Multi-objective reinforcement learning approach for trip recommendation

Trip recommendation is an intelligent service that provides personalized itinerary plans for tourists in unfamiliar cities. It aims to construct a series of ordered POIs that maximizes user travel experiences with temporal and spatial constraints. When appending a candidate POI to the recommended trip, it is critical to capture users' dynamic preferences according to real-time context. Meanwhile, the diversity and popularity of the POIs in the personalized trip play an important role in users' selections. To address these challenges, in this article, we propose a MORL-Trip (short for Multi -Objective Reinforcement Learning for Trip Recommendation) approach. MORL-Trip models the personalized trip recommendation as a Markov Decision Process (MDP), and implements it upon the Actor-Critic framework. MORL-Trip enhances the state representation with sequential information, geographic information and order information to learn user's context from real-time location. In addition, MORL-Trip augments the standard Critic component by designing a composite reward function to enforce three principal objectives: accuracy, popularity and diversity. We conduct extensive experiments on the public datasets and compare the performance of MORL-Trip with the most advanced methods to verify its superiority, and show the importance of reinforcing popularity and diversity as complementary objectives in the personalized trip recommendation.