Attention Please! A Hybrid Resource Recommender Mimicking Attention-Interpretation Dynamics
0 likes678 views
The document discusses a novel hybrid resource recommender that combines collaborative filtering and a computational model of human category learning to enhance informal learning technologies in SME clusters. An evaluation of this approach shows improvements in recommendation accuracy across multiple datasets compared to standard collaborative filtering methods. Future work will explore dynamic recommendation logic and a user-specific sustain network, as well as address computational costs involved in the model.
Attention Please! A Hybrid Resource Recommender Mimicking Attention-Interpretation Dynamics
- 1. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Learning Layers Scaling up Technologies for Informal Learning in SME Clusters Attention Please! A Hybrid Resource Recommender Mimicking Attention-Interpretation Dynamics Paul Seitlinger, Dominik Kowald, Simone Kopeinik, Ilire Hasani-Mavriqi, Tobias Ley, Elisabeth Lex 1 Austrian Science Fund: P 25593-G22
- 2. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu What will this talk be about? • Resource RecommendaBon (user-‐based CollaboraBve Filtering) • A computaBonal model of human category learning (SUSTAIN) • A novel hybrid recommender approach that combines both to further personalize and improve CF 2
- 3. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Why? • Recommender research exploits digital traces of social acBons and interacBons – E.g. CF suggests resources of most similar users • EnBBes of different quality (e.g., users, resources, tags) are related to each other • In CF, users just another enBty • Structuralist simplificaBon • Neglects nonlinear, user-‐resource dynamics that shape a"enBon and interpretaBon • No ranking of resources in CF 3
- 4. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu SUSTAIN (Love et al., 2004) 4 • Resource represented by features • Cluster(s) H – Vector of values along the n feature dimensions – Fields of interest • A"enBonal weights wi: – Importance of feature for user • Training (for each resource R) – Start with one cluster – Form new cluster if sim(R,H) < T – AdjusBng Hj and wi aYer each run • TesBng – Compare features of candidate to best cluster
- 5. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Our Approach: SUSTAIN+CFU • Step 1: Create candidate set Cu for target user u (top 100 resources of CFU • Step 2: Train SUSTAIN network of target user u • Step 3: Apply each candidate c of Cu to network • Step 4: Hybrid approach 5
- 6. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Evaluation: Datasets • Social tagging systems – Freely available for scienBfic purposes – Topics can be easily derived from tagging data (e.g., Krestel et al., 2010) à Latent Dirichlet AllocaBon (LDA) with 500 topics • No p-‐core pruning but deleted unique resources 6
- 7. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Evaluation: Method and Metrics • Training and test-‐set splits • Per user: 20% most recent for tesBng, 80% for training • Retains chronological order à predict future based on the past • Comparison of top-‐20 recommended resources with relevant resources from test-‐set • Metrics – nDCG@20 – MAP@20 – Precision / Recall plots (k = 1 – 20) 7
- 8. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Baseline Algorithms • Most Popular (MP) • User-‐Based CollaboraBve Filtering (CFU) • Resource-‐Based CollaboraBve Filtering (CFR) • Content-‐based Filtering using Topics (CBT) • SUSTAIN+CFU à Available in the open source TagRec framework • Weighted Regularized Matrix FactorizaBon (WRMF) à MyMediaLite 8
- 9. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Results 9 • SUSTAIN+CFU improves CFU on all three datasets • CiteULike: High average topic similarity per user, CFR wins • Delicious: Mutual-‐fan crawling strategy, WRMF wins
- 10. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Evaluation: Open Issues • Datasets – Other Delicious dataset, LastFM, MovieLens – External/other feature (not dependent on LDA) • Other metrics – Diversity, Serendipity, Coverage • ComputaBonal Costs – Our experiments showed that our approach is much faster than CFR and especially WRMF • Although LDA is needed – RunBme experiment is needed + computaBonal complexity • Online evaluaBon – Learning Layers field study 10
- 11. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Future Work • Technical – CF-‐independent variant • RecommendaBons solely based on user-‐specific SUSTAIN network – Detailed analysis of computaBonal costs • Conceptual – Dynamic recommendaBon logic • Exploring relaBonship between a"enBonal focus and novelty seeking and use this for recommendaBon 11
- 12. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Take Away Messages • Our approach SUSTAIN + CFU can improve CF predicBons – More robust in terms of accuracy esBmates – From our observaBon: less complex in terms of computaBonal efforts • User-‐resource dynamics, if modelled with a connecBonist approach, can help gain a deeper understanding of Web interacBons in terms of a"enBon, categorizaBon and decision making 12
- 13. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Code and Framework 13 • TagRec framework • hAps://github.com/learning-‐layers/TagRec/ • Framework for developing and evaluaBng new recommender algorithms in folksonomies • Contains our approach, the baseline algorithms and the evaluaBon protocol and metrics • Capable of tag, resource and user recommendaBons • Used as recommender engine in the Learning Layers EU project • Links to the datasets we used: – BibSonomy (2013-‐07-‐01): h"p://www.kde.cs.uni-‐ kassel.de/bibsonomy/dumps/ – CiteULike (2013-‐03-‐10): h"p://www.citeulike.org/faq/data.adp – Delicious (2011-‐05-‐01): h"p://files.grouplens.org/datasets/hetrec2011/ hetrec2011-‐ delicious-‐2k.zip
- 14. h"p://Learning-‐Layers-‐eu – Scaling up Technologies for Informal Learning in SME Clusters – layers@learning-‐layers.eu Thank you for your attention! QuesJons? Elisabeth Lex elisabeth.lex@tugraz.at Ass. Prof. at Graz University of Technology (Austria) Head of Social CompuBng at Know-‐Center (Austria) 14