Yukihiro Tagami

CTR prediction for contextual advertising: learning-to-rank approach

Contextual advertising is a textual advertising displayed within the content of a generic web page. Predicting the probability that users will click on ads plays a crucial role in contextual advertising because it influences ranking, filtering, placement, and pricing of ads. In this paper, we introduce a click-through rate prediction algorithm based on the learning-to-rank approach. Focusing on the fact that some of the past click data are noisy and ads are ranked as lists, we build a ranking model by using partial click logs and then a regression model on it. We evaluated this approach offline on a data set based on logs from an ad network. Our method is observed to achieve better results than other baselines in our three metrics.

Modeling User Activities on the Web using Paragraph Vector

Modeling user activities on the Web is a key problem for various Web services, such as news article recommendation and ad click prediction. In this paper, we propose an approach that summarizes each sequence of user activities using the Paragraph Vector, considering users and activities as paragraphs and words, respectively. The learned user representations are used among the user-related prediction tasks in common. We evaluate this approach on two data sets based on logs from Web services of Yahoo! JAPAN. Experimental results demonstrate the effectiveness of our proposed methods.

Learning Extreme Multi-label Tree-classifier via Nearest Neighbor Graph Partitioning

Web scale classification problems, such as Web page tagging and E-commerce product recommendation, are typically regarded as multi-label classification with an extremely large number of labels. In this paper, we propose GPT, which is a novel tree-based approach for extreme multi-label learning. GPT recursively splits a feature space with a hyperplane at each internal node, considering approximate k-nearest neighbor graph on the label space. We learn the linear binary classifiers using a simple optimization procedure. We conducted evaluations on several large-scale real-world data sets and compared our proposed method with recent state-of-the-art methods. Experimental results demonstrate the effectiveness of our proposed method.

Translation method of contextual information into textual space of advertisements

Contextual advertising has a key problem to determine how to select the ads that are relevant to the page content and/or the user information. We introduce a translation method that learns a mapping of contextual information to the textual features of ads by using past click data. This method is easy to implement and there is no need to modify an ordinary ad retrieval system because the contextual feature vector is simply transformed into a term vector with the learned matrix. We applied our approach with a real ad serving system and compared the online performance in A/B testing.

Multimodal Content-Aware Image Thumbnailing

News article recommendation has the key problem of needing to eliminate the redundant information in a ranked list in order to provide more relevant information within a limited time and space. In this study, we tackle this problem by using image thumbnailing, which can be regarded as the summarization of news images. We propose a multimodal image thumbnailing method considering news text as well as images themselves. We evaluate this approach on a real data set based on news articles that appeared on Yahoo! JAPAN. Experimental results demonstrate the effectiveness of our proposed method.

Article De-duplication Using Distributed Representations

In news recommendation systems, eliminating redundant information is important as well as providing interesting articles for users. We propose a method that quantifies the similarity of articles based on their distributed representation, learned with the category information as weak supervision. This method is useful for evaluation under tight time constraints, since it only requires low-dimensional inner product calculation for estimating similarities. The experimental results from human evaluation and online performance in A/B testing suggest the effectiveness of our proposed method, especially for quantifying middle-level similarities. Currently, this method is used on Yahoo!\ JAPANs front page, which has millions of users per day and billions of page views per month.

Filling context-ad vocabulary gaps with click logs

Contextual advertising is a form of textual advertising usually displayed on third party Web pages. One of the main problems with contextual advertising is determining how to select ads that are relevant to the page content and/or the user information in order to achieve both effective advertising and a positive user experience. Typically, the relevance of an ad to page content is indicated by a tf-idf score that measures the word overlap between the page and the ad content, so this problem is transformed into a similarity search in a vector space. However, such an approach is not useful if the vocabulary used on the page is expected to be different from that in the ad. There have been studies proposing the use of semantic categories or hidden classes to overcome this problem. With these approaches it is necessary to expand the ad retrieval system or build new index to handle the categories or classes, and it is not always easy to maintain the number of categories and classes required for business needs. In this work, we propose a translation method that learns the mapping of the contextual information to the textual features of ads by using past click data. The contextual information includes the users demographic information and behavioral information as well as page content information. The proposed method is able to retrieve more preferable ads while maintaining the sparsity of the inverted index and the performance of the ad retrieval system. In addition, it is easy to implement and there is no need to modify an existing ad retrieval system. We evaluated this approach offline on a data set based on logs from an ad network. Our method achieved better results than existing methods. We also applied our approach with a real ad serving system and compared the online performance using A/B testing. Our approach achieved an improvement over the existing production system.