Hoiin

About me

I am currently machine learning research engineer at Blux.
I want to innovate buisness decision making process with machine learning.

interests

• Recommender System

  Sequential Recommendation
  Graph Neural Network
  Causal Inference

• Business Intelligence

  User Lifetime Management
  Intelligent push notification
  User segmentation

• Development

  Backend Engineering
  MLOps

Education and Careers

• 

  Seoul National University

  2017 ~ 2023

  Bachelor of Economics
  (Minor : Computer Science and Engineering)

• 

  Blux

  2017 ~ Until now

  Machine Learning Research Engineer

Skills

Enthusiastically learn new area and open to use new tools

Research Works

• HEIST
• RiDAR
• CaseQ
• SASQN

• HEIST stands for Hybrid recommender system for predicting Event timing and Item with Stochastic Temporal process. This model is developed for managing user lifetime cycle , which is inspired by survival analysis.
  Its input is user history data(history item, event and timestamp) and output is user recommendation and probability of event occuring after given timestamp. Especially, user is represented by transformer output and probability distribution of event occurrence can be flexibly modeled through exponential distribution, weibull distribution, etc.

• RiDAR stands for Re-interaction Dynamics Aware Recommender system. This model is used for Intelligent push notification. The difference from regular recommender system is that it predicts the re-interaction cycle of items.
  The distribution that models the re-interaction cycle of items is mixture of exponential and gaussian distribution, which means exponential distribution represents the short term intest of items and gaussian distribution represents the lifetime interest of items. Since its parameters is not represented by analytic form, the deep learning helps estimate parmeters of given distribution.

• CaseQ is for decompounding the effect of external contexts(popularity, season or advertising) on user behavior data. This model alleviates bias in user behavior data through causal inference in sequential recommender system, which is vulnerable to biased user behavior. Since the confounders are unknown, CaseQ defines them as latent variables and is trained with modified objective function. By causal inference, recommendation output is robust to user behavior and outperforms the vanilla seqnential recommendation models.

• SASQN combines sequential recommendation with Q-learning to fine-tune the effects of user feedback. By using Q-learning, modeler can model user's implicit feedback(click, like or purchase, etc.) with reward function and the objective function of model is modified to the user's long term satifaction (cumulative rewards). SASQN uses reward as prompt to alleviate the instability of learning that comes from the absence of neagative feedback and also controls the exploration by adjusting prompt rewards.