Community detection agotihms in graph database model for a recommender system

Sustavi preporuke temeljeni na kolaborativnom filtriranju (engl. collaborative filtering recommender system) za identifikaciju sličnih proizvoda koriste algoritme za otkrivanje zajednica (engl. community detection algorithms) u grafovima. S tim ciljem je u ovom radu provedena analiza tih algoritama za pronalaženje zajednica bez preklapanja. Analizirani algoritmi su: Louvain, Leiden, širenje oznaka, infomap i Constant Potts Model. Algoritmi su provedeni nad stvarnim grafovima od kojih su neki pohranjeni u graf bazu podataka, a to su karate klub, firentinske obitelji, glavni gradovi svijeta i Amazon proizvodi te nad sintetičkim LFR grafovima. Dobivene zajednice uspoređene su po unutrašnjim metrikama modularnosti i provodnosti te vanjskim metrikama normalizirane F1 vrijednosti i normaliziranom uzajamnom sadržaju informacija.Collaborative filtering recommender systems use algorithms for community detection in graphs to identify similar products. This study analyzes community detection algorithms for crisp communities for this purpose. The analyzed algorithms include Louvain, Leiden, label propagation, infomap, and Constant Potts Model. These algorithms were applied to real-world graphs, some of which were stored in a graph database, including the karate club, Florentine families, world capitals, Amazon products, as well as synthetic LFR graphs. The obtained communities were compared using internal metrics such as modularity and conductance, as well as external metrics such as normalized F1 score and normalized mutual information