Clustering users' po is visit trajectories for next poi recommendation
Download as PPTX, PDF0 likes36 views
This document describes a system for clustering users' points of interest (POI) visit trajectories and using the clusters to generate personalized next-POI recommendations. It clusters trajectories into topics to learn generalized tourist behavior patterns. The system models user behavior as a Markov decision process and uses inverse reinforcement learning to estimate reward functions from observed behavior. It generates two types of recommendations: cluster behavior-based recommendations that leverage cluster policies when individual data is sparse, and a hybrid approach that combines cluster and individual user models. An evaluation shows it outperforms baseline methods on metrics like reward and novelty.
Clustering users' po is visit trajectories for next poi recommendation
- 1. David Massimo Free University of Bolzano damassimo@unibz.it Clustering Users' POIs Visit Trajectories for Next-POI Recommendation Francesco Ricci Free University of Bolzano fricci@unibz.it
- 2. 2
- 3. RECOMMENDER SYSTEMS Software tools that ease human decision making process 3 Recommendations are based on RATINGS REVIEWS
- 4. Motivation SUNNY AFTERNOON MUSEUM 4 Adapting recommendation to the learnt user behavior Understanding User preferences in Context Support tourists in finding points of interest (POIs) Huge variety of different POIs Leverage the power of sensors like GPS
- 5. G. Shani, D. Heckerman, and R. I. Brafman, “An mdp-based recommender system“ (2005) SEQUENTIAL RECOMMENDATIONS OBSERVATION MODEL PATTERNS Related Works O. Moling, L. Baltrunas, and F. Ricci, “Optimal radio channel recommendations with explicit and implicit feedback” (2012) Pattern-discovery (B. Mobasher et al., D. Jannach) Reinforcement Learning (G. Shani et al., O. Moling et al.) ACTIONS SEQUENCES DM POLICIES ADAPTED CONTENT B. Mobasher, H. Dai, T. Luo, and M. Nakagawa. “Using Sequential and Non Sequential Patterns in Predictive Web Usage Mining Tasks” (2002) D. Jannach, I. Kamehkhosh, and L. Lerche. “Leveraging multi-dimensional user models for personalized next-track music recommendation” (2017) 5
- 6. 6 WHAT A USER WILL DO ANYWAY
- 9. COMPLETE PICTURE 9 STRATEGIES Actions POI Features Context BEHAVIOURAL MODEL RECOMMENDATIONS Technical Approach
- 10. 10 STRATEGIES Actions POI Features Context BEHAVIOURAL MODEL RECOMMENDATIONS C. I. Muntean, F. M. Nardini, F. Silvestri, and R. Baraglia. 2015. On Learning Prediction Models for Tourists Paths. (2015) Complete Picture CASE STUDY Context POI Features Geolocalized Temporally ordered Weather summary Temperature Daytime Category Historic period Historic related person User POI-visit trajectories 1635 (Muntean et al.)
- 12. 12 PROBLEM SETUP Markov Decision Process (MDP) S State space A Action space T Transition model Z Users observations r Reward 𝛄 Discount factor Technical Approach s a T(s’ | s, a) s1 s2 +50 -10 POLICY
- 13. 13 Reward Action-selection policy THE BEHAVIOURAL MODEL IS LEARNT IN TERMS OF Inverse Reinforcement Learning (IRL) Apprenticeship Learning IRL tries to learn the (unknown) reward function, defined as: 1 0 1 .... 0 ....
- 14. 14 Often there is not enough user specific behavioural data
- 15. 15 DOC-Like representation TOPICS DOC-Like representation POI-visit trajectory TOPICS Documents all the POI-Visit trajectories NMF Clustering
- 16. 16 Term Cluster A Cluster B Cluster C Cluster D Cluster E 1 morning hot cloudy warm freezing 2 cold afternoon cold cloudy cloudy 3 square 16th century church 14th century afternoon 4 palace palace square church 14th century 5 15th century church 13th century square palace 6 13th century square palace building building 7 church 19th century rain palace 13th century 8 night 13th century museum bridge church 9 Dante museum Brunelleschi 13th century Foggini 10 10th century Brunelleschi F. Tadda 19th century 19th century #Traj. 368 339 341 297 153
- 17. 17 Clusters A,D Cluster E 15th 16th 13th Interesting Differences
- 18. 18
- 19. 19 Minimizes the impact of suboptimal behaviour Allows to deal with small sized databases with small amount of interactions per user Estimate the global preferences of different Users’ Segments Understand the mobility patterns of different Users’ Segments Benefits Clustering Generalized Tourists Behaviour Behaviour Learning
- 21. 21 Recommendations VALUE OF TAKING AN ACTION s a T(s’ | s, a) s1 s2 Technical Approach Policy Reward Q 𝜋(s,a) Expressed by the quantity
- 22. 22 CBR - Cluster Behaviour Based Recommendations1 When only few observation of a user are available the general user behaviour of the cluster the user belong to is used to suggest optimal actions. Optimizing for a Segment of Users Cluster action-selection policy optimal action
- 23. 23 2 When more observations for a user are available, recommendations are generated by combining the scores of CBR recommendations with next-POI visit scores computed on the base of a user specific preference model. Hybrid Optimization: Segment & Individual CBHR - Cluster Behaviour Hybrid Recommendations VISIT ACTIONS NEW EXHIBITION ART USER PREFERENCE MODEL + Cluster action-selection policy
- 24. 24 HOW CBR AND CBHR COMPARE TO THE STATE OF THE ART?
- 25. 25 Reward: measures the average increase of the reward of the recommended actions compared to the observed one. Precision: percentage of recommended visits that match the observed one. Dissimilarity: of the recommended actions from the observed one (train set). Novelty: how unpopular are the recommended actions. We evaluated our RS across 4 different dimensions
- 26. 26 CBR CBHR kNN Reward@5 0.8791 0.7788 0.4204 Diss.@5 0.8923 0.8706 0.8578 Precision@5 0.0834 0.0514 0.1518 Novelty@5 0.0002 0.1878 0.0000 Reward@10 0.8304 0.7695 0.3863 Diss.@10 0.8901 0.8694 0.8693 Precision@10 0.0751 0.0464 0.1285 Novelty@10 0.0503 0.2640 0.0000 Baseline Next-item recommendation
- 27. 27 Online Testing Smart Fiemme In collaboration with Ectrl Solutions South Tyrol Guide In collaboration with IDM & Ectrl Solutions
- 28. 28 Allows to deal with small sized databases with small amount of interactions per user Learn a generalized tourist behaviour Provision interesting recommendations Can we design sustainable tourism policies by leveraging the insight given by tourists behaviour analytics? Can we promote these policies through proper recommendation strategies? Q Wrap Up
- 29. THANK YOU! David Massimo Free University of Bolzano damassimo@unibz.it Francesco Ricci Free University of Bolzano fricci@unibz.it