Arvin's Bloghttp://yoursite.com/icon.pngPoem & Algorithm2023-08-19T17:10:19.584Zhttp://yoursite.com/YuyangZhangFTDHexoAn Empirical Study of Selection Bias in Pinterest Ads Retrievalhttp://yoursite.com/2023/08/20/An-Empirical-Study-of-Selection-Bias-in-Pinterest-Ads-Retrieval/2023-08-19T17:09:03.000Z2023-08-19T17:10:19.584Z<h1 id="摘要">摘要</h1>
<blockquote>
<p>Data selection bias has been a long-lasting challenge in the machine learning domain, especially inStreaming CTR Prediction: Rethinking Recommendation Task for Real-World Streaming Datahttp://yoursite.com/2023/08/06/Streaming-CTR-Prediction-Rethinking-Recommendation-Task-for-Real-World-Streaming-Data/2023-08-06T13:21:28.000Z2023-08-06T13:22:51.905Z<h1 id="摘要">摘要</h1>
<blockquote>
<p>The Click-Through Rate (CTR) prediction task is critical in industrial recom- mender systems, whereFresh Content Needs More Attention- Multi-funnel Fresh Content Recommendationhttp://yoursite.com/2023/07/02/Fresh-Content-Needs-More-Attention-Multi-funnel-Fresh-Content-Recommendation/2023-07-02T02:47:37.000Z2023-07-02T02:54:46.960Z<h1 id="摘要"><a href="#摘要" class="headerlink" title="摘要"></a>摘要</h1><blockquote>
<p>Recommendation system serves as a conduit connectingAttention is all you needhttp://yoursite.com/2023/07/01/Attention-is-all-you-need/2023-07-01T10:24:02.000Z2023-07-01T10:28:22.547Z<h1 id="摘要">摘要</h1>
<blockquote>
<p>The dominant sequence transduction models are based on complex recurrent or convolutional neuralOn the Factory Floor: ML Engineering for Industrial-Scale Ads Recommendation Modelshttp://yoursite.com/2022/10/07/On-the-Factory-Floor-ML-Engineering-for-Industrial-Scale-Ads-Recommendation-Models/2022-10-07T13:37:16.000Z2022-10-30T09:18:37.123Z<p>写在最前面,</p>
<p>本来是想简单摘抄一下这篇文章中的精华,写到一半觉得这篇文章不应如此,本文应该是一篇可以比肩Wide&Deep的文章。如果说Wide&Deep告诉业界推荐就是要搞Embedding,E2E,那么本文可能就是告诉大家CTR模型就是要搞Impression Pacing for Jobs Marketplace at LinkedInhttp://yoursite.com/2022/09/15/Impression-Pacing-for-Jobs-Marketplace-at-LinkedIn/2022-09-15T02:51:22.000Z2022-09-18T05:47:39.586Z<p>TL;DR</p>
<p>本文是LinkedIn发表在CIKM 2020上的文章,主要内容是基于曝光行为来做Pacing,控制预算消耗,帮助广告主触达更广泛的人群。</p>
<h1What Makes Forest-Based Heterogeneous Treatment Effect Estimators Work?http://yoursite.com/2022/07/03/What-Makes-Forest-Based-Heterogeneous-Treatment-Effect-Estimators-Work/2022-07-03T04:39:13.000Z2022-07-03T08:42:53.239Z<p>TL;DR,</p>
<p>本文感觉像是一篇大作业文章,应该还不是终稿,有很多typo而且写得不那么易懂,但Stefan Wager是作者之一。主要讨论两个常见的用于估计“异质处理效应”(HTE)的forest模型,causal forest和model-based1 Year in HangZhouhttp://yoursite.com/2022/06/18/1-Year-in-HangZhou/2022-06-18T14:51:20.000Z2022-06-19T01:21:32.359Z<figure>
<img src="1year.jpeg" alt="一年香" /><figcaptionA Unified Solution to Constrained Bidding in Online Display Advertisinghttp://yoursite.com/2022/06/18/A-Unified-Solution-to-Constrained-Bidding-in-Online-Display-Advertising/2022-06-18T03:26:45.000Z2022-06-25T07:40:43.458Z<p>TL;DR</p>
<p>本文提出一种通用的智能出价框架,可以满足不同广告主关于竞价目标和约束的诉求,具体出价调控算法的实现为DDPG,目前该算法已经在淘宝广告平台部署使用。</p>
<h1Ads Allocation in Feed via Constrained Optimizationhttp://yoursite.com/2022/06/10/Ads-Allocation-in-Feed-via-Constrained-Optimization/2022-06-10T14:38:43.000Z2022-06-11T10:55:54.057Z<p>TL;DR</p>
<p>LinkedIn发表在KDD2020上的文章,主要想解决在信息流中,如何将自然推荐物品和广告进行混排的问题,文中将该问题建模为带约束的优化问题,并提出了一种归并排序的方式。</p>
<p>目前这种方法及其衍生版本、简化版本,在日常工作中都非常常见,Bid Optimization by Multivariable Control in Display Advertisinghttp://yoursite.com/2022/06/06/Bid-Optimization-by-Multivariable-Control-in-Display-Advertising/2022-06-06T13:38:19.000Z2022-06-06T14:37:48.289Z<p>TL;DR,</p>
<p>预算约束+效率成本约束下,最优广告出价策略,双PID在线调节超参数(对偶变量)</p>
<h1 id="摘要">摘要</h1>
<blockquote>
<p>Real-Time Bidding (RTB) is an importantOptimized Cost per Click in Taobao Display Advertisinghttp://yoursite.com/2022/05/29/Optimized-Cost-per-Click-in-Taobao-Display-Advertising/2022-05-29T13:07:35.000Z2022-05-29T13:11:43.782Z<p>TL;DR,</p>
<p>很早期的阿里妈妈在广告出价方向的实践,细节比较多,很适合做baseline。</p>
<p>但很多做法以及解法在今天来看都比较过时了,后续会逐渐写近几年怎么做出价问题的。</p>
<h1Smart Pacing for Effective Online Ad Campaign Optimizationhttp://yoursite.com/2022/05/21/Smart-Pacing-for-Effective-Online-Ad-Campaign-Optimization/2022-05-21T07:04:53.000Z2022-05-22T08:42:41.526Z<p>这篇文章是广告预算控制平滑的第二篇经典文章解读,来自Yahoo,KDD2015。</p>
<p>与之前LinkedIn那篇文章类似,这篇文章也是用概率节流的方式进行预算控制,同时这篇文章考虑有无效率保障情况下,如何平滑预算消耗。</p>
<h1Budget Pacing for Targeted Online Advertisements at LinkedInhttp://yoursite.com/2022/05/21/Budget-Pacing-for-Targeted-Online-Advertisements-at-LinkedIn/2022-05-21T04:48:17.000Z2022-06-10T14:30:42.901Z<p>最近恶补了一波广告相关文章,本文是预算控制平滑方面很经典的文章之一,由LinkedIn团队发表在KDD2014上。</p>
<h1 id="摘要">摘要</h1>
<blockquote>
<p>Targeted online advertising is a primeField-aware Calibration- A Simple and Empirically Strong Method for Reliable Probabilistic Predictionshttp://yoursite.com/2022/05/08/Field-aware-Calibration-A-Simple-and-Empirically-Strong-Method-for-Reliable-Probabilistic-Predictions/2022-05-08T05:19:30.000Z2022-06-10T13:40:10.925Z<p>最近要做一些和广告出价相关的工作,恶补了一下广告相关的知识,这篇文章是WWW2020 腾讯的文章,解决广告场景下的概率校准问题。</p>
<h1A Deep Probabilistic Model for Customer Lifetime Value Predictionhttp://yoursite.com/2022/04/26/A-Deep-Probabilistic-Model-for-Customer-Lifetime-Value-Prediction/2022-04-26T11:41:12.000Z2022-04-30T11:11:27.000Z<p>TL;DR</p>
<p>提出一种新的回归Loss来建模用户的长期价值,解决LTV分布并非高斯分布,而是一部分为0和一部分服从log normal的问题。</p>
<h1ate_decompositionhttp://yoursite.com/2022/04/24/ate-decomposition/2022-04-24T10:42:43.000Z2022-04-24T12:22:00.000Z<p>今天在公司内部技术论坛上看到一个帖子,关于用Tree方法做ATE估计的,本着一个严(zhao)谨(cha)的态度,认认真真看了一遍,发现里面一个式子长的比较奇怪,随手推了推感觉还挺有意思的。在这里记一下。</p>
<p>POM框架就不多写了,这篇文章的目的是估计平均因果效应Beyond Customer Lifetime Valuation: Measuring the Value of Acquisition and Retention for Subscription Serviceshttp://yoursite.com/2022/03/18/Beyond-Customer-Lifetime-Valuation-Measuring-the-Value-of-Acquisition-and-Retention-for-Subscription-Services/2022-03-17T16:01:34.000Z2022-03-17T16:05:39.448Z<p>TL;DR</p>
<p>这篇文章核心是在回答,如何评估营销活动期间的拉新、留存和流失召回的增量LTV。</p>
<h1 id="introduction">Introduction</h1>
<blockquote>
<p>Understanding thePersonalized Treatment Selection using Causal Heterogeneityhttp://yoursite.com/2021/11/30/Personalized-Treatment-Selection-using-Causal-Heterogeneity/2021-11-30T15:44:24.000Z2021-12-12T16:13:31.461Z<h1Large-Scale Data-Driven Airline Market Influence Maximizationhttp://yoursite.com/2021/08/21/Large-Scale-Data-Driven-Airline-Market-Influence-Maximization/2021-08-21T15:33:55.000Z2021-08-28T16:05:03.746Z<p>这篇文章是KDD2021的一篇文章,讲的是做航空路线的资源分配的。</p>
<p>我发现自己好像就喜欢看这个类型的文章,偏实际的业务问题+充斥着骚操作的解决方案,对于那种well-defined问题的文章好像看的不是那么多。</p>
<p>最近也被若干老板diss了说写东西