從GPS軌跡以遞迴類神經網絡預測個人活動意圖

Abstract

本論文研究活動意圖類型預測方法，以遞迴類神經網路架構為基礎，用建立群組模型的概念，比較四種建立模型的架構。第一種是全體資料模型，以所有GPS軌跡資料擷取特徵作為模型輸入，建立全體資料模型。第二種是群組模型，本論文提出兩種分群方法，分別為使用者為單位進行分群的使用者分群法，及以序列為單位進行分群的序列分群法，再以群組資料建立群組模型。第三種是遷移學習模型，以全體資料進行訓練，將全體資料訓練好的參數設置為初始參數，以群組資料作為訓練資料，只對模型中部份層的參數進行調整。第四種是合成模型，將全體資料模型和群組模型預測結果，學習調和參數將兩個預測結果進行比重相加。實驗評估顯示，遷移學習模型在OSM資料集的預測結果優於全體資料模型和群組模型，合成模型在大部分情況下可良好地結合兩模型，正確預測出使用者的活動意圖。在Geolife資料集中，合成模型在Accracy@5最高可達89.31%的準確率，在OSM資料集中，則可達74.12%的準確率。In this paper, we study the problem of predicting activity intentions based on the recurrent neural network architecture. We constructed four learning models based on the various combinations of training data sets and recurrent neural network architectures. The first one is the global model, which uses all activity sequences as the training data set. The second one is the group model, in which two clustering methods: user-based clustering and sequence-based clustering methods are proposed to separate the data into groups. Accordingly, a prediction model is constructed respectively for each group of training data. The third one is the transfer-learning model, in which the parameters learned from all training data set are set as the initial parameters. Then the training data in each group is used to adjust the parameters from the middle layer of the RNN architecture to construct the predicting model for each group. The last one is the ensemble model, which concatenates the predicting results of the global model and the group model to learn the ensemble parameters to get a properly weighted sum of the two predicted results. The results of experiments show that the transfer-learning model on the OSM dataset has better performance than the global model and the group model. Furthermore, the ensemble model can combine the results of two models well in most cases and provide the highest accuracy. In the Geolife dataset, the accuracy@5 of the ensemble model achieves 89.31%, and gets 74.12% on the OSM dataset.