Γενικό Σεμινάριο Τμήματος Ομιλία 04/12/24

Ομιλητής: Μωϋσής Μπουντουρίδης (Πρώην Μέλος ΔΕΠ, Τμήμα Μαθηματικών,
Πανεπιστήμιο Πατρών)

Τίτλος: Hyperlink Prediction in the Hypergraphs of Network Terms and Citations in Harrison C. White’s “Identity and Control”

Περίληψη: Using standard NLP techniques, the text of Harrison C. White’s major theoretical work, ‘Identity and Control: A Structural Theory of Social Action’ (Princeton University Press, 1992), was segmented into paragraphs within the book’s eight chapters. In each paragraph, two types of words were identified: (1) network terms, as listed in the book’s index, and (2) in-text citations, as catalogued in the references section. This segmentation generates two hypergraphs: Hterms = (Vterms, Eterms) and Hcitations = (Vcitations, Ecitations). Here, Vterms and Vcitations are the sets of nodes containing all terms and citations, respectively, found in the book, while Eterms and Ecitations are the sets of hyperlinks representing subsets of Vterms and Vcitations, respectively, corresponding to the book’s paragraphs. Essentially, the cardinality of both Eterms and Ecitations equals the number of paragraphs in the book. For obvious reasons, elements of Eterms and Ecitations will be called ‘term hyperlinks’ and ‘citation hyperlinks.’ To understand the consistency among network terms and citations in H.C. White’s book, we are addressing the problem of hypergraph reconstruction as a subtask of hyperlink prediction for the previous two hypergraphs. Specifically, we aim to reconstruct (infer) these hypergraphs using a subset of observed paragraphs, which correspond to subsets of term hyperlinks and citation hyperlinks. This involves predicting the likelihood that the unobserved (hidden) paragraphs might form hyperlinks in the hypergraphs. To evaluate the coherence of book chapters, we apply the hyperlink prediction model to all paragraphs within the current chapter as well as all preceding chapters. Naturally, the efficiency of hyperlink prediction, when performed cumulatively across chapters, is expected to be minimal for the first chapter and to progressively improve, reaching its peak for the final chapter. This trend is assessed using standard hyperlink prediction metrics such as precision, recall, F1-score, and AUC-ROC, among others.The particular method of hyperlink prediction that we are following here is a deep learning-based method, that of the Chebyshev spectral hyperlink predictor (CHESHIRE). This approach allows us to systematically analyze and understand the structural coherence among network terms and citations within H.C. White’s book.

Μετά το πέρας της ομιλίας θα υπάρχουν καφές, χυμοί και μπισκότα για τους παρευρισκόμενους.

Εκ μέρους της επιτροπής σεμιναρίου,

Δημήτρης