LATTE: Learning Adaptive Segmentation for Efficient and Effective Trajectory Similarity Learning

This repository contains the implementation of our paper accepted at KDD 2026: "LATTE: Learning Adaptive Segmentation for Efficient and Effective Trajectory Similarity Learning"

Overview

Data Preprocessing

First download the trajectory dataset used in our paper, or your own dataset into ./data/{dataset_name}. You need to extend config.py if it is your own dataset.

The preprocessing pipeline is organized under preprocess/ and should be run in the following order:

Step 1. Trajectory map matching (map_matching/): osm_convert.py → osm2roadnetwork.py → HMMM_mm.py → filter.py Step 2. Meta Feature extraction (meta_feature/): meta_feature.py Step 3. Trajectory interpolation (interpolation/): interpolation.py Step 4. Dataset split (dataset_split/): dataset_split.py Step 5. Feature extraction (feature_extraction/): segment_graph.py → node2vec.py, feature_tokenization.py, feature_extract.py (last three are independent and can run in parallel) Step 6. TSL label generation (similarity_ground_truth/): initialization.py → dataset_preparation.py

Training

First train the teacher model:

python train_teacher.py -root_dir YOUR_DATA_DIR -dataset_name porto

Then train LATTE:

python train_LATTE.py -root_dir YOUR_DATA_DIR -dataset_name porto

Key configuration arguments:

Argument	Description	Default
-root_dir	Path to data directory	"your/project"
-dataset_name	Dataset name (porto, chengdu)	porto
-split_marker	Trajectory length split (short, long)	short
-distance_type	Similarity metric (TP, NetERP)	TP
-epochs	Number of training epochs	20
-batch_size	Batch size	32
-num_companion	Number of positive/negative companions per anchor	5
-target_ratio	Target segmentation ratio for AdaSeg	0.2