Yoon-Sik Cho


My research interests lie in the area of machine learning and data mining. Specifically, I am interested in modeling social networks and real-world phenomena. My researches during my doctoral studies are focused on developing efficient inference algorithm that infers hidden attributes of social actors, estimating missing information, and predicting the future events. 

  • Complex Dynamic Network Modeling 

In this work, we studied how node attributes and link structure evloves under mutual influence. Our contribution, Co-evolving MMSB, augments Mixed Membership Stochastic Blockmodels (MMSB) with a model of node dynamics in the latent space. In contrast to other dynamic models, Co-evolving MMSB explicitly accounts the inter-coupled dynamics of nodes and network topology, and can be used to describe the co-evolution of selection and influence in real world dynamic networks. We applied Co-evolving MMSB to a real world dataset consisting of the bill co-sponsorship network among U.S. senators, and obtained reasonable results in our experiments. In particular, Co-evolving MMSB was able to detect increasing polarization in the Senate as reported by other sources that analyze individual voting records of the senators.


  • Behavior Modeling using Point Processes 

Many of the events or the behaviors of users can be described using point processes where points often represent time and location of the events/behaviors. In this line of work, I have considered a particular type of spatial-temporal point process - sequence of events describing inter-gang violence in Los Angeles. In this dataset, each data item describes an incident of attack and is characterized by a victim gang, the perpetrator gang, the date, time, and the location of the incident. One of the problems with such datasets is the information is often missing or corrupted. To deal with these issues, I have developed a preliminary generative model that captures the spatial and temporal patterns of incidents between gangs. The preliminary results indicate that the performance of our generative model on the task reconstructing the missing data and predicting the future incident are superior to other existing methods.

Another part of the work involves check-in dataset in Location-Based Social Networks (LBSN). In this dataset, each user establishes friendship with other users and checks-ins to favorite places. Using this dataset, we are interested in clustering venues of similar types and finding factors that describe the temporal dynamics of check-ins on each venue. Our work provides insight into understanding of the venues: i.e., finding how the venues are similar each other or finding why users check in to given venues. We have shown better performance on predicting the relationship between users using their check-in records (or behavior in general) as well as better performance on predictions of time of the check-ins.     


  • Collective Behavior Modeling

Both the behaviors and the networks strongly reflect users' hidden attributes. In this work, we project users social links and their behaviors onto same latent space. Our work is focused on specific type of behavior: check-ins in LBSN, but one can easily generalize this problems by consider various on-line behaviors such as records of videos or music that users watched or listened to, records of products that users bought or reviewed. We've found that many of the users are inactive in making friends on-line or leaving their records on-line. By considering the network and their behavior on joint-space, we can build more effective recommendation system even with no previous records on the side we want to predict. For instance, we can predict venues that users might visit by only using their social links or predict edges by only using their venue records. This work also provides better understanding of the latent attributes by connecting the two types of observations: network and behavior.