Browsing by Subject "Topic models"

Now showing 1 - 3 of 3

Appropriate, accessible and appealing probabilistic graphical models
(2017-05-04) Inouye, David Iseri; Dhillon, Inderjit S.; Ravikumar, Pradeep; Mooney, Raymond J; Huang, Qixing; Wallace, Byron C
Appropriate - Many multivariate probabilistic models either use independent distributions or dependent Gaussian distributions. Yet, many real-world datasets contain count-valued or non-negative skewed data, e.g. bag-of-words text data and biological sequencing data. Thus, we develop novel probabilistic graphical models for use on count-valued and non-negative data including Poisson graphical models and multinomial graphical models. We develop one generalization that allows for triple-wise or k-wise graphical models going beyond the normal pairwise formulation. Furthermore, we also explore Gaussian-copula graphical models and derive closed-form solutions for the conditional distributions and marginal distributions (both before and after conditioning). Finally, we derive mixture and admixture, or topic model, generalizations of these graphical models to introduce more power and interpretability. Accessible - Previous multivariate models, especially related to text data, often have complex dependencies without a closed form and require complex inference algorithms that have limited theoretical justification. For example, hierarchical Bayesian models often require marginalizing over many latent variables. We show that our novel graphical models (even the k-wise interaction models) have simple and intuitive estimation procedures based on node-wise regressions that likely have similar theoretical guarantees as previous work in graphical models. For the copula-based graphical models, we show that simple approximations could still provide useful models; these copula models also come with closed-form conditional and marginal distributions, which make them amenable to exploratory inspection and manipulation. The parameters of these models are easy to interpret and thus may be accessible to a wide audience. Appealing - High-level visualization and interpretation of graphical models with even 100 variables has often been difficult even for a graphical model expert---despite visualization being one of the original motivators for graphical models. This difficulty is likely due to the lack of collaboration between graphical model experts and visualization experts. To begin bridging this gap, we develop a novel "what if?" interaction that manipulates and leverages the probabilistic power of graphical models. Our approach defines: the probabilistic mechanism via conditional probability; the query language to map text input to a conditional probability query; and the formal underlying probabilistic model. We then propose to visualize these query-specific probabilistic graphical models by combining the intuitiveness of force-directed layouts with the beauty and readability of word clouds, which pack many words into valuable screen space while ensuring words do not overlap via pixel-level collision detection. Although both the force-directed layout and the pixel-level packing problems are challenging in their own right, we approximate both simultaneously via adaptive simulated annealing starting from careful initialization. For visualizing mixture distributions, we also design a meaningful mapping from the properties of the mixture distribution to a color in the perceptually uniform CIELUV color space. Finally, we demonstrate our approach via illustrative visualizations of several real-world datasets.
Infinite-word topic models for digital media
(2014-05) Waters, Austin Severn; Miikkulainen, Risto
Digital media collections hold an unprecedented source of knowledge and data about the world. Yet, even at current scales, the data exceeds by many orders of magnitude the amount a single user could browse through in an entire lifetime. Making use of such data requires computational tools that can index, search over, and organize media documents in ways that are meaningful to human users, based on the meaning of their content. This dissertation develops an automated approach to analyzing digital media content based on topic models. Its primary contribution, the Infinite-Word Topic Model (IWTM), helps extend topic modeling to digital media domains by removing model assumptions that do not make sense for them -- in particular, the assumption that documents are composed of discrete, mutually-exclusive words from a fixed-size vocabulary. While conventional topic models like Latent Dirichlet Allocation (LDA) require that media documents be converted into bags of words, IWTM incorporates clustering into its probabilistic model and treats the vocabulary size as a random quantity to be inferred based on the data. Among its other benefits, IWTM achieves better performance than LDA while automating the selection of the vocabulary size. This dissertation contributes fast, scalable variational inference methods for IWTM that allow the model to be applied to large datasets. Furthermore, it introduces a new method, Incremental Variational Inference (IVI), for training IWTM and other Bayesian non-parametric models efficiently on growing datasets. IVI allows such models to grow in complexity as the dataset grows, as their priors state that they should. Finally, building on IVI, an active learning method for topic models is developed that intelligently samples new data, resulting in models that train faster, achieve higher performance, and use smaller amounts of labeled data.
New topic detection in microblogs and topic model evaluation using topical alignment
(2014-05) Rajani, Nazneen Fatema; Baldridge, Jason
This thesis deals with topic model evaluation and new topic detection in microblogs. Microblogs are short and thus may not carry any contextual clues. Hence it becomes challenging to apply traditional natural language processing algorithms on such data. Graphical models have been traditionally used for topic discovery and text clustering on sets of text-based documents. Their unsupervised nature allows topic models to be trained easily on datasets meant for specific domains. However the advantage of not requiring annotated data comes with a drawback with respect to evaluation difficulties. The problem aggravates when the data comprises microblogs which are unstructured and noisy. We demonstrate the application of three types of such models to microblogs - the Latent Dirichlet Allocation, the Author-Topic and the Author-Recipient-Topic model. We extensively evaluate these models under different settings, and our results show that the Author-Recipient-Topic model extracts the most coherent topics. We also addressed the problem of topic modeling on short text by using clustering techniques. This technique helps in boosting the performance of our models. Topical alignment is used for large scale assessment of topical relevance by comparing topics to manually generated domain specific concepts. In this thesis we use this idea to evaluate topic models by measuring misalignments between topics. Our study on comparing topic models reveals interesting traits about Twitter messages, users and their interactions and establishes that joint modeling on author-recipient pairs and on the content of tweet leads to qualitatively better topic discovery. This thesis gives a new direction to the well known problem of topic discovery in microblogs. Trend prediction or topic discovery for microblogs is an extensive research area. We propose the idea of using topical alignment to detect new topics by comparing topics from the current week to those of the previous week. We measure correspondence between a set of topics from the current week and a set of topics from the previous week to quantify five types of misalignments: \textit{junk, fused, missing} and \textit{repeated}. Our analysis compares three types of topic models under different settings and demonstrates how our framework can detect new topics from topical misalignments. In particular so-called \textit{junk} topics are more likely to be new topics and the \textit{missing} topics are likely to have died or die out. To get more insights into the nature of microblogs we apply topical alignment to hashtags. Comparing topics to hashtags enables us to make interesting inferences about Twitter messages and their content. Our study revealed that although a very small proportion of Twitter messages explicitly contain hashtags, the proportion of tweets that discuss topics related to hashtags is much higher.