Meta Ensembling for Recommender Systems
Authors
Kristensen, Frederik Kirk ; Klemmensen, Mathias Rosgaard
Term
4. term
Education
Publication year
2019
Submitted on
2019-05-30
Pages
8
Abstract
This thesis investigates how to improve ensemble recommender systems by customizing linear stacking weights using only metadata derived from the rating matrix. Building on collaborative filtering and matrix factorization, we motivate ensembling after prior attempts to add SPPMI-based co-occurrence signals to factorization showed signs of a capacity limit. We introduce two meta-ensembling methods: Neural Network Meta Stacking (NNMS), which learns per-user weights for constituent recommenders from user metadata, and Meta Clustered Stacking (MCS), which clusters users by metadata and assigns per-cluster weights via constrained grid search. As part of the baselines, we also propose two KNN-based recommenders, User Positive Co-occurrence (UPC) and Item Positive Co-occurrence (IPC), that use SPPMI as similarity; in our final ensembles we combine UPC, an item-based KNN, and matrix factorization. Across multiple datasets, our ensembles generally outperform individual recommenders, but MCS typically does not surpass a naive averaging ensemble, except on one dataset where the average is hindered by a particularly weak constituent model. NNMS fails to yield useful gains despite extensive tuning of architectures and loss functions. Overall, the results suggest that metadata-driven meta-ensembling with simple matrix-derived signals is feasible and can improve over single models, yet it is challenging to outperform straightforward averaging without further advances.
Denne afhandling undersøger, hvordan ensemblinger af anbefalingssystemer kan forbedres ved at tilpasse vægte i lineær stabling ud fra metadata, der udelukkende kan udledes af selve vurderingsmatricen. Med udgangspunkt i kollaborativ filtrering og matrixfaktorisering motiveres ensemblinger som alternativ, efter at tidligere forsøg på at tilføje ekstra SPPMI-baserede samforekomstdata til faktoriseringsmodeller gav tegn på en kapacitetsgrænse. Vi præsenterer to metoder til meta-ensembling: Neural Network Meta Stacking (NNMS), der lærer at mappe brugermetadata til per-bruger vægte for de sammensatte anbefalere, samt Meta Clustered Stacking (MCS), der klyngeopdeler brugere efter metadata og tildeler per‑klynge vægte via grid-søgning under vægtsum-begrænsning. Som en del af baseline-sættet introducerer vi to KNN-baserede anbefalere, User Positive Co-occurrence (UPC) og Item Positive Co-occurrence (IPC), der anvender SPPMI som lighedsmål; i de endelige ensembles kombineres UPC, en item-baseret KNN og matrixfaktorisering. I eksperimenter på flere datasæt slår vores ensembles typisk de individuelle anbefalere, men MCS overgår som hovedregel ikke en naiv middelværdi-ensemble, bortset fra ét datasæt hvor den naive kombination hæmmes af en særligt svag basismodel. NNMS leverer ikke brugbare forbedringer trods omfattende arkitektur- og tabsfunktionsafprøvning. Resultaterne indikerer, at meta-ensembling med simple matrixafledte metadata er gennemførligt og kan give moderate gevinster over enkeltmodeller, men at det er vanskeligt at slå en simpel gennemsnits-ensemble uden yderligere forbedringer.
[This apstract has been generated with the help of AI directly from the project full text]
Keywords