Method of Accurately Predicting a ClickThrough Rate (CTR) Using a Topic Based Co-Relevance Model (TCRM)
Publication Date: 2010-Oct-22
The IP.com Prior Art Database
Yu Zou: INVENTOR [+5]
AbstractA method of accurately predicting a ClickThrough Rate (CTR) using a topic based co-relevance model (TCRM) is disclosed.
A method of accurately predicting a ClickThrough Rate (CTR) using a topic based co-relevance model (TCRM) is disclosed.
Disclosed is a method of accurately predicting a ClickThrough Rate (CTR) using a topic based co-relevance model (TCRM). The TCRM includes a latent topic model to map documents including both pages and advertisements into a latent semantic space and a co-relevance model to measure the co-relevancy of a topic in a webpage and a topic associated with advertisement candidates.
The latent topic model models documents with a finite mixture model of k topics and estimates the model parameters by fitting the data with the finite mixture model. Here, a Probabilistic Latent Semantic Analysis (PLSA) model is used to deduce probability distribution of topics for documents including both web pages and advertisement. The PLSA model projects a webpage and the corresponding advertisement into the same topic space, wherein each webpage and advertisement is a vector in the probability topic space. Alternatively, the topic space of the webpage and the topic space of the advertisements may be trained separately.
The co-relevance model is a matrix of CTRs of all (page_topic, ad_topic) pairs, which describes the probability of a user clicking on one or more advertisements corresponding to a topic after the user has read pages about another topic. Additionally, the topic co-relevance model may be trained using machine learning methods from click or impression data.
To predict a CTR, the CTR of a given (page, ad) pair is re-composed in the cross topic space by the following formulae:
if written in matrix form, where
In this formula, the corresponding CTR on topic level is used as CTR of <page, ad> pair, because page and ad are children of topics in the concept hierarchy and children inherit statistical data of a parent according to s...