[reading notes] computer advertising MOOC

Time:2020-3-25

By logm

This article was originally published at https://segmentfault.com/u/logm/articles and is not allowed to be reproduced~

This paper is the reading notes of MOOC, a computational advertising science. MOOC talks about it in a macro way and does not involve too many specific technologies


1. Basic knowledge of advertising

1.1 purpose of advertisement

  • Brand advertising: memory
  • Effect advertising: Transformation

1.2 effectiveness model of advertisement

  • Choice:

    • Exposure: advertising space
    • Attention: do not interrupt users’ tasks, clearly reveal the reasons for recommendation, and meet users’ interests and needs
  • Interpretation:

    • Understanding: the understanding threshold related to the degree of attention that users can understand (long time to understand TV shopping, short time to understand Internet shopping)
    • Information acceptance: recognition of advertisers and advertising space (media)
  • Attitude:

    • Keeping: the memory effect of artistry
    • Purchase: within the price sensitive acceptance range of users

1.3 differences between advertising and marketing

  • The advertising goal is to reach potential users, and the sales goal is to reach users with clear needs

1.4 uniqueness of online advertising

  • Technology and computing orientation
  • Measurability: Click to measure the advertising effect, but not exactly
  • Standardization: different people have different advertisements, so the size of advertisements should be standardized
  • “Media” concept differentiation: portal, search engine and Taobao belong to different “media”, which are in different stages from supply to demand

1.5 online advertising market

1.6 core issues and challenges of computing advertising

  • Large scale: millions of pages, billions of users; high concurrency; delay requirements
  • Dynamic: changes in user interest
  • Rich query information
  • Exploration and discovery: actively launch some exploration advertisements to obtain data

1.7 comparison of advertising, search and recommendation

  • Search: relevance; vertical domain; billion level; less personalized demand; centralized retrieval signal
  • Search advertising: ROI; quality and safety; million to ten million levels; less personalized needs; more centralized search signals
  • Display ads: ROI; quality and safety; millions of users; personalization of billions of users; rich retrieval signals
  • Recommendation: user interest; diversity and freshness; millions to billions of users; personalization of billions of users; rich retrieval signals; downstream optimization (optimization for a whole set of clicks, rather than just one click)

1.8 ROI analysis

  • Input = purchase price
  • Return = CTR, click through rate * click value = ECPM (expected CPM, pay per thousand presentations)
  • CPM Market: ECPM fixed, common in brand advertising
  • CPC (pay per click) market: dynamic CTR, fixed click value
  • CPA (pay per action) / CPS (pay per sale) / ROI Market: dynamic CTR, dynamic click value; only applicable to special scenarios

1.9 structure of online advertising system

  • Query – > recall – > sort
  • Audience oriented platform (offline): Data Mining
  • Ad server (online): high concurrency; 10 mm real-time decision-making; 10 billion times / day
  • Data highway: internal and external TB level data real-time collection and processing, collection of online end to offline end
  • Streaming computing platform: quasi real time mining and feedback of logs, anti cheating and pricing

2. Contract advertising system

2.1 common open source tools of advertising system

2.2 contract advertisement introduction

Media (supply) guarantees the display or conversion of advertisements

  • Direct media buying: for example, buying a column ad from a newspaper; this is to introduce the following two contract ads
  • Guaranteed delivery (GD): Based on the contract, compensation shall be made for failing to reach the display quantity; quantity takes precedence over quality; CPM billing
  • Ad server: audience orientation, CTR prediction, traffic prediction

2.3 online distribution

  • adwords problem -> display problem

2.4 introduction to Hadoop

3. Audience orientation

3.1 concept of audience orientation

  • Label a (AD) U (user) C (context)

    • Redirection: if a user has visited an advertiser’s website before, tag it
    • Geographical / demographic attributes (men, women, income)
    • context
    • Behavior orientation: including inside and outside the station
    • Website / Channel: finance, automobile, etc
    • Hyper local: very fine regional orientation, fine to a building (cafe, etc.)
    • Look like: mining potential users similar to historical consumers

3.2 behavior orientation

  • Importance sorting: transaction (direct transaction behavior), pre transaction (such as commodity browsing), paid search click (click of search advertisement), ad Click (click of general advertisement), search click, search, share, page view, ad view

3.3 context orientation

  • Semi online tagging

3.4 theme model

  • PLSI
  • LDA
  • GaP

3.5 data processing and trading

4. Competitive advertising system

4.1 location auction theory

  • VCG Theory: the charge of an object should be equal to the value damage to others; in practice, it is difficult to calculate
  • Generalized second high price: higher price than the next one; more fees will be charged to advertisers compared with VCG mechanism; online advertising is widely used

4.2 advertising network concept

4.3 advertisement retrieval (recall)

  • Boolean representation retrieval: two layers inverted
  • Long query: wand search algorithm

4.4 flow forecast

  • Consider a as query, recall (U, c); estimate the traffic of an advertiser

4.5 introduction to zookeeper

  • Distributed synchronization service

4.6 click through rate prediction and logistic regression

  • Regression is better than ranking, because CTR should be estimated
  • Cold boot
  • Dynamic features: quickly adjusting features; online learning: quickly adjusting models

4.7 introduction to logistic regression method

  • BFGS
  • ADMM

4.8 dynamic characteristics

5. Search advertising and advertising network

5.1 exploration and utilization

  • Create display opportunities for long tail advertising
  • Basic method: small flow random exploration

    • UCB: calculate the expected revenue of each advertisement through the previous observation values, select the advertisement with the largest expected revenue to launch, and use the observation results to adjust the next calculation; with the increase of observation times, it will gradually converge

5.2 search advertising

  • Features: weak user orientation and strong context (short-term user behavior)
  • Query term extension: Based on recommendation (vector similarity), semantic (subject model), historical data statistics

5.3 flow computing platform

5.4 advertising purchase platform

  • Help advertisers buy media and optimize ROI

6. Advertising market

6.1 advertising market

6.2 real time bidding

  • ADX (ad exchange) inquires for price from DSP in real time

6.3 Cookie Mapping

6.4 SSP

  • Supply side platform

6.5 DSP

  • Demand side platform

6.6 DSP traffic prediction

  • It is very important to only use historical bidding information, but there is no better solution

6.7 DSP click value estimation

  • Relatively difficult: the data is sparse, and the click value of different types of advertisements varies greatly (e-commerce, games)

6.8 DSP redirection

  • For example: after purchasing or browsing a product on Taobao, Taobao ads will appear on other websites

6.9 recommended methods at the demand side

6.10 advertising traffic transaction mode

Recommended Today

On the reference count in PHP string type

Author: Wang Shu Background introduction String type is also a commonly used type. Due to the characteristics of strings, in order to save memory, the same string variables usually share a block of memory space. By reference counting, multiple variables are marked to use this memory. However, after GDB tracking, it is found that not […]