A large-scale data-driven co-moving pattern discovery method

Abstract

Embodiments of the present disclosure relate to a large-scale data-driven co-movement pattern discovery method. The method comprises: scanning a database to construct complete trajectories of all moving objects; mapping all moving objects into corresponding bins according to spatial locations based on the complete trajectories; refining clusters of moving objects in each bin according to similarity in the time dimension to obtain a co-movement pattern dataset; constructing a trajectory model based on a Transformer; training the trajectory model using the co-movement pattern dataset; and performing co-movement pattern discovery on a test dataset. Embodiments of the present disclosure assign all moving objects to corresponding bins according to the proximity of the moving objects in the spatial trajectory. Then, a more refined intra-bin refinement is used to check the similarity in the time dimension in each bin. The need to track consecutive timestamps is avoided, thereby being lower in time complexity than the state-of-the-art method.

Description

Technical Field

[0001] This disclosure relates to the field of traffic data detection technology, and more particularly to a large-scale data-driven method for discovering common mobility patterns. Background Technology

[0002] With the widespread use of rail travel, a wealth of trajectory data describing passenger movement is available. This paper explores the potential of this rail trajectory data to reveal patterns of simultaneous movement among passengers, termed co-movement patterns. Discovering co-movement patterns in the vast rail data has significant application value. First, it helps track potential infected individuals during major disease outbreaks and enhances the identification of cases of overlapping infections, providing valuable support for health authorities in developing disease transmission prevention policies. Second, this data can reveal strangers with similar travel itineraries through similar tourist or commuting routes, an insight that fosters opportunities for social interaction and shared travel experiences. Third, identifying co-movement patterns helps accurately pinpoint congested areas in public transport, allowing transport managers to adjust departure frequencies at specific rail stations to alleviate congestion. Finally, since individuals with similar itineraries are often relatives or friends, policymakers can consider offering group passenger discounts on rail, a strategy that can attract more customers by catering to group travel needs.

[0003] Discovering common movement patterns using large-scale railway data is a key challenge in urban computing. Existing solutions primarily focus on revealing continuous common movement patterns through the continuous timestamps of moving objects. The problem with these methods is that all objects need to be constantly present, which is unsuitable for railway scenarios. Railway data typically records discrete entry and exit timestamps, rather than continuous timestamps, making these methods less applicable in this context. Furthermore, these methods require continuous tracking of object timestamps, which incurs additional computational costs for objects with discrete timestamps. While continuous trajectory-based methods have achieved success in urban computing, mining common movement patterns from discrete trajectories in large-scale railway data remains an unsolved problem. Summary of the Invention

[0004] To avoid the shortcomings of existing technologies, this application provides a large-scale data-driven common movement pattern discovery method to solve the problem that existing technologies have not solved the problem of mining common movement patterns from discrete trajectories in large-scale railway data.

[0005] According to embodiments of this disclosure, a large-scale data-driven common mobility pattern discovery method is provided, the method comprising:

[0006] Scan the database to construct the complete trajectories of all moving objects;

[0007] Based on the complete trajectory, all moving objects are mapped to the corresponding boxes according to their spatial location;

[0008] The clusters of moving objects in each box are refined based on the similarity in the time dimension to obtain a common movement pattern dataset.

[0009] Construct a Transformer-based trajectory model; wherein, the Transformer-based trajectory model includes a Transformer-based trajectory encoder and decoder;

[0010] Train a Transformer-based trajectory model using a common movement pattern dataset;

[0011] We used the trained Transformer-based trajectory model to discover common movement patterns on the test dataset.

[0012] Further steps, such as scanning the database to construct the complete trajectories of all moving objects, include:

[0013] Traverse the database D of records and move the object o by connecting. i Each moving object is obtained by observing the trajectory points in the railway data within the observation time window W. i The complete trajectory.

[0014] Furthermore, the step of mapping all moving objects to their corresponding bins based on their spatial location, according to the complete trajectory, includes:

[0015] Based on each moving object o i The complete trajectory t i Through combined operations Select from the complete trajectory t i All journeys t of size n n ;

[0016] The spatial positions of each combination in n trips form a box label. b The object will be mapped to the corresponding box;

[0017] The function is defined as follows:

[0018] l b =f1(t n (1)

[0019] Here, f1() is the mapping function for extracting spatial labels.

[0020] Furthermore, the step of scanning the database to construct the complete trajectories of all moving objects also includes:

[0021] The time complexity is Where k1 is the average number of records for each moving object, and k2 corresponds to... |O| represents the number of objects being moved.

[0022] Furthermore, the step of refining the clusters of moving objects in each box based on similarity in the time dimension to obtain a common movement pattern dataset also includes:

[0023] For each moving object o in the box i Retrieve moving object o i The complete trajectory t i ;

[0024] Get the tagged box label l b Time information

[0025] The object will be moved based on its time information. i Mapped to fine-grained clusters;

[0026] The function is defined as follows:

[0027]

[0028] Among them, l′ b f2() is the label for the fine cluster and the mapping function.

[0029] Furthermore, the processing steps of the Transformer-based trajectory encoder include:

[0030] Embedding the complete trajectory point P Embedding of time information T Using the first weight parameter respectively Second weight parameter Perform the conversion;

[0031] The first weight parameter Second weight parameter The converted embedded connection is e O ;

[0032] The transformed complete trajectory embedding of the moving object O is expressed as:

[0033] e O =e P ·W P ||e T ·W T (3)

[0034] Among them, e O Embed the complete trajectory of the moving object;

[0035] The key spatial and temporal information of the complete trajectory of a moving object is encoded using an attention mechanism, and the process is defined as follows:

[0036]

[0037] Where W1 is the first learnable parameter in the attention mechanism, W2 is the second learnable parameter in the attention mechanism, and W3 is the third learnable parameter in the attention mechanism. Encodes the complete trajectory of a moving object, where softmax is the activation function and d is the dimension of the input;

[0038] Leveraging multi-head attention mechanisms to improve learning robustness:

[0039]

[0040] in, The output of the multi-head attention mechanism is given by H, where H is the number of multi-heads and W is the number of multi-heads. h is the first trainable parameter, and || is the connection operation.

[0041] Furthermore, the decoder's processing steps include:

[0042] The multilayer perceptron layer extracts a deep representation from the complete trajectory of a moving object, as shown below:

[0043]

[0044] Among them, e′ O W is a deep representation of the complete trajectory encoding. m This is the second trainable parameter;

[0045] The sigmoid function is used to map the deep representation of an object to its corresponding common movement cluster:

[0046] y′ O =sigmoid(e′) O (7)

[0047] Where, y′ O For predicting clusters of moving objects.

[0048] Furthermore, the loss function for the Transformer-based trajectory model is expressed as:

[0049]

[0050] in, p is the total number of clusters. c Let c be the weight of the positive samples in the cluster, x be the input sample, and y be the sample label. c For the sample corresponding to cluster c, yc These are the sample labels corresponding to cluster c.

[0051] The technical solutions provided by the embodiments of this disclosure may include the following beneficial effects:

[0052] In the embodiments of this disclosure, the aforementioned large-scale data-driven co-movement pattern discovery method, on the one hand, assigns all moving objects to corresponding bins based on their proximity in spatial trajectories. Then, finer-grained in-bin refinement is used to examine the temporal similarity within each bin. This avoids the need to track consecutive timestamps, resulting in lower time complexity than state-of-the-art methods. On the other hand, a Transformer-based estimation model is proposed, which uses depth representation to encode passenger spatial and temporal trajectory information, efficiently decoding co-movement patterns for rapid discovery. Attached Figure Description

[0053] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure. It is obvious that the drawings described below are merely some embodiments of this disclosure, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0054] Figure 1 A step diagram illustrating a large-scale data-driven common mobility pattern discovery method in an exemplary embodiment of this disclosure is shown.

[0055] Figure 2 A diagram illustrating the prior art's discovery of common movement patterns in railway data in an exemplary embodiment of this disclosure;

[0056] Figure 3 A schematic diagram illustrating the binning-based precision algorithm in an exemplary embodiment of this disclosure is shown;

[0057] Figure 4 A schematic diagram of a Transformer-based trajectory model is shown in an exemplary embodiment of this disclosure;

[0058] Figure 5 A comparison diagram showing the scalability of the move-Approx algorithm and other methods in exemplary embodiments of this disclosure is provided. Figure 5 (a) The training time of the algorithm increases linearly as the number of passengers increases; Figure 5 (b) A comparison of the move-Approx algorithm and other methods as the number of passengers increases; Figure 5 (c) The training time increases linearly with the number of days. Figure 5(d) is a comparison chart of other methods of the move-Approx algorithm as the number of days increases;

[0059] Figure 6 This shows the average runtime of all four methods in the exemplary embodiments of this disclosure on a single dataset; Figure 6 (a) represents the training time of Move-Exact on four datasets. Figure 6 (b) Comparison of detection time between Move-Exact and several other methods;

[0060] Figure 7 The results of ablation experiments in exemplary embodiments of this disclosure are shown; Figure 7 (a) For Move-Approx and Move-Approx on the MTR-1 week dataset na and Move-Approx nw Performance comparison Figure 7 (b) For Move-Approx and Move-Approx on the MTR-2 week dataset na and Move-Approx nw Performance comparison;

[0061] Figure 8 g is shown in an exemplary embodiment of this disclosure m The impact of d on detection performance; Figure 8 (a) for different g m Impact on detection performance; Figure 8 (b) shows the impact of different values ​​of d on detection performance;

[0062] Figure 9 A schematic diagram illustrating a common movement pattern captured in an exemplary embodiment of this disclosure; Figure 9 (a) is a schematic diagram of the common movement pattern captured when n=1. Figure 9 (b) is a schematic diagram of the common movement pattern captured when n=3. Detailed Implementation

[0063] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that this disclosure will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0064] Furthermore, the accompanying drawings are merely illustrative diagrams of embodiments of this disclosure and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities.

[0065] This example implementation provides a large-scale data-driven method for discovering common mobility patterns. (See reference...) Figure 1 As shown, the large-scale data-driven common mobility pattern discovery method may include steps S101 to S106.

[0066] Step S101: Scan the database to construct the complete trajectory of all moving objects;

[0067] Step S102: Based on the complete trajectory, map all moving objects to the corresponding boxes according to their spatial location;

[0068] Step S103: Refine the clusters of moving objects in each box based on the similarity in the time dimension to obtain a common movement pattern dataset;

[0069] Step S104: Construct a Transformer-based trajectory model; wherein, the Transformer-based trajectory model includes a Transformer-based trajectory encoder and decoder;

[0070] Step S105: Train a Transformer-based trajectory model using a common movement pattern dataset;

[0071] Step S106: Use the trained Transformer-based trajectory model to discover common movement patterns on the test dataset.

[0072] The aforementioned large-scale data-driven co-movement pattern discovery method, on the one hand, assigns all moving objects to corresponding bins based on their proximity in spatial trajectories. Then, finer-grained in-bin refinement is used to examine the temporal similarity within each bin. This avoids the need to track consecutive timestamps, resulting in lower time complexity than state-of-the-art methods. On the other hand, a Transformer-based estimation model is proposed, which uses deep representations to encode passengers' spatial and temporal trajectory information, efficiently decoding co-movement patterns for rapid discovery.

[0073] Below, we will refer to Figures 1 to 9 The steps of the large-scale data-driven common mobility pattern discovery method described in this example embodiment will be explained in more detail.

[0074] In one embodiment, the purpose of this application is to develop a method to discover common movement patterns in large-scale railway data. These patterns are crucial for understanding traffic behavior, alerting to traffic congestion, and identifying similar strangers. The complexity of this task stems from irregular time intervals and massive amounts of railway data. This application proposes two solutions: Move-Exact and Move-Approx. Move-Exact is an exact algorithm that employs a multi-granularity binning structure to effectively group moving objects and manage data with uneven time intervals. Move-Approx, on the other hand, provides a fast, approximate algorithm for real-time pattern recognition on large-scale datasets. Extensive evaluation using railway data provided by a train service provider in Hong Kong, China, shows that the method in this application achieves an average speedup of two orders of magnitude compared to six other methods and can efficiently handle millions of trajectories.

[0075] Problem definition:

[0076] The symbols used are summarized in Table 1:

[0077] Table 1: Symbols

[0078] O Moving objects G Clusters that move together with moving objects P Complete trajectory points of object O in railway data T timestamp collection of object O n Minimum number of times objects move together

[0079] Definition 1 (Trajectory Point in Railway Data): Trajectory Point It is a triple Including timestamp t i The station number s at the current location i and the current station Category s i ,in

[0080] For example, trajectory points This indicates that the passenger entered Beijing West Railway Station at 8:00 AM on June 1, 2023. This indicates that the passenger left Beijing Zoo Station at 8:18 on June 1, 2023.

[0081] Definition 2 (Complete trajectory in railway data): Object o i The complete trajectory is an ordered sequence of points. Where t1 <… <t i <… <t w , usually there is t i -t i-1 ≠t j -t j-1 , 1≤i, j≤w, and Where W represents the observation window.

[0082] Unlike vehicle trajectories on road networks, which typically have a consistent sampling rate (e.g., once per second), time series in railway trajectory data have variable time intervals, i.e., t usually exists in Definition 2. i -t i-1 ≠t j -t j-1 This variability is a unique characteristic of railway data, where trajectory points are collected only upon arrival and departure from stations. Therefore, the intervals between these time series tend to be long, typically an hour or even longer, reflecting the different nature of data collection between railway systems and highway networks.

[0083] Definition 3 (Travel in Railway Data): A travel in railway data is a pair of data, formatted as {entry station s}. i timestamp t i →Exit station j timestamp t j}, where t i <t j , and s i ≠s j .

[0084] Next, we will define the common movement pattern problem in railway data based on the concept of complete trajectory.

[0085] Question 1 (Common Mobility Patterns in Railway Data): Common mobility patterns in railway data reveal that, within a given observation window, passengers on at least n trips simultaneously share the same trajectory.

[0086] For example, when the observation time window is set to one day and the minimum number of trips n is set to 2, such as Figure 2 As shown, objects {O1, O2, O5} form a common movement pattern because they travel together twice, namely {8:10 at station A → 8:50 at station B, 18:00 at station B → 19:00 at station C}.

[0087] I. Move-Exact: A precise algorithm based on binning

[0088] 1. Framework

[0089] The framework of binning-based exact algorithms is as follows: Figure 3 As shown, this method involves mapping moving objects to specific bins based on their spatial location. For example, objects O1 and O2 are assigned to bins labeled AB-BC because they move from station A at 8:10, to station B at 8:50, then from station B at 18:00, and to station C at 19:00. The algorithm then refines this by considering the similarity of objects in each bin across the time dimension.

[0090] 2. Coarse Bin-Based Mapping

[0091] In railway big data, the time intervals between journeys are typically large, such as one hour, making traditional methods of tracking all consecutive timestamps computationally expensive. Furthermore, the sheer volume of railway big data necessitates efficient processing for real-time pattern discovery; for example, the Hong Kong MTR carries over four million passengers daily.

[0092] To address these challenges, a binning-based, efficient algorithm is proposed to discover common movement patterns. Instead of tracking all consecutive timestamps, the algorithm maps objects to bins based on spatial similarity, and then verifies their temporal similarity within each bin.

[0093] Specifically, first, the database D containing records is traversed, and then the connection object o is used to... i Each object is obtained by observing trajectory points in railway data within the observation time window W. i The complete trajectory. Then, for each object o i The complete trajectory t i Through combined operations Select from the complete trajectory t i All possible journeys t of size n n The spatial positions of each possible combination of n trips form a box label l. b The object will be mapped to the corresponding box. For example, in Figure 2 In the diagram, the journey from station A to station B and the journey from station B to station A form the label AB-BC. The function is defined as:

[0094] l b =f1(t n (1)

[0095] Here, f1() is the mapping function for extracting spatial labels.

[0096] 3. More refined interior details

[0097] For each moving object o in the box j Search object o i The complete trajectory t j Then, obtain the spatial location l of the marker. b Time information Next, based on its time information, o i This maps to a more granular cluster. Since objects may have different timestamps, a time interval is used here to measure whether some timestamps should be considered similar. For example, in... Figure 2In this scenario, if the time interval is set to ten minutes, objects O1 and O2 should be mapped to the same finer-grained cluster because they are very close in both spatial and temporal dimensions. Additionally, check if the size of the finer-grained cluster exceeds the minimum cluster size g. m In this way, only objects falling into the same, more refined cluster are considered to represent common movement patterns in railway data.

[0098]

[0099] Here, f2() is the mapping function.

[0100] Algorithm 1 outlines the overall process of the binning-based exact algorithm. It first scans the database once to construct the complete trajectory of all moving objects. Then, it maps all objects to their corresponding bins according to their spatial location. b Finally, it refines the clusters within each box based on similarity over time.

[0101]

[0102] 4. Complexity Analysis

[0103] The time complexity of the exact algorithm based on binning is... Where k1 is the average number of records per object, and k2 corresponds to... Because in railway data |t i | and n are usually not very large, therefore It is usually less than |O|.

[0104] II. Move-Approx: An approximation algorithm based on Transformer

[0105] As the minimum number of trips (n) required for objects to move together increases, the number of bins in the binning-based exact algorithm also increases, potentially putting pressure on main memory capacity. Furthermore, considering the rapid generation of railway data, timely detection of patterns of objects moving together requires a fast solution.

[0106] An approximate algorithm for solving these problems is proposed. Specifically, a semi-supervised framework is proposed, in which training labels are obtained by performing exact solutions on the trajectories of some objects that can be quickly found in main memory. Once the model is trained, it is used to reveal patterns in the trajectories of the remaining objects.

[0107] 1. Framework

[0108] Figure 4This paper presents the framework of the approximate model (Transformer-based trajectory model) proposed in this application. The idea is to extract key information from the spatial and temporal dimensions of the complete trajectory of a moving object, and then map them to corresponding clusters based on the learned deep representations. As introduced, the blurred circles represent spatial information, i.e., station IDs in a railway, while the blue rectangles represent the temporal information of each object.

[0109] 2. Transformer-based trajectory encoder

[0110] The sheer volume of real-time railway data poses a significant speed challenge to main memory-based solutions for achieving instant processing. Furthermore, rapid discovery is crucial for effective real-time pattern recognition in this context. To address this challenge, this application proposes an approximate solution to learn the latent representation behind the complete trajectory of moving objects, rather than using bins to map the moving objects in main memory. These objects are then mapped to corresponding clusters that move together.

[0111] First, embed the complete trajectory point P Embedding of time information T Using weight parameters respectively and Perform the transformation. Then, concatenate the two transformed embeddings into e. O The transformed complete trajectory embedding of the moving object O is expressed as:

[0112] e O =e P ·W P ||e T ·W T (3)

[0113] Next, an attention mechanism is used to encode key spatial and temporal information about the complete trajectory of the moving object. Through this attention mechanism, spatial and temporal information interact, and during the optimization process, key information about the object's shared movement—namely, station and temporal information—will receive more attention. This process is defined as follows:

[0114]

[0115] Where W1 is the first learnable parameter in the attention mechanism, W2 is the second learnable parameter in the attention mechanism, and W3 is the third learnable parameter in the attention mechanism. The complete trajectory of the moving object is encoded using softmax as the activation function and d as the input dimension. Then, a multi-head attention mechanism is applied to improve the robustness of the learning process.

[0116]

[0117] Among them, W h is the first trainable parameter, and || is the connection operation.

[0118] 3. Discovering common movement patterns via decoder

[0119] Specifically, the Multilayer Perceptron (MLP) layer extracts a deep representation from the complete trajectory of a moving object, as shown below:

[0120]

[0121] Among them, W m It is the second trainable parameter.

[0122] Next, the sigmoid function is used to map the deep representations of the objects to their corresponding common movement clusters:

[0123] y′ O =sigmoid(e′) O (7)

[0124] Where, y′ O It is the prediction cluster for moving object O.

[0125] 4. Optimization

[0126] The loss function of the Transformer-based approximation algorithm is expressed as:

[0127]

[0128] in, pc represents the total number of clusters, and pc is the weight of the positive samples in the cluster, used to solve the problem of imbalance between positive and negative samples in each cluster.

[0129] The overall process of the Transformer-based approximation model is shown in Algorithm 2. During training, the Transformer-based encoder and decoder architectures are optimized. We learn the model parameters. These learned parameters are then used for common movement pattern detection in real-time scenarios. Because the required parameters have already been learned during the training phase, pattern detection is very fast, as shown in Section 3.

[0130]

[0131] 5. Complexity Analysis

[0132] The detection time complexity of the Transformer-based approximation algorithm is O(n). Where H is the number of heads, B is the batch size, and |p|m This is the maximum length of the input trajectory. Since H and B are constants, the time complexity of the approximate solution is O(n log n). Where |p| m Typically, the frequency is small (e.g., about 6 times per week in the evaluation dataset).

[0133] Here, the properties of the algorithm are discussed as follows:

[0134] The Transformer-based approximation algorithm demonstrates its relatively low time complexity. It is quadratic with respect to the trajectory length. This proves the efficiency of the algorithm.

[0135] Its training process exhibits significant scalability, maintaining near-linear complexity as the size of the mobile object dataset increases. This scalability is further validated in experimental results across four datasets, such as... Figure 5 As shown. Figure 5 (a) The training time of the algorithm increases linearly as the number of passengers increases; Figure 5 (b) A graph showing the comparison results of move-Approx with several other methods as the number of passengers increases; Figure 5 (c) The training time increases linearly with the number of days. Figure 5 (d) is a graph showing the comparison between move-Approx and several other methods as the number of days increases.

[0136] There is a trade-off between effectiveness and efficiency. When the minimum number of trips n that objects should move together is small and main memory can accommodate the boxes, an exact solution can be used to generate an exact cluster. However, when computational resources are limited, an approximate algorithm can be chosen.

[0137] The following explanation will be further illustrated with simulation experiments.

[0138] 1. Experimental setup

[0139] (1) Datasets: Methods evaluated on real-world datasets. These datasets consist of anonymous passenger entry and exit records from the MTR Corporation (Hong Kong), the sole MTR service provider in Hong Kong, which handles 40% of the daily passenger flow. MTR data was extracted from November 1st to November 7th, 2020, November 14th, 2020, November 21st, 2020, and November 28th, 2020, resulting in MTR-1, MTR-2, MTR-3, and MTR-4 datasets. Statistical information for these datasets is shown in Table 2.

[0140] Table 2: Trajectory data for all datasets

[0141]

[0142]

[0143] (2) Preprocessing: Extract the station and timestamp information for all passengers' entry and exit points. For trips with missing entry and exit records, filter them out during the evaluation.

[0144] (3) Other methods: The model of this application is compared with six other algorithms.

[0145] ObjectGrowth: A standard approach to discovering cluster patterns in shared mobility scenarios.

[0146] CMC: A classic method focused on detecting formation patterns between moving objects.

[0147] Flock: A traditional method designed to identify clustering patterns among moving objects.

[0148] PlatoonMiner: An efficient method specifically designed for formation pattern discovery.

[0149] Buddy-based Clustering (Buddy): A well-known method for discovering common movement patterns by using partner clustering and merging.

[0150] EvolvingClusters: An advanced approach designed specifically for discovering evolutionary clusters.

[0151] Six competing algorithms were employed to discover common movement patterns in railway data to ensure adjusted accuracy. Specifically, for methods designed to discover clusters of continuously moving objects, clusters were continuously tracked across all past time intervals, and it was checked whether objects appeared simultaneously in these discrete time intervals. The efficiency of the model proposed in this application was then compared with these adjusted competing algorithms.

[0152] (4) Performance metrics: The standard metric Hamming Loss (HL) is used to evaluate performance.

[0153] Hamming Loss represents the proportion of clusters that make prediction errors, and is defined as follows:

[0154]

[0155] Where N represents the number of objects and M represents the number of clusters, if we predict cluster y i ′ and the actual cluster y i If they are different Returns 1 otherwise returns 0.

[0156] (5) Hyperparameter settings: A grid search was applied to find the optimal hyperparameters for the method across all datasets. The number of heads H was searched in {2,3,4,5}, and the minimum cluster size g was searched in {3,4,5,6}. m Search for the hidden state dimension d in {256,512,1024,2048}.

[0157] 2. Effectiveness assessment

[0158] The effectiveness of the proposed Move-Approx was evaluated by varying the training ratio r and the minimum number of runs n.

[0159] (1) Changing the training ratio r: When n=4, evaluate the effectiveness of Move-Approx with respect to r in Table 3, where r=0.5 indicates that only 50% of the data was observed during the training phase. The following observations were made:

[0160] The evaluation metric HL increases with a higher training ratio. This suggests that more observations are beneficial for encoder-decoder based methods. The HL value peaks when a larger proportion of data is observed during training, particularly when r = 0.8.

[0161] Compared to other datasets, Move-Approx achieved the best results on the MTR-4-week dataset with r=0.5 or 0.6. This is because longer datasets provide clearer common movement patterns, which are easier for the model to detect.

[0162] Table 3: Average results of five runs with varying training ratio r

[0163] 0.50 0.15 0.18 0.22 0.15 0.60 0.12 0.14 0.13 0.07 0.70 0.09 0.08 0.05 0.06 0.80 0.00 0.07 0.00 0.05

[0164] (2) Variation of minimum number of passes n: When r = 0.7, the effectiveness of Move-Approx on all datasets was evaluated by varying the minimum number of passes n. Table 4 shows that for larger n values, i.e., n = 5 and 6, the HL value exhibited the smallest error in pattern detection, at 0.00, meaning all predicted clusters were perfectly correct. This indicates that using more passes for pattern detection is more accurate. The basic principle is that a larger n, compared to a lower n, makes patterns moving together more distinct.

[0165] Table 4: Average results of five runs when n is changed

[0166] 3 0.04 0.03 0.09 0.08 4 0.09 0.08 0.05 0.06 5 0.00 0.00 0.00 0.00 6 0.00 0.00 0.00 0.00

[0167] 3. Efficiency Evaluation

[0168] Evaluations were conducted using a GeForce GTX 1080 Ti 11 GB GPU. Average runtimes for all methods are reported, based on five runs.

[0169] (1) Efficiency comparison: such as Figure 6 As shown, the average runtime of all four methods on the dataset is displayed; Figure 6 (a) represents the training time of Move-Exact on four datasets; Figure 6 (b) Comparison of detection time between Move-Exact and other methods. The x-axis represents the dataset for MTR-i weeks, i∈{1,2,3,4}. Experiments show that Move-Exact runs faster than other methods. This verifies that binning-based solutions designed for discrete timestamps outperform other methods designed for continuous timestamps. Since model training can be performed offline, Move-Approx achieves an average speedup of two orders of magnitude in detection time compared to the other six methods.

[0170] (2) Scalability of the number of objects: Figure 5 (a) and Figure 5 (b) describes the scalability of the model and other methods in this application with respect to the number of passengers, demonstrating a linear relationship between runtime and the number of passengers.

[0171] (3) Scalability of time interval size: Figure 5 (c) and Figure 5 (d) illustrates the scalability of the model and other methods in this application when the time dimension is expanded from 7 days to 365 days. The trend shows linear scalability, indicating the efficiency of the model in handling different time intervals.

[0172] 4. Ablation test

[0173] Ablation experiments were conducted to evaluate the impact on key components in the model. Since other settings are similar to those for MTR-1 and MTR-2 weeks when n = 3, 4, only the latter setting is shown. Two variants are considered:

[0174] Move-Approx na (No attention mechanism): This variant excludes the attention mechanism from the model in this application (Equation 4).

[0175] Move-Approx nw (No weight): This variant excludes the weights of positive samples in the cluster from the loss function (Equation 8).

[0176] Figure 7 The results of the ablation experiment are shown. Figure 7(a) For Move-Approx and Move-Approx on the MTR-1 week dataset na and Move-Approx nw Performance comparison Figure 7 (b) For Move-Approx and Move-Approx on the MTR-2 week dataset na and Move-Approx nw Performance comparison. It can be observed that:

[0177] (1) Move-Approx outperforms Move-Approx on all datasets. na The performance gap is even more significant on the MTR-2 week dataset, which demonstrates the importance of attention mechanisms in improving performance.

[0178] (2) Move-Approx outperforms Move-Approx on all datasets. nw This is particularly evident on the MTR-2 week dataset. This confirms the importance of incorporating the weights of positive samples into the loss function for effectively discovering common movement patterns.

[0179] 5. Hyperparameters

[0180] This application proposes a hyperparameter analysis to examine the impact of key hyperparameters on model performance. Since the trends of the MTR-2, MTR-3, and MTR-4 week datasets are similar to those of the MTR-1 week dataset, they are omitted here. Specifically, two hyperparameters are proposed:

[0181] (1) Minimum cluster size g m The minimum cluster sizes of different scales were evaluated, including 3, 4, 5 and 6. Figure 8 As shown in (a), different g m The impact on detection performance. Figure 8 (a) It can be seen that performance increases with g m The performance improves with increasing cluster size until it reaches a certain point (i.e., 5), after which it reaches its optimal level. This is because larger cluster sizes result in more passengers moving together, making them easier to detect.

[0182] (2) Dimension d of the hidden state. The effects of changing the dimension of the hidden state are proposed, with values ​​selected from 256, 512, 1024 and 2048. Figure 8 (b) shows the effect of different values ​​of d on detection performance. From... Figure 8 (b) It can be seen that beyond a certain point (i.e. 512), increasing d will lead to a performance decrease due to overfitting.

[0183] These observations provide insights into selecting appropriate hyperparameters to achieve optimal model performance.

[0184] 6. Case Study

[0185] This application provides a case study to illustrate the insights gained from common movement patterns discovered by the proposed model. Specifically, it compares the results of the model in this application when n=3 and n=1. Figure 9 It provides visualization of the common movement pattern. Figure 9 (a) is the common movement pattern captured when n=1. Figure 9 (b) is the common movement pattern captured when n=3. Passengers O1, O2, O3, O4, O5

[0186] Their journey begins at station A, departing at 8:10 and arriving at station B at 8:50. Subsequently, O1, O2, and O5 depart from station B at 18:00 and arrive at station C at 19:00. Finally, O1 and O5 depart from station C at 19:00 and arrive at station D at 19:40.

[0187] The model in this application demonstrates the ability to capture relationships and time information among passengers when n=3, generating more accurate and meaningful co-movement patterns, i.e., O1 and O5 appearing simultaneously in three trips. On the other hand, when n=1, the detected pattern of passengers O1, O2, O3, O4, and O5 moving together in at least one trip appears less coherent and informative.

[0188] This case demonstrates the superior performance of the model proposed in this application in discovering co-movement patterns in railway big data, showcasing its effectiveness in capturing meaningful co-movement among passengers, especially when the minimum number of trips n for which passengers move together is large.

[0189] In this application, we propose a precise solution for discovering common movement patterns in railway big data. This is challenging due to the non-continuous timestamps and large, non-fixed time intervals. Previous methods for tracking continuous timestamps incur additional costs because they monitor unnecessary timestamps. This challenge is addressed by introducing a bin-based precision algorithm. All objects are assigned to corresponding bins based on their proximity in the spatial trajectory. Then, finer intra-bin refinement is used to examine the similarity of the temporal dimension within each bin. The novelty lies in avoiding the need to track continuous timestamps, thus achieving lower time complexity than state-of-the-art methods.

[0190] Furthermore, due to the massive volume of railway data and its real-time updates in real-world scenarios, this application proposes a method to accelerate the processing time of large-scale railway data. Additionally, common movement patterns must be computed within very short time intervals to facilitate real-time discovery. To overcome these challenges, this application proposes a Transformer-based approximation algorithm. It uses a depth representation to encode the spatial and temporal trajectory information of passengers, and then efficiently decodes common movement patterns for rapid discovery.

[0191] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, "a plurality of" means two or more, unless otherwise explicitly specified.

[0192] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0193] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the appended claims.

Claims

1. A large-scale data-driven method for discovering common mobility patterns, characterized in that, The method includes: Scan the database to construct the complete trajectory of all moving objects; specifically, this includes traversing the database of records. By connecting moving objects During the observation time window Use the trajectory points in the railway data to obtain each moving object The complete trajectory; Based on the complete trajectory, all moving objects are mapped to corresponding bins according to their spatial location; specifically, this includes: based on each moving object... Complete trajectory Through combined operations Select from the complete trajectory medium size All itineraries ; The spatial position of each combination in the next trip forms a box label. The object will be mapped to the corresponding box; The function is defined as: （1） in, A mapping function for extracting spatial labels; The time complexity is ,in, The average number of records per moved object. Corresponding to , The number of objects to be moved; The clusters of moving objects within each box are refined based on similarity over time to obtain a dataset of common movement patterns; specifically including: For the moving objects in each box Retrieve moving objects Complete trajectory ; Get the tagged box label Time information ; Move the object based on its time information. Mapped to fine-grained clusters; The function is defined as: （2） in, For fine-grained clustering labels, For mapping functions; Construct a Transformer-based trajectory model; wherein, the Transformer-based trajectory model includes a Transformer-based trajectory encoder and decoder; Train a Transformer-based trajectory model using a common movement pattern dataset; We used the trained Transformer-based trajectory model to discover common movement patterns on the test dataset.

2. The large-scale data-driven common mobility pattern discovery method according to claim 1, characterized in that, The processing steps of a Transformer-based trajectory encoder include: Complete trajectory points Embedded and time information Embedded Using the first weight parameter respectively Second weight parameter Perform the conversion; The first weight parameter Second weight parameter The converted embedded connection is ; Move object The transformed complete trajectory embedding is expressed as: （3） in, Embed the complete trajectory of the moving object; The key spatial and temporal information of the complete trajectory of a moving object is encoded using an attention mechanism, and the process is defined as follows: （4） in, This is the first learnable parameter in the attention mechanism. This is the second learnable parameter in the attention mechanism. This is the third learnable parameter in the attention mechanism. Encoding the complete trajectory of a moving object, For activation function, d The dimension of the input; Leveraging multi-head attention mechanisms to improve learning robustness: （5） in, For the output of the multi-head attention mechanism, H For the number of long positions, The first trainable parameter, For connection operations.

3. The large-scale data-driven common mobility pattern discovery method according to claim 2, characterized in that, The decoder's processing steps include: The multilayer perceptron layer extracts a deep representation from the complete trajectory of a moving object, as shown below: （6） in, A deep representation of the complete trajectory encoding. This is the second trainable parameter; The sigmoid function is used to map the deep representation of an object to its corresponding common movement cluster: （7） in, For predicting clusters of moving objects.

4. The large-scale data-driven common mobility pattern discovery method according to claim 3, characterized in that, The loss function of the Transformer-based trajectory model is expressed as: （8） in, The total number of clusters, The weights of the positive samples in the cluster. For clusters, For the input sample, For sample labels, For clusters The corresponding samples, For clusters The corresponding sample labels.

Images