In the area of student knowledge assessment, knowledge tracing is a model that has been used for over a decade to predict student knowledge and performance. Many modifications to this model have been proposed and evaluated, however, the modifications are often based on a combination of intuition and experience in the domain. This method of model improvement can be difficult for researchers without high level of domain experience and furthermore, the best improvements to the model could be unintuitive ones. Therefore, we propose a completely data driven approach to model improvement. This alternative allows for researchers to evaluate which aspects of a model are most likely to result in model performance improvement. Our results suggest a variety of different improvements to knowledge tracing many of which have not been explored.

Computational indices related to n-gram production were developed in order to assess the potential for n-gram indices to predict human scores of essay quality. A regression analyses was conducted on a corpus of 313 argumentative essays. The analyses demonstrated that a variety of n-gram indices were highly correlated to essay quality, but were also highly correlated to the number of words in the text (although many of the n-gram indices were stronger predictors of writing quality than the number of words in a text). A second regression analysis was conducted on a corpus of 88 argumentative essays that were controlled for text length differences. This analysis demonstrated that n-gram indices were still strong predictors of essay quality when text length was not a factor.

Are the words that people use indicative of their personality type preferences? In this paper, it is hypothesized that word-usage is not independent of personality type, as measured by the Myers-Briggs Type Indicator (MBTI) personality assessment tool. In-class writing samples were taken from 40 graduate students along with the MBTI. The experiment utilizes naïve Bayes classifiers and Support Vector Machines (SVMs) in an attempt to guess an individual’s personality type based on their word-choice. Classification is also attempted using emotional, social, cognitive, and psychological dimensions elicited by the analysis software, Linguistic Inquiry and Word Count (LIWC). The classifiers are evaluated with 40 distinct trials (leave-one-out cross validation), and parameters are chosen using leave-one-out cross validation of each trial’s training set. The experiment showed that the naïve Bayes classifiers (word-based and LIWC-based) outperformed the SVMs when guessing Sensing-Intuition (S-N) and Thinking-Feeling (T-F).

In this thesis I develop a variety of techniques to train, evaluate, and sample from intractable and high dimensional probabilistic models. Abstract exceeds arXiv space limitations -- see PDF.

Machine learning research is often conducted in vitro, divorced from motivating practical applications. A researcher might develop a new method for the general task of classification, then assess its utility by comparing its performance (such as accuracy or AUC) to that of existing classification models on publicly available datasets. In terms of advancing machine learning as an academic discipline, this approach has thus far proven quite fruitful. However, it is our view that the most interesting open problems in machine learning are those that arise during its application to real-world problems. We illustrate this point by reviewing two of our interdisciplinary collaborations, both of which have posed unique machine learning problems, providing fertile ground for novel research.

Collaborative filtering is a rapidly advancing research area. Every year several new techniques are proposed and yet it is not clear which of the techniques work best and under what conditions. In this paper we conduct a study comparing several collaborative filtering techniques -- both classic and recent state-of-the-art -- in a variety of experimental contexts. Specifically, we report conclusions controlling for number of items, number of users, sparsity level, performance criteria, and computational complexity. Our conclusions identify what algorithms work well and in what conditions, and contribute to both industrial deployment collaborative filtering algorithms and to the research community.

This paper studies parallelization schemes for stochastic Vector Quantization algorithms in order to obtain time speed-ups using distributed resources. We show that the most intuitive parallelization scheme does not lead to better performances than the sequential algorithm. Another distributed scheme is therefore introduced which obtains the expected speed-ups. Then, it is improved to fit implementation on distributed architectures where communications are slow and inter-machines synchronization too costly. The schemes are tested with simulated distributed architectures and, for the last one, with Microsoft Windows Azure platform obtaining speed-ups up to 32 Virtual Machines.

Learning the parameters of a (potentially partially observable) random field model is intractable in general. Instead of focussing on a single optimal parameter value we propose to treat parameters as dynamical quantities. We introduce an algorithm to generate complex dynamics for parameters and (both visible and hidden) state vectors. We show that under certain conditions averages computed over trajectories of the proposed dynamical system converge to averages computed over the data. Our "herding dynamics" does not require expensive operations such as exponentiation and is fully deterministic.

Item recommendation is the task of predicting a personalized ranking on a set of items (e.g. websites, movies, products). In this paper, we investigate the most common scenario with implicit feedback (e.g. clicks, purchases). There are many methods for item recommendation from implicit feedback like matrix factorization (MF) or adaptive knearest-neighbor (kNN). Even though these methods are designed for the item prediction task of personalized ranking, none of them is directly optimized for ranking. In this paper we present a generic optimization criterion BPR-Opt for personalized ranking that is the maximum posterior estimator derived from a Bayesian analysis of the problem. We also provide a generic learning algorithm for optimizing models with respect to BPR-Opt. The learning method is based on stochastic gradient descent with bootstrap sampling. We show how to apply our method to two state-of-the-art recommender models: matrix factorization and adaptive kNN. Our experiments indicate that for the task of personalized ranking our optimization method outperforms the standard learning techniques for MF and kNN. The results show the importance of optimizing models for the right criterion.

In this paper we build on previous work which uses inferences techniques, in particular Markov Chain Monte Carlo (MCMC) methods, to solve parameterized control problems. We propose a number of modifications in order to make this approach more practical in general, higher-dimensional spaces. We first introduce a new target distribution which is able to incorporate more reward information from sampled trajectories. We also show how to break strong correlations between the policy parameters and sampled trajectories in order to sample more freely. Finally, we show how to incorporate these techniques in a principled manner to obtain estimates of the optimal policy.