Loading...
| Friend's email | |
| Your name | |
| Your email | |
| enter code | |
This page was sent successfuly
- Type of Document: M.Sc. Thesis
- Language: Farsi
- Document No: 54731 (19)
- University: Sharif University of Technology
- Department: Computer Engineering
- Advisor(s): Rabiee, Hamid Reza
- Abstract:
- Online advertisements have captured a major part of websites and mobile applications nowadays. In this kind of advertisements, as soon as a user interacts with a website or a mobile application, we have to decide which ad to show in a fraction of a second. In online advertising systems, revenue increases only when a user clicks on an ad banner or when they interact with the ad. So in order to select the best (most profitable) ad banner, the probability of clicking on or interacting with the ad banner must be estimated and then the most profitable ad will be selected.So one of the most essential problems in online display advertising, is estimating the probability of the click on a certain ad, by a certain user. This problem is well regarded in researches as Click Through Rate prediction (CTR prediction). An accurate estimation of CTR would increase the relevance of the shown ads to the user in one hand and increase the revenue of the online ad systems on the other.Previous researches in CTR prediction transform the problem into binary classification (click or not click) and try to estimate the clicking probability using history data in three domains: user side, advertiser side and publisher side.Class imbalance, data sparsity, high dimensionality and cold start are the challenges which distinct this problem from classical machine learning problems. Current methods can be divided into two categories of shallow methods and deep methods. The simplicity and speed of shallow models often gives them an advantage when it comes to implement them in an enterprise display advertising system.In this study, we take a close look at the problem, the mentioned challenges and then we propose a new method to solve this problem. To design our proposed method, we adopt and introduce a diverse set of ideas. We design our proposed method regarding the aforementioned challenges and therefore we expect our method can outperform previous methods in most realistic scenarios.We evaluate our method using Intersection Over Union (IOU), Area Under ROC curve (AUC) and also classical measures like Precision and Recall. We then report these metrics on two benchmark datasets. We conclude our proposed method has achieved acceptable results and is ready to be tested on online scenarios
- Keywords:
- Display Advertisting ; Interaction Between Features ; Embedding Vectors ; Interaction Probability ; Users Interaction Probability
Digital Object List
محتواي کتاب
Bookmark