Yunyi Liao

Experience

• 

  CSCI Innovative Team Of Guangzhou University

  Jul 2016 - Sept 2016

  Software Developer

  Developed a social mobile application called "IM" like WeChat, What's up or Messenger, with a variety of functions like register, login, chat, etc. In 2017, won Third Prize in Guangdong Province Campus Selection of "Zhongxing Cup" Undergraduate Computer Works Competition.

• 

  CSCI Innovative Team Of Guangzhou University

  Feb 2017 - Apr 2017

  Software Developer

  Developed an application called "KoreanLearning", which inspires users to learn Korean not only by traditional memorization but also through seven little games. In 2017, won both the First Prize in Guangdong Province Campus Selection of "Zhongxing Cup" Undergraduate Computer Works Competition, and the Second Prize of "Huazi Cup" Guangdong Undergraduate Computer Works Competition.

• 

  Guangzhou Dayang Education Technology Co., Ltd.

  Sept 2017 - Oct 2017

  Software Engineer

  • Collaborated within an 8-person development team and developed a scientific experiment android application “aikexue” , which allows users to watch and conduct both free and paid scientific experiments.• It is being used in China now, serving the Education Bureau of Huadu District in Guangzhou and Guangzhou Dongfeng East Road Primary School. Shown as http:// itdayang.com/class/57.html

• 

  USC Viterbi School of Engineering

  Mar 2020 - May 2020

  • Using Item Based Collaborative Filtering algorithm, built a recommendation system on Yelp business, which recommended products that the customer might like by predicting the customer’s ratings of unknown products.• In the training process, generated business pairs, calculated the Pearson Correlation of them, and recorded them into a model file.• The Model file recorded the relational weight between two businesses, and it could be used for prediction for multiple times.• In the prediction process, got a list of neighbor businesses for a target business of a target user, and used at most 15 neighbor businesses with their weights to predict the target rating. Evaluated the performance by RMSE on the predictions and ground truth.• Used python3.6 and spark2.3 as the programming tools, and Vocareum as the running cloud platform. The final rank is top 20 among 60+ participants.• Project link: https://github.com/Ueny/Recommendation_System_ItemBasedCF Show less • By using the BFR algorithm, I splitted the big data into 5 small data files so I could load them chunk by chunk.• In the first load, I randomly took 20% of data, initialized centroids by using kmeans++ algorithm, and assigned them as the Discard Set(DS). The rest data would be assigned to the Discard Set, Compression Set(CS) or Retained Set(RS) by calculating the Mahalanobis Distance.• In all of the left loads, data was assigned to the three different sets by Mahalanobis Distance as well, but at the end of each load, the CS and RS would be checked if they were close enough to the near DS or CS by using Mahalanobis Distance.• In the end after all loads, all the left CS and RS were merged to DS by using Mahalanobis Distance. If it was above the threshold, the points would be classified as outliers.• Project link: https://github.com/Ueny/BFR Show less

  • Recommendation system designer

    Apr 2020 - May 2020

  • Data Analyst

    Mar 2020 - Apr 2020
• 

  USC Viterbi School of Engineering

  Sept 2020 - Nov 2020

  Machine Learning Researcher

  • Worked in a 6-person team, and did the forecast for the El Niño-Southern Oscillation, using machine learning algorithms like Linear Regression and CNN as well as the classification based on these algorithms.• Established Linear Regression and CNN models, and used model data to train them. Get the best test result by adjusting different variables and model parameters.• The Classification was extended based on the Linear Regression and CNN models, and the same test process was completed to obtain the best model.• Using Bokeh to demonstrate our testing results with line charts, compared the Pearson correlation coefficient and prediction accuracy of different models to determine which model had the best prediction effect.• The results predicted by the CNN algorithm were the most accurate, and could be predicted up to the next nine months.• Project link: https://github.com/DS-560/ENSO_forecast Show less