基于貝葉斯網(wǎng)的廣告點擊率預(yù)測方法及實現(xiàn)
發(fā)布時間：2019-07-08 13:12
【摘要】：隨著互聯(lián)網(wǎng)產(chǎn)業(yè)的發(fā)展,互聯(lián)網(wǎng)廣告逐漸成為推動互聯(lián)網(wǎng)產(chǎn)業(yè)健康良性發(fā)展的重要力量,點擊率(CTR, Click-Through Rate)預(yù)測為廣告的精確化投放提供了依據(jù),而且可以提高用戶對所展示的廣告的滿意度,促使用戶點擊自己最感興趣的廣告,不僅提高了廣告主和廣告媒介的收入,而且推動了第三方付費模式的發(fā)展,促進(jìn)了互聯(lián)網(wǎng)產(chǎn)業(yè)的發(fā)展。考慮到廣告投放的精確化和個性化的要求,.需要針對用戶進(jìn)行廣告的精準(zhǔn)化推薦,然而對于沒有歷史記錄的用戶,仍需為其推薦廣告并預(yù)測所推薦廣告的點擊率,是計算廣告領(lǐng)域研究的關(guān)鍵問題之一。 本文以從用戶歷史數(shù)據(jù)中發(fā)現(xiàn)用戶行為的相似性為切入點,基于發(fā)現(xiàn)的用戶間的相似關(guān)系,為缺少數(shù)據(jù)的用戶預(yù)測CTR。由于用戶的行為具有不確定性,需要一種框架來表達(dá)用戶行為中的不確定性,因此本文以貝葉斯網(wǎng)(BN, Bayesian Network)這一重要的概率圖模型作為發(fā)現(xiàn)用戶行為相似性的模型中不確定性知識表達(dá)和推理的基本框架,通過分析用戶歷史數(shù)據(jù)來構(gòu)建貝葉斯網(wǎng),反映用戶間的直接相似關(guān)系及相似關(guān)系的不確定性,進(jìn)而基于貝葉斯網(wǎng)的推理機(jī)制挖掘用戶間的間接相似關(guān)系,從而為沒有歷史點擊記錄的用戶預(yù)測其對廣告的點擊率。 本文的主要工作及貢獻(xiàn)可概括如下： ■為了構(gòu)建反映用戶間在廣告搜索行為方面相似關(guān)系的模型,稱為相似貝葉斯網(wǎng)(SBN, Similarity Bayesian Network),本文針對SBN有向無環(huán)圖(DAG, Directed Acyclic Graph)結(jié)構(gòu)構(gòu)建這一關(guān)鍵和難點,通過對用戶搜索廣告的歷史記錄進(jìn)行統(tǒng)計計算,給出構(gòu)建網(wǎng)絡(luò)結(jié)構(gòu)的方法,進(jìn)而發(fā)現(xiàn)直接相似用戶。 ■利用貝葉斯的概率推理機(jī)制,給出基于Gibbs采樣算法的SBN推理,高效地發(fā)現(xiàn)SBN中的具有間接相似關(guān)系的用戶。進(jìn)而利用用戶間的這種相似關(guān)系,給出預(yù)測CTR的算法。 ■通過建立在KDD Cup2012Track2的訓(xùn)練數(shù)據(jù)集上的實驗,測試了方法的有效性,并設(shè)計開發(fā)了基于本文方法的“基于貝葉斯網(wǎng)的精確化廣告投放仿真軟件”。
[Abstract]:With the development of Internet industry, Internet advertising has gradually become an important force to promote the healthy and healthy development of Internet industry. Click rate (CTR, Click-Through Rate) prediction provides the basis for the accurate delivery of advertising, and can improve the satisfaction of users with the advertisements displayed, and urge users to click on the advertisements they are most interested in, which not only improves the revenue of advertisers and advertising media. And promote the development of third-party payment model, promote the development of the Internet industry. Considering the precision and individualization of advertising,. It is necessary to make accurate recommendation for users, but for users with no history, it is one of the key problems in the field of computational advertising that users still need to recommend advertisements and predict the click rate of recommended advertisements. In this paper, the similarity of user behavior is found from the historical data of users, and the CTR. is predicted for users who lack data based on the similar relationship between discovered users. Because the behavior of users is uncertain, a framework is needed to express the uncertainty in user behavior. Therefore, this paper takes (BN, Bayesian Network), an important probability graph model, as the basic framework of uncertain knowledge representation and reasoning in the model of discovering the similarity of user behavior, and constructs the Bayes network by analyzing the historical data of users, which reflects the direct similarity relationship between users and the uncertainty of similarity relationship. Then, based on the reasoning mechanism of Bayes net, the indirect similarity relationship between users is excavated, and the click rate of advertising is predicted for users who have no historical click record. The main work and contributions of this paper can be summarized as follows: in order to construct a model reflecting the similar relationship between users in advertising search behavior, called similar Bayes net (SBN, Similarity Bayesian Network), this paper aims at the key and difficult point of constructing (DAG, Directed Acyclic Graph) structure of SBN directed acyclic graph, and gives the method of constructing network structure by statistical calculation of the historical records of users' search advertising. Furthermore, it is found that the users are directly similar. By using the probabilistic reasoning mechanism of Bays, the SBN reasoning based on Gibbs sampling algorithm is given, and the users with indirect similarity relationship in SBN are found efficiently. Based on the similarity relationship between users, an algorithm for predicting CTR is given. Through the experiment based on the training data set of KDD Cup2012Track2, the effectiveness of the method is tested, and the "accurate advertising simulation software based on Bayes net" based on this method is designed and developed.
【學(xué)位授予單位】：云南大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2013
【分類號】：TP18;TP393.09
本文編號：2511606
[Abstract]:With the development of Internet industry, Internet advertising has gradually become an important force to promote the healthy and healthy development of Internet industry. Click rate (CTR, Click-Through Rate) prediction provides the basis for the accurate delivery of advertising, and can improve the satisfaction of users with the advertisements displayed, and urge users to click on the advertisements they are most interested in, which not only improves the revenue of advertisers and advertising media. And promote the development of third-party payment model, promote the development of the Internet industry. Considering the precision and individualization of advertising,. It is necessary to make accurate recommendation for users, but for users with no history, it is one of the key problems in the field of computational advertising that users still need to recommend advertisements and predict the click rate of recommended advertisements. In this paper, the similarity of user behavior is found from the historical data of users, and the CTR. is predicted for users who lack data based on the similar relationship between discovered users. Because the behavior of users is uncertain, a framework is needed to express the uncertainty in user behavior. Therefore, this paper takes (BN, Bayesian Network), an important probability graph model, as the basic framework of uncertain knowledge representation and reasoning in the model of discovering the similarity of user behavior, and constructs the Bayes network by analyzing the historical data of users, which reflects the direct similarity relationship between users and the uncertainty of similarity relationship. Then, based on the reasoning mechanism of Bayes net, the indirect similarity relationship between users is excavated, and the click rate of advertising is predicted for users who have no historical click record. The main work and contributions of this paper can be summarized as follows: in order to construct a model reflecting the similar relationship between users in advertising search behavior, called similar Bayes net (SBN, Similarity Bayesian Network), this paper aims at the key and difficult point of constructing (DAG, Directed Acyclic Graph) structure of SBN directed acyclic graph, and gives the method of constructing network structure by statistical calculation of the historical records of users' search advertising. Furthermore, it is found that the users are directly similar. By using the probabilistic reasoning mechanism of Bays, the SBN reasoning based on Gibbs sampling algorithm is given, and the users with indirect similarity relationship in SBN are found efficiently. Based on the similarity relationship between users, an algorithm for predicting CTR is given. Through the experiment based on the training data set of KDD Cup2012Track2, the effectiveness of the method is tested, and the "accurate advertising simulation software based on Bayes net" based on this method is designed and developed.
【學(xué)位授予單位】：云南大學(xué)
【學(xué)位級別】：碩士
【學(xué)位授予年份】：2013
【分類號】：TP18;TP393.09
【參考文獻(xiàn)】
相關(guān)期刊論文 前2條
1 周傲英;周敏奇;宮學(xué)慶;;計算廣告:以數(shù)據(jù)為核心的Web綜合應(yīng)用[J];計算機(jī)學(xué)報;2011年10期
2 張少中;高飛;;一種基于小世界網(wǎng)絡(luò)和貝葉斯網(wǎng)絡(luò)的混合推薦模型[J];小型微型計算機(jī)系統(tǒng);2010年10期
，本文編號：2511606
本文鏈接：http://sikaile.net/wenyilunwen/guanggaoshejilunwen/2511606.html
