Speeding up tensor based recommenders with clustered tag space and improving quality of recommendations with Non-negative tensor factorization
Authors
Leginus, Martin ; Žemaitis, Valdas
Term
2. term
Education
Publication year
2011
Submitted on
2011-05-31
Pages
30
Abstract
Social tagging hjælper brugere med at mærke og kategorisere weblinks, billeder, film m.m. Disse tags afspejler brugernes præferencer og kan bruges til at anbefale nye elementer. Moderne sociale tagging‑systemer repræsenterer forbindelserne mellem tags, brugere og elementer som en tredjeordens tensor (en tredimensionel datastruktur). Ved at faktorisere denne tensor med metoder som Higher Order Singular Value Decomposition (HOSVD) og Canonical Decomposition (CD) kan man finde skjulte mønstre og relationer. Disse systemer samler dog meget store datamængder, hvilket gør tensorfaktorisering langsom og kræver meget hukommelse. Vi foreslår tre forbedringer: (1) reducere tag‑rummet ved at klustre semantisk lignende tags, så beregningerne bliver hurtigere og bruger mindre hukommelse, samtidig med at anbefalingskvaliteten bevares; (2) inddrage personlig forhåndsviden for at øge præcisionen i tensor‑baserede anbefalere; og (3) bruge ikke‑negativ tensorfaktorisering (NTF) for at undgå negative værdier, som er svære at fortolke. Til sidst kombinerer vi disse tilgange for at forbedre både anbefalingskvalitet og beregningstid og præsenterer lovende eksperimentelle resultater.
Social tagging lets people label and organize web links, photos, movies, and more. These tags reflect user preferences and help recommenders suggest new items. Modern social tagging systems represent the connections among tags, users, and items as a third‑order tensor (a three‑dimensional data structure). By factoring this tensor with methods such as Higher Order Singular Value Decomposition (HOSVD) and Canonical Decomposition (CD), we can uncover hidden patterns and relationships. However, these systems collect very large datasets, which make tensor factorization slow and memory‑intensive. We propose three improvements: (1) reduce the tag space by clustering semantically similar tags, so computations run faster and use less memory while keeping recommendation quality; (2) incorporate personal prior knowledge to increase the precision of tensor‑based recommenders; and (3) use Non‑negative Tensor Factorization (NTF) to avoid negative values that are hard to interpret. Finally, we combine these approaches to improve both recommendation quality and computation time, and we present promising experimental results.
[This abstract was generated with the help of AI]
Keywords
Documents
Other projects by the authors