Trajectory Similarity Learning with Auxiliary Supervision and Optimal Matching

Authors: Hanyuan Zhang, Xinyu Zhang, Qize Jiang, Baihua Zheng, Zhenbang Sun, Weiwei Sun, Changhu Wang

IJCAI 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The comprehensive experiments on real-world datasets demonstrate that our model substantially outperforms all existing approaches.
Researcher Affiliation Collaboration 1School of Computer Science, Fudan University 2Shanghai Key Laboratory of Data Science, Fudan University 3Shanghai Institute of Intelligent Electronics & Systems 4Singapore Management University 5Byte Dance AI Lab, Beijing, China
Pseudocode No No pseudocode or algorithm blocks were found.
Open Source Code No The paper provides a link to the source code of a competitor model (NEUTRAJ) but does not provide concrete access to the source code for the methodology (Traj2Sim Vec) described in this paper.
Open Datasets Yes We use two real trajectory datasets in our experimental study, namely Porto that was extracted from open source dataset available at https://kaggle.com/c/pkdd15-predict-taxi-service-trajectory-i and Shanghai.
Dataset Splits Yes We split each dataset into training set, validation set and test set in the ratio 3:1:6.
Hardware Specification Yes We test the time cost on an AMD Ryzen Threadripper 2950X processor with 32 threads.
Software Dependencies No The paper mentions using LSTM models and the Adam optimizer but does not specify version numbers for any software dependencies or libraries.
Experiment Setup Yes We set the length of simplified trajectory T to 5, and the number of elements sampled from distance matrix D and E to 10 (i.e., k = 5, r = 10). The tunable parameter α is the reciprocal of the maximum distance of training sample distances for Fr echet and Hausdorff distance, and the mean plus three times the variance of training sample distances for DTW. The margin value ξ of point matching loss is 0.01. The sampled triple number of point matching loss is 10 (i.e., r = 10). The trajectory representation vector has a dimensionality of 128 (i.e., d = 128). We use the LSTM model as the RNN encoder and matching space mapping unit. The hidden unit is 128. We train the model using Adam algorithm [Kingma and Ba, 2014] with an initial learning rate at 0.001. All the weights are uniformly initialized to ( 1/128).