计算机工程与应用 ›› 2023, Vol. 59 ›› Issue (11): 302-311.DOI: 10.3778/j.issn.1002-8331.2203-0168

• 工程与应用 • 上一篇    下一篇

融合卷积、注意力和MLP的出租车目的地预测

余丹青,邬群勇,姚江涛,邝嘉恒   

  1. 1.福州大学 数字中国研究院(福建),福州 350003
    2.福州大学 空间数据挖掘与信息共享教育部重点实验室,福州 350108
    3.福州大学 卫星空间信息技术综合应用国家地方联合工程研究中心,福州 350108
  • 出版日期:2023-06-01 发布日期:2023-06-01

Taxi Destination Prediction Based on Convolution, Attention and Multilayer Perceptron

YU Danqing, WU Qunyong, YAO Jiangtao, KUANG Jiaheng   

  1. 1. The Academy of Digital China(Fujian), Fuzhou University, Fuzhou 350003, China
    2. Key Lab of Spatial Data Mining and Information Sharing of Ministry of Education, Fuzhou University, Fuzhou 350108, China
    3. National & Local Joint Engineering Research Center of Satellite Geospatial Information Technology, Fuzhou University, Fuzhou 350108, China
  • Online:2023-06-01 Published:2023-06-01

摘要: 出租车目的地预测是基于位置服务的重要内容,对城市交通合理规划具有重要意义。基于出租车历史轨迹数据进行出行目的地预测,存在数据稀疏性问题与轨迹数据的特征单一性问题,影响目的地预测的精度。针对数据稀疏问题,利用出租车原始轨迹数据(出租车速度、行驶方向角和时间)结合轨迹截断方法确定模型输入特征,克服轨迹稀疏性问题。针对多层感知机具有参数过多、训练困难和大量参数导致较快的过拟合等问题,提出利用卷积的参数共享机制解决参数冗余。进一步提出采用注意力机制使神经网络把更多的计算资源分配给更重要的任务,聚焦重要信息提升模型预测性能。基于此构建了一种融合卷积、注意力模块和多层感知机的出租车目的地预测方法(CCMLP),在克服MLP和卷积各自结构不足的同时,对目的地实现了更为准确的预测,通过不同数据与实验验证CCMLP模型的可靠性。实验结果表明,选取轨迹前[m]点与末[n]点及其对应行驶方向角、速度、车牌号和时间特征作为模型输入特征有效提高目的地预测精度;提出的CCMLP方法具有较好特征学习能力,相比于基于MLP预测模型的距离误差下降了10%,相比于基于集成学习算法的距离误差下降了19.6%;基于工作日与非工作日不同数据分布数据集划分,基于CCMLP目的地预测模型距离损失分别为2.25?km与2.23?km,验证了CCMLP对于不同数据分布的泛化能力;基于轨迹前十点,40%、60%、80%不同完整度轨迹得到的距离损失分别为2.23?km、1.80?km、0.97?km、0.68?km,验证给定不同轨迹完整度对目的地预测的影响。

关键词: 出租车轨迹, 目的地预测, 厦门岛, 神经网络, 注意力模块, 多层感知机

Abstract: Taxi destination prediction based on location services is an important aspect for reasonable planning of urban transportation. However, there exist problems of data sparsity and single feature of trajectory data, which affect the accuracy of destination prediction. To address the problem of data sparsity, the original taxi trajectory data(i. e. taxi speed, direction angle, and time) is combined with trajectory truncation method to determine the model input features, thus overcoming data sparsity. To address the problems of excessive parameters and overfitting caused by a large number of parameters of the multilayer perceptron(MLP), a parameter sharing mechanism based on convolution is proposed to solve parameter redundancy. Furthermore, an attention mechanism is used to allocate more computing resources to more important tasks, focusing on important information to improve the prediction performance of the model. Based on this, a taxi destination prediction method(CCMLP) that combines convolution, attention modules, and MLP is proposed, which achieves more accurate destination prediction while overcoming the structural deficiencies of MLP and convolution. The reliability of the CCMLP model is verified through different data and experiments. The experimental results show that selecting the trajectory features of the first m points and the last n points, as well as their corresponding direction angles, speeds, license plate numbers, and time features as the input features of the model can effectively improve the accuracy of destination prediction. The proposed CCMLP method has good feature learning ability. Compared with the MLP prediction model, the distance error is reduced by 10%, and compared with the ensemble learning algorithm, the distance error is reduced by 19.6%. The CCMLP model’s generalization ability for different data distributions is verified by dividing the data set based on weekdays and weekends, with distance losses of 2.25 km and 2.23 km respectively. The effect of different trajectory completeness on destination prediction is verified based on the first ten points of the trajectory, with distance losses of 2.23 km, 1.80 km, 0.97 km, and 0.68 km for completeness rates of 40%, 60%, 80%, and complete trajectories respectively.

Key words: taxi trajectory, destination prediction, Xiamen Island, neural network, attention module, multilayer perceptron