[Paper Review] Efficient and Effective Similar Subtrajectory Search with Deep Reinforcement Learning

Zheng Wang, Cheng Long+2 more|arXiv (Cornell University)|Mar 5, 2020

Data Management and Algorithms47 references4 citations

TL;DR

This paper introduces the Similar Subtrajectory Search (SimSub) problem, proposing a novel framework that uses deep reinforcement learning (RL) to efficiently and effectively identify the most similar subtrajectory from a trajectory database to a query trajectory. The RL-based algorithms, RLS and RLS-Skip, outperform heuristic and non-learning baselines in both accuracy and speed, achieving up to 20x faster performance than exact methods on DTW and Frechet metrics while maintaining high effectiveness.

ABSTRACT

Similar trajectory search is a fundamental problem and has been well studied over the past two decades. However, the similar subtrajectory search (SimSub) problem, aiming to return a portion of a trajectory (i.e., a subtrajectory) which is the most similar to a query trajectory, has been mostly disregarded despite that it could capture trajectory similarity in a finer-grained way and many applications take subtrajectories as basic units for analysis. In this paper, we study the SimSub problem and develop a suite of algorithms including both exact and approximate ones. Among those approximate algorithms, two that are based on deep reinforcement learning stand out and outperform those non-learning based algorithms in terms of effectiveness and efficiency. We conduct experiments on real-world trajectory datasets, which verify the effectiveness and efficiency of the proposed algorithms.

Motivation & Objective

• To address the gap in research on fine-grained trajectory similarity by introducing the Similar Subtrajectory Search (SimSub) problem, which identifies the most similar portion of a data trajectory to a query trajectory.
• To develop a suite of algorithms—exact, approximate, and splitting-based—that balance effectiveness and efficiency for SimSub.
• To explore the use of deep reinforcement learning to learn optimal splitting policies for data trajectories, improving search performance over heuristic methods.
• To evaluate the proposed methods on real-world datasets, demonstrating superior performance in both effectiveness and efficiency compared to existing approaches.

Proposed method

• The paper formulates the subtrajectory splitting process as a Markov Decision Process (MDP), where each point in a data trajectory is a state, and the agent decides whether to split at that point to form candidate subtrajectories.
• Two deep RL-based algorithms, RLS and RLS-Skip, are proposed: RLS uses a policy network to predict split decisions, while RLS-Skip incorporates a skip-connection mechanism to improve training stability and efficiency.
• The algorithms are trained using a reward function that encourages the selection of subtrajectories with high similarity to the query trajectory, using similarity measures such as t2vec, DTW, and Frechet distance.
• An exact algorithm is designed with incremental similarity computation to reduce time complexity from O(nm) to O(n), enabling baseline comparison.
• A controllable approximate algorithm is introduced that restricts subtrajectory size to be close to the query trajectory size, balancing efficiency and effectiveness.
• The RL models are trained on trajectory pairs, with hyperparameters tuned for each similarity measure, and inference is performed in O(n) time per trajectory.

Experimental results

Research questions

• RQ1 Can deep reinforcement learning be effectively applied to the problem of finding the most similar subtrajectory in a trajectory database?
• RQ2 How does the performance of RL-based subtrajectory search compare to heuristic and non-learning-based methods in terms of effectiveness and efficiency?
• RQ3 To what extent can the proposed splitting-based algorithms achieve a favorable trade-off between computational cost and retrieval accuracy?
• RQ4 How do the RL-based algorithms generalize across different trajectory similarity measures such as t2vec, DTW, and Frechet distance?

Key findings

• The RLS and RLS-Skip algorithms significantly outperform non-learning baselines: RLS-Skip achieves up to 20x speedup over the exact algorithm on DTW and Frechet metrics while maintaining high effectiveness.
• RL-based methods achieve higher Relative Rank (RR) than heuristic and sampling-based approaches, with RLS-Skip outperforming Random-S even at small sample sizes (e.g., 10–20).
• The training time for RLS and RLS-Skip models ranges from 4.4 to 104.1 hours depending on the dataset and similarity measure, with RLS-Skip training faster due to architectural efficiency.
• RLS-Skip dominates UCR and Spring in both efficiency and effectiveness under most settings, particularly for R ∈ [0.2, 0.3], where it achieves better trade-offs.
• The performance of UCR is insensitive to the R parameter, with RR remaining around 60% across R = 0 to 1, indicating limited sensitivity to accuracy control.
• The exact algorithm with incremental similarity computation reduces time complexity by O(n), demonstrating the benefit of algorithmic optimization even before learning-based methods are applied.