Pseudo-document simulation for comparing LDA, GSDMM and GPM topic models on short and sparse text using Twitter data.

Topic models are a useful and popular method to find latent topics of documents. However, the short and sparse texts in social media micro-blogs such as Twitter are challenging for the most commonly used Latent Dirichlet Allocation (LDA) topic model. We compare the performance of the standard LDA topic model with the Gibbs Sampler Dirichlet Multinomial Model (GSDMM) and the Gamma Poisson Mixture Model (GPM), which are specifically designed for sparse data. To compare the performance of the three models, we propose the simulation of pseudo-documents as a novel evaluation method. In a case study with short and sparse text, the models are evaluated on tweets filtered by keywords relating to the Covid-19 pandemic. We find that standard coherence scores that are often used for the evaluation of topic models perform poorly as an evaluation metric. The results of our simulation-based approach suggest that the GSDMM and GPM topic models may generate better topics than the standard LDA model.

Pseudo-document simulation for comparing LDA, GSDMM and GPM topic models on short and sparse text using Twitter data

Topic models are a useful and popular method to find latent topics of documents. However, the short and sparse texts in social media micro-blogs such as Twitter are challenging for the most commonly used Latent Dirichlet Allocation (LDA) topic model. We compare the performance of the standard LDA topic model with the Gibbs Sampler Dirichlet Multinomial Model (GSDMM) and the Gamma Poisson Mixture Model (GPM), which are specifically designed for sparse data. To compare the performance of the three models, we propose the simulation of pseudo-documents as a novel evaluation method. In a case study with short and sparse text, the models are evaluated on tweets filtered by keywords relating to the Covid-19 pandemic. We find that standard coherence scores that are often used for the evaluation of topic models perform poorly as an evaluation metric. The results of our simulation-based approach suggest that the GSDMM and GPM topic models may generate better topics than the standard LDA model.

An iterative topic model filtering framework for short and noisy user-generated data: analyzing conspiracy theories on twitter.

Conspiracy theories have seen a rise in popularity in recent years. Spreading quickly through social media, their disruptive effect can lead to a biased public view on policy decisions and events. We present a novel approach for LDA-pre-processing called Iterative Filtering to study such phenomena based on Twitter data. In combination with Hashtag Pooling as an additional pre-processing step, we are able to achieve a coherent framing of the discussion and topics of interest, despite of the inherent noisiness and sparseness of Twitter data. Our novel approach enables researchers to gain detailed insights into discourses of interest on Twitter, allowing them to identify tweets iteratively that are related to an investigated topic of interest. As an application, we study the dynamics of conspiracy-related topics on US Twitter during the last four months of 2020, which were dominated by the US-Presidential Elections and Covid-19. We monitor the public discourse in the USA with geo-spatial Twitter data to identify conspiracy-related contents by estimating Latent Dirichlet Allocation (LDA) Topic Models. We find that in this period, usual conspiracy-related topics played a marginal role in comparison with dominating topics, such as the US-Presidential Elections or the general discussions about Covid-19. The main conspiracy theories in this period were the ones linked to "Election Fraud" and the "Covid-19-hoax." Conspiracy-related keywords tended to appear together with Trump-related words and words related to his presidential campaign.

Gradient boosting for linear mixed models.

Gradient boosting from the field of statistical learning is widely known as a powerful framework for estimation and selection of predictor effects in various regression models by adapting concepts from classification theory. Current boosting approaches also offer methods accounting for random effects and thus enable prediction of mixed models for longitudinal and clustered data. However, these approaches include several flaws resulting in unbalanced effect selection with falsely induced shrinkage and a low convergence rate on the one hand and biased estimates of the random effects on the other hand. We therefore propose a new boosting algorithm which explicitly accounts for the random structure by excluding it from the selection procedure, properly correcting the random effects estimates and in addition providing likelihood-based estimation of the random effects variance structure. The new algorithm offers an organic and unbiased fitting approach, which is shown via simulations and data examples.