Paper read records

This article is used to record the reading status of the paper during the study period, and give my own understanding of each article. If necessary, I will make comments .... it mainly records my "reading paper career" and will be able to recall my papers read by those elders in the future..."

Article 1st: Question: Analysis of AD click intentions queried by search engine users

Author: Yan Yanqin, Zhang Min, Liu yiqun, and Ma shaoping; Unit: State Key Laboratory of intelligent technology and system, Tsinghua University

This article describes the historical click information of the user query log to predict whether the user intends to click an advertisement in a new query. To improve the effectiveness and accuracy of advertising. For a new query, the model determines that if there is no AD click intention for this query, then there will be less advertising or no advertising (this is assumed not to be advertising ); if this query determines that there is an ad click intention, when the search result is returned,
Advertising. This reduces the blindness of advertising (each query sends an advertisement) and reduces the cost of advertising engines.

First, the user gives historical data to calculate the CTR (ad clicks/total clicks) of the AD and the query ratio of the ad clicks triggered each day;

Secondly, the author puts forward two methods to analyze and predict the ad click intention queried by search engine users, namely the click Prediction Model Based on query word content matching and the prediction model based on Bayesian classification.

Click Prediction Model Based on query word content matching: how can we determine whether a word item can trigger ad clicks? In this article, all the queries in the log are divided into two types: query that triggers ad clicks and query that does not trigger ad clicks. Word items in all queries are ranked based on their frequency in the two types of queries. If the former ranking is higher than (the value is smaller than) in the latter ranking, it has certain ad click intention. whether a single word item has an ad click intention ---------------->
Maps to the advertisement intent of the complete query. finally, given any user's query Q, perform Chinese word segmentation, get a word item set S, and define a ing g, that is, S = segment (q) = {T1, t2 ,..., tn} g (q) = g (T1, T2 ,..., tn); If G (q) is greater than a certain threshold value, it is determined that Q has a tendency to trigger ad clicks. Otherwise, no.

Prediction Model Based on Naive Bayes classification: All queries are classified into two types, excluding queries with AD click intention C1 and AD click intention C2.

Prior probability P (CI): calculate the ratio of queries with and without ad click intentions in all queries respectively. Conditional Probability p (t | CI) of each word item ): you can estimate the frequency of word item t in C1 and C2 queries.

According to the Bayesian formula, we can obtain P (CI | Q)

Assume that the word items in the query are independent: S = segment (q) = {T1, T2,..., tn };

P1 = P (C1 | S) = P (C1) * 1_p (T | C1), P2 = P (C2 | S) = P (C2) * ∏ P (t | C2)

If P1> P2, Q belongs to C1 and does not contain ad click intention. This reduces the number of advertising or even does not serve ads. If P1 <P2, Q belongs to C2, q. I agree to include the intention of clicking an advertisement, and should serve the relevant advertisement.

Summary: This article uses the historical click information to determine whether a new query has an ad click intention. If yes, the ads (CTR and click value) will be put. If no, the ads will be less or no. This article only predicts whether a given query has an ad click intention (binary classification problem ), user information is not used as the feature, and the user's Historical query or browsing behavior is used to determine whether the user has a click intention (with a single feature ), the query user is ignored (assuming that all users are the same ).
Advantage: by predicting whether a query has a click intention, you can determine whether to put an advertisement. This reduces the blindness of advertising and improves the accuracy and effectiveness of advertising. (Note that this is not the precise targeting of advertisements and the amount of information is insufficient). Disadvantages: ① no clear CTR is provided, and the real serving engine still needs to be considered as CTR; ② it only reduces the cost of the advertising engine. In fact, it does not matter to the advertising engine ....