Data prediction method, apparatus, device, storage medium, and computer program product

Abstract

This application discloses a data prediction method, apparatus, device, storage medium, and computer program product, relating to the fields of artificial intelligence and data processing technology. Embodiments of this application can be applied to fields such as mapping and transportation. The method includes: acquiring a universal causal graph to characterize causal knowledge prevalent in different regions, whereby the causal knowledge characterizes causal relationships between attributes of different regions; based on the universal causal graph and observation data of a target region, acquiring the full range of regional attributes corresponding to each region in the target region, including unobserved regional attributes; and based on the full range of regional attributes corresponding to each region in the target region, acquiring predicted values ​​for object flows between regions in the target region. This application solves the problem of inaccurate object flow prediction in some regions due to a lack of observation data, thus improving the accuracy of object flow prediction.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and data processing technology, and in particular to a data prediction method, apparatus, device, storage medium, and computer program product. Background Technology

[0002] With the development of artificial intelligence technology, machine learning techniques can be used to predict the movement of objects in cities.

[0003] Related technologies use neural networks to obtain predictive models that can be used to predict object flows between different regions. However, these predictive models require a large amount of observational data for training. For cities where it is difficult to obtain large amounts of observational data, there is a problem of data sparsity. In this case, the object flow predictions obtained by these technologies are not accurate enough. Summary of the Invention

[0004] This application provides a data prediction method, apparatus, device, storage medium, and computer program product, which can improve the prediction accuracy of object flow in areas with sparse observation data. The technical solution is as follows:

[0005] According to one aspect of the embodiments of this application, a data prediction method is provided, the method comprising:

[0006] Acquire observation data of the source region and observation data of the target region; wherein the observation data of the source region is greater than the observation data of the target region, and the observation data includes the regional attributes corresponding to each area in the region;

[0007] Based on the observation data of the source region, a universal causal graph is obtained. The universal causal graph is used to represent the causal knowledge that is prevalent in different regions. The causal knowledge is used to represent the causal relationship between different regional attributes.

[0008] Based on the universal causal graph and the observation data of the target region, the full regional attributes corresponding to each region in the target region are obtained, and the full regional attributes include unobserved regional attributes;

[0009] Based on the full-area attributes corresponding to each region in the target region, the predicted values ​​of object flow between each region in the target region are obtained.

[0010] According to one aspect of the embodiments of this application, a data prediction apparatus is provided, the apparatus comprising:

[0011] The observation data acquisition module is used to acquire observation data of the source region and observation data of the target region; wherein, the observation data of the source region is greater than the observation data of the target region, and the observation data includes the regional attributes corresponding to each area in the region;

[0012] The causal graph acquisition module is used to acquire a universal causal graph based on the observation data of the source region. The universal causal graph is used to represent causal knowledge that is common in different regions, and the causal knowledge is used to represent the causal relationship between attributes of different regions.

[0013] The full attribute acquisition module is used to acquire the full regional attributes corresponding to each region in the target region based on the universal causal graph and the observation data of the target region. The full regional attributes include unobserved regional attributes.

[0014] The prediction value acquisition module is used to acquire the predicted value of object flow between regions in the target region based on the full region attributes corresponding to each region in the target region.

[0015] According to one aspect of the embodiments of this application, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the above-described data prediction method.

[0016] The computer device is a terminal or a server.

[0017] According to one aspect of the embodiments of this application, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored in the storage medium, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the above-described data prediction method.

[0018] According to one aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the data prediction method described above.

[0019] The technical solutions provided in this application embodiment may have the following beneficial effects:

[0020] By acquiring a universal causal graph based on a large amount of observation data from the source region, and then acquiring the full regional attributes corresponding to each region in the target region based on the universal causal graph and a small amount of observation data from the target region, the flow of objects between the regions in the target region is predicted based on the full regional attributes corresponding to each region in the target region. The universal causal graph makes up for the defects caused by the sparsity of observation data and solves the problem of inaccurate prediction of object flow in some regions due to the lack of observation data in related technologies, thereby improving the prediction accuracy of object flow.

[0021] Furthermore, the solution provided by the embodiments of this application only requires a small amount of observation data of the target area to predict the flow of objects between different areas in the target area, thereby reducing the amount and difficulty of obtaining observation data of the target area and improving the prediction efficiency of object flow. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a data prediction method provided in one embodiment of this application;

[0024] Figure 2 This is a schematic diagram of a causal discovery model provided in one embodiment of this application;

[0025] Figure 3 This is a schematic diagram of a data recovery model provided in one embodiment of this application;

[0026] Figure 4 This is a flowchart of a method for obtaining object flow prediction values ​​according to an embodiment of this application;

[0027] Figure 5 This is a schematic diagram of a training method for a first prediction model provided in one embodiment of this application;

[0028] Figure 6 This is a schematic diagram of a training method for a second prediction model provided in one embodiment of this application;

[0029] Figure 7 This is a schematic diagram of a method for obtaining object flow prediction values ​​according to an embodiment of this application;

[0030] Figure 8 This is a block diagram of a data prediction apparatus provided in one embodiment of this application;

[0031] Figure 9 This is a block diagram of a data prediction apparatus provided in another embodiment of this application;

[0032] Figure 10 This is a block diagram of a computer device provided in one embodiment of this application. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0034] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0035] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0036] Machine learning (ML) is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learn-by-doing.

[0037] With the research and advancement of artificial intelligence (AI) technology, AI is being studied and applied in various fields, such as smart homes, smart wearable devices, virtual assistants, smart speakers, smart marketing, autonomous driving, drones, robots, smart healthcare, and smart customer service. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0038] The technical solution provided in this application relates to machine learning technology in artificial intelligence. It utilizes machine learning technology to obtain a prediction model for predicting object flow, a data recovery model for obtaining full-area regional attributes, and a causal discovery model for obtaining a universal causal graph. Based on the prediction model, the data recovery model, and the causal discovery model, the prediction of object flow is achieved.

[0039] The method provided in this application can be executed by a computer device, which refers to an electronic device with data computing, processing, and storage capabilities. This computer device can be a terminal such as a PC (Personal Computer), tablet computer, smartphone, desktop computer, wearable device, or intelligent robot; or it can be a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0040] The technical solutions provided in this application are applicable to any scenario requiring object flow prediction, such as object flow prediction, urban development planning, transportation planning, and object flow research. The technical solutions provided in this application can improve the accuracy and efficiency of object flow prediction in areas with sparse observation data.

[0041] For example, taking the object flow prediction scenario as an example. Based on a large amount of observation data of the source region, a universal causal graph is obtained. Then, based on the universal causal graph and a small amount of observation data of the target region, the full regional attributes corresponding to each region in the target region are obtained. The full regional attributes include the observed regional attributes and the unobserved regional attributes corresponding to the region. Finally, based on the full regional attributes corresponding to each region in the target region, the predicted values ​​of object flow between the regions in the target region are obtained.

[0042] Please refer to Figure 1 The diagram shows a flowchart of a data prediction method provided in an embodiment of this application. The execution subject of each step of the method can be a computer device, such as a terminal or a server. The method may include the following steps (101-104).

[0043] Step 101: Obtain observation data from the source region and observation data from the target region; wherein, the observation data from the source region is more than the observation data from the target region, and the observation data includes the regional attributes corresponding to each area in the region.

[0044] In this application embodiment, the hierarchical level of the region can be either the national level or the city level; this application embodiment does not limit the hierarchical level of the region. For example, when the region is at the national level, it can be divided according to the cities within the region to obtain the various areas corresponding to that region. When the region is at the city level, it can be divided into grids according to latitude and longitude to obtain the various areas corresponding to that region, or it can be divided into blocks according to the road network within the city to obtain the various areas corresponding to that region; this application embodiment does not limit this approach. Here, a region can refer to the smallest unit after spatially dividing a region according to some method (such as latitude and longitude, road network, etc.).

[0045] Optionally, the source region refers to a region where a large amount of observational data can be obtained, while the target region refers to a region where only a small amount of observational data can be obtained. For example, the source region can be a region where obtaining observational data is relatively easy, while the target region can be a region where obtaining observational data is more difficult. For instance, taking the city level as an example, for developed cities (i.e., source cities), the city structure is relatively mature, and the data collection equipment within the city is relatively complete, so a large amount of observational data can be obtained for developed cities. For developing cities (i.e., target cities), the city structure is still developing, and the data collection equipment within the city is incomplete, so obtaining observational data for developing cities is difficult, and only a small amount of observational data can be obtained without spending a lot of time and money. The observational data can at least include: the regional attributes corresponding to each area within the region, and the movement of objects between different areas (i.e., population commuting).

[0046] Regional attributes refer to the characteristics of a region, which can be used to characterize certain features of the region itself. Regional attributes include, but are not limited to, the regional population composition distribution, the regional POI (Point of Interest) distribution, the regional economy, the regional transportation structure, and regional policies. A POI is a geographical location, typically a location that is useful and of interest to the problem (such as predicting object movement) (e.g., the geographical locations corresponding to train stations, subway stations, shopping malls, hotels, etc.). The distribution of POI types within a region is closely related to the region's land use functions.

[0047] Optionally, since the source region has a large amount of observational data, the full range of regional attributes corresponding to each area in the source region can be obtained based on the observational data of the source region. This full range of regional attributes is used to characterize all the regional attributes corresponding to the area. However, since the target region only has a small amount of observational data, and some areas (such as areas planned for future construction) even lack observational data, the observational data of the target region can only obtain some of the regional attributes corresponding to each area in the target region. That is, compared with the true regional attributes of each area in the target region, there are missing regional attributes, such as regional attributes that cannot be reflected by the observational data of some target regions (i.e., unobserved regional attributes).

[0048] Step 102: Based on the observation data of the source region, obtain a universal causal graph. This universal causal graph is used to represent the causal knowledge that is common in different regions. The causal knowledge is used to represent the causal relationship between different regional attributes.

[0049] A causal graph is a causal representation in causal inference theory. It is a DAG (Directed Acyclic Graph) that satisfies the directed acyclic constraint. A causal graph includes nodes and causal edges. In this embodiment, the causal graph uses regional attributes as nodes. Causal edges are used to characterize whether a causal relationship exists between regional attributes. For example, for a first regional attribute and a second regional attribute in the causal graph, if there is a causal edge from the first regional attribute to the second regional attribute, then the first regional attribute can be called the cause of the second regional attribute; that is, intervention or regulation of the first regional attribute will affect the second regional attribute (i.e., a causal effect). In some examples, the causal graph can be referred to as a Structural Causal Model.

[0050] Optionally, a universal causal graph refers to a causal graph constructed based on causal knowledge that is universally present in a region. For example, taking the city level as an example, a universal causal graph is used to represent universally unchanging causal knowledge within a city. This causal knowledge is the essential mechanism behind things, and this mechanism exhibits consistency across cities; that is, the causal knowledge in the universal causal graph corresponding to a city applies to any city. For example, if the universal causal graph indicates that regional attribute A is the cause of regional attribute B, then for any city, there exists a situation where regional attribute A is the cause of regional attribute B.

[0051] In one example, reinforcement learning methods can be used to mine prevalent causal knowledge in source regions with abundant observational data, in order to construct a universal causal graph. The specific process can be as follows:

[0052] The regional attributes corresponding to each region in the source region are transformed to obtain the state vectors corresponding to each regional attribute. For the first regional attribute and the second regional attribute, the state vectors corresponding to the first regional attribute and the second regional attribute are calculated to obtain a first predicted value between the first regional attribute and the second regional attribute. This first predicted value is used to characterize the possibility that there is a directed acyclic causal edge between the first regional attribute and the second regional attribute. Based on the first predicted value between each regional attribute, candidate causal graphs are obtained. The universal score corresponding to the candidate causal graph is obtained. The universal score is used to measure the degree of matching between the candidate causal graph and the observed data. Based on the universal score, a universal causal graph is selected from the candidate causal graphs.

[0053] Here, the state vector refers to the vector obtained by mapping the region attributes to the state space.

[0054] For example, refer to Figure 2 A universal causal graph can be obtained through the causal discovery model 201. The causal discovery model 201 includes a graph generator, which comprises an encoder 202 and a decoder 203. The encoder 202 transforms the region attributes corresponding to each region in the source region to obtain the state vector corresponding to each region attribute. Optionally, the encoder 202 can use a neural network with a Transformer structure, specifically an encoder based on a self-attention mechanism within the Transformer architecture.

[0055] Decoder 203 is used to obtain the first predicted values ​​between the attributes of each region. Optionally, decoder 203 can use a bilinear model and a sigmoid activation function, that is, the range of the first predicted values ​​is 0-1. Then, based on the first predicted values ​​between the attributes of each region, a candidate causal graph is constructed.

[0056] Alternatively, the reward function of the causal discovery model can be determined by the weighted sum between the Bayesian Information Criteria (BIC) (i.e., the scoring function used to obtain a universal score) in the field of causal inference and the directed acyclic penalty term of the graph structure.

[0057] The causal discovery model is optimized by maximizing the reward function. During the iterative optimization of the causal discovery model, all generated directed acyclic graphs (i.e., candidate causal graphs) are recorded. The candidate causal graph corresponding to the best reward function (i.e., the one with the largest reward function value) is determined as the universal causal graph.

[0058] Due to the complexity of factors and the numerous regional attributes within a region, as well as the highly complex relationships between these attributes, the solution space discovered by universal cause-effect graphs is extremely large (0 (2 n This application utilizes reinforcement learning for solution search to find the causal graph that best matches the data generation pattern (i.e., the universal causal graph), which can greatly improve search efficiency.

[0059] Step 103: Based on the universal causal graph and the observation data of the target region, obtain the full regional attributes corresponding to each region in the target region. The full regional attributes include unobserved regional attributes.

[0060] This application embodiment, based on causal knowledge in a universal causal graph and combined with a small number of regional attributes corresponding to the target region, can obtain all regional attributes of the regions within the target region, thereby restoring a comprehensive characterization of the regions within the target region. This enables the prediction of object flow between regions in a target region with sparse observation data. Because this application embodiment can predict object flow in target regions with sparse observation data, and is not limited to regions with abundant observation data, it improves the versatility of object flow prediction.

[0061] Optionally, embodiments of this application construct causal constraints between regional attributes through causal knowledge to obtain the full set of regional attributes corresponding to each region in the target region based on a small number of regional attributes corresponding to each region. The specific process can be as follows: a data recovery model is trained based on a universal causal graph and observation data of the source region. This data recovery model is used to obtain the unobserved regional attributes corresponding to the region. For the first region in the target region, the full set of regional attributes corresponding to the first region is obtained through the data recovery model based on the observation data of the first region.

[0062] The full set of regional attributes corresponding to the first region includes both observed and unobserved regional attributes. The first region can refer to any region within the target area. Optionally, the network structure of the data recovery model can be a VAE (Variational Auto-encoder). For example, by inputting a small number of regional attributes corresponding to the first region into the data recovery model, the full set of regional attributes corresponding to the first region can be obtained.

[0063] In one example, the training process of the data recovery model can be as follows: Based on the observation data of the source region, construct training samples; where the training samples are any regional attributes from the regional attributes corresponding to each region in the source region; obtain the predicted missing regional attributes corresponding to the training samples using the data recovery model; obtain the inferred missing regional attributes corresponding to the training samples based on the universal causal graph; obtain the training loss of the data recovery model based on the predicted missing regional attributes, the inferred missing regional attributes, and the training samples; adjust the parameters of the data recovery model based on the training loss of the data recovery model to obtain the trained data recovery model.

[0064] Optionally, predicting missing regional attributes refers to the missing regional attributes predicted through a data recovery model, while inferring missing regional attributes refers to inferring missing regional attributes based on the causal constraints between regional attributes in a universal causal graph. For example, for regional attribute A, each regional attribute causally related to regional attribute A is sequentially recovered based on the causal links between regional attribute A and other regional attributes in the universal causal graph. For instance, if regional attribute A is the cause of regional attribute B, and regional attribute B is the cause of regional attribute C, then regional attribute A can be recovered first, followed by regional attribute B, and finally regional attribute C. Therefore, for regional attribute A, regional attributes B and C can be identified as the inferred missing regional attributes.

[0065] This application embodiment learns causal knowledge in a universal causal graph through a data recovery model, and trains the data recovery model using observation data from the source region to obtain a data recovery model with full regional attributes that can be used to recover the region.

[0066] In one example, based on the above motivation, the process of obtaining the training loss of the data recovery model can be as follows: Obtain the missing region loss based on the difference between the predicted missing region attributes and the inferred missing region attributes; obtain the spatial distribution of latent variables based on the predicted missing region attributes and the training samples; sample the first latent variable from the spatial distribution of the latent variables; obtain the output full region attributes corresponding to the training samples based on the first latent variable; obtain the reconstruction loss based on the difference between the region attributes corresponding to the training samples and the output full region attributes; obtain the constraint loss based on the spatial distribution of the latent variables; and obtain the training loss of the data recovery model based on the missing loss, reconstruction loss, and constraint loss.

[0067] Optionally, a Gaussian distribution, or a unit Gaussian distribution, or a Gaussian mixture distribution can be used to constrain the spatial distribution of the latent variables. For example, the distance between the spatial distribution of the latent variables and the Gaussian distribution (or a unit Gaussian distribution, or a Gaussian mixture distribution) can be calculated using KL divergence to generate a constraint function. The mean squared error function can be used to obtain the missing loss and reconstruction loss, and the sum of the missing loss, reconstruction loss, and constraint loss can be used to determine the training loss of the data recovery model.

[0068] By minimizing the training loss of the data recovery model, the data recovery model learns causal knowledge in the universal causal graph and the ability to reconstruct attributes for the entire region.

[0069] Optionally, the likelihood expectation of the training samples can be constructed based on a Gaussian distribution, a unit Gaussian distribution, or a mixture of Gaussian distributions. The difference between the likelihood expectation and the reconstruction loss and missing loss is determined as the training loss of the data recovery model. By maximizing the training loss of the data recovery model, the data recovery model learns the causal knowledge in the universal causal graph and learns the reconstruction ability for the attributes of the entire region.

[0070] Optionally, the training loss of the data recovery model can be obtained directly based on the reconstruction loss and constraint loss, or it can be obtained directly based on the likelihood expectation and constraint loss. This application embodiment does not limit this.

[0071] For example, refer to Figure 3 The data recovery model 301 includes an encoder 302 and a decoder 303. The encoder 302 predicts the missing region attribute Y based on the training sample X, then concatenates X and Y to obtain the connection region attribute. Based on the connection region attribute, the encoder 302 obtains the spatial distribution Z of the latent variables. The first latent variable is sampled from Z, and the decoder 303 recovers the output training sample based on the first latent variable. Then, based on the training samples, decoder 303... Recover the missing output region attributes using the first hidden variable. Will and A connection is made to obtain the full output region attributes. Based on the difference between the full output region attributes and the region attributes corresponding to the training sample X, the reconstruction loss is determined. Based on Z, the constraint loss is obtained. Based on the reconstruction loss and the constraint loss, the training loss of the data recovery model is obtained. The data recovery model 301 is trained using the training loss of the data recovery model to obtain the trained data recovery model 301.

[0072] Optionally, by inputting a small number of regional attributes corresponding to each region into the trained data recovery model 301, the full range of regional attributes corresponding to each region in the target region can be obtained.

[0073] Step 104: Based on the full regional attributes corresponding to each region in the target region, obtain the predicted values ​​of object flow between each region in the target region.

[0074] Object flow is typically represented as Origin-Destination flow (OD flow). Object flow refers to a detailed representation of population movement from a group aggregation perspective, encompassing both the size and direction of the flow. In this embodiment, the origin and destination are regions within a region. The predicted object flow value is used to characterize the predicted object flow between different regions.

[0075] In one example, step 104 may include the following sub-steps (104a-104c):

[0076] Step 104a: Based on the observation data of the source region, a first prediction model is trained, which is used to predict the flow of objects between different regions in the source region.

[0077] Optionally, the first prediction model can be constructed using GNNs (Graph Neural Networks). The first prediction model is trained using the regional attributes corresponding to each region in the source region. For example, the first prediction model can be trained in a supervised manner based on a large amount of observation data corresponding to a city (which may include regional attributes of each region in the city, object flow between regions, distances between regions, etc.). In one example, the training process of the first prediction model can be as follows:

[0078] 1. Using the first prediction model, obtain the starting point vector representation and ending point vector representation of each region in the source region.

[0079] Here, the starting point vector representation refers to the vector representation of the region as the starting point of the flow, and the ending point vector representation refers to the vector representation of the region as the ending point of the flow. For example, for a target region in a source region, if the target region is the starting point of the flow, the first prediction model can obtain the corresponding generation features (i.e., the starting point vector representation) of the target region; if the target region is the ending point of the flow, the first prediction model can obtain the corresponding attraction features (i.e., the ending point vector representation) of the target region.

[0080] In one example, the first prediction model includes a starting point feature extraction network and an ending point feature extraction network; the starting point feature extraction network is used to obtain the starting point vector representation of the region, and the ending point feature extraction network is used to obtain the ending point vector representation of the region. Optionally, the network structures of both the starting point feature extraction network and the ending point feature extraction network can be GNNs.

[0081] The process of obtaining the starting point vector representation and the ending point vector representation can be as follows: Based on the observation data of the source region, a first structure graph is constructed; wherein, the first structure graph uses regions in the source region as nodes, region attributes as node representations, and adjacency relationships between regions as relation edges; through the starting point feature extraction network, based on the first structure graph, the starting point vector representation corresponding to each region in the source region is obtained; through the ending point feature extraction network, based on the first structure graph, the ending point vector representation corresponding to each region in the source region is obtained.

[0082] For example, GAT (Graph Attention Networks) can be used to extract topological information about regions and spatial dependencies between regions. (See reference...) Figure 5 The first prediction model includes a starting point feature extraction network 501 and an ending point feature extraction network 502. Both the starting point feature extraction network 501 and the ending point feature extraction network 502 have the GAT network structure. Based on the first structure graph, the starting point feature extraction network 501 obtains the starting point vector representation corresponding to each region in the source region, and the ending point feature extraction network 502 obtains the ending point vector representation corresponding to each region in the source region, also based on the first structure graph.

[0083] 2. For the second and third regions in the source region, based on the starting point vector representation of the second region and the ending point vector representation of the third region, as well as the distance between the second and third regions, the predicted value of object flow from the second region to the third region is obtained.

[0084] The second region can refer to any region in the source region, and the third region can refer to any region in the source region other than the first region.

[0085] For example, refer to Figure 5 The first prediction model includes a predictor 503, which can be a simple linear model or any other predictor; this embodiment does not limit this. Based on the starting point vector representation corresponding to the second region and the ending point vector representation corresponding to the third region, as well as the distance between the second and third regions, the predictor 503 can obtain the predicted value of object flow from the second region to the third region.

[0086] Optionally, the distance between the second and third regions can be obtained from the distance matrix corresponding to the source region.

[0087] 3. Based on the starting point vector representation of the third region and the ending point vector representation of the second region, as well as the distance between the second and third regions, obtain the predicted value of object flow from the third region to the second region.

[0088] For example, refer to Figure 5 The predictor 503 can obtain the predicted value of the object flow from the third region to the second region based on the starting vector representation of the third region and the ending vector representation of the second region, as well as the distance between the second region and the third region.

[0089] 4. Based on the predicted object flow values ​​from the second region to the third region and from the third region to the second region, obtain the predicted object flow values ​​between the second region and the third region.

[0090] Optionally, the sum of the object flow forecast values ​​from the second region to the third region and the object flow forecast values ​​from the third region to the second region can be determined as the object flow forecast value between the second region and the third region.

[0091] 5. Based on the difference between the predicted object flow values ​​between the second and third regions and the actual object flow values ​​in the observed data between the second and third regions, determine the sub-prediction loss between the second and third regions.

[0092] 6. Sum the sub-prediction losses among the regions to obtain the prediction loss of the first prediction model.

[0093] 7. Based on the prediction loss of the first prediction model, train the first prediction model to obtain the trained first prediction model.

[0094] Optionally, with the goal of minimizing the prediction loss of the first prediction model, the model parameters of the first prediction model are adjusted to obtain the trained first prediction model.

[0095] Step 104b: Based on the observation data of the target area and the first prediction model, a second prediction model is trained to predict the flow of objects between different areas in the target area.

[0096] Optionally, the network structure of the second prediction model can be the same as that of the first prediction model, but the network parameters of the second prediction model are different from those of the first prediction model.

[0097] In one example, a second prediction model can be obtained based on observation data from the target region using a knowledge distillation method. The specific process is as follows: For regions four and five in the target region, a first prediction model is used to obtain the first predicted object flow value between regions four and five; a second prediction model is used to obtain the second predicted object flow value between regions four and five; based on the difference between the first and second predicted object flow values, a first sub-training loss between regions four and five is obtained; the first sub-training losses between each region in the target region are summed to obtain the first training loss of the second prediction model; based on the difference between the second predicted object flow value and the actual object flow value between regions four and five, a second sub-training loss between regions four and five is obtained; the second sub-training losses between each region in the target region are summed to obtain the second training loss of the second prediction model; based on the first and second training losses, the second prediction model is trained to obtain the trained second prediction model.

[0098] The fourth region can refer to any region within the target region, and the fifth region can refer to any region within the target region other than the fourth region. Optionally, the sum of the first training loss and the second training loss can be determined as the training loss of the second prediction model. By minimizing the training loss of the second prediction model, the model parameters of the second prediction model are adjusted to obtain the trained second prediction model.

[0099] For example, refer to Figure 6 Based on the observation data of the target region, a structure graph of the target region is constructed. This structure graph uses regions in the target region as nodes, regional attributes in the observation data as node representations, and adjacency relationships between regions as relation edges. A first prediction model 601 obtains a first predicted value of object flow based on the structure graph of the target region, and a second prediction model 602 obtains a second predicted value of object flow based on the structure graph of the target region. Based on the difference between the first and second predicted values ​​of object flow, a first training loss for the second prediction model is obtained. Based on the difference between the second predicted value of object flow and the actual value of object flow, a second training loss for the second prediction model is obtained. The second prediction model is then trained based on the first and second training losses to obtain the trained second prediction model.

[0100] This application employs a knowledge distillation method to adjust the parameters of a second prediction model based on a first prediction model and incorporating limited prediction data from the target region. This results in a second prediction model for predicting object flows between different areas within the target region. This solves the problem that a highly accurate prediction model cannot be trained using limited observation data from the target area, thus improving the prediction accuracy of the second model and further enhancing the accuracy of object flow predictions. Simultaneously, it reduces the training difficulty of the second prediction model, thereby increasing the efficiency of acquiring the second prediction model and further improving the efficiency of object flow predictions.

[0101] Step 104c: Using the second prediction model, based on the full regional attributes corresponding to each region in the target region, obtain the predicted values ​​of object flow between regions in the target region.

[0102] Optionally, a second structural graph is constructed based on the full regional attributes corresponding to each region in the target region. The second structural graph uses regions in the target region as nodes, full regional attributes as node representations, and adjacency relationships between regions as relation edges. Using a second prediction model, based on the second structural graph, the starting point vector representation and ending point vector representation corresponding to each region in the target region are obtained. Based on the starting point vector representation and ending point vector representation corresponding to each region in the target region, and the distance between each region in the target region, the predicted value of object flow between each region in the target region is obtained.

[0103] For example, for the fourth and fifth regions in the target region, the starting vector representation and ending vector representation of the fourth region, as well as the starting vector representation and ending vector representation of the fifth region, are obtained by the second prediction model based on the second structure diagram.

[0104] Using the second prediction model, based on the starting vector representation of the fourth region and the ending vector representation of the fifth region, as well as the distance between the fourth and fifth regions, the predicted object flow value from the fourth region to the fifth region is obtained.

[0105] The sum of the predicted object flow values ​​from the fourth region to the fifth region and from the fifth region to the fourth region is used to determine the object flow prediction model between the fourth and fifth regions.

[0106] In summary, the technical solution provided in this application obtains a universal causal graph based on a large amount of observation data from the source region. Then, based on the universal causal graph and a small amount of observation data from the target region, it obtains the full regional attributes corresponding to each region in the target region. Based on the full regional attributes corresponding to each region in the target region, it predicts the flow of objects between regions in the target region. The universal causal graph compensates for the defects caused by the sparsity of observation data and solves the problem of inaccurate object flow prediction caused by the lack of observation data in related technologies, thereby improving the accuracy of object flow prediction.

[0107] Furthermore, the solution provided by the embodiments of this application only requires a small amount of observation data of the target area to predict the flow of objects between different areas in the target area, thereby reducing the amount and difficulty of obtaining observation data of the target area and improving the prediction efficiency of object flow.

[0108] Furthermore, by utilizing reinforcement learning for solution search to find universal causal graphs, the search efficiency of universal causal graphs can be greatly improved.

[0109] Furthermore, by employing a knowledge distillation method, the model parameters of the second prediction model are adjusted based on the first prediction model and combined with a limited amount of prediction data from the target region. This yields a second prediction model for predicting object flows between different areas within the target region. This solves the problem that a highly accurate prediction model cannot be trained using limited observation data from the target area, thus improving the prediction accuracy of the second prediction model and further enhancing the accuracy of object flow predictions. Simultaneously, it reduces the training difficulty of the second prediction model, thereby increasing the efficiency of acquiring the second prediction model and further improving the efficiency of object flow predictions.

[0110] In an exemplary embodiment, taking an object flow prediction scenario as an example, refer to... Figure 7 The method for obtaining the predicted values ​​of object flow between different areas in the target region can be as follows:

[0111] 1. Obtain a large amount of observation data from the source region and a small amount of observation data from the target region. The observation data includes the regional attributes corresponding to each area in the region and the actual values ​​of object flow between each area.

[0112] 2. Using the causal discovery model 701, a universal causal graph is obtained based on the regional attributes corresponding to each region in the source region.

[0113] Optionally, a universal causal graph is used to represent causal knowledge prevalent in different regions, which in turn represents causal relationships between different regional attributes. The causal discovery model 701 refers to a model used to extract causal knowledge from regions with abundant observational data.

[0114] 3. Based on the universal cause-effect graph and the regional attributes corresponding to each region in the source region, a data recovery model 702 is trained. The data recovery model 702 is used to obtain all regional attributes.

[0115] 4. Using data recovery model 702, based on a small number of regional attributes corresponding to each region in the target region, obtain the full set of regional attributes corresponding to each region in the target region. Regional attributes are used to characterize the features of a region. The full set of regional attributes includes both observed and unobserved regional attributes, that is, regional attributes that can be reflected by the observed data and regional attributes that cannot be reflected by the observed data.

[0116] 5. Based on the regional attributes corresponding to each region in the source region and the adjacency relationships between the regions, a first prediction model 703 is constructed. The first prediction model 703 is used to predict the object flow between regions in the source region. The first prediction model 703 can be constructed based on GNN.

[0117] 6. Using a knowledge distillation method, a second prediction model 704 is trained based on the first prediction model 703, the regional attributes corresponding to each region in the target region, and the distances between each region in the target region. The second prediction model 704 is used to predict object flow between regions in the target region. Optionally, the network structure of the second prediction model 704 is the same as the network structure of the first prediction model 703.

[0118] 7. Using the second prediction model 704, based on the full regional attributes corresponding to each region in the target region and the distance between each region in the target region, obtain the predicted value of object flow between each region in the target region.

[0119] For technical details not described in the embodiments of this application, please refer to the above embodiments, which will not be repeated here.

[0120] To verify the performance of the technical solution provided in this application, experiments were conducted using publicly available observation data from three cities, denoted as A, B, and C. City A is designated as the source city with abundant observation data, while cities B and C are designated as target cities with sparse observation data. The spatial granularity of the region is at the census tract level. The total regional attributes are set to 60, which may include population structure distribution, the number distribution of different types of POIs, etc. The actual value of object mobility is the statistical result of the annual employment relationship census, specifically the actual value of object mobility in 2015.

[0121] The baseline algorithms for price comparison are the traditional gravity model and the ordinary GAT. The number of missing region attributes is set to 40. The evaluation metric chosen is RMSE (Root Mean Square Error), which is the RMSE of each model based on the number of predicted missing region attributes and 40. Referring to Table 1, which shows the RMSE corresponding to the gravity model, the ordinary GAT, and this application, respectively. For cities A and C, the RMSE values ​​corresponding to this application are lower than those corresponding to the gravity model and the ordinary GAT. In terms of the RMSE evaluation metric, this application can achieve at least an 18% improvement over the gravity model and the ordinary GAT. This demonstrates that the technical solution provided by this application can significantly improve the prediction accuracy of object flow in target areas with sparse observation data.

[0122]

[0123] Table 1

[0124] In summary, the technical solution provided in this application obtains a universal causal graph based on a large amount of observation data from the source region. Then, based on the universal causal graph and a small amount of observation data from the target region, it obtains the full regional attributes corresponding to each area in the target region. Based on the full regional attributes corresponding to each area in the target region, it predicts the flow of objects between the various areas in the target region. The universal causal graph compensates for the deficiencies caused by the sparsity of observation data and solves the problem of inaccurate object flow prediction caused by the lack of observation data in some areas in related technologies, thereby improving the accuracy of object flow prediction.

[0125] Furthermore, the solution provided by the embodiments of this application only requires a small amount of observation data of the target area to predict the flow of objects between different areas in the target area, thereby reducing the amount and difficulty of obtaining observation data of the target area and improving the prediction efficiency of object flow.

[0126] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0127] refer to Figure 8 This diagram illustrates a block diagram of a data prediction apparatus according to an embodiment of this application. The apparatus has the functionality to implement the method example described above; this functionality can be implemented in hardware or by hardware executing corresponding software. The apparatus can be the computer device described above, or it can be located within a computer device. Figure 8 As shown, the device 800 includes: an observation data acquisition module 801, a causal graph acquisition module 802, a full attribute acquisition module 803, and a predicted value acquisition module 804.

[0128] The observation data acquisition module 801 is used to acquire observation data of the source region and observation data of the target region; wherein, the observation data of the source region is greater than the observation data of the target region, and the observation data includes the regional attributes corresponding to each area in the region.

[0129] The causal graph acquisition module 802 is used to acquire a universal causal graph based on the observation data of the source region. The universal causal graph is used to characterize the causal knowledge that is common in different regions, and the causal knowledge is used to characterize the causal relationship between different regional attributes.

[0130] The full attribute acquisition module 803 is used to acquire the full regional attributes corresponding to each region in the target region based on the universal causal graph and the observation data of the target region. The full regional attributes include unobserved regional attributes.

[0131] The prediction value acquisition module 804 is used to acquire the predicted value of object flow between regions in the target region based on the full region attributes corresponding to each region in the target region.

[0132] In one exemplary embodiment, such as Figure 9 As shown, the full attribute acquisition module 803 includes: a first model acquisition submodule 803a and a full attribute acquisition submodule 803b.

[0133] The first model acquisition submodule 803a is used to train a data recovery model based on the universal causal graph and the observation data of the source region. The data recovery model is used to acquire the unobserved regional attributes corresponding to the region.

[0134] The full attribute acquisition submodule 803b is used to acquire the full regional attributes corresponding to the first region in the target region based on the observation data of the first region and through the data recovery model; wherein, the full regional attributes corresponding to the first region include the observed regional attributes and the unobserved regional attributes corresponding to the first region.

[0135] In an exemplary embodiment, the first model acquisition submodule 803a is used for:

[0136] Based on the observation data of the source region, training samples are constructed; wherein, the training samples are any regional attribute among the regional attributes corresponding to each region in the source region;

[0137] The data recovery model is used to obtain the predicted missing region attributes corresponding to the training samples;

[0138] Based on the universal causal graph, the attributes of the inferred missing regions corresponding to the training samples are obtained;

[0139] Based on the predicted missing region attributes, the inferred missing region attributes, and the training samples, the training loss of the data recovery model is obtained;

[0140] Based on the training loss of the data recovery model, the parameters of the data recovery model are adjusted to obtain the trained data recovery model.

[0141] In an exemplary embodiment, the first model acquisition submodule 803a is further configured to:

[0142] Based on the difference between the predicted missing region attributes and the inferred missing region attributes, the missing loss is obtained;

[0143] Based on the predicted missing region attributes and the training samples, the spatial distribution of latent variables is obtained;

[0144] The first latent variable is obtained by sampling from the spatial distribution of the latent variables;

[0145] Based on the first latent variable, the output full-area region attributes corresponding to the training sample are obtained;

[0146] Based on the difference between the regional attributes of the region corresponding to the training sample and the attributes of the entire output region, the reconstruction loss is obtained;

[0147] Based on the spatial distribution of the latent variables, the constraint loss is obtained;

[0148] The training loss of the data recovery model is obtained based on the missing loss, the reconstruction loss, and the constraint loss.

[0149] In one exemplary embodiment, the cause-effect graph acquisition module 802 is configured to:

[0150] The regional attributes corresponding to each region in the source region are transformed to obtain the state vectors corresponding to each regional attribute.

[0151] For the first regional attribute and the second regional attribute in each regional attribute, the state vector corresponding to the first regional attribute and the state vector corresponding to the second regional attribute are calculated to obtain a first predicted value between the first regional attribute and the second regional attribute. The first predicted value is used to characterize the possibility that there is a directed acyclic causal edge between the first regional attribute and the second regional attribute.

[0152] Based on the first predicted values ​​among the attributes of each region, candidate causal graphs are obtained.

[0153] Obtain the universal score corresponding to the candidate causal graph; wherein the universal score is used to measure the degree of matching between the candidate causal graph and the observation data;

[0154] Based on the universal score, the universal causal graph is selected from the candidate causal graphs.

[0155] In one exemplary embodiment, such as Figure 9 As shown, the prediction value acquisition module 804 includes: a second model acquisition submodule 804a, a third model acquisition submodule 804b, and a prediction value acquisition submodule 804c.

[0156] The second model acquisition submodule 804a is used to train a first prediction model based on the observation data of the source region. The first prediction model is used to predict the flow of objects between different regions in the source region.

[0157] The third model acquisition submodule 804b is used to train a second prediction model based on the observation data of the target area and the first prediction model. The second prediction model is used to predict the flow of objects between different regions in the target area.

[0158] The prediction value acquisition submodule 804c is used to obtain the predicted value of object flow between regions in the target region based on the full region attributes corresponding to each region in the target region through the second prediction model.

[0159] In one exemplary embodiment, the second model acquisition submodule 804a is used for:

[0160] Using the first prediction model, the starting point vector representation and the ending point vector representation corresponding to each region in the source region are obtained respectively; wherein, the starting point vector representation refers to the vector representation of the region as the starting point of the flow, and the ending point vector representation refers to the vector representation of the region as the ending point of the flow.

[0161] For the second and third regions in the source region, based on the starting point vector representation corresponding to the second region and the ending point vector representation corresponding to the third region, as well as the distance between the second and third regions, the predicted value of object flow from the second region to the third region is obtained;

[0162] Based on the starting point vector representation corresponding to the third region and the ending point vector representation corresponding to the second region, as well as the distance between the second region and the third region, the predicted value of object flow from the third region to the second region is obtained;

[0163] Based on the predicted object flow values ​​from the second region to the third region and the predicted object flow values ​​from the third region to the second region, the predicted object flow values ​​between the second region and the third region are obtained.

[0164] Based on the difference between the predicted object flow values ​​between the second region and the third region and the actual object flow values ​​in the observed data between the second region and the third region, the sub-prediction loss between the second region and the third region is determined;

[0165] The prediction loss of the first prediction model is obtained by summing the sub-prediction losses among the various regions.

[0166] Based on the prediction loss of the first prediction model, the first prediction model is trained to obtain the trained first prediction model.

[0167] In an exemplary embodiment, the first prediction model includes a starting point feature extraction network and an ending point feature extraction network; the second model acquisition submodule 804a is further configured to:

[0168] Based on the observation data of the source region, a first structural graph is constructed; wherein, the first structural graph uses regions in the source region as nodes, region attributes as the representation of nodes, and the adjacency relationships between regions as relation edges;

[0169] Using the starting point feature extraction network, based on the first structure graph, the starting point vector representations corresponding to each region in the source region are obtained;

[0170] The endpoint feature extraction network obtains the endpoint vector representations corresponding to each region in the source region based on the first structure graph.

[0171] In one exemplary embodiment, the third model acquisition submodule 804b is used for:

[0172] For the fourth and fifth regions in the target area, the first object flow prediction value between the fourth and fifth regions is obtained through the first prediction model;

[0173] The second prediction model is used to obtain the predicted value of the second object flow between the fourth region and the fifth region;

[0174] Based on the difference between the first object flow prediction value and the second object flow prediction value, a first sub-training loss is obtained between the fourth region and the fifth region;

[0175] Summing the first sub-training losses among the regions in the target region yields the first training loss of the second prediction model.

[0176] Based on the difference between the second object flow prediction value and the actual object flow value between the fourth region and the fifth region, a second sub-training loss between the fourth region and the fifth region is obtained;

[0177] Summing the second sub-training losses among the regions in the target region yields the second training loss of the second prediction model.

[0178] Based on the first training loss and the second training loss, the second prediction model is trained to obtain the trained second prediction model.

[0179] In one exemplary embodiment, the predicted value acquisition submodule 804c is used for:

[0180] Based on the full range of regional attributes corresponding to each region in the target region, a second structural graph is constructed; wherein, the second structural graph uses regions in the target region as nodes, full range of regional attributes as node representations, and adjacency relationships between regions as relation edges;

[0181] Using the second prediction model and based on the second structure diagram, the starting point vector representation and ending point vector representation corresponding to each region in the target region are obtained respectively;

[0182] Based on the starting point vector representation and ending point vector representation corresponding to each region in the target region, and the distance between each region in the target region, the predicted value of object flow between each region in the target region is obtained.

[0183] In summary, the technical solution provided in this application obtains a universal causal graph based on a large amount of observation data from the source region. Then, based on the universal causal graph and a small amount of observation data from the target region, it obtains the full regional attributes corresponding to each area in the target region. Based on the full regional attributes corresponding to each area in the target region, it predicts the flow of objects between the various areas in the target region. The universal causal graph compensates for the deficiencies caused by the sparsity of observation data and solves the problem of inaccurate object flow prediction caused by the lack of observation data in some areas in related technologies, thereby improving the accuracy of object flow prediction.

[0184] Furthermore, the solution provided by the embodiments of this application only requires a small amount of observation data of the target area to predict the flow of objects between different areas in the target area, thereby reducing the amount and difficulty of obtaining observation data of the target area and improving the prediction efficiency of object flow.

[0185] It should be noted that the apparatus provided in the above embodiments is only illustrated by the division of the above functional modules when implementing its functions. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process can be found in the method embodiments, which will not be repeated here.

[0186] Please refer to Figure 10 This diagram illustrates the structural block diagram of a computer device according to an embodiment of this application. This computer device can be used to implement the data prediction method provided in the above embodiments. Specifically, it may include the following:

[0187] The computer device 1000 includes a central processing unit (such as a CPU, GPU, or FPGA) 1001, a system memory 1004 including RAM (Random-Access Memory) 1002 and ROM (Read-Only Memory) 1003, and a system bus 1005 connecting the system memory 1004 and the central processing unit 1001. The computer device 1000 also includes a basic input / output system (I / O system) 1006 to facilitate information transfer between various devices within the server, and a large-capacity storage device 1007 for storing the operating system 1013, application programs 1014, and other program modules 1015.

[0188] The basic input / output system 1006 includes a display 1008 for displaying information and an input device 1009 for user input, such as a mouse or keyboard. Both the display 1008 and the input device 1009 are connected to the central processing unit 1001 via an input / output controller 1010 connected to the system bus 1005. The basic input / output system 1006 may also include the input / output controller 1010 for receiving and processing input from multiple other devices such as a keyboard, mouse, or electronic stylus. Similarly, the input / output controller 1010 also provides output to a display screen, printer, or other types of output devices.

[0189] The mass storage device 1007 is connected to the central processing unit 1001 via a mass storage controller (not shown) connected to the system bus 1005. The mass storage device 1007 and its associated computer-readable media provide non-volatile storage for the computer device 1000. That is, the mass storage device 1007 may include computer-readable media (not shown) such as a hard disk or a CD-ROM (Compact Disc Read-Only Memory) drive.

[0190] Without loss of generality, the computer-readable medium may include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented using any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer storage media include RAM, ROM, EPROM (Erasable Programmable Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other solid-state storage technologies, CD-ROM, DVD (Digital Video Disc) or other optical storage, magnetic tape cassettes, magnetic tape, disk storage, or other magnetic storage devices. Of course, those skilled in the art will recognize that the computer storage medium is not limited to the above-mentioned types. The system memory 1004 and mass storage device 1007 described above can be collectively referred to as memory.

[0191] According to an embodiment of this application, the computer device 1000 can also be connected to a remote computer on a network, such as the Internet. That is, the computer device 1000 can be connected to the network 1012 via the network interface unit 1011 connected to the system bus 1005, or the network interface unit 1011 can be used to connect to other types of networks or remote computer systems (not shown).

[0192] The memory further includes at least one instruction, at least one program, code set, or instruction set, wherein at least one instruction, at least one program, code set, or instruction set is stored in the memory and configured to be executed by one or more processors to implement the above data prediction method.

[0193] In one exemplary embodiment, a computer-readable storage medium is also provided, the storage medium storing at least one instruction, at least one program, code set, or instruction set, the at least one instruction, the at least one program, the code set, or the instruction set being executed by a processor to implement the above-described data prediction method.

[0194] Optionally, the computer-readable storage medium may include: ROM (Read-Only Memory), RAM (Random-Access Memory), SSD (Solid State Drives), or optical disc, etc. The random access memory may include ReRAM (Resistance Random Access Memory) and DRAM (Dynamic Random Access Memory).

[0195] In one exemplary embodiment, a computer program product or computer program is also provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the data prediction method described above.

[0196] It should be noted that the specific embodiments of this application involve user information and other related data. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant national and regional laws, regulations, and standards. For example, the acquisition of user data and the acquisition of object flow involved in this application are all carried out with the user's permission and in compliance with relevant laws, regulations, and standards.

[0197] It should be understood that "multiple" as used herein refers to two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. Furthermore, the step numbers described herein are merely illustrative of one possible execution order. In some other embodiments, the steps may not be executed in numerical order, such as two steps with different numbers being executed simultaneously, or two steps with different numbers being executed in the reverse order of the illustration. This application does not limit this.

[0198] The above description is merely an exemplary embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A data prediction method, characterized in that, The method includes: Acquire observation data of the source region and observation data of the target region; wherein, the observation data of the source region is more than the observation data of the target region, and the observation data includes regional attributes corresponding to each area in the region, and the regional attributes include the population composition distribution of the region; Based on the first predicted value between the various regional attributes corresponding to the source region, a candidate causal graph is obtained. The first predicted value is used to characterize the possibility that there is a directed acyclic causal edge between the regional attributes. Based on the universality score of the candidate causal graphs, a universal causal graph is selected from the candidate causal graphs. The universality score is used to measure the degree of matching between the candidate causal graphs and the observed data. The universal causal graph is used to represent causal knowledge that is prevalent in different regions. The causal knowledge is used to represent the causal relationship between different regional attributes. For the first region in the target region, based on the observation data of the first region, the full regional attributes corresponding to the first region are obtained through the data recovery model. The full regional attributes corresponding to the first region include the observed regional attributes and the unobserved regional attributes corresponding to the first region. Using a second prediction model, based on the full regional attributes corresponding to each region in the target region, the predicted value of object flow between each region in the target region is obtained. The second prediction model is used to predict object flow between each region in the target region, and the object flow is used to indicate population flow from the perspective of group aggregation. The data recovery model is trained based on the universal causal graph and the observation data of the source region. The training process of the data recovery model includes: Based on the observation data of the source region, training samples are constructed; wherein, the training samples are any regional attribute among the regional attributes corresponding to each region in the source region; The data recovery model is used to obtain the predicted missing region attributes corresponding to the training samples; Based on the universal causal graph, the attributes of the inferred missing regions corresponding to the training samples are obtained; Based on the predicted missing region attributes, the inferred missing region attributes, and the training samples, the training loss of the data recovery model is obtained; Based on the training loss of the data recovery model, the parameters of the data recovery model are adjusted to obtain the trained data recovery model.

2. The method according to claim 1, characterized in that, The step of obtaining the training loss of the data recovery model based on the predicted missing region attributes, the inferred missing region attributes, and the training samples includes: Based on the difference between the predicted missing region attributes and the inferred missing region attributes, the missing loss is obtained; Based on the predicted missing region attributes and the training samples, the spatial distribution of latent variables is obtained; The first latent variable is obtained by sampling from the spatial distribution of the latent variables; Based on the first latent variable, the output full-area region attributes corresponding to the training sample are obtained; Based on the difference between the regional attributes of the region corresponding to the training sample and the attributes of the entire output region, the reconstruction loss is obtained; Based on the spatial distribution of the latent variables, the constraint loss is obtained; The training loss of the data recovery model is obtained based on the missing loss, the reconstruction loss, and the constraint loss.

3. The method according to claim 1, characterized in that, The method further includes: The regional attributes corresponding to each region in the source region are transformed to obtain the state vectors corresponding to each regional attribute. For the first regional attribute and the second regional attribute among the various regional attributes, the state vector corresponding to the first regional attribute and the state vector corresponding to the second regional attribute are calculated to obtain the first predicted value between the first regional attribute and the second regional attribute.

4. The method according to claim 1, characterized in that, The method further includes: Based on the observation data of the source region, a first prediction model is trained, which is used to predict the object flow between various regions in the source region. The second prediction model is trained based on the observation data of the target area and the first prediction model.

5. The method according to claim 4, characterized in that, The first prediction model, trained based on the observation data from the source region, includes: Using the first prediction model, the starting point vector representation and the ending point vector representation corresponding to each region in the source region are obtained respectively; wherein, the starting point vector representation refers to the vector representation of the region as the starting point of the flow, and the ending point vector representation refers to the vector representation of the region as the ending point of the flow. For the second and third regions in the source region, based on the starting point vector representation corresponding to the second region and the ending point vector representation corresponding to the third region, as well as the distance between the second and third regions, the predicted value of object flow from the second region to the third region is obtained; Based on the starting point vector representation corresponding to the third region and the ending point vector representation corresponding to the second region, as well as the distance between the second region and the third region, the predicted value of object flow from the third region to the second region is obtained; Based on the predicted object flow values ​​from the second region to the third region and the predicted object flow values ​​from the third region to the second region, the predicted object flow values ​​between the second region and the third region are obtained. Based on the difference between the predicted object flow values ​​between the second region and the third region and the actual object flow values ​​in the observed data between the second region and the third region, the sub-prediction loss between the second region and the third region is determined; The prediction loss of the first prediction model is obtained by summing the sub-prediction losses among the various regions. Based on the prediction loss of the first prediction model, the first prediction model is trained to obtain the trained first prediction model.

6. The method according to claim 5, characterized in that, The first prediction model includes a starting point feature extraction network and an ending point feature extraction network; The step of obtaining the starting point vector representation and ending point vector representation corresponding to each region in the source region through the first prediction model includes: Based on the observation data of the source region, a first structural graph is constructed; wherein, the first structural graph uses regions in the source region as nodes, region attributes as the representation of nodes, and the adjacency relationships between regions as relation edges; Using the starting point feature extraction network, based on the first structure graph, the starting point vector representations corresponding to each region in the source region are obtained; The endpoint feature extraction network obtains the endpoint vector representations corresponding to each region in the source region based on the first structure graph.

7. The method according to claim 4, characterized in that, The second prediction model is trained based on the observation data of the target region and the first prediction model, including: For the fourth and fifth regions in the target area, the first object flow prediction value between the fourth and fifth regions is obtained through the first prediction model; The second prediction model is used to obtain the predicted value of the second object flow between the fourth region and the fifth region; Based on the difference between the first object flow prediction value and the second object flow prediction value, a first sub-training loss is obtained between the fourth region and the fifth region; Summing the first sub-training losses among the regions in the target region yields the first training loss of the second prediction model. Based on the difference between the second object flow prediction value and the actual object flow value between the fourth region and the fifth region, a second sub-training loss between the fourth region and the fifth region is obtained; Summing the second sub-training losses among the regions in the target region yields the second training loss of the second prediction model. Based on the first training loss and the second training loss, the second prediction model is trained to obtain the trained second prediction model.

8. The method according to claim 1, characterized in that, The step of obtaining predicted object flow values ​​between regions in the target region using a second prediction model, based on the full region attributes corresponding to each region in the target region, includes: Based on the full range of regional attributes corresponding to each region in the target region, a second structural graph is constructed; wherein, the second structural graph uses regions in the target region as nodes, full range of regional attributes as node representations, and adjacency relationships between regions as relation edges; Using the second prediction model and based on the second structure diagram, the starting point vector representation and ending point vector representation corresponding to each region in the target region are obtained respectively; Based on the starting point vector representation and ending point vector representation corresponding to each region in the target region, and the distance between each region in the target region, the predicted value of object flow between each region in the target region is obtained.

9. A data prediction device, characterized in that, The device includes: The observation data acquisition module is used to acquire observation data of the source region and observation data of the target region; wherein, the observation data of the source region is more than the observation data of the target region, and the observation data includes regional attributes corresponding to each area in the region, and the regional attributes include the population composition distribution of the region; The causal graph acquisition module is used to acquire candidate causal graphs based on first predicted values ​​between various regional attributes corresponding to the source region, wherein the first predicted value is used to characterize the possibility of directed acyclic causal edges between the regional attributes; and to select a universal causal graph from the candidate causal graphs according to the universality score of the candidate causal graphs, wherein the universality score is used to measure the degree of matching between the candidate causal graphs and the observed data, wherein the universal causal graph is used to characterize causal knowledge that is prevalent in different regions, and the causal knowledge is used to characterize the causal association between different regional attributes. The full attribute acquisition module is used to acquire the full regional attributes corresponding to the first region in the target region based on the observation data of the first region and through a data recovery model. The full regional attributes corresponding to the first region include the observed regional attributes and the unobserved regional attributes corresponding to the first region. The prediction value acquisition module is used to obtain the predicted value of object flow between regions in the target region based on the full region attributes corresponding to each region in the target region through the second prediction model. The second prediction model is used to predict the object flow between regions in the target region, and the object flow is used to indicate population flow from the perspective of group aggregation. The data recovery model is trained based on the universal causal graph and the observation data of the source region. The full attribute acquisition module is further used to construct training samples based on the observation data of the source region. The training samples are any regional attributes from the regional attributes corresponding to each region in the source region. The data recovery model is used to obtain the predicted missing region attributes corresponding to the training samples. Based on the universal causal graph, the inferred missing region attributes corresponding to the training samples are obtained. Based on the predicted missing region attributes, the inferred missing region attributes, and the training samples, the training loss of the data recovery model is obtained. Based on the training loss of the data recovery model, the parameters of the data recovery model are adjusted to obtain the trained data recovery model.

10. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one instruction, which is loaded and executed by the processor to implement the data prediction method as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement the data prediction method as described in any one of claims 1 to 8.

12. A computer program product, characterized in that, The computer program product includes computer instructions that are executed by a processor to implement the data prediction method as described in any one of claims 1 to 8.

Images