Top k alternative itineraries for trip recommendation

Abstract

Trip itinerary recommendation finds an ordered sequence of Point-of-Interests (POIs) from a large number of candidate POIs in a city. Users visiting a new city can use such a recommendation system to easily find an itinerary to follow, given a source POI from where they want to start to a destination POI where they want to reach. But users may often want more than one itinerary to choose from. Also in many cases users adopt more than one trend of route when travelling from a particular source POI to a particular destination POI. Furthermore, users may also provide some constraints that they want the recommended itineraries to follow, for example visiting a must-see POI, a budget limit etc. In this thesis, we propose a deep learning-based framework, called DeepAltTrip, that learns to recommend top-k alternative itineraries a given source and destination POI. These alternative itineraries would not only be popular given the historical routes adopted by past users but also dissimilar (or diverse) to each other. Existing learning based trip recommendationsystemslearntorecommendonlyonesinglemostpopularitinerary. Search based techniques that recommend multiple itineraries fail to capture the semanticsofroutesadoptedbypastusers,andthesetechniquesalsorequireexplicit modeling of diversity which can vary greatly depending upon user query. To the best of our knowledge, we are the first to learn from historical trips and provide a setofalternativeitinerariestotheuserthatwouldbebothpopularanddiverse.Our proposed system does not require any explicit modeling of popularity or diversity during training or inference. DeepAltTrip consists of two major components: (i) Itinerary Net (ITRNet) which estimates the likelihood of POIs on an itinerary by usinggraphautoencodersandtwo(forwardandbackward)LSTMs;and(ii)Aroute generationproceduretogeneratekdiverseitinerariespassingthroughrelevantPOIs obtainedusingITRNet.Fortheroutegenerationstep,weproposeanovelsampling algorithmthatcanseamlesslyhandleawidevarietyofuser-definedconstraints.We conductedextensiveexperimentsoneightpopularreal-worlddatasets.Wecompared our method against the baselines in terms of popularity and diversity individually, and also in terms of a combined measurement that considers both popularity and diversity.ThetwovariantsofDeepAltTripwepropose,namelyDeepAltTrip-LSTM and DeepAltTrip-Samp, outperform the best performing baseline by upto 29.24% and 25.34% respectively, in the combined popularity and diversity measure under the default settings. We also performed experiments under various settings and the results show the effectiveness and efficacy of ourapproach.