2B. Statistical methods for textual data

14:30 - 15:40, Aula 10

Organizer: Michelangelo Misuraca

Chair: Michelangelo Misuraca

PROCSIMA: Probability Distribution Clustering Using Similarity Matrix Analysis

Marco Ortu

Abstract: This study presents PROCSIMA, a methodological approach to document clustering, that defines a similarity metric derived from Jensen-Shannon divergence, for measuring similarities between topic probability distributions obtained from Topic Modeling techniques, such as Latent Dirichlet Allocation (LDA). Unlike conventional approaches that allocate documents to a singular, most pertinent topic, PROCSIMA allocates the clustering of documents by considering their comprehensive topic distribution. By transforming the similarity matrix into an adjacency matrix and subsequently applying community detection algorithms it defines document clusters. Empirical validation on both synthetic and real-world datasets is performed by PROCSIMA by bootstrapping the optimal number of network communities to outperform traditional clustering methods.

Click here to view the abstract.

Causal inference from texts: a random-forest approach

Chiara Di Maria, Alessandro Albano, Mariangela Sciandra and Antonella Plaia

Abstract: This paper employs causal random forests to analyse textual reviews in an e-commerce context, specifically investigating the causal impact of sentiment on the Positive Feedback Count (PFC). The PFC denotes the number of users who found the review helpful. The results uncover a negative causal effect, indicating that transitioning from negative to positive sentiment reduces the count of users perceiving a review as helpful. The analysis further explores heterogeneity, highlighting the nuanced influence of specific words and variations in treatment effects. This research underscores the efficacy of causal inference in elucidating the intricate dynamics between sentiment and the perceived utility of reviews.