Summary:This paper is an article published by Huawei cloud database Innovation Lab in conjunction with the data and Intelligence Laboratory of the University of Electronic Science and technology in CIKM’21. This article proposes the first trajectory recovery model to overcome the multi-level periodicity, periodic offset and data sparsity commonly existing in human movement trajectory data.
This article is shared from Huawei cloud community《Interpretation of CIKM’21 periodicmove paper》, author: cloud database innovation lab.
This paper (periodicmove: shift aware human mobility recovery with graph neural network) is an article published by Huawei cloud database Innovation Lab in conjunction with the data and Intelligence Laboratory of University of Electronic Science and technology at the summit CIKM’21. This article proposes the first trajectory recovery model to overcome the multi-level periodicity, periodic offset and data sparsity commonly existing in human movement trajectory data. CIKM is one of the top academic conferences in the field of information retrieval and data mining. A total of 1251 contributions were received, including 271 accepted papers, with an acceptance rate of about 21.7%. This paper is one of the key technical achievements of cloud database Innovation Lab at the track analysis level.
With the introduction of various location-based services, it is particularly important to restore and complete the sparse human movement trajectory data to improve the accuracy of these downstream tasks. However, the recovery of human movement trajectory data faces great challenges:
• There are complex transfer modes between track points in the track
• Multi-level periodicity and periodic offset are common in human movement trajectory data
• At present, the trajectory data collected is relatively sparse
In this paper, we propose a graph neural network based human behavior trajectory recovery model called periodicmove. In this model, we first construct a directed graph for each historical trajectory, and use the graph neural network to capture the complex transfer patterns between locations; Then, we designed two attention mechanisms to capture the multi-level periodicity and periodic offset of human behavior trajectory respectively; Finally, we design a spatial aware loss function to introduce the spatial proximity information of location into the model, which alleviates the problem of data sparsity to a certain extent. We have done a lot of experiments on two representative human trajectory data sets. The experimental results show that our model periodicmove has achieved a significant performance improvement of 2.9% – 9% compared with the current SOTA model.
2.1 model architecture
Our model mainly includes five parts:Figure neural network layer, timing embedding layer, two attention mechanism layers and the final fusion recovery layer
Figure 2.2 neural network layer
In order to capture the complex spatial transfer relationship between trajectory points in the trajectory, we first build the graph of each trajectory as shown in the figure, and then use the graph neural network to learn the complex spatial transfer mode between trajectory nodes in the directed graph
2.3 timing embedding layer
We use the relative phase in the trigonometric function mentioned in attention is all you need to describe the relative order relationship in the trajectory sequence, and then we splice the results of the graph neural network layer and the temporal embedding layer to form an embedded vector representation containing complex spatio-temporal dependencies
2.4 attention mechanism layer
The cross attention layer is mainly used to solve the periodic offset phenomenon in human movement track data. We compare the movement mode at the current time t with the movement mode at all times in each historical track, and aggregate the relevant historical information at time t of the historical track based on a similarity weight to solve the periodic offset phenomenon
After passing through the cross attention layer, the track point representation of each time of each historical track can be understood as the offset calibration according to each time of the current track to be completed. Next, in the soft attention layer, we perform an attention operation between the t-th time of the current track and the t-th time of each historical track to form a multi-level periodic historical track representation including the historical track
2.5 fusion recovery layer
In the final fusion restoration layer, we use the historical trajectory enhancement sequence containing complex spatio-temporal dependencies, multi-level periodicity and overcoming the phenomenon of periodic offset to assist the current trajectory for the final completion restoration
2.6 design distance loss
In the scene with highly sparse trajectory data, the cross entropy loss can not well capture the spatial proximity, which is an important feature of human movement restoration. Therefore, we designed a distance loss function to incorporate the model of spatial proximity information, and used noise contractual estimation (NCE) to accelerate the training of the model. The visualization results show that adding distance loss can effectively help the model capture spatial proximity information
3.1 experimental results
Compared with the current SOTA model (2021-aaai), our model periodicmove has achieved a significant performance improvement of 2.9% – 9%
3.2 Ablation Experiment
We conducted ablation experiments on the five parts of the model respectively. From the experimental results, we can see that each module has a certain contribution to our task. Among them, after the soft attention layer module is removed, the effect of the model decreases the fastest, indicating that the multi-level periodicity plays a very important role in the task of human movement trajectory data recovery
3.3 robustness test
We conduct a robustness experiment between this model and the latest SOTA model (2021-aaai) under different deletion rates. From the experimental results, we can see that both models have good robustness, but the effect of our model under each deletion rate is improved on attnmove
In various location-based services, such as personalized geographic location recommendation, urban intelligent traffic scheduling, trajectory anomaly detection and many other scenarios, as long as the collected trajectory data is sparse, it will affect the accuracy of these downstream tasks. The purpose of our paper is to recover the sparse trajectory data into dense and fine trajectory data, so as to improve the accuracy of these downstream tasks
Huawei cloud database Innovation Lab official website:https://www.huaweicloud.com/l…