Spring

In this project, we present several algorithms that infers a Bayesian network that explains the observed data naturally (maximizing p(Model | Data)) by using Stochastic Gradient Langevin Dynamics MCMC, thermodynamic integration and generalized importance sampling. Here, we assumed that Bayesian network has a fixed graph topology and unobserved random variables (interestingly this assumption makes the model more general).

The motivation behind this project is to be able to make a sentiment analysis on Turkish financial documents. Classify-ing financial news as negative or positive helps predicting the change in stock price for a particular company. In our work, we process more than 20,000 documents obtained from financial institutions in Turkey. We translate Loughran & McDonald's word list into Turkish and label the news based on the number of negative words they contain. We then construct a Logistic Regression model with Ll penalization using bag-of-words ap-proach to predict the labels of news.

Time series is anything that you can represent with numbers and extract statistical information. Stochastic Volatility models are the models that the variance itself changes with time. Inspecting and implementing a solution for such models, often faces the problem of unaffordable complexity with the increasing dimensions. In this project we have tried out different models and mixed solution, as well as we have done attempts in order to increase the performance of algorith in higher dimensions.

Morphological analysis is a very important sub-task of natural language processing. It is used for tokenization, stemming, lemmatization and normalization. For the NLP task in which machine learning approach plays a crucial role, pre-processing the data is vital and the success rate is highly dependent on the pre-processing methodologies. This project proposes a tool for morphologically analyzing words in Turkish which is an agglutinative language in order to constitute a baseline for the further NLP projects.

Age estimation is one research field of face analysis. Image-based age estimation can be defined as determination of a person's age from facial images. Estimating the age of the elderly from the images attempt came up from the interesting approach of the doctors of Alzheimer’s patients. It seems like most of the Alzheimer’s patients look younger than their actual age according to doctors’ observations. Until today, video data has been gathered from up to 20 people, each one older than the age of 65.