Research and Development

The Z Holdings Group aims to contribute to the development of information technology society through the promotion of R&D and open innovation including the development of open source software (OSS).

R&D in the Z Holdings Group Companies

Yahoo! JAPAN Research

In line with Yahoo Japan Corporation‘s (“Yahoo! JAPAN”) mission statement, “UPDATE JAPAN” and its aspiration to make Japan more convenient with cutting-edge technologies, Yahoo! JAPAN Research works on the research and development of the next-generation of Internet technologies.
Today, various devices are used in different everyday life situations, and a variety of data is collected from them. By further deepening its understanding of individuals and trends in the world, Yahoo! JAPAN Research works to create value that will solve the problems of the future.
Yahoo! JAPAN Research broadly collaborates with service divisions of Yahoo! JAPAN, and numerous universities and research institutes, in order to achieve maximum results from the valuable environment in which enormous amount of data and feedback from the customers using Yahoo! JAPAN’s services are collected every day. Yahoo! JAPAN Research also hosts an internship program for students in the hope of providing them with firsthand experience that will broaden and further their future research and development activities.
In order to contribute to the development of the information technology society, some of the outcomes of these research and development efforts are made available to the public in the form of open source software (OSS), etc. and are utilized in various situations.

LINE Corporation (“LINE”)

LINE positions AI business as one of its strategic businesses and is actively pursuing business expansion in the area. By making the most of its massive cross-sectoral data, the company creates new AI-related services and features. With a view to further enhancing the user experience, it also promotes R&D to deliver new user experience that will form “daily life standards of tomorrow” in the digital society. The focus R&D areas include recognition/generation technologies for media content, such as speech processing, language processing, and image processing, and the company works to realize human-friendly AI using machine learning.
In addition, LINE has a collaborative framework for conducting joint research with numerous universities in Japan. In April 2018, the company established a joint research unit in the National Institute of Informatics (NII), which houses Center for Robust Intelligence and Social Technology as the joint research facility. Going ahead, LINE aims to expand the collaboration network to research institutions overseas.

ZOZO Research – ZOZO’s R&D wing

ZOZO Technologies, Inc., a member of the ZOZO Group, operates an R&D team named “ZOZO Research.” The mission of the team is to “quantify fashion.” Leveraging the information asset possessed by the ZOZO Group, it aims to present a scientific answer to the fascinating and puzzling question of fashion: “What does it mean to be ‘beautiful,’ ‘cool,’ and ‘cute’?”

Case in Which Research Led to New Functions in Our Service

Yahoo! JAPAN

Certified with “FIDO 2,” a standard that enables safe login through use of biometrics authentication devices, etc.

Yahoo! JAPAN is certified with “FIDO2,” a new standard established by an industry association called FIDO Alliance, whose mission is to promote the standardization of simple and robust authentications. The company acquired this certification in the world’s first FIDO2 certification test held in August 2018, and it is the only FIDO2-certified company in Japan. From October 2018, users of Android devices can log into Yahoo! JAPAN’s services on the Google Chrome browser by using biometric authentication, such as fingerprint authentication, f, and iOS device users on the Safari browser from December 2020. This became the world’s first ever case of applying FIDO2-certified authentication to commercial services for consumers.

Auto-generation of candidate topic headlines for Yahoo! JAPAN News

Yahoo! JAPAN developed a model to auto-generate candidate topic headlines for Yahoo! JAPAN News and introduced it into the editing assistance tool.

Congestion Prediction Feature (Yahoo! JAPAN MAP, Yahoo! JAPAN Search, Yahoo! JAPAN Car Navigation Apps)

The results of research that uses location history data to "predict" population anomalies several days in advance have been implemented as a function to seek "usual congestion" in various applications.

Spelling Correction for Yahoo! JAPAN Search Queries

Using a large amount of search history data, Yahoo! JAPAN developed a machine learning model that auto-corrects spelling errors in search queries.

Judgment of Positive/Negative Tweets for Yahoo! JAPAN Realtime Search

In the renewal of the positive/negative judgment of tweets in Yahoo! JAPAN Realtime Search, Yahoo! JAPAN introduced an efficient machine learning approach to achieve a highly accurate judgment feature.

ZOZO, Inc. (“ZOZO”)

Application to Workforce Management in Customer Support

ZOZO Research has been working on how to achieve more efficiency in the back-end operations with various technologies. One specific example is workforce management (WFM) implemented in the ZOZOTOWN customer support.* The task allocation problem required in WFM is formulated as a mixed integer optimization problem, a kind of mathematical optimization problem, and the optimal task allocation is calculated.

  • * A term used to describe overall efforts to achieve more efficient and higher performance through the appropriate allocation of human resources.
Release of an AI-Powered Feature for Fashion Styling App “WEAR” That Enables Outfit Idea Searches Based on Hairstyles

The feature enables users to find outfit ideas based on 14 hairstyle categories, such as long hair, medium-length hair, and bobs. The feature employs ZOZO Research’s image analysis technology that makes use of AI to extract the face from the image and judge the hairstyle. Although hairstyle recognition with AI was difficult in the past depending on the direction of the face, the technology resolved the problem by taking the face angle into account and succeeded in achieving high accuracy for the feature.

Active Communication of Research Results

Yahoo! JAPAN

Yahoo! JAPAN Research actively presents research results at world-leading international conferences etc. (※)

Communication of research results FY2013 DOMESTIC:7, INTERNATIONL:4, FY2014 DOMESTIC:17, INTERNATIONL:15, FY2015 DOMESTIC:21, INTERNATIONL:11, FY2016 DOMESTIC:32, INTERNATIONL:28, FY2017 DOMESTIC:25, INTERNATIONL:34, FY2018 DOMESTIC:37, INTERNATIONL:35, FY2019 DOMESTIC:38, INTERNATIONL:53, FY2020 DOMESTIC:35, INTERNATIONL:67
Communication of Research Results
  • *Major conferences in 2020
  • ・AAAI(Association for the Advancement of Artificial Intelligence): World-leading international conference on artificial intelligence
  • ・PerCom (Annual IEEE International Conference on Pervasive Computing and Communications): World-leading international conference on ubiquitous computing
  • ・KDD(International Conference on Knowledge Discovery and Data Mining):Top conference in the fields of machine learning/data mining
  • ・UIST (ACM Symposium on User Interface Software and Technology): Premier forum for innovations in user interfaces

LINE

LINE has been promoting presentation of papers to top-level conferences since 2019. The number of peer-reviewed papers adopted in international conferences amounted to six in 2019 and 25 in 2020. At the International Conference on Acoustics, Speech, & Signal Processing (ICASSP), the premier conference in the field of speech and audio signal processing, 11 papers and seven papers from LINE were adopted in 2020 and 2021, respectively. This is the top-class level in Japan. In addition, LINE participated and topped the world in Task 4 of DCASE2020 Challenge, the largest workshop in the field of ambient sound recognition.

ZOZO

ZOZO Research presented two papers in Japan and overseas in FY2020.
・ECCV2020
・ICML2020

Research Papers by Yahoo! JAPAN Research That Led to Patent Acquisitions

Numerous patents have been obtained from research papers published by Yahoo! JAPAN Research. Please refer to the following;

Year Title of research paper, name of OSS* Classification
2020 ScraTouch: Extending Interaction Technique Using Fingernail on Unmodified Capacitive Touch Surfaces INTERNATIONAL
2020 LightTouch: Passive Gadgets for Extending Interactions on Capacitive Touchscreens by Automating Touch Inputs INTERNATIONAL
2020 A Simple Text-based Relevant Location Prediction Method using Knowledge Base INTERNATIONAL
2020 Diverse and Non-redundant Answer Set Extraction on Community QA based on DPPs INTERNATIONAL
2020 Context-Free TextSpotter for Real-Time and Mobile End-to-End Text Detection and Recognition (external site) INTERNATIONAL
2020 PredicTaps: Latency Reduction Technique for Single-taps Based on Recognition for Single-tap or Double-tap INTERNATIONAL
2020 ScraTouch: Extending Touch Interaction Technique Using Fingernail on Capacitive Touch Surfaces INTERNATIONAL
2020 Extracting interpersonal relationships from Wi-Fi Connection Data (Japanese only) DOMESTIC
2020 Association Analysis of Pet Foods with Diseases in dogs Using E-Commerce Data DOMESTIC
2020 Estimation of Relevant Locations from News Articles using Knowledge Base DOMESTIC
2020 News Headline Generation Based on Reinforcement Learning Considering Human Evaluation (external site) DOMESTIC
2020 Click-Through Rate Prediction with Confidence for News Headlines Using Natural Gradient Boosting (external site) (Japanese only) DOMESTIC
2020 Parasitic Location Logging: Estimating Users’ Location from Context of Passersby INTERNATIONAL
2020 SheetKey: Generating Touch Events by a Pattern Printed with Conductive Ink for User Authentication INTERNATIONAL
2020 Fast Convergence Algorithm for Adaptive Noise Cancellers with SNR-Based Stepsize Control INTERNATIONAL
2020 Fast Start-Up Algorithm for Adaptive Noise Cancellers with Novel SNR Estimation and Stepsize Control INTERNATIONAL
2020 Yosegi: Columnar format for efficient nested data processing by schema conversion INTERNATIONAL
Paper before 2019
2019 Search parameter adjustment
2019 Performance measurement for the specified accuracy range
2019 Time management for building a graph-based index
2019 PredicTaps: Latency Reduction Technique for Single-taps Based on Recognition for Single-tap or Double-tap INTERNATIONAL
2019 ScraTouch: Extending Touch Interaction Technique Using Fingernail on Capacitive Touch Surfaces INTERNATIONAL
2019 Weakly Supervised Multilingual Causality Extraction from Wikipedia DOMESTIC
2019 Real-World Product Deployment of Adaptive Push Notification Scheduling on Smartphones INTERNATIONAL
2019 Simultaneous Detection and Localization of a Wake-Up Word using Multi-Task Learning of the Duration and Endpoint INTERNATIONAL
2019 Parasitic Location Logging: Estimating Users’ Location from Context of Passersby DOMESTIC
2019 Context-Aware Authentication Using Co-Located Devices INTERNATIONAL
2019 incremental skip-gram model with negative sampling
2019 A Case Study on Neural Headline Generation for Editing Support INTERNATIONAL
2019 Dataset Creation for Ranking Constructive News Comments INTERNATIONAL
2018 Application of Neural Title Generation as News Editing Assistant (Japanese only) DOMESTIC
2018 SPEAKER SELECTIVE BEAMFORMER WITH KEYWORD MASK ESTIMATION INTERNATIONAL
2018 Optimization of Indexing Based on k-Nearest Neighbor Graph for Proximity Search in High-dimensional Data INTERNATIONAL
2018 Symbiotic Construction of Individual’s Rich Location Dataset INTERNATIONAL
2018 FIDO Authentication and Its Technology (Japanese only) INTERNATIONAL
2018 Robust ASR against background speech using DNN-mask specializes in wakeup word (Japanese only) DOMESTIC
2018 Continuous Authentication System Using Online Activities INTERNATIONAL
2018 Pretraining Sentiment Classifiers with Unlabeled Dialog Data INTERNATIONAL
2018 Semi-supervised Sentiment Classification with Dialog Data DOMESTIC
2018 Steering through Successive Objects INTERNATIONAL
2017 Identity Verification Using Personal Device’s Context Information(Japanese only) DOMESTIC
2017 Towards authentication using multi-modal online activities INTERNATIONAL
2017 Embedding-based News Recommendation for Millions of Users INTERNATIONAL
2017 AnnexML: Approximate Nearest Neighbor Search for Extreme Multi-label Classification INTERNATIONAL
2017 Extreme Multi-label Learning via k-nearest Neighbor Graph Partitioning(Japanese only) DOMESTIC
2017 Learning Extreme Multi-label Tree-classifier via Nearest Neighbor Graph Partitioning INTERNATIONAL
2016 Quantifying Query Ambiguity with Topic Distributions INTERNATIONAL
2016 Pruned Bi-directed K-nearest Neighbor Graph for Proximity Search INTERNATIONAL
2016 big3store INTERNATIONAL
2016 Ensure User Identity by Using Biometrics(Japanese only) DOMESTIC
2016 Multimodal Content-Aware Image Thumbnailing INTERNATIONAL
2016 Quantifying Query Ambiguity with Topic Distributions(Japanese only) DOMESTIC
2015 The Optimization of Budget Consumption for Internet Targeting Ads.(Japanese only) DOMESTIC
2015 The Agitation Detection in Stock Bulletin Board(Japanese only) DOMESTIC
2014 A Distributed Query Execution Method for RDF Storage Managers INTERNATIONAL
2014 Filling Context-Ad Vocabulary Gaps with Click Logs INTERNATIONAL
2014 Translation Method of Contextual Information into Textual Space of Advertisements INTERNATIONAL
2013 Applying a Graph-Structured Index to Product Image Search(Japanese only) DOMESTIC
2013 Extendable Dialog Script Description Language for Natural Language User Interfaces INTERNATIONAL
2013 Primitive Operation Aggregation Algorithms for Improving Taxonomies for Large-Scale Hierarchical Classifiers INTERNATIONAL
2012 Toward personal experience management in a socially networked world INTERNATIONAL
2012 Organization and exploration of heterogeneous personal data collected in daily life INTERNATIONAL
2012 Policy Provisioning and Its Access Control Beyond Administrative and Collaborative Domains INTERNATIONAL
2012 Authentication Trust Metric and Assessment for Federated Identity Management Systems INTERNATIONAL
2011 Access Control Model and Design for Delegation Using Authorization Tokens INTERNATIONAL
2011 Establishing Authentication Trust in Open Environment Using Social Approach INTERNATIONAL
2011 Dynamic Identity Delegation Using Access Tokens in Federated Environments INTERNATIONAL
2011 Data and Access Management Using Access Tokens for Delegating Authority to Persons and Software INTERNATIONAL
2011 Proximity Search using Approximate K Nearest Neighbor Graph with a Tree Structured Index(Japanese only) DOMESTIC
2010 Policy Provisioning for Distributed Identity Management Systems INTERNATIONAL
2010 Improving Taxonomies for Large-Scale Hierarchical Classifiers of Web Documents INTERNATIONAL
2010 Tag Clouds in the Cloud: A Visual Interface for the Display of Tag Clusters(Japanese only) DOMESTIC
2010 An Authentication Trust Metric for Federated Identity Management Systems INTERNATIONAL
2010 A Persistent Data Tracking Mechanism for User-centric Identity Governance INTERNATIONAL
2010 Doughnut Crumbs:Visual Navigation for Data Hierarchies INTERNATIONAL
2010 Proximity Search in Metric Spaces Using Approximate K Nearest Neighbor Graph(Japanese only) DOMESTIC
2009 User-Centric Identity Governance across Domain Boundaries INTERNATIONAL
2009 Identity Information Exchange and Control with Access Token(Japanese only) DOMESTIC
2009 Web contents navigation guided by surrounding information and navigation route(Japanese only) DOMESTIC
2008 Building an Autopoietic Knowledge Structure for Natural Language Conversational Agents INTERNATIONAL
2008 Determination of importance of phrases included in a product title using statistical information of “shopping category” and “handling store”(Japanese only) DOMESTIC
  • *Linked to the webpage of Yahoo Japan Corporation

Promoting Open Innovations

Yahoo! JAPAN

In order to contribute to the development of the information technology society, Yahoo! JAPAN Research’s development results and intellectual properties created in Yahoo! JAPAN are made available to the public in the form of OSS, etc. Many of them are being utilized in various situations. Yahoo! JAPAN believes that encouraging to create a healthy market will also lead to the future development of Yahoo! JAPAN.

OSS Developer Certification Program

This is a program that certifies selected employees as developers for OSS which is strategically adopted by Yahoo! JAPAN. Time used for developing specific OSS is regarded as work hours and a maximum of JPY 1 million budget per year is granted for the certified employees to work as OSS committers.

NGT(Neighborhood Graph and Tree for Indexing)

NGT is an OSS developed by Yahoo! JAPAN. This is a software that allows users to quickly search and identify data with multiple characteristics (high-dimensional data), such as text, image, and product/user data, from a large database. NGT is used in Yahoo! JAPAN Shopping, YAHUOKU!, and the Yahoo! JAPAN browser.

LINE

LINE released Resembla, a library for similar text string search (approximate string matching), to the public as open source software. It was developed with a focus on improved quality and speed. The quality goal was achieved by the right selection and combination of the scales according to the properties of the text string, and the speed, through candidate refinement and preprocessing of the corpus.
It is built for high-speed processing and adopts a highly scalable architecture.
Since it is implemented with C++ and provides a gRPC interface, it can easily fit into various implementation requirements.

ZOZO

The ZOZO Group makes active use of its vast amount of data and technologies. In particular, ZOZO Research is engaged in research and development to build a more convenient website and improve services for customers, by adopting AI-powered recommendation and search technologies into its products.

Release of ZOZOTOWN Fashion Recommendation Data and Algorithm R&D Platform to the Public as an OSS

ZOZO Research has released the large-scale fashion recommendation data and R&D platform “Open Bandit Data & Pipeline” to the public as an OSS. The project, named "Open source development and implementation infrastructure for social decision-making algorithms," was conducted by a joint research team consisting of Hanjuku-kaso, Inc. and ZOZO Research researchers, and received the Prime Minister's Award at Japan Open Innovation Prize, an award program held by the Cabinet Office.

Collaborations with External Partners

The Z Holdings Group collaborates with external partners from diverse fields, including enterprises, universities, and research institutions, to make the lives of users more convenient and more enriching. It also aims to solve various social problems through these collaborative initiatives.

SEIBU RAILWAY Co., Ltd. and Yahoo! JAPAN conduct demonstration experiment using big data and AI for prediction of train congestion

The congestion pattern of each station is projected thorough AI (machine learning)-powered analytics of big data collected from the Yahoo! JAPAN Transit Navigation service; the data utilized includes search histories of future train/transit schedules, and is pre-processed into anonymized statistical data (removing personal identifiers). Combining this with the SEIBU RAILWAY’s data on the number of people getting off the train by stations and by time, we aim to realize even more accurate congestion prediction.

Yahoo! JAPAN Conducts Demonstration Experiments for AI-Powered Big Data Collaboration between Companies

Yahoo! JAPAN has conducted demonstration experiments with various companies and municipalities, including Nissan Motor Co., Ltd., Ezaki Glico Co., Ltd., Japan Professional Football League, Kobe City, and Fukuoka City, for AI-powered big data collaboration between organizations.

Yahoo! JAPAN Conducts Joint Demonstration Experiment with Tokyo Metropolitan Government for Prevention of Heatstroke Using AI

Yahoo! JAPAN conducted a joint demonstration experiment with the Tokyo Metropolitan Government for predicting heatstroke risks at crowded places such as event venues, with high accuracy. In the experiment, the WBGT (wet bulb globe temperature; heat stress index) open data released by the Ministry of Environment for the purpose of preventing heatstroke was combined with the congestion information derived by Yahoo! JAPAN using its big data on location information. Congestion is considered to be one of the factors pushing up the heat stress index, and the heatstroke risks were predicted through the AI-powered analytics of the combined big data.

LINE and NAVER to Jointly Develop the World’s First Super-Giant Language Model Specialized in the Japanese Language – Enabling the Generation of Various Japanese-Language AI, Such as Dialogues and Translations without New Development Work

LINE and NAVER jointly made an announcement on the development of the world’s first super-giant language model specialized in the Japanese language and the construction of the infrastructure necessary for its processing. Using AI, the super-giant language model (general-purpose language model generated from a huge amount of data) enables a more natural language processing and language expression. This is the first attempt in the world to develop a language model of this scale specialized in the Japanese language.
In developing a general-purpose language model specialized in Japanese, LINE plans to use more than 175 billion parameters and over 10 billion pages of Japanese data as learning data. This by far exceeds the amount of parameters and learning used for other existing Japanese-based language models in the world. The company plans to further expand the amount of parameters and learning in the future.

ZOZO Technologies Collaborates with The University of Tokyo and HOSOO Co., Ltd. in a Joint Research on Expanding Design Potential for Functional Textiles

The joint research brings together the strengths of the three partners: knowledge in fashion, cutting-edge materials, and technologies from ZOZO Technologies; expertise in materials and interaction technologies and know-how on material-driven experience design from the research team led by Associate Professor Yasuaki Kakehi, the Graduate School of Interfaculty Initiative in Information Studies, The University of Tokyo; and innovative textile design capabilities backed by long-established traditional Nishijin textile technology and artistry from HOSOO Co., Ltd.

ZOZO Inc. Launches Joint Research with Yale University to Apply Economics to Understanding of Customer Incentives and Behavioral Psychology and Optimization of Customer Communication

The objective of this joint research is to understand customer incentives and psychology behind their purchasing behavior and to optimize communications tailored for each user, utilizing both the expertise of Dr. Kosuke Uetake, Associate Professor of Marketing, Yale School of Management, who specializes in quantitative marketing and empirical industrial organization, and ZOZOTOWN’s large-scale and diverse data. Through this initiative, we aim to further enhance our services in ZOZOTOWN, and in the long run, to maximize the value of use customer lifetime value for ZOZOTOWN users.