[ad_1]
The creation of a problem
Travel activity zones are aggregated from individual travel direction footprints (U) graph contains:
$$\begin{aligned} G_{u} = (V_{u},E_{u}), \end{aligned}$$
(1)
where \(V_{u}\) Represents a set of graph nodes:
$$\begin{aligned} V_{u}=\{v_{u,i}|i \in [0,n]\}, \end{fixed}$$
(2)
and each node (\(v_{u,i}\)) represents an individual’s representative travel activity zone (i.e., activity node) by aggregating individual travel footprints. Nodes of an individual activity are connected through a set of graph edges:
$$\begin{aligned} E_{u}=\{e_{u,j}|j \in [0,m]\} \end{fixed}$$
(3)
Each end (\(e_{u,j}\)) represents a linear journey between two final movement nodes. Aggregate statistics representing the spatial characteristics of travel directions are stored as node attributes ((X_{v}\)) and edge weight (\(Y_{e}\)). Distributions of surrounding POIs are also set as node attributes. The conversion method is explained in the next two subsections. Each active node has two types of neighbors namely neighbors (\(N_{in}\)) and neighbors (\(N_{out}\)). Crossroads for movement \(v’\) As a neighbor of the node of activity No,
$$\begin{aligned}&{if (v’,v) \in E, v’ \in N_{in}}, \end{aligned}$$
(4)
$$\begin{aligned}&{if (v,v’) \in E, v’ \in N_{out}}. \end{fixed}$$
(5)
Based on the travel activity graph, Gstp2Vec is shown in Figure 1. Basically, two sets of fully connected convolutional neural networks (NN) are formed by combining the weights with feature embedding to distribute the weights from the nodes’ neighbors through the graph structure.15:37. A set of NNs is wrapped as a multihop (ie, K-hop) collectors to collect neighborhood embeddings from sampled neighboring nodes and edges within them. K Hops specifically, each collector function (eg, AGG1, AGG2) It consists of a fully connected NN layer with an offline activation function \(\sigma\). Node embeddings (original cross-section properties) and edge weights are binned by the assembler to generate low-dimensional embeddings.
Graph-based representation learning for individual travel activity type identification.
Another set of NNSs are constructed using modified cross-sections with input as cross-section embeddings with integrated neighborhood embeddings. The current cross-sectional embeddings are considered as approximate representations, which are fed into another activation function to estimate the types of travel motions. This process is called forward expansion15Iterates over all activity nodes in an epoch to train the model. In this way, the information of the activity nodes farthest away from the current one is propagated to it through the multihop neighbors and contributes to the identification of the activity type. The weight matrix and summation parameters are adjusted by reducing the graph-based cross-entropy loss through Adam’s optimizer.38Hence Gstp2Vec supervised learning modeling.
Movement nodes
Individual travel trajectories are represented as sequences of travel footprints, each footprint representing an individual’s presence at each location and time point, and timestamped as a pair of geographic coordinates. Travel time points at a speed of less than 1300 m/h39 They are derived based on spatial proximity based on the density of sound (DBSCAN) application.40. Functional zones are generated from the resulting spatial clusters (i.e., waiting regions) as convex hulls so that spatial boundaries are defined for the travel activities represented.41. Activity zones are then used to produce activity type identifiers.
Characteristics of the cross section
The topological relationship between each node and its neighboring nodes and the distribution of nodes in the neighborhood are coded and transmitted to identify the type of activity represented by each node. Aggregate statistics indicating the distribution of pedestrians in time and space and the distribution of surrounding POIs are calculated for each activity zone and clustered as nodes.
Specifically, the total or average track numbers in each 24-hour period on both weekdays or weekends are considered to represent time characteristics (T) the activity zone of each individual, where T denotes matrix transposition:
$$\begin{aligned} t_{weekday}= and {} [n_{1},n_{2},\ldots ,n_{24}]^T, \end{fixed}$$
(6)
$$\begin{aligned} {\overline{t}}_{weekday}= and {} [{\overline{n}}_{1},{\overline{n}}_{2},\ldots ,{\overline{n}}_{24}]^T, \end{fixed}$$
(7)
$$\begin{aligned} t_{weekend}= and {} [n_{1}’,n_{2}’,\ldots ,n_{24}’]^T, \end{fixed}$$
(8)
$$\begin{aligned} {\overline{t}}_{weekend}= and {} [{\overline{n}}_{1}’,{\overline{n}}_{2}’,\ldots ,{\overline{n}}_{24}’]^T, \end{fixed}$$
(9)
$$\begin{aligned} t= and {} \left[ t_{weekday},{{\overline{t}}_{weekday}}, t_{weekend}, {{\overline{t}}_{weekend}}\right] . \end{fixed}$$
(10)
In addition, the average durations spent in each activity zone during each week were calculated and combined to generate an additive representation.42 (\(t^{+}\)Temporary patterns:
$$\begin{aligned} \overline{\Delta t}_{dow}= and {} \left[\overline{\Delta t}_{1}’,\overline{\Delta t}_{2}’,\ldots ,\overline{\Delta t}_{7}’\right]^T, \end{fixed}$$
(11)
$$\begin{aligned} t^{+}= and {} [t,\overline{\Delta t}_{dow}]. \end{fixed}$$
(12)
The maximum and average values of the past period (\(\delta t_{max}\) And \(\overline{\Delta t}\)) or distance (\(\delta d_{max}\) And \(\overline{\Delta d}\)) all footprints within the movement range of the next travel footprint are also computed as spatial features (S)::
$$\begin{aligned} s=\left[ \Delta t_{max},\overline{\Delta t},\Delta d_{max},\overline{\Delta d}\right] . \end{fixed}$$
(13)
To hide the characteristics of the POI distribution by simulating with natural language processing43Each separate POI feature class (for example, bedroom, cafe, bar, hospital, etc.) is considered as a word.44. All possible POI feature classes are considered as a dictionary, and POI feature classes overlapping with the activity zone are considered as a corpus. A total of 335 different POI feature classes (i.e., terms) were collected from the OSM dictionary. Then, the occurrences of each POI feature class are counted for the motion zone to produce a sparse POI feature vector (Page).
In addition, 335 POI feature classes were integrated into 18 different types of places (e.g., home, eating, school) based on their functionality (e.g., cafe) in urban areas. \(\right arrow\) to eat)25. Then a small dictionary is constructed to produce a dense vector, which is concatenated with Page To generate an improved POI feature vector (\(p^{+}\)). Next, \(p^{+}\) It is concatenated with the above spatiotemporal feature vector to produce a cross-sectional feature matrix ((X_v\)) for each activity zone;
$$\begin{aligned} {X_v}=\left[ {t^{+}},s,{p^{+}}\right] . \end{fixed}$$
(14)
Edge weights
A trip is defined as a transition for an individual from one travel activity (ie, origin) to another (ie, destination), usually indicating the individual’s spatial transition. In our proposed Gstp2Vec framework, travel directions are consistent with edge directions. In addition to characteristics of travel direction and last movement nodes, travel characteristics also include statistics that measure individual transitions over space and time, such as travel frequency (fAverage trip duration (\({\overline{t}}\)) and average travel distance (\({\overline{d}}\)are denoted as edge weights (\(Y_e\)).
In particular, f It is calculated by counting the travel events from each origin activity zone to the corresponding destination zone for each individual, traversing all travel paths in the origin zone. and then \({\overline{t}}\) It is measured by the average time spent on those trips, and \({\overline{d}}\) It is measured by averaging their straight-line distances in 2D space. These statistics measure different aspects of travel properties and are combined to represent them. Y:
$$\begin{aligned} {Y_e}=\left[ f,{\overline{t}},{\overline{d}}\right] . \end{fixed}$$
(15)
Collectors
As shown in Table 1 and Figure 2 , aggregators (i.e. aggregation functions) in Gstp2Vec receive feature embeddings of sampled neighboring nodes, which are integrated with their corresponding edge weights as nodes. In our proposed framework, since neighborhood nodes are inherently unordered, summation functions must be balanced to operate on randomly ordered nodes. In addition, they should be easy and trainable15. The maximal clustering classifier is balanced and trainable, and is therefore implemented in our proposed framework.45.
Assembler Architecture and Supervised Learning.
Specifically, a single-layer perceptron implements a fully connected NN in a single collector. At each iteration of forward propagation, a fixed number (eg, 2) of neighboring nodes are sampled for each active node. Then, the perceptron is applied to the nesting matrix of each sample neighbor node to compute a series of features, and an element-wise maximum value between all sample neighbor nodes is generated for each computed feature and transmitted to the current nodes. In this way, the model successfully captures the various aspects of the environment15.
Supervised education
Model weights are iteratively adjusted through end-to-end supervised learning.14, 15, 31. First, graphs consisting of activity zones and trips are divided into training, validation and test sets based on individuals (Figure 3). Therefore, the activity zones of an individual do not appear in different sets (eg, both training and test sets). For example, the graph (\(G_{u_{2}}\) Figure 3) composed of individual activity zones \(u_{2}\) It is divided into test set and two other graphs (ie, (G_{u_{1}}\) And (G_{u_{3}}\)) in the training set and the remainder (i.e., (G_{u_{4}}\) It is in the confirmation set.
Based on individuals with randomly distributed activity zone graphs into training, validation, and test sets.
The training process consists of two steps, which are forward propagation and parameter learning. Multiply forward (Z by number (16)46) first combines the cross sections with the neighboring nests to generate the cross sections (Figure 2) and this is done through the assemblers as described above. The combined node features are then combined by a single-layer NN with an activation function that produces an embedding of the modified node features and finally produces approximate representations (ie, \(h^{k-1}_v\)) ( Table 2 ) and then another softmax function is applied to the resulting representations to predict travel activity types in multiple categories (argmax in Eq. (16)). For parameter learning, graph-based cross-entropy loss is applied to the previously estimated output to adjust the weight matrices.
$$\begin{aligned} {Z=f\bigg (h_{v}^{k-1},h_{v’}^{k}\bigg )=argmax\bigg (softmax\bigg (\sigma \bigg ) (W^k\cdot CONCAT\bigg (h_{v}^{k-1},h_{N(v)}^{k}\bigg )\bigg )\bigg )\bigg )}. \end{fixed}$$
(16)
[ad_2]
Source link
