Hayato Kobayashi

Summarization Based on Embedding Distributions

In this study, we consider a summarization method using the document level similarity based on embeddings, or distributed representations of words, where we assume that an embedding of each word can represent its “meaning.” We formalize our task as the problem of maximizing a submodular function defined by the negative summation of the nearest neighbors’ distances on embedding distributions, each of which represents a set of word embeddings in a document. We proved the submodularity of our objective function and that our problem is asymptotically related to the KL-divergence between the probability density functions that correspond to a document and its summary in a continuous space. An experiment using a real dataset demonstrated that our method performed better than the existing method based on sentence-level similarity.

Autonomous Learning of Ball Trapping in the Four-Legged Robot League

This paper describes an autonomous learning method used with real robots in order to acquire ball trapping skills in the four-legged robot league. These skills involve stopping and controlling an oncoming ball and are essential to passing a ball to each other. We first prepare some training equipment and then experiment with only one robot. The robot can use our method to acquire these necessary skills on its own, much in the same way that a human practicing against a wall can learn the proper movements and actions of soccer on his/her own. We also experiment with two robots, and our findings suggest that robots communicating between each other can learn more rapidly than those without any communication.

Transductive Classification on Heterogeneous Information Networks with Edge Betweenness-based Normalization

This paper proposes a novel method for transductive classification on heterogeneous information networks composed of multiple types of vertices. Such networks naturally represent many real-world Web data such as DBLP data (author, paper, and conference). Given a network where some vertices are labeled, the classifier aims to predict labels for the remaining vertices by propagating the labels to the entire network. In the label propagation process, many studies reduce the importance of edges connecting to a high-degree vertex. The assumption is unsatisfactory when reliability of a label of a vertex cannot be implied from its degree. On the basis of our intuition that edges bridging across communities are less trustworthy, we adapt edge betweenness to imply the importance of edges. Since directly applying the conventional edge betweenness is inefficient on heterogeneous networks, we propose two additional refinements. First, the centrality utilizes the fact that networks contain multiple types of vertices. Second, the centrality ignores flows originating from endpoints of considering edges. The experimental results on real-world datasets show our proposed method is more effective than a state-of-the-art method, GNetMine. On average, our method yields 92.79 ± 1.25% accuracy on a DBLP network even if only 1.92% of vertices are labeled. Our simple weighting scheme results in more than 5 percentage points increase in accuracy compared with GNetMine.

On Approximately Searching for Similar Word Embeddings.

We discuss an approximate similarity search for word embeddings, which is an operation to approximately find embeddings close to a given vector. We compared several metric-based search algorithms with hash-, tree-, and graphbased indexing from different aspects. Our experimental results showed that a graph-based indexing exhibits robust performance and additionally provided useful information, e.g., vector normalization achieves an efficient search with cosine similarity.

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.

Two Step graph-based semi-supervised Learning for Online Auction Fraud Detection

We analyze a social graph of online auction users and propose an online auction fraud detection approach. In this paper, fraudsters are those who participate in their own auction in order to drive up the final price. They tend to frequently bid in auctions hosted by fraudulent sellers, who work in the same collusion group. Our graph-based semi-supervised learning approach for online auction fraud detection is based on this social interaction of fraudsters. Auction users and their transactions are represented as a social interaction graph. Given a small set of known fraudsters, our aim was to detect more fraudsters based on the hypothesis that strong edges between fraudsters frequently exist in online auction social graphs. Detecting fraudsters who work in collusion with known fraudsters was our primary goal. We also found that weighted degree centrality is a distinct feature that separates fraudsters and legitimate users. We actively used this fact to detect fraud. To this end, we extended the modified adsorption model by incorporating the weighted degree centrality of nodes. The results, from real world data, show that by integrating the weighted degree centrality to the model can significantly improve accuracy.

Weighted Micro-Clustering: Application to Community Detection in Large-Scale Co-Purchasing Networks with User Attributes

We propose a simple and scalable method for soft community detection that makes use of both graph structures and vertex attributes. Our method is based on micro-clustering, which is a scalable and efficient clique-based method for detecting overlapping communities in unweighted graphs. We extend this method to graphs with vertex attributes so that we can make use of information supplied by vertex attributes. Our method still requires the same time complexity as micro-clustering. We confirm the validity and efficiency of our method by applying it to a large-scale co-purchasing network of real online auction data.

Effects of Game on User Engagement with Spoken Dialogue System

In this study, we examine the effects of using a game for encouraging the use of a spoken dialogue system. As a case study, we developed a word-chain game, called Shiritori in Japanese, and released the game as a module in a Japanese Android/iOS app, Onsei-Assist, which is a Siri-like personal assistant based on a spoken dialogue technology. We analyzed the log after the release and confirmed that the game can increase the number of user utterances. Furthermore, we discovered a positive side effect, in which users who have played the game tend to begin using non-game modules. This suggests that just adding a game module to the system can improve user engagement with an assistant agent.

Perplexity on Reduced Corpora

This paper studies the idea of removing low-frequency words from a corpus, which is a common practice to reduce computational costs, from a theoretical standpoint. Based on the assumption that a corpus follows Zipf’s law, we derive tradeoff formulae of the perplexity of k-gram models and topic models with respect to the size of the reduced vocabulary. In addition, we show an approximate behavior of each formula under certain conditions. We verify the correctness of our theory on synthetic corpora and examine the gap between theory and practice on real corpora.

Development of an Augmented Environment and Autonomous Learning for Quadruped Robots

This paper describes an interactive experimental environment for autonomous soccer robots, which is a soccer field augmented by utilizing camera input and projector output. This environment, in a sense, plays an intermediate role between simulated environments and real environments. We can simulate some parts of real environments, e.g., real objects such as robots or a ball, and reflect simulated data into the real environments, e.g., to visualize the positions on the field, so as to create a situation that allows easy debugging of robot programs. As an application in the augmented environment, we address the learning of goalie strategies on real quadruped robots in penalty kicks. Our robots learn and acquire sophisticated strategies in a fully simulated environment, and then they autonomously adapt to real environments in the augmented environment.