CN115051925B - Time-space sequence prediction method based on transfer learning - Google Patents
Time-space sequence prediction method based on transfer learning Download PDFInfo
- Publication number
- CN115051925B CN115051925B CN202210683711.8A CN202210683711A CN115051925B CN 115051925 B CN115051925 B CN 115051925B CN 202210683711 A CN202210683711 A CN 202210683711A CN 115051925 B CN115051925 B CN 115051925B
- Authority
- CN
- China
- Prior art keywords
- space
- time
- learning
- network
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/14—Network analysis or design
- H04L41/147—Network analysis or design for predicting network behaviour
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
- H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
- H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W24/00—Supervisory, monitoring or testing arrangements
- H04W24/06—Testing, supervising or monitoring using simulated traffic
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/12—Messaging; Mailboxes; Announcements
- H04W4/14—Short messaging services, e.g. short message services [SMS] or unstructured supplementary service data [USSD]
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/20—Services signaling; Auxiliary data signalling, i.e. transmitting data via a non-traffic channel
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/20—Services signaling; Auxiliary data signalling, i.e. transmitting data via a non-traffic channel
- H04W4/203—Services signaling; Auxiliary data signalling, i.e. transmitting data via a non-traffic channel for converged personal network application service interworking, e.g. OMA converged personal network services [CPNS]
Abstract
The invention discloses a time-space sequence prediction method based on transfer learning, belonging to the field of time-space sequence prediction and the field of transfer prediction; the method comprises the following steps: firstly, collecting wireless traffic of cities to be predicted as space-time sequence source data, and learning space-time characteristics based on a local convolution network and a residual error network of a space-time prediction model; extracting external information source data at the same time, and learning external information features based on local convolution; then, inputting the space-time characteristics into a countermeasure generation network, and learning to obtain space-time commonality characteristics; obtaining a predicted value of a space-time prediction model by fusing space-time characteristics, space-time commonality characteristics and external information characteristics; and optimizing parameters of the predictive model by minimizing a predictive loss function. And finally, migrating the space-time prediction model to target domain data with the same space-time common characteristics as the source data to obtain target domain model parameters, and further predicting a target domain. The invention improves the prediction accuracy and stability.
Description
Technical Field
The invention relates to the field of space-time sequence prediction and migration prediction, in particular to a space-time sequence prediction method based on migration learning.
Background
In real-world applications, spatio-temporal sequential data has a situation where data is scarce in some tasks due to the varying degree of robustness of the data collection mechanism. Data scarcity can increase the training difficulty of the model, resulting in reduced prediction accuracy. Transfer learning is a main method for solving the prediction performance of the time-space sequence under the condition of data scarcity. The transfer learning firstly trains the model to learn knowledge in the source domain with abundant data, and then transfers the knowledge to the target domain with scarce data to further scarce, thereby improving the prediction accuracy of the model. As in document 1: wang L, geng X, ma X.Cross-City Transfer Learning for Deep Spatio-technical prediction.In: proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligent. (2019).
Therefore, under the condition that the target domain samples are rare, a more accurate prediction result can be obtained by utilizing the migration learning method. However, if the space-time distribution of the source domain data and the space-time distribution of the target domain data are too different, the migration of knowledge from the source domain to the target domain may cause poor prediction performance, even form negative migration, and conversely reduce the prediction accuracy of the model. Therefore, researches suggest that compared with the method that knowledge is migrated from a single source domain, the accuracy and stability of model space-time migration prediction can be effectively improved by performing migration learning through a plurality of source domains; as in document 2: the Huaxiu Yao, YIding Liu, YIng Wei, et al 2019.Learning from Multiple Cities: A Meta-Learning Approach for Spatial-technical prediction. In: the World Wide Web Conference (WWW' 19), (2019) multisource migration algorithm improves the predictive performance of neural networks in data-scarce datasets.
The model based on the single source improves the migration stability and accuracy of space-time prediction by calculating the similarity of space-time distribution between different areas of the source area and the target area and by learning the correlation between the areas. As in document 3: cross-City Transfer Learning for Deep Spatio-Temporal Prediction, firstly calculating the space-time similarity between each region of a source domain and a target domain, then matching the regions with the highest space-time similarity, and then designing a region matching function based on which knowledge is migrated from a source domain city to a target domain city, and obtaining a prediction result after a small number of iterations of the target domain.
The model based on multiple sources directly learns space-time dependency relationships from multiple source domain data, and migration stability and accuracy of space-time prediction are improved through multiple data sources. As in document 4: model-diagnostic meta-learning for fast adaptation of deep networks, by combining a plurality of source domain data training initial parameters with different time-space distributions, migrates the initial parameters to a target domain, quickly learns the time-space characteristics of a new task through a small amount of gradient update and a small amount of data, and can be used for improving the prediction performance of time-space sequence migration learning.
For a spatiotemporal sequence migration prediction model, as in document 5: the On first-order meta-learning algorithms provides a meta-learning method Reptile based On first-order update, which is faster to train, and a meta-learning method MAML based On second-order update is further researched to carry out migration prediction through multi-source data; document 6: learning from Multiple Cities A Meta-Learning Approach for Spatial-Temporal Prediction provides a multi-source traffic flow migration prediction model MetaST based on a Meta learning method MAML, and a global-based space-time information mode storage mechanism is designed for extracting and migrating long-term space-time characteristics.
However, the existing multi-source space-time sequence migration prediction method needs rich source domain sample data to train model parameters, so that the learned model parameters migrate to a target domain with scarce data to have a good prediction effect. When the number of samples in the source domain is also relatively small, the accuracy of the spatio-temporal migration prediction is low.
Moreover, the existing multi-source space-time sequence migration prediction method is mainly aimed at migration prediction of multi-source data of the same type under different space-time conditions, and the migration prediction of the multi-source data of different types under the space-time distribution is almost not available. The data of different types of data under the simultaneous air condition has the space-time common characteristic, and the prior method lacks the study of the space-time common characteristic.
For the problem of space-time sequence migration prediction of simultaneous air-conditioning multisource different types of data, which exhibit partial temporal commonality characteristics, no study has been made on such characteristics. In some real applications, source domain data of different service types under the same space-time condition as that of target domain data can be obtained, and the space-time distribution of the source domain and the target domain is different due to the difference of the service types. However, since the different types of data are all under the same space-time distribution condition, the data have strong correlation in partial space regions, and space-time commonality characteristics exist.
Disclosure of Invention
Aiming at the situation that data of a source domain and data of a target domain are relatively scarce, the invention provides a space-time sequence prediction method based on transfer learning.
The space-time sequence prediction method based on transfer learning comprises the following specific steps:
firstly, dividing grids of cities to be predicted, collecting wireless service flow of each grid area according to fixed time intervals, taking the wireless service flow as space-time sequence source data, and simultaneously extracting external information source data for preprocessing and taking the external information source data as flow data;
the wireless traffic flow includes: user short message data (SMS), call service data (Call), and mobile Internet data (Internet).
The external information source data includes: point of interest data (POI), base station information data (BS), social data (Social) for each region.
The point of interest data includes the number of bus subway stations, shops, churches, parks, banks, bars, etc. for each area;
the base station information data includes the number of base stations per area;
social data is derived from the twitter information.
Step two, local convolution network and residual error network based on space-time prediction model for space-time sequence source data, and learning space-time characteristics
The method comprises the following steps:
local convolution network-based short-distance spatial feature X learning Near ：
Wherein the method comprises the steps ofInput representing a layer I short-range spatial module, < >>Representing the output of the first layer close space module, < ->Parameter matrix representing a local convolutional network, +.>A bias vector representing a local convolutional network;
remote space feature X based on residual network learning Dis ：
Input representing layer I remote space module, < >>Representing the output of the layer i telespace module,parameter matrix representing residual network,/->Offset vector, k representing residual network l Representing the expansion factor of the hole convolution network of the first layer.
By stitching short-distance spatial features X Near And a long-distance spatial feature X Dis Obtaining a complete spatial feature representation X S 。
Then, respectively learning sequential characteristics and periodic characteristics of the space-time sequence through a time network module formed by two parallel ConvLSTM models, and splicing the two parts of characteristics to obtain space-time characteristic representation
Step three, local convolution learning external information characteristics of external information source data based on space-time prediction model
The calculation formula for learning the external information features is as follows:
X f is the extracted external information data;and->Representing a three-layer convolution operation;
step four, the space-time characteristicsChallenge generation network input with space-time prediction model, learning to obtain space-time commonality characteristic>
The calculation formula is as follows:
fifthly, fusing the space-time characteristics, the space-time commonality characteristics and the external information characteristics to obtain a predicted value of the space-time prediction model;
the fusion formula is:
W C parameter matrix representing convolutional network, b C A bias vector representing a convolutional network;
and step six, enabling the predicted value of the space-time prediction model to approach to the true value by minimizing the prediction loss function, so as to optimize the parameters of the whole prediction model.
The predictive loss function includes: predicting loss and distinguishing loss;
the predicted loss is expressed as follows:X t representing a true value of supervised learning;
the discrimination loss is expressed as follows:
wherein L is C () Representing a binary cross entropy loss function,indicating the discriminating type of the first type source data input through the network,/->Representing the discrimination type of the second type source data input through the network;
the objective function of the corresponding predictive network is:alpha represents a discrimination loss weight value;
the objective function of the discrimination network is expressed as follows:
and step seven, migrating the source data pre-trained space-time prediction model to target domain data with the same space-time common characteristics as the source data, obtaining target domain model parameters by only a small number of iterative training steps, outputting a prediction result of a target domain test set based on the trained target domain model parameters, and performing prediction performance evaluation.
The invention has the advantages that:
1) According to the space-time sequence prediction method based on transfer learning, based on the same-space-time multi-type data, knowledge is transferred to a target domain for iterative updating through learning space-time characteristics and space-time commonality characteristics in a source domain training model, and prediction accuracy and stability are improved.
2) According to the space-time sequence prediction method based on transfer learning, the external characteristic information based on the environment information is added to conduct transfer prediction, so that the space-time transfer prediction performance when the source domain and the target domain are relatively rare is improved.
Drawings
FIG. 1 is a schematic diagram of a time-space sequence prediction method based on transfer learning;
FIG. 2 is a diagram of a GD-STNet model architecture according to the present invention;
FIG. 3 is a flow chart of a time-space sequence prediction method based on transfer learning;
FIG. 4 is a graph showing the effect of the external features of the present invention on SMS traffic prediction performance;
FIG. 5 is a graph showing the effect of external features of the present invention on Call flow prediction performance
FIG. 6 illustrates the impact of external features of the present invention on Internet traffic prediction performance.
Detailed Description
The following description of specific embodiments of the present invention will be given in further detail with reference to the accompanying drawings and examples. The following examples or figures are illustrative of the invention and are not intended to limit the scope of the invention.
The invention provides a space-time sequence prediction method based on simultaneous space conditions and multi-source different types of data migration learning, as shown in figure 1, firstly preprocessing data to obtain flow data, then learning space-time characteristics based on source domain space-time sequences, learning external information characteristics based on external information data, simultaneously generating network learning space-time commonality characteristics based on antagonism, and finally fusing characteristic prediction and performing time-space migration prediction; according to the invention, the type of the source domain is classified by designing the countermeasure generation network and utilizing the discrimination network training model, the learning of the time-space characteristics and the time-space commonality characteristics is realized by utilizing the basic time-space prediction network and the generation network, so that the discrimination network is difficult to accurately judge the type of the source domain through the time-space commonality characteristics, and finally, the prediction result is obtained by combining the time-space characteristics and the time-space commonality characteristics.
As shown in fig. 2, the whole process is to collect two source domain data, train respective space-time characteristic information, generate network learning space-time commonality characteristics by using antagonism, and combine the two parts of characteristics to predict. And migrating the network to the target domain, and finally evaluating the effect on the test set of the target domain through a small amount of iteration.
The space-time sequence prediction method based on transfer learning is shown in fig. 3, and comprises the following specific steps:
firstly, dividing grids of cities to be predicted, collecting wireless service flow of each grid area according to fixed time intervals, taking the wireless service flow as space-time sequence source data, extracting external information source data, and preprocessing a plurality of space-time data sources through a data preprocessing module to obtain flow values of each area in space in different time intervals.
Dividing the city into a grid map according to longitude and latitude by a data preprocessing module, wherein each grid represents an area, and taking wireless service flow of each area in each time interval as space-time sequence source data which are respectively user short message data (SMS), call service data (Call) and mobile Internet data (Internet).
In addition, the model also adopts external information data to reflect the influence of users on wireless communication in different areas; including point of interest data (POI), base station information data (BS), social data (Social) for each region.
The point of interest data includes the number of bus subway stations, shops, churches, parks, banks, bars, etc. for each area;
the base station information data includes the number of base stations per area;
social data is derived from the twitter information.
Step two, local convolution network and residual error network based on space-time prediction model for space-time sequence source data, and learning space-time characteristics
The method specifically comprises the following steps: and learning multi-distance scale features of the space-time sequence based on the local convolution network and the cavity convolution network.
Learning close-range spatial dependency relationship X using local convolution-based network Near ：
Wherein the method comprises the steps ofInput representing a layer I short-range spatial module, < >>Representing the output of the first layer close space module, < ->Parameter matrix representing a local convolutional network, +.>A bias vector representing a local convolutional network;
the remote space characteristics are learned by using a hole convolution network based on a residual structure, the parameter and the layer number of the space dependence relationship between the distance of a learning volume can be reduced by the hole convolution, the degradation phenomenon of the neural network during training can be reduced by the residual network, and the training difficulty of a model is reduced; the method comprises the following steps:
input representing layer I remote space module, < >>Representing the output of the layer i telespace module,parameter matrix representing residual network,/->Offset vector, k representing residual network l Representing the expansion factor of the hole convolution network of the first layer.
Finally, by stitching the short-distance spatial features X Near And a long-distance spatial feature X Dis Obtaining a complete spatial feature representation X S 。
The time network module consists of two parallel ConvLSTM models which are respectively used for learning the sequential characteristics and the periodic characteristics of the space-time sequence and splicing the two parts of characteristics to obtain the space-time characteristic representation
Step three, learning external information features by an external feature module based on local convolution of the space-time prediction model for external information source data
The external network module consists of three layers of convolutional neural networks and is identical to learning external information characteristics of adjacent areas; the calculation formula is as follows:
X f is the extracted external information data;and->Representing a three-layer convolution operation;
step four, the space-time characteristicsChallenge generation network input with space-time prediction model, learning to obtain space-time commonality characteristic>
Learning the space-time commonality characteristics of different types of source domain data under the same space-time condition through antagonism generation network; the goal of generating the network is to train the network to learn the space-time commonality characteristics from a plurality of different source domains under the simultaneous air, and improve the performance of model space-time migration prediction. When the discrimination network has difficulty in realizing correct classification of the types of the source domains through the space-time commonality characteristics of the generation network, the characteristics learned by the generation network are shown to have space-time commonality.
In particular, in space-time characteristicsFor input, a three-layer convolution network is adopted to obtain a generated characteristic representation:
the discrimination network is used to maximize the type of real source domain that distinguishes the spatio-temporal commonality features. The discriminant network designed by the invention comprises three layers of convolutional neural networks and a full-connection layer, and finally, the output is mapped into a source domain type through a Sigmoid function:
and fifthly, fusing the space-time characteristics, the space-time commonality characteristics and the external characteristics to obtain a predicted value of the space-time sequence.
The fusion formula is:
W C parameter matrix representing convolutional network, b C A bias vector representing a convolutional network;
and step six, enabling the predicted value of the space-time prediction model to approach to the true value by minimizing the prediction loss function, so as to optimize the parameters of the whole prediction model.
In model training, the error of the predicted value and the true value and the structural difference of the space-time distribution characteristics of a plurality of source domains are optimized simultaneously. Therefore, in the design of the loss function, it is necessary to minimize the prediction error and the discrimination error; the objective function of the generated network consists of two parts, namely prediction loss and discrimination loss.
The predicted loss is expressed as follows:X t representing a true value of supervised learning;
the discrimination loss is expressed as follows:
wherein L is C () Representing a binary cross entropy loss function, expressed as: l (L) C (x,y)＝∑ i [-[xlog(y)+(1-x)log(1-y)]]；
Indicating the discriminating type of the first type source data input through the network,/->Representing the discrimination type of the second type source data input through the network;
the objective function of the corresponding predictive network is:alpha represents a discrimination loss weight value;
during source domain training, the generation network enables the predicted value of the prediction model to approach the real value by minimizing the prediction loss function. Meanwhile, the generation network minimizes discrimination loss, so that the discrimination network is difficult to correctly distinguish the source domain attribution type of the generation characteristic, and the prediction model captures the space-time commonality characteristic.
The objective function of the discrimination network is expressed as follows:
the goal of discriminating the network is to correctly determine the type of source domain by generating features.
And step seven, migrating the source data pre-trained space-time prediction model to target domain data with the same space-time common characteristics as the source data, obtaining target domain model parameters by only a small number of iterative training steps, outputting a prediction result of a target domain test set based on the trained target domain model parameters, and performing prediction performance evaluation.
On the source domain, after multiple training iterations, parameters of a prediction model GD-STNet are obtained and used as initial parameters theta of the target domain 0 . To improve the predictive performance of the predictive model in the target domain where the data is scarce, further training is required to migrate the initial parameters to the target domain dataset.
During the training of the target domain, a batch of data is sampled and selected from the target domain training setBased on initial parameter theta 0 Batch data->Loss functionL P And the optimization algorithm Adam performs a small number of iterative updates. Finally, the parameter after iteration update is adopted to evaluate the prediction performance of the prediction model on the test set of the target domain.
The migration aims at multi-source migration learning under the simultaneous air condition, so that the source domain data and the target domain data have space-time commonality characteristics, and a small amount of iterative training is performed based on the training set of the target domain after the migration.
The invention uses three types of wireless service flow prediction to test the effect of entity representation, and the three types of data are respectively:
(1) SMS traffic data set: and each user in the area receives or sends a short message, and the SMS flow in the area is recorded to be increased once.
(2) Call traffic data set: and each user in the area accepts or sends a Call request once, and the Call flow of the area is recorded to be increased once.
(3) Internet traffic data set: each user in the area continues to connect to the network for 15 minutes or uses more than 5MB of traffic, then record that the area Internet traffic has increased once.
The GD-STNet of the invention is superior to the existing space-time migration prediction method in space-time data set (wireless communication traffic data constructed according to user behaviors of each region), as shown in fig. 4, 5 and 6, and embodies the analysis of the influence of external information characteristics on model performance: the method for carrying out migration prediction by adding external information provided by the invention truly improves the prediction performance of space-time migration prediction.
Claims (4)
1. A time-space sequence prediction method based on transfer learning is characterized by comprising the following specific steps:
firstly, meshing cities to be predicted, collecting wireless traffic flow of each mesh area according to fixed time intervals as space-time sequence source data, and learning space-time characteristics based on a local convolution network and a residual error network of a space-time prediction model
Simultaneously extracting external information source data for preprocessing, and learning external information characteristics based on local convolution of space-time prediction modelThe calculation formula for learning the external information features is as follows:
X f is the extracted external information data;and->Representing a three-layer convolution operation;
then, the space-time characteristicsChallenge generation network input with space-time prediction model, learning to obtain space-time commonality characteristic>
The learning space-time commonality featureThe calculation formula of (2) is as follows:
then, the space-time characteristics, the space-time commonality characteristics and the external information characteristics are fused to obtain a predicted value of a space-time prediction model;
the fusion formula is:
W C parameter matrix representing convolutional network, b C A bias vector representing a convolutional network;
further, the prediction value of the space-time prediction model is close to the true value by minimizing the prediction loss function, so that the parameters of the whole prediction model are optimized;
and finally, migrating the source data pre-trained space-time prediction model to target domain data with the same space-time common characteristics as the source data, obtaining target domain model parameters through limited times of training, predicting a target domain test set, and performing performance evaluation on a prediction result.
2. The method for predicting time-space sequence based on transfer learning according to claim 1, wherein the wireless traffic comprises: user short message data SMS, call service data Call and mobile Internet data Internet;
the external information source data includes: the interest point data POI, the base station information data BS and the Social data Social of each area;
the point of interest data includes the number of bus subway stations, shops, churches, parks, banks or bars for each area;
the base station information data includes the number of base stations per area;
social data is derived from the twitter information.
3. The method for predicting spatiotemporal sequences based on transfer learning of claim 1, wherein said learning spatiotemporal featuresThe specific process of (2) is as follows:
local convolution network-based short-distance spatial feature X learning Near ：
Wherein the method comprises the steps ofInput representing a layer I short-range spatial module, < >>Representing the output of the layer i close-space module,parameter matrix representing a local convolutional network, +.>A bias vector representing a local convolutional network;
remote space feature X based on residual network learning Dis ：
Input representing layer I remote space module, < >>Representing the output of the layer I remote space module, < >>Parameter matrix representing residual network,/->Offset vector, k representing residual network l Representing an expansion factor of the hole convolution network of the first layer;
by stitching short-distance spatial features X Near And a long-distance spatial feature X Dis Obtaining a complete spatial feature representation X S ；
Then, respectively learning sequential characteristics and periodic characteristics of the space-time sequence through a time network module formed by two parallel ConvLSTM models, and splicing the two parts of characteristics to obtain space-time characteristic representation
4. The method for predicting a spatiotemporal sequence based on transfer learning according to claim 1, wherein said prediction loss function comprises: predicting loss and distinguishing loss;
the predicted loss is expressed as follows:X t representing a true value of supervised learning;
the discrimination loss is expressed as follows:
wherein L is C () Representing a binary cross entropy loss function,indicating the discriminating type of the first type source data input through the network,/->Representing the discrimination type of the second type source data input through the network;
target function of corresponding prediction networkThe number is as follows:alpha represents a discrimination loss weight value;
the objective function of the discrimination network is expressed as follows:
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110678113 | 2021-06-18 | ||
| CN2021106781137 | 2021-06-18 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115051925A CN115051925A (en) | 2022-09-13 |
| CN115051925B true CN115051925B (en) | 2023-07-18 |
Family
ID=83162201
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210683711.8A Active CN115051925B (en) | 2021-06-18 | 2022-06-16 | Time-space sequence prediction method based on transfer learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115051925B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116597167B (en) * | 2023-06-06 | 2024-02-27 | 中国人民解放军92942部队 | Permanent magnet synchronous motor small sample demagnetization fault diagnosis method, storage medium and system |
| CN117575111B (en) * | 2024-01-16 | 2024-04-12 | 安徽农业大学 | Agricultural remote sensing image space-time sequence prediction method based on transfer learning |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2001236337A (en) * | 2000-02-22 | 2001-08-31 | Fuji Electric Co Ltd | Predicting device using neural network |
| WO2002080458A1 (en) * | 2001-03-30 | 2002-10-10 | Nokia Corporation | Method for configuring a network by defining clusters |
| WO2020010717A1 (en) * | 2018-07-13 | 2020-01-16 | 南京理工大学 | Short-term traffic flow prediction method based on spatio-temporal correlation |
| WO2020064134A1 (en) * | 2018-09-28 | 2020-04-02 | Nokia Technologies Oy | Radio-network self-optimization based on data from radio network and spatiotemporal sensors |
| CN111585783A (en) * | 2020-03-18 | 2020-08-25 | 宁波送变电建设有限公司永耀科技分公司 | Flow dynamic scheduling method based on space-time flow matrix prediction |
| CN111860949A (en) * | 2020-06-12 | 2020-10-30 | 深圳市新产业眼科新技术有限公司 | Prediction method and device based on time sequence image |
| CN112291807A (en) * | 2020-10-15 | 2021-01-29 | 山东科技大学 | Wireless cellular network traffic prediction method based on deep migration learning and cross-domain data fusion |
| CN112532439A (en) * | 2020-11-24 | 2021-03-19 | 山东科技大学 | Network flow prediction method based on attention multi-component space-time cross-domain neural network model |
| WO2021082809A1 (en) * | 2019-10-29 | 2021-05-06 | 山东科技大学 | Training optimization method for foreign exchange time series prediction |
| CN112863182A (en) * | 2021-01-13 | 2021-05-28 | 南京航空航天大学 | Cross-modal data prediction method based on transfer learning |
| CN112905807A (en) * | 2021-03-25 | 2021-06-04 | 北京邮电大学 | Multi-source space-time knowledge fusion method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102110365B (en) * | 2009-12-28 | 2013-11-06 | 日电（中国）有限公司 | Road condition prediction method and road condition prediction system based on space-time relationship |
| US10361926B2 (en) * | 2017-03-03 | 2019-07-23 | Nec Corporation | Link prediction with spatial and temporal consistency in dynamic networks |
-
2022
- 2022-06-16 CN CN202210683711.8A patent/CN115051925B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2001236337A (en) * | 2000-02-22 | 2001-08-31 | Fuji Electric Co Ltd | Predicting device using neural network |
| WO2002080458A1 (en) * | 2001-03-30 | 2002-10-10 | Nokia Corporation | Method for configuring a network by defining clusters |
| WO2020010717A1 (en) * | 2018-07-13 | 2020-01-16 | 南京理工大学 | Short-term traffic flow prediction method based on spatio-temporal correlation |
| WO2020064134A1 (en) * | 2018-09-28 | 2020-04-02 | Nokia Technologies Oy | Radio-network self-optimization based on data from radio network and spatiotemporal sensors |
| WO2021082809A1 (en) * | 2019-10-29 | 2021-05-06 | 山东科技大学 | Training optimization method for foreign exchange time series prediction |
| CN111585783A (en) * | 2020-03-18 | 2020-08-25 | 宁波送变电建设有限公司永耀科技分公司 | Flow dynamic scheduling method based on space-time flow matrix prediction |
| CN111860949A (en) * | 2020-06-12 | 2020-10-30 | 深圳市新产业眼科新技术有限公司 | Prediction method and device based on time sequence image |
| CN112291807A (en) * | 2020-10-15 | 2021-01-29 | 山东科技大学 | Wireless cellular network traffic prediction method based on deep migration learning and cross-domain data fusion |
| CN112532439A (en) * | 2020-11-24 | 2021-03-19 | 山东科技大学 | Network flow prediction method based on attention multi-component space-time cross-domain neural network model |
| CN112863182A (en) * | 2021-01-13 | 2021-05-28 | 南京航空航天大学 | Cross-modal data prediction method based on transfer learning |
| CN112905807A (en) * | 2021-03-25 | 2021-06-04 | 北京邮电大学 | Multi-source space-time knowledge fusion method |
Non-Patent Citations (6)
| Title |
|---|
| Deep spatial temporal networks for crowd flows prediction by dilated convolutions and region-shifting attention mechanism;chujie Tian等;《Applied Intelligence》;第50卷(第10期);1-14页 * |
| Learning from Multiple Cities:A Meta-Learning Approach for Spatial-Temporal Prediction;Huaxiu Yao等;《Learning from Multiple Cities:A Meta-Learning Approach for Spatial-Temporal Prediction》;2181-2191 * |
| Retail Time Series Prediction Based on EMD and Deep Learning;Chujie Tian等;《2018 International Conference on Network Infrastructure and Digital Content (IC-NIDC)》;全文 * |
| 城市环境下的用户移动行为建模概述;徐丰力;李勇;;通信学报;41(第07期);全文 * |
| 基于时空融合的缺失值填补算法;苏思凡;竹翠;朱文军;赵枫朝;;计算机工程与设计(第01期);全文 * |
| 时空序列预测方法综述;黎维;陶蔚;周星宇;潘志松;;计算机应用研究(第10期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115051925A (en) | 2022-09-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115051925B (en) | Time-space sequence prediction method based on transfer learning | |
| CN109142171B (en) | Urban PM10 concentration prediction method based on feature expansion and fusing with neural network | |
| CN109636049B (en) | Congestion index prediction method combining road network topological structure and semantic association | |
| CN104462190B (en) | A kind of online position predicting method excavated based on magnanimity space tracking | |
| CN112561191B (en) | Prediction model training method, prediction device, prediction apparatus, prediction program, and program | |
| CN111368881A (en) | Low-frequency GPS track road network matching method based on multidimensional data fusion analysis | |
| CN112911626B (en) | Wireless network flow prediction method based on multi-graph convolution | |
| CN111343650B (en) | Urban scale wireless service flow prediction method based on cross-domain data and loss resistance | |
| CN111784041A (en) | Wind power prediction method and system based on graph convolution neural network | |
| CN110826244B (en) | Conjugated gradient cellular automaton method for simulating influence of rail transit on urban growth | |
| CN113139140B (en) | Tourist attraction recommendation method based on space-time perception GRU and combined with user relationship preference | |
| Cui et al. | Improved genetic algorithm to optimize the Wi-Fi indoor positioning based on artificial neural network | |
| CN116108984A (en) | Urban flow prediction method based on flow-POI causal relationship reasoning | |
| Xu et al. | Short‐term traffic flow prediction based on whale optimization algorithm optimized BiLSTM_Attention | |
| CN115099328A (en) | Traffic flow prediction method, system, device and storage medium based on countermeasure network | |
| Panovski et al. | Real-time public transportation prediction with machine learning algorithms | |
| CN116052427A (en) | Inter-city inter-regional mobility prediction method and device based on private car travel track data | |
| CN115589369A (en) | Base station flow prediction method and system based on double-attention machine system | |
| Ma et al. | Research and application of second-hand housing price prediction model based on LSTM | |
| ABBAS | A survey of research into artificial neural networks for crime prediction | |
| CN113051474B (en) | Passenger flow prediction method and system integrating multi-platform multi-terminal search indexes | |
| CN115048530A (en) | Graph convolution recommendation system fusing neighbor importance and feature learning | |
| CN115034478A (en) | Traffic flow prediction method based on domain self-adaptation and knowledge migration | |
| CN113240219A (en) | Land utilization simulation and prediction method | |
| Luan et al. | An efficient target tracking approach through mobile crowdsensing |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |