Machine Learning 2020 Fall
Yitao Zhou
Video Head Detection with YOLO (paper)(presentation)
Common images and videos primarily focus on people. Indeed, about 35% of pixels in movies and YouTube videos as well as about 25% of pixels in photographs belong to people (Laptev, 2013). Therefore, person detection in videos as well as photographs is a key problem for computer vision and object detection. While face detection has reached maturity, detecting people under full variation of camera view-points, human poses, lighting conditions and occlusions is still a difficult challenge. Especially in surveillance video data, diverse angles and poses are involved, which is valuable to analyze and worthwhile information can be extracted. In our school, all the classrooms has limits for the number of people allowed at the same time and they are equipped with video surveillance cameras. However, no detection of overrunning the people limit has been made. In this paper, I will apply the prevailing YOLO model, train it on customized data, to conduct a real-time detection of head number in videos.
Classification of Leukemia Types Based on Patient Gene Expression (paper)(presentation)
This project shows a successful approach to diagnosing acute myeloid leukemia (AML) or acute lymphoblastic leukemia (ALL), and classifying patients with ALL into two subcategories using logistic regression and lasso regression on data of gene expression monitoring.
Yumo Yao and Wanyi Yang
COVID-19 Cases Prediction – A Time Series Prediction Using Machine Learning Approach (paper)(presentation)
Our project aims to predict the number of cases that will break out in the future based on the number of COVID-19 cases already available. We study the impact of case base, population density, GDP level, economy, political constraints, and policies on the spread of the virus in several major countries to predict the number of future case growth. It provides information for the government to issue restraint policies during the epidemic and alleviates the fear and anxiety about the unknown of the virus. We find data sets that quantify various policies and apply multiple machine learning models to make predictions. From a broad perspective, we applied two approaches to solve the time series problem, one is to add time as features inside the attributions of the prediction, and the other is to use a neural network structure for the memory of historical data to solve the problem. The first data approach uses five models, and for the second data approach, we use two types of RNN. finally, we use the part of the divided data for validation.
Chengxi Wang, Tianyu Zhang and Vivian Lu
Default Risk Prediction Problem: Application of Machine Learning Models in P2P Online Lending (paper)(presentation)
Our project aims at training a suitable machine learning model for the Peer-to-Peer lending companies to identify the potential default risks in each loan. Initially, we focused on a classification problem with multiple objects to distinguish different types of badness, but as we visualized the features and applied several models, the result was far from satisfactory. Then, we restarted and simplified it into a binary classification problem. Since the dataset is not balanced enough, we tried different methods of re-sampling under various types and models. After tuning the hyper-parameters, Random Forest yields the best Recall of 0.7220, while AdaBoost and Gradient Boosting perform better in Precision. This result reveals that each model has its strength in a certain evaluation metric, and we are able to choose the best one according to specific demand of the company.
Hanyuan Zhang, Liqiang Luo, Tinglong Liao
Sentiment Extraction Of Twitter Messages (paper)(presentation)
In the current decade, we witness the booming of social media, people all around the world express their feelings and emotions on social media. In this project, we analyzed 31026 twitter messages and focused on capturing the sentiments of those tweets. We first discuss different feature extraction methods, including One-hot, TF-IDF, Word2Vec and GloVe. Then, we apply various models with above feature extraction methods, including Naive Bayes, Logistic Regression, Ada-Boosting, SVM and LSTM. The result was not perfect, we reached the highest accuracy rate of 75.43% by using LSTM with Glove. This comparatively low accuracy could attribute to the small size of our dataset. And we believe that in the future, when a larger amount of data is collected, a significant higher accuracy rate could be obtained.
In addition, we find that in traditional non-neural-network models, feature extraction methods have a more significant impact on performance than different models. This could be indicative to our future work that more effort could be put on finding better feature extraction methods.
Haobo Li and Ruyi Liu (Ada)
TransformNet: A Generative Approach to Image Style Transfer(paper)(presentation)
In recent years, Convolutional Neural Networks has endowed computers with the capacity to extract more abstract features of images. This development in computer vision has triggered the whole study of image style transfer. In this paper, we follow one of the most popular approaches, the generative method, and implemented an algorithm for Neural Style Transfer. We also applied instance normalization to our model in order to speed up the training process as well as produce more aesthetic output. After training, our model is capable of performing style transfer within a second; compared to iterative approaches, our implementation preserves the high visual quality of output images and is over 20 times faster than iterative methods.
Junting Chen, Yiyang Luo and Peiyang Wen
Lifestyle Data Analysis and Stress Level Prediction(paper)(presentation)
This project aims to help people predict their stress level based various aspects of their lives, including professional, lifestyle, behavior, etc. We first implement Exploratory Data Analysis, Agglomerative Hierarchical Clustering and Correlation Analysis to analyze the distribution pattern of our dataset. Based on the analysis, we apply certain adjustments to our dataset, such as feature selection, data cleaning and data vectorization. Then we apply different machine learning models, include but not limited to Linear Regression, Random Forest, SVM, Neural Network and Recurrent Neural Network, to predict the stress level based on the data input. After that, we compare the performances of different model, and find that SVM has the relatively best performance because it handles high dimensionality input and outliers well. This result reveals that we can help people to analyze their stress level based on their living, and may further provide advice to them so that people would have a work-stress balanced life.
Zihan Dai, Juncheng Dong and You Xu
Machine Learning Approach to Movie Rating Prediction (paper)(presentation)
The project aims to find a Machine Learning approach to predict the public rating of a movie given its characteristics. The project applies four categories of models - Linear Regression, KNN, Decision Tree, and Neural Network to train the regressor. The data are collected from the online rating platform Douban. The prediction performance is good with MSE around 0.1.
Yancheng He, Liyang Zhu and Ding Wang
Network Traffic Protocol Analysis with Machine Learning (paper)(presentation)
In this project, we try to use machine learning techniques to predict the application type of network traffic from the traffic shape. We used PCA, SMOTE, NearMiss and other techniques to pre-process the dataset. Then, we tried to solve this problem by training three difference models: KNN, Random Forest, Artificial Neural Network. Finally, we evaluate the effectiveness of each model and analyse the potential improvements of our process.
Yuxin Feng
Painters by Numbers---Does every painter leave a fingerprint? (paper)(presentation)
The project mainly focuses on building a CNN model to learn the style of artworks painted by different artists and determine whether pairs of paintings are by the same artist. The algorithm will help identify whether an unknown painting is an authentic work or a forgery in the high-stakes art authentication industry. Related models including t-SNE and Siamese Networks are also presented in the paper. The project was originally a Playground Prediction Competition on Kaggle.
Chengxun Wu, Yuxuan Li and Xingjian Gao
Mathematical Symbol Identification: A CNN Implementation(paper)(presentation)
Convolutional Neural Network (CNN) is one of the most robust tools in dealing with images in machine learning and deep learning tasks. This project focuses on the implementation of hand-written mathematical symbol classification based on Convolutional Neural Networks. After processing the original images into tensor-based data, this project constructs the CNN structure based on pytorch and improves the model using different optimizers. After 25 epochs of training, the CNN-based classifier obtained a performance with a test accuracy of 98.7% and is successful in identifying personally-written samples.
Cinny Lin, Yingrong Mao, and Zhihan Yang
Improving Book Recommender Systems Using Reading Tags and Book Covers(paper)(presentation)
Our project aims to use the available features about users and books to improve recommendation accuracy for books. Initially, we planned to recommend books based on user features. However, after implementing k-means clustering, we realized that the customers purchase habits do not vary significantly according to demographics. Then, we learned about item-based collaborative filtering, but also its limitation regarding scalability and cold-starts. Therefore, to offer more useful suggestions, we turned our task to a supervised learning problem. In terms of dimensionality reduction, we tested stochastic gradient descent (SDG), alternating least squares (ALS) and singular vector decomposition (SVD) matrix factorization methods. While for the cold-start problem, we used popularity model and content-based filtering as complement. Regarding content-based filtering, we built two models. We used Term Frequency-inverse Document Frequency (TF-iDF) on book tags. We also employed neural network models, incorporating the categorical and numerical book features as well as the images of book covers. The aggregated performance of our models turned out pretty interesting. To our surprise, SVD did not yield very good results. We discovered that neural based model gives the best performance, reducing RMSE by 94% than the basic k-nearest neighbors (kNN) model. This result reveals that we could efficiently reach a larger customer base since we are able to recommend books accurately even when we know very little about our new users. More specifically, considering cover images when making book recommendations can yield better results than only using text or numbers.
Chenqi Jiang, Yufei Xue and Zhiheng Yang
Study of Speech Emotion Recognition using Classification Models (paper)(presentation)
Our object of research is to apply machine learning and deep learning methods for Speech Emotion Classification (SER). The research is conducted on 1440 sound files in the Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS), including 8 common emotions. Nonetheless, having difficulty in improving the testing accuracy, we restrict the research to 6 emotion labels after analyzing the confusion matrix. Having extracted features using python library librosa and implemented models such as SVM, LSTM, CNN, we obtained a 72.84% testing accuracy using LSTM.
Liu Han
Room Neatness Detection with Two DifferentModels: SVM, and CNN Using Pre-trained ResNet(paper)(presentation)
The goal of this project is to evaluate the messy probability of a given room picture. For an input image, our model will output a messy probability, indicating how messy the room is. I used two different machine learning models, which are the support vector machine, Convolutional neural network using the pre-trained ResNet. Different parameters are tested to achieve the best test probability compared to ground truth. Finally, I achieved 93.3% accuracy for the test data set.
Yifan Zhuo
Machine Learning Project Report(paper)(presentation)
We aim to solve the problem of mine sweeping using Deep Q Learning (DQN).