Structural Similarity-Based Matrix Factorization Using Type Extension Trees for Item Cold-Start
Authors
Vorm, Casper ; Eilers, Thorbjørn Leonard ; Rasmussen, Daniel Drejer
Term
4. term
Education
Publication year
2021
Submitted on
2021-06-10
Pages
16
Abstract
This thesis addresses the item cold-start problem in recommender systems by combining structural similarity learned from relational item metadata with matrix factorization. The proposed pipeline first uses Type Extension Trees (TETs) to capture structural similarities between items from graph-structured metadata and summarizes them in an item–item similarity matrix; second, this matrix is injected into Learning Local Collective Embeddings (LCE) to jointly embed items and users in a shared low-dimensional space. The method, termed Structural Similarity Factorization (LCE-SSF), aims to strengthen the content-based signal so that new items with no interactions can be recommended to likely interested users. We formalize the problem, describe how TET-based similarities are computed and integrated with LCE, and evaluate the approach in item cold-start settings on three publicly available datasets against LCE and other hybrid baselines. Experiments indicate a slight overall improvement over the original LCE in item cold-start scenarios; the method outperforms LCE on some datasets and, when used as a graph regularization term, yields consistent gains. These results suggest that exploiting structural relationships among items can enhance matrix factorization models for cold-start recommendation.
Dette speciale adresserer item‑koldstart‑problemet i anbefalingssystemer ved at kombinere strukturel lighed lært fra relationelle metadata om items med matrixfaktorisering. Den foreslåede pipeline anvender først Type Extension Trees (TET’er) til at indfange strukturelle ligheder mellem items ud fra grafstrukturerede metadata og opsummerer dem i en item‑item‑lighedsmatrix; derefter integreres denne matrix i Learning Local Collective Embeddings (LCE) for samlet at indlejre items og brugere i et fælles, lavdimensionelt rum. Metoden, benævnt Structural Similarity Factorization (LCE‑SSF), har til formål at styrke det indholdsbaserede signal, så nye items uden interaktioner kan anbefales til sandsynligt interesserede brugere. Vi formaliserer problemet, beskriver hvordan TET‑baserede ligheder beregnes og integreres i LCE, og evaluerer tilgangen i item‑koldstart på tre offentligt tilgængelige datasæt mod LCE og andre hybride baselines. Eksperimenterne viser en mindre samlet forbedring i forhold til den oprindelige LCE i item‑koldstart; metoden overgår LCE på nogle datasæt og giver, når den bruges som grafregularisering, konsekvente gevinster. Resultaterne indikerer, at udnyttelse af strukturelle relationer mellem items kan styrke matrixfaktoriseringsmodeller til koldstart‑anbefaling.
[This apstract has been generated with the help of AI directly from the project full text]