1. Designed and implemented a Hill Climbing Auto Focus (AF) algorithm for camera using data generated after photoelectric effect to reduce 50% computational cost and achieve 10% better focus value (FV) score2. Evaluated output image quality using PQ Tools (with better white balance, less noise, etc.) and presented to other teams3. Constructed virtual machine installation module for Fusion Compute (Huawei Cloud Platform) using Flask and NodeJS and tested the feasibility on frontend

1. Utilized Graph Signal Processing and eigenvalue decomposition to detect corresponding filter types of synthetically generated graph signals2. Generating a classification model to detect the filter type driven by an unknown input and unknown underlying graph, thereby generalizing the classical blind multi-channel identification problem to graphs

1. Pre-processed employer liability insurance transaction datasets in SQL, imputed missing label with PU learning and predicted the purchase probability with 93% accuracy using Random Forest2. Further analyzed the most significant indicators selected in the model above and generated report using Seaborn and Excel3. Used Chi-square test and Pearson correlation to select variables for the XGBoost model to predict the customer loyalty score and provide a reference for user recommendations… Show more 1. Pre-processed employer liability insurance transaction datasets in SQL, imputed missing label with PU learning and predicted the purchase probability with 93% accuracy using Random Forest2. Further analyzed the most significant indicators selected in the model above and generated report using Seaborn and Excel3. Used Chi-square test and Pearson correlation to select variables for the XGBoost model to predict the customer loyalty score and provide a reference for user recommendations and profiles4. Constructed business scope labels by mapping the National Industry Segmentation Labels with the phrases generated by NLP methods such as parsing, POS tagging and word frequency of the local dataset using Jieba5. Designed and constructed prototypes of group data analysis platform and intelligent linking platform using Axure Show less

Optimized similarity search based on text-data in time complexity from O(mn) to O(n) by constructing a multi-layer convnet to map real-value vectors to d-dimensional space to use Euclidean distance to represent the Edit distance using PyTorch

Created data regression models to predict generate revenue with 20% more accuracy than historical average using LSTM model