Jiahao Wang Email and Phone Number

Project: Crowd CountingOverview: To address the challenge of manual counting as well as costly sensor installation, we utilized Keras to build two deep learning models, CNN and Transfer learning model, to count customers in malls captured by surveillance cameras. Compared the performance of the two models, and found the CNN to be the better model. Improved customer traffic management in malls and contributed to cost savings for businesses.Roles & Responsibilities: Data augmentation, built and tested CNN models with different architectures, built dense neural network on top of VGG16, model evaluation• Used the ImageDataGenerator class from Keras to perform data augmentation on the input images, including random rotations, flips, shears, and zooms.• Built the first model, the CNN model, with convolutional layers, pooling layers, and dense layers, and conducted experiments to evaluate the impact of different components of the models on their performance, such as number of layers, number of filters, and use of regularizations and dropout. • Built a second model by implementing transfer learning with VGG16 for feature selection, and constructed a dense neural network for further learning on selected features. • Achieved better performance with a 4-layer CNN model, resulting in a testing MAE error of 1.5 Packages & Techniques: tensorflow, keras.applications, sklearn, pandas, numpy; computer vision techniques, deep learning, CNNs, edge detection, data augmentation, transfer learning

Project: Customer Churning Prediction in the Telecommunications SectorOverview: Analyzed a large dataset to determine the significant factors that are responsible for customer churn in the Telecommunications Sector. Developed and found well-performing models that can predict customer churn with high accuracy. These findings can be used to retain customers and reduce the churn rate in the telecommunications sector.Roles & Responsibilities: Data cleaning, EDA, data visualization, imbalance data oversampling, feature selection, model construction, hyperparameter tuning, model evaluation• Conducted exploratory data analysis on the dataset and handled the imbalanced target variable by using SMOTE from Imblearn to oversample the minority class.• Used feature selection techniques such as correlation matrix and recursive feature elimination to select the most important features that contribute to customer churn.• Trained and tested various basic models such as logistic regression, decision tree, and SVM, as well as ensemble models of Random Forest and XGBoost. The hyperparameters of these models were tuned to increase their performance and prevent overfitting.• Evaluated each model on a holdout test set to measure its performance, including metrics such as accuracy, precision, recall, and F1 score. Visualized the performance of the models using confusion matrices and ROC curves.• The XGBoost model was identified as the best-performing model with an accuracy of 96%, precision of 96%, and recall of 97%.• The most important features that contribute to customer churn were identified as customer service calls, total day charge, and total eve charge.Packages & Techniques: Sklearn, Imblearn, XGBoost, Numpy, Pandas, Seaborn; EDA, SMOTE, PCA, feature selection, ensemble learning

Project: California Housing Price PredictionOverview: The goal of the project was to build a machine-learning model that predicted the median house value in California based on various features such as location, size, number of rooms, etc. This model could be used by real estate agents, homebuyers, and sellers to make more informed decisions about pricing and buying/selling properties.Roles & Responsibilities: Data cleaning, missing value imputing, EDA, data visualization, pipeline building, model construction, hyperparameter tuning, cross-validation, model evaluation Conducted exploratory data analysis to gain insights into the dataset and identified potential outliers and missing values. Applied KNN to impute missing values Used data visualization techniques to identify correlations between variables and features that have a strong influence on the target variable. Applied feature engineering techniques such as one-hot encoding and scaling to prepare the data for machine learning models, and built pipelines to orchestrate the flow of data into and output from the machine learning models. Built several machine learning models including Linear Regression, Polynomial Regression, Decision Tree, SVM, and Random Forest. Implemented cross-validation techniques to validate each model and ensure their generalizability.  Chose the best-performing model, the random forest model. Optimized the performance of the random forest model by fine-tuning hyperparameters such as the number of estimators and maximum depth using random search. The best-performing random forest model achieved an R2 value of 0.82 and MSE of around 41000 on testing data.Packages & Techniques: Sklearn, Numpy, Pandas, ydata_profiling, Matplotlib, Seaborn; EDA, Hyperparameter tuning, Pipeline building, feature engineering, Cross-validation

• Developed back-testing systems in Python to evaluate the effectiveness of multiple trading models in the Chinese Bond market. Utilized data analysis and data visualization techniques to explore financial data, evaluate model performance, and provide investment strategies.• Designed and implemented algorithms to generate two scores for mutual funds, measuring independence and spread of investment. Analyzed scores and provided valuable insights for selecting high-performing funds to guide investment decisions.• Built a quantitative model in Python using entanglement theory to identify promising stocks based on historical trading data. Collaborated with other analysts and stakeholders to support investment decision-making processes.