Data-Driven Destination Recommender Systems Linus W. Dietz Technical University of Munich Department of Informatics Chair of Connected Mobility (I11) UMAP’18, Singapore July 10, 2018
Introduction Problem Recommend composite trips of global travel destinations “I want to travel to South-East Asia for six weeks in summer to experience culture, good food and go hiking in the mountains. I have a budget of $1500.” Motivation Independent travel planning is complex , information is scattered , outdated , and of uncertain quality Challenges � c Created by Freepik � Find distinct touristic regions � Classification of tourist destinations � User modeling with little user effort � Recommendation algorithm Linus W. Dietz (TUM) 2
Overview Items User Algorithms � Item discovery � Effective and effortless � Content-based filtering preference elicitation � Constraint satisfaction � Diversity of activities � Item classification � Traveler clustering � Durations of item consumption Linus W. Dietz (TUM) 3
Data Mining & Domain Modeling Recommendation items: set of travel regions Combination of heterogeneous data sources Item discovery � Mismatch between political regions and tourist destinations! � Employ hierarchical region tree � Make the granularity of destinations dependent on the query area Item characterization � What are the characteristics , attractions , and activities of a destination? � Aggregation based on single points of interests � What is the typical duration of stay at a destination? � Analyze tourist mobility patterns for domain understanding Evaluation: Offline comparison with crowdsourcing and expert knowledge Contribution: framework for data-driven recommender systems Linus W. Dietz (TUM) 4
User Modeling Preference elicitation � Which activities are best suited for a traveler? � How can traveling preferences be elicited effectively with little effort ? Traveler clustering � What are the relevant features to characterize travelers ? � What distinct types of travelers are there? � How can the pace of the travel itinerary be personalized based on past trips? Evaluation: Controlled lab experiments Contribution: Novel approaches for domain-specific user modeling Linus W. Dietz (TUM) 5
Recommendation Algorithms Content-based recommendation under constraints . Knapsack problem! Personalize item consumption durations Ensure sufficient diversity within a trip Measure the benefits of explaining recommendations and critiquing Evaluation: Measure user satisfaction in an online field study Contribution: Constraint-based algorithms for composite trips Linus W. Dietz (TUM) 6
Current Progress Investigated traveler mobility patterns Linus W. Dietz (TUM) 7
Preliminary investigation of durations of stay Current Progress Linus W. Dietz (TUM) Days 0 3 6 9 12 15 18 21 Kuwait Dominican Rep. Cyprus Azerbaijan Costa Rica Ghana Kazakhstan Paraguay Australia Belarus Saudi Arabia Turkey Martinique Brazil United States Russia Colombia Mexico Philippines Canada New Zealand Indonesia Japan Argentina China United Kingdom Malaysia Venezuela Thailand Spain Singapore Un. Arab Emirates Italy Ireland Germany Sweden observations mean days France Czech Rep. Finland Netherlands Switzerland Belgium Denmark Hungary Estonia 0 500 1500 2500 3500 Observations 8
Recommend
More recommend
