Project Experience
Course Project and
Essay for Master Graduation, advised by Sungjin Ahn
- Applied pre-trained ALBERT model on the problem Named Entity Recognition
- Implemented multiple semi-supervised techniques to reduce labeled data required
Seq-to-Seq Thematic Music Comments Generation (Research)
As Researcher Assitant, in Laboratory of Sematic Computing and Data Mining, USTC, advised by Linli Xu
- Collected music and comments from Web to construct well-paired music and text datasets
- Built a Seq2Seq model to generate text descriptions of music, with the dilated convolutional layer as the encoder and a memory based recurrent neural network as the decoder.
- Improved the model by fine-tuning it with a discriminator to progress 13.5 on Bleu than MLE baseline
- Proposed two new GAN-based evaluation metrics more aligned with human evaluation than traditional metrics such as BLEU
Improving the Evolutionary Search Algo (Research)
As Research Assistant, in USTC-Birmingham Joint Research Institute, advised by Ke Tang
- Added asymmetry to the distance measurement between the searching vectors to improve the efficiency of Negative Correlated Search for more than 10 times
- Improved effectiveness and surpassed all results of IEEE CEC-11 competition in application
Gross Merchandise Volume Prediction Modeling for Audience Targeting (Industrial)
As Machine Learning Engineer in E-commerce Marketing Group, in Kuaishou Company
- Designed and implemented a zero-inflated model to model both the probability and amount of user payment
- Developed the real-time API system to adapt the experiment
Ranking Model Improvement for Landing Page of Advertisement (Industrial)
As Machine Learning Engineer in Search Engine Marketing Group, in Kuaishou Company
- Applied an MTL model with ESMM structure on the ranking model in recommendation system based on DeepFM
- Reconstructed the offline data pipeline on Hive-sql for better sample selection and more flexible time window
