统的个性化内容推荐任务中表现惊人，在要求极高的商业化广告排序任务中同样展现出了巨大的性能提升。 - **核心研究洞察**：推荐系统同样遵循模型参数量与数据量的 Scaling Law。 - **统一建模视角**：用户交互行为即 Token，以生成式架构统一表征召回与精排目标。 --- ### [17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) 从基础 AI 到广告排序的研究传导链与工程约束前沿 AI 技术在工业界存在一条清晰的传导路径：基础研究成果通常最先诞生于 NLP、CV 和图表示学习领域，随后被个性化推荐系统借鉴并吸收，最终才逐步流动并应用到广告排序系统之中。 ```mermaid flowchart LR accTitle: Research Flow to Ads Ranking subgraph S1["基础 AI 研究"] NLP["自然语言处理 (NLP)"] CV["计算机视觉 (CV)"] GNN["图表征学习 (Graphs)"] end subgraph S2["应用领域"] REC["推荐系统 (RecSys) 用户兴趣建模 / 序列召回"] ADS["广告排序 (Ads Ranking) 极致低延迟 / 单槽位成本敏感"] end NLP --> REC CV --> REC GNN --> REC REC --> ADS ``` 之所以广告排序不会直接从基础 AI 研究中同步落地，而是存在阶段性滞后，主要受制于广告系统极为严苛的物理约束： 1. **极致的低延迟要求**：广告竞价请求必须在几十毫秒的严苛窗口内返回结果，无法直接运行计算开销过大的深层超大模型。 2. **巨大的计算与推理成本**：在信息流（Feed）推荐中，通常一次性完成全页面的内容流生成即可；但对于商业广告而言，页面上出现的每一个广告位（Ad Slot）都需要实时触发独立的复杂特征抽取、候选过滤与打分竞价，其累积的算力成本呈指数级上升。 --- ### [18:07](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1087s) 工业界落地经验：避开“锤子与钉子”陷阱与重构的意外红利在先进模型架构层出不穷的时代，工程师很容易陷入“手里拿着先进架构的锤子，看什么问题都像钉子”的误区。盲目堆砌复杂模型并不可取，深入分析模型在哪些特定数据分段（Data Segments）上表现欠佳（Underperforming），并针对这些薄弱环节定向提出解决方案，往往能带来更为可观且扎实的业务收益。此外，在实际应用机器学习中，训练与推理工程底座（Serving & Training Infrastructure）往往会成为算法迭代的最大瓶颈，必须持续对基础设施进行重投入。最后，分享一条多年沉淀的个人工程心得：定期对特征工程流水线与模型在线服务（Serving）代码进行重构与迁移，常常会带来意想不到的线上正向收益。因为算法研究与模型迭代的速度极快，历史代码中不可避免会累积各种微小、隐蔽的非预期 Bug，而推倒重构正是扫除这些技术负债、修复隐藏缺陷最彻底的手段。 - **警惕唯模型论**：关注模型在不同切片数据上的表现差异，比盲目上复杂模型更有价值。 - **基础设施决定上限**：训练与在线推理工程架构必须持续投资优化，避免成为算法落地的瓶颈。 - **代码重构的红利**：快速迭代期留下的特征与服务 Bug，往往通过代码重写和架构迁移得以根除，从而带来线上指标的意外跃升。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=4s) 本次分享的内容纯属个人在各家公司从事应用 AI（Applied AI）实践的经验总结，并不代表 Reddit 的官方立场。 — Vishal Gupta > [01:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=117s) 在严苛的时间延迟约束内，挑出最合适展示给该用户的广告。 — Vishal Gupta > [01:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=117s) 广告系统最大的区别在于拍卖机制（Auction）——在这里，广告主真金白银地花钱竞价，并拥有直接影响哪个广告能够展现在用户眼前的商业杠杆。 — Vishal Gupta > [03:59](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=239s) 校准（Calibration）是广告排序的生命线——因为预估概率将直接与真金白银的出价相乘，一旦跨多种互动类型的概率输出失准，下游整条竞价链路就会彻底失衡。 — Vishal Gupta > [06:57](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=417s) 在正确的时间，将正确的广告展示给正确的用户（Right ad to the right user at the right time）。 — Vishal Gupta > [11:33](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=693s) 用模型内部的自动化机制彻底取代工程师人工设计特征交叉。 — Vishal Gupta > [13:00](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=780s) 构建 2 到 3 层的特征交叉层通常就已经绰绰有余；同时，并行交叉网络（Parallel Cross Networks）在实际效果和性能表现上普遍优于层叠式交叉网络（Stacked Cross Networks）。 — Vishal Gupta > [16:40](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1000s) 在大语言模型（LLM）中被充分验证的扩展定律（Scaling Laws）同样适用于推荐与广告系统。 — Vishal Gupta > [17:09](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1029s) 基础研究成果通常最先诞生于 NLP、CV 和图表示学习领域，随后被个性化推荐系统借鉴并吸收，最终才逐步流动并应用到广告排序系统之中。 — Vishal Gupta > [18:07](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1087s) 在先进模型架构层出不穷的时代，工程师很容易陷入‘手里拿着先进架构的锤子，看什么问题都像钉子’的误区。 — Vishal Gupta > [18:07](https://www.youtube.com/watch?v=CbTWx11S9wQ&t=1087s) 定期对特征工程流水线与模型在线服务（Serving）代码进行重构与迁移，常常会带来意想不到的线上正向收益。 — Vishal Gupta ## Mentioned Vishal Gupta, Reddit, Google, DeepMind, Twitter / X, Elon Musk, Meta / Facebook, OpenAI, GBDT + LR, Wide & Deep, DCN-v2 (Deep & Cross Network V2), DLRM (Deep Learning Recommendation Model), CLIP, Graph Neural Networks (GNN), Transformer, Actions Speak Louder than Words ## Video description Vishal Gupta, Engineering Manager, Machine Learning, Reddit Inc Abstracts How advanced AI can improve Ads Ranking? AI models have been helping Ads platforms to show relevant Ads to their users since last couple of decades. The advancement in Ads ranking models has accelerated with the advancement in fields like NLP, Computer Vision, Recommender Systems. In this talk you will learn how Ads ranking models have improved by drawing inspiration from Transformers, LLMs, Graph neural networks etc. Lets dive deep into concepts that can improve Ads ranking and other related domains like Recommender Systems, e-commerce etc. [Source](https://www.youtube.com/watch?v=CbTWx11S9wQ)

Source quotes

[0:04] Vishal Gupta: 本次分享的内容纯属个人在各家公司从事应用 AI（Applied AI）实践的经验总结，并不代表 Reddit 的官方立场。 [1:57] Vishal Gupta: 在严苛的时间延迟约束内，挑出最合适展示给该用户的广告。 [1:57] Vishal Gupta: 广告系统最大的区别在于拍卖机制（Auction）——在这里，广告主真金白银地花钱竞价，并拥有直接影响哪个广告能够展现在用户眼前的商业杠杆。 [3:59] Vishal Gupta: 校准（Calibration）是广告排序的生命线——因为预估概率将直接与真金白银的出价相乘，一旦跨多种互动类型的概率输出失准，下游整条竞价链路就会彻底失衡。 [6:57] Vishal Gupta: 在正确的时间，将正确的广告展示给正确的用户（Right ad to the right user at the right time）。 [11:33] Vishal Gupta: 用模型内部的自动化机制彻底取代工程师人工设计特征交叉。 [13:00] Vishal Gupta: 构建 2 到 3 层的特征交叉层通常就已经绰绰有余；同时，并行交叉网络（Parallel Cross Networks）在实际效果和性能表现上普遍优于层叠式交叉网络（Stacked Cross Networks）。 [16:40] Vishal Gupta: 在大语言模型（LLM）中被充分验证的扩展定律（Scaling Laws）同样适用于推荐与广告系统。 [17:09] Vishal Gupta: 基础研究成果通常最先诞生于 NLP、CV 和图表示学习领域，随后被个性化推荐系统借鉴并吸收，最终才逐步流动并应用到广告排序系统之中。 [18:07] Vishal Gupta: 在先进模型架构层出不穷的时代，工程师很容易陷入‘手里拿着先进架构的锤子，看什么问题都像钉子’的误区。 [18:07] Vishal Gupta: 定期对特征工程流水线与模型在线服务（Serving）代码进行重构与迁移，常常会带来意想不到的线上正向收益。

Source quotes

[0:04] Vishal Gupta: kind of fun [7:00] Vishal Gupta: the right ad to the right user at the right time [9:46] Vishal Gupta: memorize [16:40] Vishal Gupta: Actions Speak Louder Than Words [17:09] Vishal Gupta: hammer [17:09] Vishal Gupta: every problem starts looking like nails

Original transcript

[0:04] So today we will be talking about [0:06] advanced AI for ads ranking. Uh this is [0:10] a short talk so has a lot of content. I [0:13] might need to skim through some details. [0:15] So we'll do we'll see how we are doing [0:17] doing on timing and then adjust [0:19] accordingly. [0:22] Um yeah I'll start. [0:25] So first of all disclaimer right like [0:28] please not note that the thoughts that [0:30] I'm sharing do not represent official [0:33] views of my company. What I'll be [0:35] sharing is what I have seen through my [0:38] journey in applied AI space across [0:41] multiple companies. I started working on [0:44] applied AI in Google and then from [0:47] Google I moved on to deep mind in 2018. [0:51] worked on a bunch of applied machine [0:52] learning problems there and in 2022 I [0:57] moved to another company called Twitter [1:00] now X worked there for like a year 3 [1:04] months under Elon that was kind of fun [1:07] and then like [snorts] from Twitter I [1:10] joined Reddit right in like Feb 2023 and [1:14] again I joined Reddit as like a tech [1:16] lead and then I basically moved on to [1:20] management. [1:25] So this is going to be the rough outline [1:28] of the talk. First we will go into like [1:31] you know introduction to ads ranking and [1:34] then we will see how ad ranking models [1:37] evolved across the industry not just [1:40] Reddit. uh this is my own classification [1:44] like first generation to fourth [1:46] generation and then in the end we will [1:48] be sharing like key takeaways. [1:52] So I would start. [1:57] So before we dive into ads ranking, we [2:00] need to understand what does adfunnel [2:03] means, right? Like this is super [2:08] simplified view of ads funnel like more [2:11] than couple of decades went into [2:13] optimizing the ads funnel. So let's [2:16] start with what does entire ad funnels [2:19] do. So given a user request [2:23] and a ad slot that needs to be filled, [2:26] the job of adfunnel is to find the best [2:30] possible ad to show to this user in this [2:33] particular slot. So it typically starts [2:36] with the retrieval stage where there are [2:38] like you know millions or for some ad [2:41] platforms hundreds of millions of ads [2:44] and then there is like ranking which [2:47] when after retrieval basically like you [2:50] know uh or let me put it this way [2:52] retrieval stage takes like you know [2:54] millions of ads and like you know narrow [2:57] it down to like you know a few hundred [2:59] or few thou thousands and then ranking [3:02] stage basically like you predicts bunch [3:05] of probabilities and then in the final [3:08] step like you know ads compete in an [3:12] auction and winner ad is basically [3:15] shown. So I think one thing is really [3:19] important to understand like you know uh [3:22] which is difference between ad auction [3:24] and recommendation system because before [3:27] I started working on ads I worked on [3:29] recommendation systems across like you [3:32] know multiple Google products and when I [3:35] started working on ads it it was very [3:37] evident to me like you know this this [3:39] retrieval and ranking is pretty common [3:42] but the key distinction here is auction [3:44] at the end of the day advertiser that [3:46] are spending money and they can [3:49] influence like you know which ad gets [3:51] shown to the user. [3:56] So [3:59] for the remainder of the talk I will [4:01] mostly focus on like the second stage of [4:03] ranking which of the adfunnel which is [4:06] basically ranking. So heavy ranker [4:09] models are typically like you know [4:12] engagement prediction models. All of [4:15] these models are predicting some [4:18] probability of engagement to be precise [4:22] like these model predict what is the [4:25] probability that user will engage with [4:29] an ad [4:31] given ad user and context. [4:37] There are multiple engagement event [4:39] types. For example, they can be like you [4:41] know firstparty events or like you know [4:44] some people in ads world call them upper [4:46] funnel events like click on an ad or [4:50] video view on a video ad and then there [4:53] are like you know lower funnel events [4:56] like web and mobile conversions, app [4:59] installs etc. and they are typically [5:01] coming from like you know third party [5:03] instead of first party. So now how [5:08] engagement/heavy [5:10] ranker models fit into this auction. So [5:13] typically like you know um estimated [5:16] cost per millie is computed by [5:19] multiplying bid with these probabilities [5:22] of ranking models and then like the [5:26] final utility is computed by taking some [5:29] function of like you know eCPM and how [5:33] relevant is to the user. [5:36] So now let's let's take a pause here [5:39] right [5:43] let's look at the probability right it's [5:45] conditioned on add user and context [5:49] context is really really important guys [5:51] right so here what context mean is time [5:54] of the day surface on which the ad is [5:58] being shown like for example Reddit has [6:01] surfaces like you know main home feed [6:04] there are subreddits there is subreddit [6:06] feed and then there are like you know [6:09] comments page where basically or [6:11] conversations page where like you know [6:13] there is one post and people have done [6:15] bunch of comments. So multiple surfaces [6:19] have like you know their own nuances and [6:23] ranking models need to like you know [6:24] model for all those nuances [6:28] and I think if anyone has worked in ads [6:31] ranking like calibration is super [6:35] important right at the end of the day [6:37] the probability your models are [6:40] predicting will be multiplied by bid [6:43] there are so many engagement types so if [6:46] your probability ilities are not well [6:48] calibrated, you will run into a lot of [6:50] problems. So calibration is super [6:53] important. [6:57] So now like why use AI in ads ranking, [7:00] right? Why can't you we just use simple [7:03] mathematical formulas or simple [7:05] statistical models? So AI powered ad [7:08] ranking platforms a AI powered ad [7:11] ranking models enable platform to show [7:14] right ad to the right user at right [7:17] time. Right time is really important. [7:21] [snorts] Uh typically like you know [7:25] if you go like you know few years back [7:28] uh like most of the ad ranking models [7:31] were based on like you know past user [7:34] behavior and they were trying to like [7:36] you know predict the future user [7:38] behavior. But if user is searching for [7:42] some information which is not at all [7:44] related to like you know their uh their [7:47] past behaviors [7:49] your uh your ad prediction will not work [7:52] out. So context again matters, right [7:55] time again matters and uh yeah so by [8:00] using more powerful models like ad [8:03] platform like you know try to maximize [8:06] both short-term and long-term value for [8:09] both advertisers and users. [8:13] So now we will quickly go through like [8:16] you know evolution of engagement [8:17] prediction models or like you know heavy [8:19] ranker models. [8:23] So so this is this is my own [8:25] classification again. So for for myself [8:28] like you know first generation models [8:29] were like you know pre-dearning models. [8:32] A typical architecture included like you [8:35] know some form of like you know logistic [8:37] regression that was basically doing [8:40] linear combination of bunch of numerical [8:42] and categorical features and people did [8:45] introduce some nonlinearity in those [8:48] models by training like you know [8:50] gradient boosted decision trees or some [8:51] other decision trees. uh a seminal paper [8:55] from like you know Facebook which I have [8:58] linked basically talks in detail about [9:00] these models rough rough roughly these [9:05] GBD models that captured nonlinearity [9:08] were trained periodically like every few [9:11] weeks or every few days and then there [9:14] was this logistic regression component [9:16] that was trained online with the most [9:18] recent data so you can capture the most [9:21] recent trends [9:24] Hundreds or thousands of engineers [9:26] worked in improving these models. It [9:29] required a lot of like you know [9:30] handcrafted features and then like you [9:33] know as ad platform scaled with more [9:36] users with more data like this model [9:39] architecture became a huge bottleneck. [9:46] Now gen two or second generation like [9:48] you know from my point of view was deep [9:52] learning. So various ad platforms [9:55] introduced deep learning at like you [9:57] know different point of times. Companies [9:59] like Google and Meta did it like you [10:02] know maybe as early as 2016 2017 and [10:06] some other ad platforms did it like you [10:09] know later. But again like this diagram [10:13] captures it very well. So if you look at [10:15] the diagram on your extreme left you [10:17] have simple logistic regression models [10:20] and on your extreme right you have like [10:23] you know deep learning models. [10:25] So the entire idea of using deep [10:28] learning was that okay these logistic [10:31] regression models are just memorizing [10:34] what happened in the past. They are not [10:36] able to generalize what will happen in [10:38] the future. let's put like you know deep [10:41] learning and then they can generalize [10:44] then at the same time like you know uh a [10:47] lot of research happening in [10:49] recommendation systems also flew into [10:51] like uh also came into like ads ranking. [10:54] So there was this paper wide and deep by [10:57] Google. So then they combined the ideas [10:59] from both logistic regression and deep [11:02] learning to create architecture that [11:04] just that does both memoization and [11:07] generalization [11:08] and all the second gen models typically [11:12] were trained offline on batch data. [11:15] Still still like a lot of manual work in [11:18] coming up with you know like good [11:20] features. Feature engineering became [11:22] important again but like with mass [11:25] massive amount of data deep learning was [11:27] able to squeeze out a lot of [11:29] performance. [11:33] Let's move on to like gen 3. So then [11:36] again like you know instead of manual [11:38] feature engineering and like you know [11:40] creating this large monolithic DNN [11:43] researchers came with idea that okay [11:45] like you know although DNN's are [11:48] universal function approximators one can [11:51] argue that a single layer DNN with large [11:54] number of neurons can basically learn [11:56] any function but handcrafting DNN [12:00] architectures suitable to particular [12:03] task were very helpful. [12:05] So here I came up with like you know [12:07] non-comprehensive list of various model [12:09] architectures across industry like I [12:12] will not go through each of these but I [12:14] will talk about like you know two of [12:16] them in like you know subsequent slides. [12:22] So one was DC and V2. I think this was [12:26] very powerful architecture. So the [12:29] entire idea behind DC and V2 was instead [12:32] of machine learning engineers to come up [12:34] with to come up with like you know [12:36] various kind of feature crosses can an [12:38] architecture capture all the feature [12:40] crosses by itself. So the magic happens [12:44] in this cross layer and the number of [12:47] cross layers means the number the number [12:51] of cross layers define the number of uh [12:54] feature interactions that happen. So for [12:57] example like you know if you put two [12:58] cross layers you mean that you can learn [13:00] feature crosses up to like you know [13:02] three degrees. [13:04] Personally I have seen like you know you [13:07] you can throw in like a lot more cross [13:08] layers but personally I have seen like [13:10] you know two to three cross layers are [13:12] more than sufficient and like you know [13:15] parallel cross networks are better than [13:17] like you know uh stacked cross networks. [13:22] Then let's look at DLRM. it was [13:25] pioneered by like you know Facebook but [13:27] some flavor of it also existed at [13:29] Google. So the the core idea here is [13:33] like you know feature crossing helps [13:36] okay and from [13:40] and from like you know like previous [13:42] generation recommendation systems it was [13:44] evident that like you know pair-wise [13:46] interactions of sparse features are [13:48] really important. If a user watched a [13:50] movie from a particular genre then and [13:53] user buys this particular product this [13:56] information is really really helpful. So [13:58] the core innovation in this paper was [14:00] that okay there are bunch of like you [14:02] know sparse features let's just embed [14:04] them into like you know same [14:06] dimensionality of embedding and then [14:08] take crosses or take dotproducts of all [14:11] these like you know feature pairs and [14:13] this also pushed the boundaries of you [14:15] know what was possible uh in both [14:17] recommendation systems and deep [14:19] learning. [14:22] Now another important thing that [14:25] happened was like you know powerful [14:27] entity representations. So like you know [14:31] powerful entity representations help [14:33] both in ranking and retrieval. Example [14:36] of entities at Reddit for example [14:38] include like you know post subreddit [14:39] ads. As advancements in NLP and vision [14:44] happened like you know people started [14:46] using those models to generate [14:48] representations. For example, vision [14:50] models helped in embedding like you know [14:52] images. LSTMs, GRUs, transformers helped [14:56] with embedding text. And then recently, [14:59] not so recently actually uh like you [15:02] know transformers are used for [15:04] everything. So for there are transformer [15:06] models for example clip from open AI [15:09] that can be used to generate [15:10] representation which are multimodel for [15:13] both like you know text and images. [15:15] Again, general recipe here was to use [15:18] off-the-shelf pre-trained models, then [15:21] use them directly or fine-tune them on [15:24] your data and then use those [15:26] representations to improve your ranking [15:29] models. [15:32] Now, one more thing happened again NLP [15:35] and graph representation learning [15:37] research like you know flew like into [15:40] ads ranking. So, I will just include [15:42] like couple of quick examples. basically [15:45] like you know LSTM's transformers like [15:49] you know started encoding sequence of [15:51] user actions not just text and then [15:54] graph representation learning started [15:56] leveraging like you know relationship [15:58] information between various entity types [16:01] to like you know learn good [16:03] representations [16:06] I'm running a bit low on time so I will [16:08] like you know quickly skim through this. [16:11] So Reddit has this powerful graph [16:13] structure where users are subscribed to [16:16] subreddits but they can also do actions [16:19] on the post on which they are not [16:21] subscribed to. So we throw this graph [16:25] into like you know powerful graph [16:26] encoders and then supervision happens [16:29] and we get representations for like you [16:32] know all the entities present in the [16:33] graph. If you're interested I can hang [16:35] out after the talk and we can talk about [16:37] this in more detail. [16:40] Then this is the fourth gen guys. Okay. [16:42] Like paper from meta came like actions [16:45] speak louder than word. The core thing [16:48] that this paper showed was like you know [16:50] scaling laws that are applicable to LLMs [16:53] are also applicable to recommendation [16:55] systems. So where instead of text user [16:58] actions are like you know input and meta [17:01] showed that this type of models work [17:03] really well for both recommendation [17:05] tasks and ads ranking tasks. [17:09] Now let's move on to key takeaways. So [17:13] this is a typical flow of like you know [17:15] research like some things happen in NLP [17:19] vision graphs they flow into like you [17:21] know recommendation systems and then [17:25] those research flows from recommendation [17:27] system to ads ranking. Now one might be [17:30] curious that why not directly to ads [17:32] ranking [17:35] like models in ads ranking are not super [17:39] super powerful because they operate [17:41] under strict latency constraints. Number [17:43] one. Number two, like no matter what, [17:48] right? Like the computation cost of like [17:51] you know predicting ad at a particular [17:54] slot is really really high as compared [17:56] to like you know your feed you do the [17:58] feed generation once and that's it. But [18:00] for every ad slot you are like you know [18:03] running these uh powerful models. [18:07] And if one thing I would like to like [18:10] you know uh tell you guys that in the [18:13] world of advanced model architectures we [18:16] know that they add value but like you [18:19] know sometimes I have seen applied ML [18:21] practitioners use them as hammer and [18:24] then every problem starts looking like [18:26] nails. There is still significant value [18:29] in analyzing on what segments of your [18:31] data your models are underperforming and [18:34] then like you know generating ideas that [18:36] can improve models on those segments and [18:39] then again the training and serving [18:41] infrastructure is critical for advancing [18:44] applied ML research. Your infrastructure [18:47] will become your bottleneck. So [18:49] definitely invest in improving that. And [18:51] then this is my personal learning across [18:53] like you know multiple years. Sometimes [18:55] rewriting and migrating the [18:57] serving/feature engineering codebase [19:00] will provide un unexpected online [19:02] improvements because like research moves [19:06] fast MLES move fast sometimes there are [19:09] unintentional bugs. So rewriting get [19:11] rids of those bugs [19:14] and these are references. And yeah thank [19:16] you we are hiring. Please apply on [19:18] careers page directly if these problems [19:20] interest you.