广告点击预测
广告系统是广告与用户流量的匹配。
定向，粗排，精排，检索，bidding，新广告，中长尾广告，排期，保量，波动分析，多广告位拍卖，DPA，素材优化，自动化审核，用户择优
转化漏斗：曝光 —> 点击 —> 转化
1. requirements
场景类
We have a bidding server which makes bids and produces logs. Also, we have information about impressions and conversions (usually with some delays). We want to have a model which using this data will predict a probability of click (conversion转化)
What types of ads are we predicting clicks for (display ads, video ads, searched ads, sponsored content)?
Are there specific user segments or contexts we should consider (demographics, location, browsing history)?
Do we have fatigue period (where ad is no longer shown to the users where there is no interest, for X days)?
What type of user-ad interaction data do we have access to can we use it for training our models?
Do we have negative feedback features (such as hide ad, block)?
How do we collect negative samples (not clicked, negative feedback)?
功能类
personalization
diversity, 不能把相似广告放一起
explicit negative feedback, multi-task ranking增加一个head, label是hide block
objective
primary business objective: maximize revenue
How will we define and measure the success of click predictions (click-through rate, conversion rate)?
personalization, diversity, Handle explicit negative feedback
constraint
scale: number of users
latency: 50ms to 100ms
Imbalanced Data: Click events are sparse relative to impressions, requiring techniques to address imbalance.
2. ML task & pipeline
召回(match/retrieval)：流量访问时，从可选的广告库全集中，筛选合适的广告候选子集
排序(rank)：对于给定的广告候选子集，给出相应的预估值。
策略(bidding&strategy)：根据预估值，通过控制广告的出价、排序公式等，影响流量的最终分配
业务过程
advertiser create ads
ads indexing (inverted index, we can use elastic search)
如何减少广告索引的latency，inverted index + db replica + cache
users search for certain keywords
recall
ranking
特点
Imbalance data
3. data collection
Data Sources
Users
demographics
Ads
category
user-ad interaction
user-user friend
Labelling
negative sampling for imbalance
4. feature
It is important for CTR prediction to learn implicit feature interactions behind user click behaviors.
Contextual Features
Time: Time of day, day of the week.
Device: Mobile vs. desktop, OS, browser.
Location: User’s current location.
User-Ad Interaction Features
Collaborative Filtering: Similar users clicked on similar ads.
Real-time Signals: Recent interactions, session data.
Implicit Feedback: Hover time, scroll depth.
5. model
广告算法主流模型（广告算法基本都是point-wise训练方式，因为广告是很少以列表的形式连续呈现）
Logistic regression (feature crossing)
GBDT
GBDT+LR
NN
Deep and Cross Network
FM (FFM)
DeepFM
DIN (DIEN)
DSSM双塔模型
ESSM
Wide and Deep
Multi-task learning
6. evaluation
Offline metrics
Log Loss
ROC-AUC
Online metrics
CTR
Overall revenue (or ROI)
Time Spend
7. deployment & serving
A/B testing
Real-Time Inference
low-latency serving framework (e.g., TensorFlow Serving) to generate predictions within 50-100ms
Cache frequently requested user-ad pairs to reduce latency
8. monitoring & maintenance
Detecting Issues
Performance Degradation: Monitor CTR, revenue, and latency in real-time.
Data Drift: Compare feature distributions (e.g., user demographics) over time.
Continuous Improvement
Retrain models periodically with fresh data.
Incorporate new features (e.g., video ad engagement signals).
Experiment with advanced architectures (e.g., transformer-based models).
9. 优化与问答
bad ads
侧重解决数据来源(人工标注), 以及数据量比较小的问题
LLM fine tune teacher, teacher做bulk inference, distill到student
calibration:
fine-tuning predicted probabilities to align them with actual click probabilities
data leakage:
info from the test or eval dataset influences the training process
target leakage, data contamination (from test to train set)
catastrophic forgetting
model trained on new data loses its ability to perform well on previously learned tasks
gdpr、dma这些rule对广告的影响
uplift: 预测增量值(lift的部分), 预测某种干预对于个体状态或行为的因果效应(识别营销敏感人群)。
reference
精读
扩展
Last updated
