Jinda Han
CS PhD @ UIUC | ML @ Apple | MS&E @ Stanford
HCI & AI Researcher in Fashion & Social Media

Mining Fashion Influence on Social Media

The rise of social media has changed the nature of the fashion industry. Influence is no longer concentrated in the hands of an elite few: social networks have distributed power across a broader set of tastemakers. To understand this new landscape of influence, we created FITNet — a network of the top 10k influencers of the larger Twitter fashion graph. To construct FITNet, we trained a content-based classifier to identify fashion-relevant Twitter accounts. Leveraging this classifier, we estimated the size of Twitter’s fashion subgraph, snowball sampled more than 300k fashion-related accounts based on the following relationships, and identified the top 10k influencers in the resulting subgraph. We use FITNet to perform a large-scale analysis of fashion influencers, and demonstrate how the network facilitates discovery, surfacing influencers relevant to specific fashion topics that may be of interest to brands, retailers, and media companies.

• Mining Fashion Influence on Social Media
  University of Illinois at Urbana-Champaign. 2023.
• FITNet: Identifying Fashion Influencers on Twitter (Dataset available at http://fashioninfluence.net/fitnet)
  Jinda Han, Qinglin Chen, Xilun Jin, Weikai Xu, Wanxian Yang, Suhansanu Kumar, Li Zhao, Hari Sundaram, and Ranjitha Kumar.
  Proceedings of the ACM on Human-Computer Interaction (PACM HCI). CSCW 2021.
• Identifying Fashion Accounts in Social Networks
  Doris Jung-Lin Lee, Jinda Han, Dana Chambourova, and Ranjitha Kumar.
  KDD Workshop on ML Meets Fashion. 2017.

Fashion Stylist Chatbot

Fashion serves as a means to not only present an enhanced version of oneself but also to actively become a better individual through its influence. Meanwhile, the rapid development of AI technology has brought more possibilities in the tech-assisted personal fashion domain. We review the literature regarding imitation theory, fashion psychology, and the changes in fashion paradigms. Leveraging these theories, we propose a future personalized fashion solution: a fashion stylist chatbot that is capable of generating inspirational fashion styles on virtual representations of our bodies. Differentiating from previous work, this solution can help us build our wardrobe starting from thinking about our psychosocial aspects.

Virtual Influencers on Social Meida

Virtual Influencers (VIs) are computer-generated characters, many of which are often visually indistinguishable from humans and interact with the world in the first-person perspective as social media influencers. They are gaining popularity by creating content in various areas, including fashion, music, art, sports, games, environmental sustainability, and mental health. Marketing firms and brands increasingly use them to capitalise on their millions of followers. Yet, little is known about what prompts people to engage with these digital beings. In this paper, we present our interview study with online users who followed different VIs on Instagram beyond the fashion application domain. Our findings show that the followers are attracted to VIs due to a unique mixture of visual appeal, sense of mystery, and creative storytelling that sets VI content apart from that of real human influencers. Specifically, VI content enables digital artists and content creators by removing the constraints of bodies and physical features. The followers not only perceived VIs' rising popularity in commercial industries, but also are supportive of VI involvement in non-commercial causes and campaigns. However, followers are reluctant to attribute trustworthiness to VIs in general though they display trust in limited domains, e.g., technology, music, games, and art. This research highlights VI's potential as innovative digital content, carrying influence and employing more varied creators, an appeal that could be harnessed by diverse industries and also by public interest organisations.

Fashion Style Recommendation Engine

Used deep learning to train the fashion model by using the latent fashion concepts jointly in two languages capturing these elements and styles. Optimized search engine backend with new data from ecommerce and ezine fashion websites. Output personalized dressing combo recommendations based on style, similarity, etc.

Predicting Fashion Choices Using Unimportant Domains: Personality vs Fashion Questionnaire Systems

Recently, in light of the current emphasis on security and privacy, people have become more aware of the data that companies gather from them. However, companies need this data to create recommendations that will be useful to their customers. Our idea is to develop knowledge of a person’s preferences using data that does not gather sensitive information such as viewing or buying history. Many people already take fun quizzes on Facebook and other web platforms to learn more about themselves, so why not take the answers collected from these quizzes instead? In this project, we aim to connect answers to fun questions such as “What would you rather do on a Saturday morning?” and “Which instrument would you like to learn?” to choice of fashion styles.

Large-scale Study on Heterogeneous Information Networks

SetSearch+: Web System 1, Web System 2

• A Constrained Maximum Likelihood Estimator for Unguided Social Sensing
  Huajie Shao, Shuochao Yao, Yiran Zhao, Chao Zhang, Jinda Han, Lance M. Kaplan, Lu Su, Tarek F. Abdelzaher.
  The IEEE International Conference on Computer Communications (INFOCOM). 2018
• SetSearch+: Entity-Set-Aware Search and Mining for Scientific Literature
  Jiaming Shen, Jinfeng Xiao, Yu Zhang, Carl Yang, Jingbo Shang, Jinda Han, Saurabh Sinha, Peipei Ping, Richard Weinshilboum, Zhiyong Lu and Jiawei Han.
  In The 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. KDD 2018 Demo Track
• HighSim: Highly Effective Similarity Measurement in Large Heterogeneous Information Networks
  Huajie Shao, Jinda Han, Sida Li.
  Manuscript. 2016.

Landuse Evolution and Impact Assessment Model

"Better tools are needed to manage regional dynamics, not just as economic systems or static inventories of resources, but as complex systems that are part of regional and global networks. However, effective management requires both that we understand the systems to be administered and that we understand the implications of our strategies. We have attempted here to outline an approach for understanding the dynamics of urban systems and the potential implications of urban policy and investment management decisions. We described one modeling approach — LEAM — that utilizes cellular automata and other technological advances in spatial simulation modeling to help improve a community’s ability to make ecologically and economically sound decisions. LEAM was intended to enable users to capture stochastic influences and view the reported probable consequences of intended events in a scenario-based format that is comprehensible by local experts, decision-makers, and stakeholders. The LEAM Model, its development, and its application to several regions within the continental United States is conducted and managed by a team of faculty, staff, and students at the University of Illinois at Urbana-Champaign. This team brings together expertise in substantive issues, modeling, high-performance computing, and visualization coming from the departments of Landscape Architecture, Urban Planning, Geography, Economics, Natural Resources and Environmental Sciences, the National Center for Supercomputing Applications (NCSA), the United States Army Corp of Engineers, and private industry. The mission of the LEAM group is to help others understand the relationships between human economic/cultural activities and biophysical cycles from a changing land use perspective. All of us must realize that these interacting systems behave in very complex and dynamic ways. Understanding the extent of how one system affects another will allow us to make better land use management decisions in the future."

Mobile Element Search Engine

[Search Engine Demo]
Identify the elements and icons on Android UI, enable the basic search function to categories these elements.

Virtual Reality For Personal Interior Design

[floor plan mining demo]
Mining the floor plans on the web, and the goal was to generate the floorplan layout design into a 3D model automatically.

Enterprise Information Update and Alert System

Many companies today require a system to gather valuable online information about products or customers. With such a system, they can improve their products or services based on this information, thereby enhancing overall company performance. Receiving a large volume of accurate information is the first crucial step in this process, followed by storing and retrieving the data. The next step involves comparing this data with the existing system database to update outdated information. Finally, future predictions and automatic updates via alerts represent the last essential stage. Currently, there is limited research on how to build such a system. In this work, I propose an Update and Alert System (EIUAS) to help companies maintain their product databases. The system includes several components: a web crawler, data extractor, data verifier, data integrator, dynamic database design, data retrieval engine, update checker, prediction module, and alert system. This system enables companies to gather relevant product or customer information from public sources. To evaluate its effectiveness, we calculate the update percentage. Additionally, a qualitative assessment by a domain expert confirmed the system's usefulness and its capability to handle dynamic, updatable information.

Intelligent Medical Diagnosis System

The work presents an intelligent online medical diagnosis system designed to help users identify possible diseases based on their symptom descriptions. Unlike general-purpose search engines, this system provides more accurate and structured health information by leveraging a curated database of over 4,000 diseases. It incorporates a custom-built search engine that uses BM25L scoring and Rocchio feedback to rank relevant diseases based on symptom similarity. Users can refine results using filters like age, gender, race, and diagnostic methods (e.g., CT, MRI), and the system includes a user account feature to track search history and feedback. The system also features a user-friendly web interface developed with HTML, CSS, PHP, and Java backend components. Evaluations using Mean Reciprocal Rank (MRR) show that the system provides highly relevant results for most queries. However, current limitations include strict keyword matching and difficulty handling misspellings or complex medical terminology. Future improvements include incorporating stemming, enhancing data quality, adding NLP features, and optimizing performance with caching and compression techniques.

Misc. & tools

Before 2014, I also built some random stuff in various areas:

• PHP Image Parser Demo
  
• PHP Server Web Editor Demo
  
• PHP Health Forum Crawler Demo
  
• BMI Health Checker Demo
  
• JDTools (Android App)
  
• Typing Zombie (macOS Game)