CN112633607B - Dynamic space-time event prediction method and system - Google Patents

Dynamic space-time event prediction method and system Download PDF

Info

Publication number
CN112633607B
CN112633607B CN202110009698.3A CN202110009698A CN112633607B CN 112633607 B CN112633607 B CN 112633607B CN 202110009698 A CN202110009698 A CN 202110009698A CN 112633607 B CN112633607 B CN 112633607B
Authority
CN
China
Prior art keywords
event
behavior
location
node
features
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
Application number
CN202110009698.3A
Other languages
Chinese (zh)
Other versions
CN112633607A (en
Inventor
刘晓明
张占伟
吴少聪
张兆晗
沈超
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Xian Jiaotong University
Original Assignee
Xian Jiaotong University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Xian Jiaotong University filed Critical Xian Jiaotong University
Priority to CN202110009698.3A priority Critical patent/CN112633607B/en
Publication of CN112633607A publication Critical patent/CN112633607A/en
Application granted granted Critical
Publication of CN112633607B publication Critical patent/CN112633607B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • Human Resources & Organizations (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Economics (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Strategic Management (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Game Theory and Decision Science (AREA)
  • General Business, Economics & Management (AREA)
  • Development Economics (AREA)
  • Tourism & Hospitality (AREA)
  • Quality & Reliability (AREA)
  • Operations Research (AREA)
  • Marketing (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The invention discloses a method and a system for predicting dynamic space-time events, which comprise the steps of embedding event behavior information, place information and interaction characteristic information thereof into event potential characteristics, and constructing an event data bipartite graph by utilizing the embedded event potential characteristics; respectively calculating the tightness between the positions of different events and the behaviors of different events to obtain a position edge weight matrix and a behavior edge weight matrix; updating the event data bipartite graph to obtain a place weighted directed graph and a behavior weighted directed graph; random walk sampling is carried out, characteristics of neighbors of the location nodes and characteristics of neighbors of the behavior nodes are respectively aggregated and spliced by using a graph rolling network, and the aggregated neighbor characteristics are input into a multi-layer perceptron prediction model for prediction, so that an event prediction result is obtained; the invention uses the graph rolling network with the attention mechanism to aggregate the behavior and the location neighbor characteristics, and predicts by using the multi-layer perceptron prediction model, so that the prediction result has higher accuracy and smaller error.

Description

Dynamic space-time event prediction method and system
Technical Field
The invention belongs to the technical field of graph data mining, and particularly relates to a dynamic space-time event prediction method and a system.
Background
Predicting future events has important significance for our daily life, especially events which seriously threaten the life and property safety of citizens, and is not slow for strengthening public safety work; how to effectively and reasonably model temporal data for dynamic event reasoning is still a challenge for the problem of predicting future events.
In recent years, many studies have deduced possible events by mining spatiotemporal data; among them, according to the research motivation, the current situation of research is divided into three categories: namely a data driving model, a characteristic regression model and a dynamic process description model; the traditional algorithm generally adopts a regression algorithm based on characteristics or a fitting method with space-time distribution to solve the problem of dynamic event reasoning; the reasoning method models are basically based on strong parameter assumptions and need to be obviously improved; in particular, they often take into account spatial and temporal effects separately, and it is difficult to dynamically describe interactions between events.
With the rapid development of internet technology, the disclosed heterogeneous data provides an opportunity to learn the dynamic process of events, and can reveal rules and interactions between events. How to construct dynamic event patterns through comprehensive research on multiple spatiotemporal data is a research focus on the event inference problem.
Disclosure of Invention
Aiming at the technical problems in the prior art, the invention provides a dynamic space-time event prediction method and a system, which are used for solving the technical problems that the prior future event reasoning prediction process is generally based on strong parameter assumption, the spatial influence and the time influence are separately considered, and the mutual influence between events cannot be dynamically described.
In order to achieve the above purpose, the invention adopts the following technical scheme:
the invention provides a dynamic space-time event prediction method, which comprises the following steps:
step 1, acquiring a multisource space-time event data set, and embedding event behaviors, event places and interaction features between the event behaviors and places into event potential features by utilizing a behavior-place co-evolution model to obtain embedded event potential features; constructing an event data bipartite graph by utilizing the embedded event potential characteristics;
step 2, calculating the tightness degree between different event places or different event behaviors respectively by using the event data bipartite graph to obtain a place edge weight matrix W and a behavior edge weight matrix S; updating the event data bipartite graph by using the place edge weight matrix W and the behavior edge weight matrix S respectively to obtain placesWeighted directed graph G 1 And behavior weighted directed graph G 2 ；
Step 3, weighting the directed graph G by the places 1 Node and behavior weighted directed graph G of (1) 2 Random walk sampling is carried out on the nodes of the node (C), and the characteristics of the neighbors of the location node and the characteristics of the neighbors of the behavior node are respectively aggregated and spliced by using a graph convolution network to obtain aggregated neighbor characteristics; and inputting the aggregated neighbor features into a multi-layer perceptron prediction model to predict, so as to obtain an event prediction result.
Further, in step 1, the method specifically includes the following steps:
step 11, acquiring a multisource space-time event data set, and extracting event behavior information, event location attributes, event occurrence time and interaction characteristics between event behaviors and locations; wherein the event location attribute comprises a location longitude and latitude;
step 12, embedding event behavior information, event location attribute, event occurrence time and interaction feature information between event behavior and location into event potential features by using a behavior-location co-evolution model to obtain an embedded event potential feature list O;
and 13, constructing an event data bipartite graph by using the embedded event potential feature list O.
Further, in step 12, the expression of the embedded event latent feature O is:
Figure BDA0002884552560000021
wherein i is event sequence number, e i For the ith event, N is the total number of events; a, a i Is the ith event behavior, where a i ∈{a 1 ,a 2 ,...,a m M is the number of event behavior types; l (L) i Is the ith event location, wherein, l i ∈{l 1 ,l 2 ,...,l n N is the number of event location categories; b i Is an interactive feature of event behavior and places.
Further, the method comprises the steps of,in step 13, the event data bipartite graph includes behavior embedding features
Figure BDA0002884552560000031
Place-embedding features
Figure BDA00028845525600000315
；
Wherein behavior embedding features
Figure BDA0002884552560000032
The expression of (2) is:
Figure BDA0002884552560000033
wherein a is the event behavior,
Figure BDA0002884552560000034
for the occurrence time of the kth time of the event behavior a, W 1 、W 2 、W 3 W and W 4 Respectively a weight matrix, wherein sigma is a sigmoid function; />
Figure BDA0002884552560000035
Interaction characteristics which are the kth time evolution of the event behavior a and are time-dependent; />
Figure BDA0002884552560000036
-the time before the kth occurrence of event behavior a; />
Figure BDA0002884552560000037
The occurrence time of the k-1 th time of the event behavior a; f (f) a Embedded features of event behavior a; q l The embedded feature of event location l;
place-embedding features
Figure BDA0002884552560000038
The expression of (2) is:
Figure BDA0002884552560000039
wherein, l is the event location,
Figure BDA00028845525600000310
embedding features for event location/location after the kth event, for->
Figure BDA00028845525600000311
For event location/kth event occurrence time, W 5 、W 6 、W 7 W and W 8 Respectively a weight matrix; />
Figure BDA00028845525600000312
For the time when the k-1 st event occurs at event location l,/time>
Figure BDA00028845525600000313
-immediately before the occurrence of the kth event for event location l, < >>
Figure BDA00028845525600000314
The kth time of interaction characteristic is the event location l.
Further, in step 2, specifically:
step 21, calculating the tightness degree of different places according to the product of the distance between different places and the traffic data to obtain a place edge weight matrix W; calculating the tightness degree of different behaviors according to the co-occurrence times among different behaviors to obtain a behavior edge weight matrix S;
step 22, updating the event data bipartite graph by using the place edge weight matrix W to obtain a place weighted directed graph G 1 The method comprises the steps of carrying out a first treatment on the surface of the Updating the event data bipartite graph by using the behavior edge weight matrix S to obtain a behavior weighted directed graph G 2 。
Further, the expression of the place edge weight matrix W is:
Figure BDA0002884552560000041
wherein w is ij For the location node v i With the site node v j Is the influence weight of the distance between event sites, beta is the influence weight of traffic flow, ρ is the distance attenuation parameter between two event sites, c ij Is the event location l i And event location l j Traffic flow of (2);
the expression of the behavior edge weight matrix S is as follows:
Figure BDA0002884552560000042
wherein s is ij For behaving as nodes s i And behavior node s j Is a side weight of (1); delta ij Is event behavior a i And event behavior a j B is the time window length, M is the number of co-occurring event behaviors within the time window, and η is the distance decay parameter.
Further, in step 3, the method specifically includes the following steps:
step 31, weighting the location-weighted directed graph G 1 Adopting a random walk sampling strategy to select the nearest neighbor of the place node; weighted behavior directed graph G 2 Selecting a neighbor closest to the behavior node by adopting a random walk sampling strategy;
step 32, aggregating the features of the neighbors of the location nodes by using a graph-coiler network based on an attention mechanism to obtain the aggregated features of the neighbors of the location nodes; splicing the aggregated site node neighbor features with the site node potential features to obtain site aggregation features;
aggregating the characteristics of the behavior node neighbors by using a graph-coiler network based on an attention mechanism to obtain aggregated behavior node neighbor characteristics; splicing the aggregated behavior node neighbor features with the behavior node potential features to obtain behavior aggregation features;
splicing the location aggregation features and the behavior aggregation features to obtain aggregated neighbor features; inputting the aggregated neighbor features into a multi-layer perceptron prediction model, establishing a corresponding loss function, and training by adopting an optimizer to minimize the loss function;
and step 34, after the training is finished, respectively updating the potential characteristics of the behavior and the place node, and circulating the steps 32-33 to carry out the next training and outputting the result.
Further, in step 31, the slave node v j Randomly walk to adjacent site node v i The probability of (2) is:
Figure BDA0002884552560000051
wherein,,
Figure BDA0002884552560000052
for the slave point node v j Randomly walk to adjacent site node v i Probability of (2); n (v) j ) For the location node v j Is a neighbor to all neighbors of (a); w (w) kj For the location node v k With the site node v j Is a side weight of (1);
slave action node s j Random walk to adjacent behavioural node s i The probability of (2) is:
Figure BDA0002884552560000053
wherein,,
Figure BDA0002884552560000054
for slave action node s j Random walk to adjacent behavioural node s i Probability of (2); n(s) j ) For behaving as nodes s j Is a neighbor to all neighbors of (a); s is(s) kj For behaving as nodes s k And behavior node s j Is a side weight of (b).
Further, aggregating the features of the neighbors of the location nodes by using a graph-coiler network based on an attention mechanism to obtain aggregated features of the neighbors of the location nodes; splicing the aggregated site node neighbor features with the site node potential features to obtain site aggregation features;
wherein the second layer of sites aggregate features Q' u The expression of (2) is:
Q′ u ＝σ([(∑ v∈N(u) α vu q v )||q u ]W 6 )
wherein alpha is vu For the location node v u Randomly walk to adjacent site node v v Probability of (2); n (u) is a location node v u Is a neighbor to all neighbors of (a); q v For the location node v v Features; q u For the location node v u Potential features;
first tier site aggregation feature Q l The expression of (2) is:
Q l ＝σ([(∑ u∈N(l) α ul Q′ u )||q l ]W 5 )
wherein alpha is ul For the location node v l Randomly walk to adjacent site node v u Probability of (2); n (l) is a place node v l Is a neighbor to all neighbors of (a); q l For the location node v l Potential features;
aggregating the characteristics of the behavior node neighbors by using a graph-coiler network based on an attention mechanism to obtain aggregated behavior node neighbor characteristics; splicing the aggregated behavior node neighbor features with the behavior node potential features to obtain behavior aggregation features;
wherein the second layer behavior aggregate feature F' v The expression of (2) is:
F′ v ＝σ([(∑ r∈N(v) α rv f r )||f v ]W 8 )
wherein alpha is rv For behaving as nodes s v Random walk to adjacent behavioural node s r Probability of (2); f (f) r For behaving as nodes s r Features; f (f) v For behaving as nodes s v Potential features; n (v) is a behavior node s v Is a neighbor to all neighbors of (a);
wherein the first layer behavior aggregates featuresF a The expression of (2) is:
F a ＝σ([(∑ v∈N(a) α va F′ v )||f a ]W 7 )
wherein N (a) is a behavior node s a Is a neighbor to all neighbors of (a); alpha va For behaving as nodes s a Random walk to adjacent behavioural node s v Probability of (2); f (f) a For behaving as nodes s a Potential features;
in step 33, the prediction model based on the multi-layer perceptron is:
E(a,l,t)＝φ(Q l (t)||F a (t))
wherein phi is a multi-layer perceptron prediction model; q (Q) l (t) is a location aggregation feature for event location l; f (F) a (t) is a behavior aggregation feature of event behavior a; e (a, l, t) is an event prediction expectation;
the expression of the loss function is:
Figure BDA0002884552560000061
wherein Y is i Represents the i-th ground truth value,
Figure BDA0002884552560000062
and representing the ith prediction result, wherein G is a ground truth value data set.
The invention also provides a dynamic space-time event prediction system which is characterized by comprising a construction module, a matrix module and a prediction module;
the construction module is used for acquiring a multisource space-time event data set, embedding event behavior information, event location information and interaction characteristic information between event behaviors and locations into event potential characteristics by utilizing a behavior-location co-evolution model, and obtaining embedded event potential characteristics; constructing an event data bipartite graph by utilizing the embedded event potential characteristics;
the computing and updating module is used for respectively computing the tightness degree between different event places and different event behaviors by utilizing the event data bipartite graph to obtainThe arrival location edge weight matrix W and the behavior edge weight matrix S; updating the event data bipartite graph by using the place edge weight matrix W and the behavior edge weight matrix S respectively to obtain a place weighted directed graph G 1 And behavior weighted directed graph G 2 ；
Prediction module for weighting the directed graph G by the locations 1 Node and behavior weighted directed graph G of (1) 2 Random walk sampling is carried out on the nodes of the node (C), and the characteristics of the neighbors of the location node and the characteristics of the neighbors of the behavior node are respectively aggregated and spliced by using a graph convolution network to obtain aggregated neighbor characteristics; and inputting the aggregated neighbor features into a multi-layer perceptron prediction model to predict, so as to obtain an event prediction result.
Compared with the prior art, the invention has the beneficial effects that:
the invention provides a dynamic space-time event prediction method and a system, which embed behavior information, place information and interaction thereof into event potential characteristics through a co-evolution model, more fully consider time and space influence in space-time data events and capture interaction relationship between behaviors and places; by extracting multi-source data, calculating the tightness degree between different behaviors and different places by using traffic flow, distance data and statistical data, acquiring an edge weight matrix, and further revealing the rules and the mutual influence between space-time events; by random walk sampling, a graph convolution network with an attention mechanism is used for aggregating behavior and location neighbor features, the aggregated features are input into a multi-layer perceptron for behavior and location type prediction and event occurrence time prediction, the prediction result accuracy is higher, the error is smaller, and the prediction result practical value is higher.
Drawings
FIG. 1 is a flow chart of a method of dynamic spatiotemporal event prediction according to an embodiment;
FIG. 2 is a schematic diagram of a random walk sampling method using a location as a node in an embodiment;
FIG. 3 is a schematic diagram of a process for feature aggregation and multi-layer perceptron prediction based on a graph roll-up neural network with an attention mechanism;
FIG. 4 is a visual result diagram of the predicted results of dynamic spatiotemporal data events.
Detailed Description
In order to make the technical problems, technical schemes and beneficial effects solved by the invention more clear, the following specific embodiments are used for further describing the invention in detail. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.
The invention provides a dynamic space-time event prediction method, which comprises the following steps:
step 1, acquiring a multisource space-time event data set, and extracting event behavior information, event location attributes, event occurrence time and interaction characteristics between event behaviors and locations; embedding event behavior information, event location attributes, event occurrence time and interaction characteristics between event behaviors and locations into event potential characteristics by using a behavior-location co-evolution model to obtain embedded event potential characteristics; the dynamic process between the events is described through the embedded features, the rule of event development is obtained, and additional information of interaction between the events which evolve together along with time is captured, so that an event data bipartite graph is obtained.
Specifically, the method comprises the following steps:
step 11, acquiring a multisource space-time event data set, and extracting event behavior information, event location attributes, event occurrence time and interaction characteristics between event behaviors and locations; the event location attribute comprises location longitude and latitude, and the distance between the locations can be calculated by using the location longitude and latitude; acquiring traffic data among event sites from a multi-source space-time event data set; wherein the traffic data includes bicycle traffic and rental traffic.
Step 12, embedding event behavior information, event location attribute, event occurrence time and interaction characteristics between event behavior and location into event potential characteristics by using a behavior-location co-evolution model to obtain an embedded event potential characteristic list O; the expression of the embedded event potential feature list O is as follows:
Figure BDA0002884552560000081
wherein i is event sequence number, e i For the ith event, N is the total number of events; a, a i Is the ith event behavior, where a i ∈{a 1 ,a 2 ,...,a m M is the number of event behavior types; l (L) i Is the ith event location, wherein, l i ∈{l 1 ,l 2 ,...,l n N is the number of event location categories; b i Is an interactive feature of event behavior and places.
In the invention, the embedded event potential feature list O is used for describing the events in the multisource spatio-temporal event data set, the embedded event potential feature list can store records of a plurality of events, and each record can clearly and intuitively represent the occurrence time, place, behavior and interaction information of a certain event.
Step 13, constructing an event data bipartite graph by using the embedded event potential feature list O; wherein the event data bipartite graph comprises event behavior embedding features
Figure BDA0002884552560000091
Event location embedded feature->
Figure BDA0002884552560000092
Both the event behavior embedded features and the event location embedded features evolve with the interaction track of time;
wherein event behavior embeds features
Figure BDA0002884552560000093
The expression of (2) is:
Figure BDA0002884552560000094
wherein a is the event behavior,
Figure BDA0002884552560000095
for the occurrence time of the kth time of the event behavior a, W 1 、W 2 、W 3 W and W 4 Respectively a weight matrix, wherein sigma is a sigmoid function; />
Figure BDA0002884552560000096
Interaction characteristics which are the kth time evolution of the event behavior a and are time-dependent; />
Figure BDA0002884552560000097
-the time before the kth occurrence of event behavior a; />
Figure BDA0002884552560000098
The occurrence time of the k-1 th time of the event behavior a; f (f) a Embedded features of event behavior a; q l And is an embedded feature of event location l.
Wherein the location embeds the feature
Figure BDA0002884552560000099
The expression of (2) is:
Figure BDA00028845525600000910
wherein, l is the event location,
Figure BDA00028845525600000911
embedding features for event location/location after the kth event, for->
Figure BDA00028845525600000912
For event location/kth event occurrence time, W 5 、W 6 、W 7 W and W 8 Respectively a weight matrix; />
Figure BDA00028845525600000913
For the time when the k-1 st event occurs at event location l,/time>
Figure BDA00028845525600000914
-immediately before the occurrence of the kth event for event location l, < >>
Figure BDA00028845525600000915
The kth time of interaction characteristic is the event location l.
In the invention, event behavior is embedded with characteristics
Figure BDA00028845525600000916
Event location embedded feature->
Figure BDA00028845525600000917
Time drift refers to the steady evolution of the properties and features of event behavior and location over time; self-motivating points refer to the fact that the current behavior or location characteristics are affected by their early behavior or location characteristics; the co-evolution characteristics of the behaviors and the places are used for mining interdependence information between the event behaviors and the places; the interaction characteristics represent additional information of inter-event interactions that co-evolve over time.
Step 2, using the event data bipartite graph to respectively calculate the tightness between different event places and different event behaviors to obtain a place edge weight matrix W and a behavior edge weight matrix S; updating the event data bipartite graph by using the place edge weight matrix W to obtain a place weighted directed graph G 1 The method comprises the steps of carrying out a first treatment on the surface of the Updating the event data bipartite graph by using the behavior edge weight matrix S to obtain a behavior weighted directed graph G 2 。
The method specifically comprises the following steps:
step 21, calculating the tightness degree of different places according to the product of the distance between different places and the traffic data to obtain a place edge weight matrix W; the expression of the place edge weight matrix W is as follows:
Figure BDA0002884552560000101
wherein w is ij For the location node v i Node with locationv j Is the influence weight of the distance between event sites, beta is the influence weight of traffic flow, ρ is the distance attenuation parameter between two event sites, c ij Is the event location l i And event location l j Traffic flow of the vehicle.
Calculating the tightness degree of different behaviors according to the co-occurrence times among different behaviors to obtain a behavior edge weight matrix S; the expression of the behavior edge weight matrix S is as follows:
Figure BDA0002884552560000102
wherein s is ij For behaving as nodes s i And behavior node s j Is a side weight of (1); delta ij Is event behavior a i And event behavior a j B is the time window length, M is the number of co-occurring event behaviors within the time window, and η is the distance decay parameter.
Step 22, updating the event data bipartite graph by using the place edge weight matrix W to obtain a place weighted directed graph G 1 The method comprises the steps of carrying out a first treatment on the surface of the The expression of the place weighted directed graph is as follows:
G 1 ＝(V 1 ,E 1 ,W)
wherein V is 1 Weighting the directed graph vertices for places V e V 1 Wherein one vertex represents an event location; e (E) 1 Weighting an edge of a directed graph for a location, E 1 One of the edges represents a connection between two event locations; w is an edge weight matrix of the compactness of two event places, and W epsilon W.
Updating the event data bipartite graph by using the behavior edge weight matrix S to obtain a behavior weighted directed graph G 2 The method comprises the steps of carrying out a first treatment on the surface of the The expression of the behavior weighted directed graph is as follows:
G 2 ＝(V 2 ,E 2 ,S)
wherein V is 2 Weighting nodes of a directed graph for behavior, V e V 2 Wherein one node represents an event behavior; e (E) 2 Weighting one edge of the directed graph, E, for behavior 2 Wherein one edge represents a connection between two event behaviors; s is an edge weight matrix of the tightness degree of two event behaviors, and S epsilon S.
Step 3, weighting the directed graph G by the places 1 Node and behavior weighted directed graph G of (1) 2 Random walk sampling is carried out on the nodes of the node (C) to obtain a place node neighbor and a behavior node neighbor; respectively aggregating the features of the neighbors of the location nodes and the features of the neighbors of the behavior nodes by using a graph rolling network, and splicing the aggregated features of the neighbors of the location nodes with the potential features of the location nodes to obtain location aggregation features; splicing the aggregated behavior node neighbor features with the behavior node potential features to obtain behavior aggregation features; and respectively inputting the place aggregation features and the behavior aggregation features into a prediction model to predict, so as to obtain an event prediction result.
The method specifically comprises the following steps:
step 31, weighting the location-weighted directed graph G 1 Adopting a random walk sampling strategy to select the nearest neighbor of the place node;
wherein the slave point node v j Randomly walk to adjacent site node v i The probability of (2) is:
Figure BDA0002884552560000111
wherein,,
Figure BDA0002884552560000112
for the slave point node v j Randomly walk to adjacent site node v i Probability of (2); n (v) j ) For the location node v j Is a neighbor to all neighbors of (a); w (w) kj For the location node v k With the site node v j Is a side weight of (1);
weighted behavior directed graph G 2 Selecting a neighbor closest to the behavior node by adopting a random walk sampling strategy;
wherein the slave behavior node s j Random walk to adjacent behavioural node s i The probability of (2) is:
Figure BDA0002884552560000121
wherein,,
Figure BDA0002884552560000122
for slave action node s j Random walk to adjacent behavioural node s i Probability of (2); n(s) j ) For behaving as nodes s j Is a neighbor to all neighbors of (a); s is(s) kj For behaving as nodes s k And behavior node s j Is a side weight of (b).
Step 32, aggregating the features of the neighbors of the location nodes by using a graph-coiler network based on an attention mechanism to obtain the aggregated features of the neighbors of the location nodes; splicing the aggregated site node neighbor features with the site node potential features to obtain site aggregation features;
in the invention, when considering the two-layer neighbor condition of the event location:
wherein, the second layer location aggregate feature Q' u The expression of (2) is:
Q′ u ＝σ([(∑ v∈N(u) α vu q v )||q u ]W 6 )
wherein alpha is vu For the location node v u Randomly walk to adjacent site node v v Probability of (2); n (u) is a location node v u Is a neighbor to all neighbors of (a); q v For the location node v v Features; q u For the location node v u Potential features;
first tier site aggregation feature Q l The expression of (2) is:
Q l ＝σ([(∑ u∈N(l) α ul Q′ u )||q l ]W 5 )
wherein alpha is ul For the location node v l Randomly walk to adjacent site node v u Probability of (2); n (l) is a place node v l Is a neighbor to all neighbors of (a); q l For the location node v l Potential features.
Aggregating the characteristics of the behavior node neighbors by using a graph-coiler network based on an attention mechanism to obtain aggregated behavior node neighbor characteristics; splicing the aggregated behavior node neighbor features with the behavior node potential features to obtain behavior aggregation features;
in the invention, when considering the two-layer neighbor condition of the event behavior:
wherein the second layer behavior aggregate feature F' v The expression of (2) is:
F′ v ＝σ([(∑ r∈N(v) α rv f r )||f v ]W 8 )
wherein alpha is rv For behaving as nodes s v Random walk to adjacent behavioural node s r Probability of (2); f (f) r For behaving as nodes s r Features; f (f) v For behaving as nodes s v Potential features; n (v) is a behavior node s v Is a neighbor to all neighbors of (a);
wherein the first layer behavior aggregate feature F a The expression of (2) is:
F a ＝σ([(∑ v∈N(a) α va F′ v )||f a ]W 7 )
wherein N (a) is a behavior node s a Is a neighbor to all neighbors of (a); alpha va For behaving as nodes s a Random walk to adjacent behavioural node s v Probability of (2); f (f) a For behaving as nodes s a Potential features.
Step 33, establishing a prediction model based on a multi-layer perceptron, and determining a loss function; splicing the location aggregation features and the behavior aggregation features to obtain aggregated neighbor features; inputting the aggregated neighbor features into a prediction model, and optimizing the prediction model by adopting an Amad optimizer to minimize a loss function value;
the prediction model based on the multi-layer perceptron is as follows:
E(a,l,t)＝φ(Q l (t)||F a (t))
wherein phi is a multi-layer perceptionA machine prediction model; q (Q) l (t) is a location aggregation feature for event location l; f (F) a (t) is a behavior aggregation feature of event behavior a; e (a, l, t) is an event prediction expectation;
the expression of the loss function is:
Figure BDA0002884552560000131
wherein Y is i Represents the i-th ground truth value,
Figure BDA0002884552560000132
and representing the ith prediction result, wherein G is a ground truth value data set.
And step 34, after the training is finished, respectively updating the potential characteristics of the behavior and the place node, and circulating the steps 32-33 to carry out the next training and outputting the result.
The invention also provides a dynamic space-time event prediction system, which comprises a construction module, a matrix module and a prediction module;
the construction module is used for acquiring a multisource space-time event data set, embedding event behavior information, event location information and interaction characteristic information between event behaviors and locations into event potential characteristics by utilizing a behavior-location co-evolution model, and obtaining embedded event potential characteristics; constructing an event data bipartite graph by utilizing the embedded event potential characteristics;
the computing and updating module is used for respectively computing the tightness degree between different event places and different event behaviors by utilizing the event data bipartite graph to obtain a place edge weight matrix W and a behavior edge weight matrix S; updating the event data bipartite graph by using the place edge weight matrix W and the behavior edge weight matrix S respectively to obtain a place weighted directed graph G 1 And behavior weighted directed graph G 2 ；
Prediction module for weighting the directed graph G by the locations 1 Node and behavior weighted directed graph G of (1) 2 Random walk sampling is carried out on the nodes of the node (C), and the characteristics of the neighborhood of the location node and the characteristics of the neighborhood of the behavior node are respectively carried out by using a graph convolution networkPolymerizing and splicing to obtain polymerized neighbor features; and inputting the aggregated neighbor features into a multi-layer perceptron prediction model to predict, so as to obtain an event prediction result.
Examples
As shown in FIG. 1, the invention provides a method and a system for predicting dynamic space-time events, which comprises the following steps:
step 1, embedding behavior information, place information and interaction thereof into event potential characteristics by utilizing a co-evolution model, describing a dynamic process between events by the embedded characteristics, and capturing interaction relationship between behaviors and places; the specific process is as follows:
step 1.1, the data set in this embodiment is derived from the new york city crime record: https:// opendata. Cityoftyyork. Us/, 331855 crime event records, 1889564 taxi flow records; calculating the distance between places of the longitude and latitude information of the places in the crime event record, and acquiring traffic flow data between the places:
for example: the longitude and latitude coordinates of the centers of the location 0 and the location 1 are (-73.9845, 40.7552) and (-73.8792, 40.6673), respectively, the center distance of the location 0 and the location 1 is 13.2081km, and the traffic flow is 53.
Step 1.2, adopt list
Figure BDA0002884552560000141
Representing criminal events in a dataset, including at time t i 30 crime types of information a at that time i ∈{a 1 ,a 2 ,…,a 30 Information of 195 places i ∈{l 1 ,l 2 ,…,l 195 ' and interaction information b i Performing one-hot coding on the crime type and the place;
for example: one e 0 The crime event= (24,10,0,0) indicates that a crime of type 10 has occurred at the point 24 at time 0, and the inter-event interaction information is 0.
Step 1.3, expressing crime event data into two graphs, wherein the crime type is taken as one part, and the crime place is taken as the other part; embedded features for crime type and locationInitially [0,0]Evolving with the interaction track of time; weight matrix W 1 、W 2 、W 3 W and W 4 The sizes of (2) are 1×5,1×1 and 1×5 respectively; w (W) 5 、W 6 、W 7 W and W 8 The sizes of (2) are 1×5,1×1 and 1×5, respectively; w (W) 1 、W 2 、W 3 、W 4 、W 5 、W 6 、W 7 W and W 8 The initial values of (2) are all 0.
For example: the update characteristics of crime type 10 occurring at location 24 are as follows:
Figure BDA0002884552560000151
feature f of crime type 10 after the end of a round 10 From [0,0]The updating is as follows:
[-5.095×10 -2 ，2.0656×10 -2 ，-4.2313×10 -2 ，1.3154×10 -1 ，-4.1675×10 -2 ]
after crime type 10 occurs at location 24, the update characteristics are as follows:
Figure BDA0002884552560000152
feature q of crime scene 24 after completion of a round 24 From [0,0]The updating is as follows:
[1.4074×10 -1 ，2.8928×10 -2 ，5.7950×10 -2 ，-5.4407×10 -2 ，1.4077×10 -1 ]
all crime types and places involved will be updated during training.
Step 2, obtaining side weight values by calculating the types of different crimes and the tightness degree between different crime places, and generating an attention matrix;
step 2.1, constructing a weighted directed graph G for crime location 1 ＝(V 1 ,E 1 W), where V e V in the graph 1 Representing a place for a vertex; e 1 Representing a connection between two sites for one edge; w epsilon W is an edge weight matrix representing the tightness degree of two places, and the size is 195 multiplied by 195; τ=1, β=1, ρ=2.
The distance between the sites is obtained according to the longitude and latitude coordinates, if the central longitude and latitude coordinates of site 0 and site 1 are (-73.9845, 40.7552) and (-73.8792, 40.6673) respectively, the central distance is 13.2081km, and the traffic flow is 53, the normalized weight of the boundary between the two sites is 6.9361 ×10 -4 。
Step 2.2, constructing a weighted directed graph G for crime types 2 ＝(V 2 ,E 2 S), where V ε V in the graph 2 Representing an action for a node; e 2 Representing a connection between two behaviors for one edge; s epsilon S is an edge weight matrix representing the tightness degree of two crime types, the size is 30 multiplied by 30, and B and eta are set to be 2 and 2.
And (3) sequencing the crime events according to time, calculating the occurrence interval time of the two crime types, namely, the occurrence interval time of the two crime types 0 and 1 in the 42 th record and the 152 th record is 1.983 hours in time at the position with the longitude and latitude (-73.9723, 40.7910) of the center of the same place, and calculating the side weight of the normalized two crime types to be 0.9844 when the historical occurrence times of the two crime types are 7641 times.
Step 3, in the co-evolution process, a graph convolution network with an attention mechanism is used for aggregating crime types and place neighbor features through random sampling as shown in fig. 2, the aggregated features are input into a multi-layer perceptron for prediction, and a required result is output as shown in fig. 3; the method specifically comprises the following steps:
step 3.1, for the graph G constructed in step 2 1 And G 2 Selecting the nearest 3 neighbors by adopting a random walk sampling strategy, wherein the neighbors after sampling the place 0 are [12, 17, 22 ]]The aggregation weights are respectively [0.47,0.25,0.16 ]]The neighbors after site 1 sampling are [11,7, 27]The aggregation weights are respectively [0.56,0.24,0.04 ]]The method comprises the steps of carrying out a first treatment on the surface of the The neighbors after crime type 0 sampling are [2,3,1]Its aggregate time weight scoreIs not [0.34,0.28,0.14 ]]The type 1 sampled neighbors are [0,2,6 ]]The aggregation weights are respectively [0.44,0.20,0.03 ]]；
Step 3.2, under the condition of considering two layers of neighbors, carrying out graph G by adopting weighted pooling based on attention mechanism 1 And G 2 In the node neighbor feature aggregation, the aggregated neighbor feature is spliced with the node potential feature, W 5 、W 6 、W 7 And W is 8 The size is set to be 5 multiplied by 10, and the initial value is 0; the aggregated crime type and location neighborhood feature size is 1 x 5, and the spliced features, such as the features of crime type 10 at location 24, are represented as:
[-5.095×10 -2 ，2.0656×10 -2 ，-4.2313×10 -2 ，1.3154×10 -1 ，-4.1675×10 -2 ，1.4074×10 -1 ，2.8928×10 -2 ，5.7950×10 -2 ，-5.4407×10 -2 ，1.4077×10 -1 ]
wherein the size becomes 1×10.
Step 3.3, inputting the spliced crime type and the spliced point node characteristics in the step 3.2 into a multi-layer perceptron, wherein the neuron numbers are (7, 3 and 1) respectively when the crime time is predicted; the number of neurons when predicting crime type is (7, 3, 30), respectively; respectively establishing corresponding objective functions according to the crime type, the place and the time, and minimizing a loss function, wherein an optimizer is selected as Adam, and the learning rate is set to be 0.001;
step 3.4, after one training is finished, the potential characteristics of the crime location and the type node are updated;
if the location 0 is updated as follows:
[-3.349×10 -3 ,6.879×10 -5 ,1.191×10 -4 ,1.737×10 -4 ,-2.748×10 -3 ]
crime type 0 is updated as:
[-3.552×10 -3 ,-1.307×10 -3 ,1.719×10 -4 ,5.392×10 -3 ,-4.766×10 -5 ]
cycling the steps 3.2 and 3.3 to perform the next training, performing 10 rounds of training by taking 70% of 331855 data as a training set, and taking 30% of 331855 data as a test set;
predicting the number of crimes at each place according to the model, wherein the weight value of the number of crimes at the places [0,1,2,3,4] is [207, 145, 141, 122, 108], outputting a thermodynamic diagram of the predicted result of the new york crime according to the longitude and latitude coordinates and the weight of the places, and as shown in fig. 4, it can be seen from fig. 4 that the predicted result of the dynamic space-time event according to the embodiment matches with the fact; in the experimental results of this embodiment, the Accuracy (AUC) of the test set was stabilized at 0.7854 and the mean absolute error value (MAE) was stabilized at 0.1006-0.1013; the experimental result shows that the dynamic space-time event deducing method based on the co-evolution graph convolution neural network provided by the embodiment can predict the crime type, the place type and the crime occurrence time, and has high accuracy, small error and low calculation complexity of the predicted result and high practical value.
The above embodiment is only one of the implementation manners capable of implementing the technical solution of the present invention, and the scope of the claimed invention is not limited to the embodiment, but also includes any changes, substitutions and other implementation manners easily recognized by those skilled in the art within the technical scope of the present invention.

Claims (6)

1. A method for predicting a dynamic spatio-temporal event, comprising the steps of:
step 1, acquiring a multisource space-time event data set, and embedding event behaviors, event places and interaction features between the event behaviors and places into event potential features by utilizing a behavior-place co-evolution model to obtain embedded event potential features; constructing an event data bipartite graph by utilizing the embedded event potential characteristics;
step 2, calculating the tightness degree between different event places or different event behaviors respectively by using the event data bipartite graph to obtain a place edge weight matrix W and a behavior edge weight matrix S; updating the event data bipartite graph by using the place edge weight matrix W and the behavior edge weight matrix S respectively to obtain a place weighted directed graph G 1 And behavior weighted directed graph G 2 ；
Step 3, weighting the directed graph G by the places 1 Node and behavior weighted directed graph G of (1) 2 Random walk sampling is carried out on the nodes of the node (C), and the characteristics of the neighbors of the location node and the characteristics of the neighbors of the behavior node are respectively aggregated and spliced by using a graph convolution network to obtain aggregated neighbor characteristics; inputting the aggregated neighbor features into a multi-layer perceptron prediction model for prediction to obtain an event prediction result;
the step 1 specifically comprises the following steps:
step 11, acquiring a multisource space-time event data set, and extracting event behavior information, event location attributes, event occurrence time and interaction characteristics between event behaviors and locations; wherein the event location attribute comprises a location longitude and latitude;
step 12, embedding event behavior information, event location attribute, event occurrence time and interaction feature information between event behavior and location into event potential features by using a behavior-location co-evolution model to obtain an embedded event potential feature list O;
step 13, constructing an event data bipartite graph by using the embedded event potential feature list O;
in step 12, the expression of the embedded event latent feature O is:
Figure FDA0004237427540000011
wherein i is event sequence number, e i For the ith event, N is the total number of events; a, a i Is the ith event behavior, where a i ∈{a 1 ,a 2 ,...,a m M is the number of event behavior types; l (L) i Is the ith event location, wherein, l i ∈{l 1 ,l 2 ,...,l n N is the number of event location categories; b i Interaction characteristics of event behaviors and places;
in step 13, the event data bipartite graph includes behavior embedding features
Figure FDA00042374275400000215
Place embedded feature->
Figure FDA00042374275400000216
Wherein behavior embedding features
Figure FDA0004237427540000021
The expression of (2) is:
Figure FDA0004237427540000022
wherein a is the event behavior,
Figure FDA0004237427540000023
for the occurrence time of the kth time of the event behavior a, W 1 、W 2 、W 3 W and W 4 Respectively a weight matrix, wherein sigma is a sigmoid function; />
Figure FDA0004237427540000024
Interaction characteristics which are the kth time evolution of the event behavior a and are time-dependent; />
Figure FDA0004237427540000025
-the time before the kth occurrence of event behavior a; />
Figure FDA0004237427540000026
The occurrence time of the k-1 th time of the event behavior a; f (f) a Embedded features of event behavior a; q l The embedded feature of event location l;
place-embedding features
Figure FDA0004237427540000027
The expression of (2) is:
Figure FDA0004237427540000028
wherein, l is the event location,
Figure FDA0004237427540000029
embedding features for event location/location after the kth event, for->
Figure FDA00042374275400000210
For event location/kth event occurrence time, W 5 、W 6 、W 7 W and W 8 Respectively a weight matrix; />
Figure FDA00042374275400000211
For the time when the k-1 st event occurs at event location l,/time>
Figure FDA00042374275400000212
-immediately before the occurrence of the kth event for event location l, < >>
Figure FDA00042374275400000213
The kth interaction characteristic of the event location l is evolved with time;
the expression of the place edge weight matrix W is:
Figure FDA00042374275400000214
wherein w is ij For the location node v i With the site node v j Is the influence weight of the distance between event sites, beta is the influence weight of traffic flow, ρ is the distance attenuation parameter between two event sites, c ij Is the event location l i And event location l j Traffic flow of (2);
the expression of the behavior edge weight matrix S is as follows:
Figure FDA0004237427540000031
wherein s is ij For behaving as nodes s i And behavior node s j Is a side weight of (1); delta ij Is event behavior a i And event behavior a j B is the time window length, M is the number of co-occurring event behaviors within the time window, and η is the distance decay parameter.
2. The method for predicting dynamic spatiotemporal events according to claim 1, wherein in step 2, specifically:
step 21, calculating the tightness degree of different places according to the product of the distance between different places and the traffic data to obtain a place edge weight matrix W; calculating the tightness degree of different behaviors according to the co-occurrence times among different behaviors to obtain a behavior edge weight matrix S;
step 22, updating the event data bipartite graph by using the place edge weight matrix W to obtain a place weighted directed graph G 1 The method comprises the steps of carrying out a first treatment on the surface of the Updating the event data bipartite graph by using the behavior edge weight matrix S to obtain a behavior weighted directed graph G 2 。
3. The method for predicting dynamic spatiotemporal events of claim 1, wherein in step 3, the method specifically comprises the following steps:
step 31, weighting the location-weighted directed graph G 1 Adopting a random walk sampling strategy to select the nearest neighbor of the place node; weighted behavior directed graph G 2 Selecting a neighbor closest to the behavior node by adopting a random walk sampling strategy;
step 32, aggregating the features of the neighbors of the location nodes by using a graph-coiler network based on an attention mechanism to obtain the aggregated features of the neighbors of the location nodes; splicing the aggregated site node neighbor features with the site node potential features to obtain site aggregation features;
aggregating the characteristics of the behavior node neighbors by using a graph-coiler network based on an attention mechanism to obtain aggregated behavior node neighbor characteristics; splicing the aggregated behavior node neighbor features with the behavior node potential features to obtain behavior aggregation features;
splicing the location aggregation features and the behavior aggregation features to obtain aggregated neighbor features; inputting the aggregated neighbor features into a multi-layer perceptron prediction model, establishing a corresponding loss function, and training by adopting an optimizer to minimize the loss function;
and step 34, after the training is finished, respectively updating the potential characteristics of the behavior and the place node, and circulating the steps 32-33 to carry out the next training and outputting the result.
4. A method of dynamic spatiotemporal event prediction according to claim 3, characterized in that in step 31, the node v is selected from the group consisting of j Randomly walk to adjacent site node v i The probability of (2) is:
Figure FDA0004237427540000041
wherein,,
Figure FDA0004237427540000042
for the slave point node v j Randomly walk to adjacent site node v i Probability of (2); n (v) j ) For the location node v j Is a neighbor to all neighbors of (a); w (w) kj For the location node v k With the site node v j Is a side weight of (1);
slave action node s j Random walk to adjacent behavioural node s i The probability of (2) is:
Figure FDA0004237427540000043
wherein,,
Figure FDA0004237427540000044
for slave action node s j Random walk to adjacent behavioural node s i Probability of (2); n(s) j ) For behaving as nodes s j Is a neighbor to all neighbors of (a); s is(s) kj For behaving as nodes s k And behavior node s j Is a side weight of (b).
5. A method for predicting dynamic spatiotemporal events according to claim 3, wherein the feature of the neighborhood of the location node is aggregated by using a graph-coiler network based on an attention mechanism to obtain the feature of the neighborhood of the location node after aggregation; splicing the aggregated site node neighbor features with the site node potential features to obtain site aggregation features;
wherein the second layer of sites aggregate features Q' u The expression of (2) is:
Q′ u ＝σ([(∑ v∈N(u) α vu q v )||q u ]W 6 )
wherein alpha is vu For the location node v u Randomly walk to adjacent site node v v Probability of (2); n (u) is a location node v u Is a neighbor to all neighbors of (a); q v For the location node v v Features; q u For the location node v u Potential features;
first tier site aggregation feature Q l The expression of (2) is:
Figure FDA0004237427540000045
wherein alpha is ul For the location node v l Randomly walk to adjacent site node v u Probability of (2); n (l) is a place node v l Is a neighbor to all neighbors of (a); q l For the location node v l Potential features;
aggregating the characteristics of the behavior node neighbors by using a graph-coiler network based on an attention mechanism to obtain aggregated behavior node neighbor characteristics; splicing the aggregated behavior node neighbor features with the behavior node potential features to obtain behavior aggregation features;
wherein the second layer behavior aggregate feature F' v The expression of (2) is:
F′ v ＝σ([(∑ r∈N(v) α rv f r )||f v ]W 8 )
wherein alpha is rv For behaving as nodes s v Random walk to adjacent behavioural node s r Probability of (2); f (f) r For behaving as nodes s r Features; f (f) v For behaving as nodes s v Potential features; n (v) is a behavior node s v Is a neighbor to all neighbors of (a);
wherein the first layer behavior aggregate feature F a The expression of (2) is:
F a ＝σ([(∑ v∈N(a) α va F′ v )||f a ]W 7 )
wherein N (a) is a behavior node s a Is a neighbor to all neighbors of (a); alpha va For behaving as nodes s a Random walk to adjacent behavioural node s v Probability of (2); f (f) a For behaving as nodes s a Potential features;
in step 33, the prediction model based on the multi-layer perceptron is:
E(a,l,t)＝φ(Q l (t)||F a (t))
wherein phi is a multi-layer perceptron prediction model; q (Q) l (t) is a location aggregation feature for event location l; f (F) a (t) is a behavior aggregation feature of event behavior a; e (a, l, t) is an event prediction expectation;
the expression of the loss function is:
Figure FDA0004237427540000051
wherein Y is i Represents the i-th ground truth value,
Figure FDA0004237427540000052
and representing the ith prediction result, wherein G is a ground truth value data set.
6. The dynamic space-time event prediction system is characterized by comprising a construction module, a matrix module and a prediction module;
the construction module is used for acquiring a multisource space-time event data set, embedding event behavior information, event location information and interaction characteristic information between event behaviors and locations into event potential characteristics by utilizing a behavior-location co-evolution model, and obtaining embedded event potential characteristics; constructing an event data bipartite graph by utilizing the embedded event potential characteristics;
the computing and updating module is used for respectively computing the tightness degree between different event places and different event behaviors by utilizing the event data bipartite graph to obtain a place edge weight matrix W and a behavior edge weight matrix S; updating the event data bipartite graph by using the place edge weight matrix W and the behavior edge weight matrix S respectively to obtain a place weighted directed graph G 1 And behavior weighted directed graph G 2 ；
Prediction module for weighting the directed graph G by the locations 1 Node and behavior weighted directed graph G of (1) 2 Random walk sampling is carried out on the nodes of the node (C), and the characteristics of the neighbors of the location node and the characteristics of the neighbors of the behavior node are respectively aggregated and spliced by using a graph convolution network to obtain aggregated neighbor characteristics; inputting the aggregated neighbor features into a multi-layer perceptron prediction model for prediction to obtain an event prediction result;
acquiring a multisource space-time event data set, and embedding event behavior information, event location information and interaction feature information between event behavior and location into event potential features by using a behavior-location co-evolution model to obtain embedded event potential features; the embedded event potential characteristics are utilized to construct an event data bipartite graph, which specifically comprises the following steps:
step 11, acquiring a multisource space-time event data set, and extracting event behavior information, event location attributes, event occurrence time and interaction characteristics between event behaviors and locations; wherein the event location attribute comprises a location longitude and latitude;
step 12, embedding event behavior information, event location attribute, event occurrence time and interaction feature information between event behavior and location into event potential features by using a behavior-location co-evolution model to obtain an embedded event potential feature list O;
step 13, constructing an event data bipartite graph by using the embedded event potential feature list O;
in step 12, the expression of the embedded event latent feature O is:
Figure FDA0004237427540000061
wherein i is event sequence number, e i For the ith event, N is the total number of events; a, a i Is the ith event behavior, where a i ∈{a 1 ,a 2 ,...,a m M is the number of event behavior types; l (L) i Is the ith event location, wherein, l i ∈{l 1 ,l 2 ,...,l n N is the number of event location categories; b i Interaction characteristics of event behaviors and places;
in step 13, the event data bipartite graph includes behavior embedding features
Figure FDA0004237427540000071
Place embedded feature->
Figure FDA0004237427540000072
Wherein behavior embedding features
Figure FDA0004237427540000073
The expression of (2) is:
Figure FDA0004237427540000074
wherein a is the event behavior,
Figure FDA0004237427540000075
for the occurrence time of the kth time of the event behavior a, W 1 、W 2 、W 3 W and W 4 Respectively a weight matrix, wherein sigma is a sigmoid function; />
Figure FDA0004237427540000076
Interaction characteristics which are the kth time evolution of the event behavior a and are time-dependent; />
Figure FDA0004237427540000077
-the time before the kth occurrence of event behavior a; />
Figure FDA0004237427540000078
The occurrence time of the k-1 th time of the event behavior a; f (f) a Embedded features of event behavior a; q l The embedded feature of event location l;
place-embedding features
Figure FDA00042374275400000716
The expression of (2) is:
Figure FDA0004237427540000079
wherein, l is the event location,
Figure FDA00042374275400000710
embedding features for event location/location after the kth event, for->
Figure FDA00042374275400000711
For event location/kth event occurrence time, W 5 、W 6 、W 7 W and W 8 Respectively a weight matrix; />
Figure FDA00042374275400000712
For the time when the k-1 st event occurs at event location l,/time>
Figure FDA00042374275400000713
-immediately before the occurrence of the kth event for event location l, < >>
Figure FDA00042374275400000715
The kth interaction characteristic of the event location l is evolved with time;
the expression of the place edge weight matrix W is:
Figure FDA00042374275400000714
wherein w is ij For the location node v i With the site node v j Is the influence weight of the distance between event sites, beta is the influence weight of traffic flow, ρ is the distance attenuation parameter between two event sites, c ij Is the event location l i And event location l j Traffic flow of (2);
the expression of the behavior edge weight matrix S is as follows:
Figure FDA0004237427540000081
wherein s is ij For behaving as nodes s i And behavior node s j Is a side weight of (1); delta ij Is event behavior a i And event behavior a j B is the time window length, M is the number of co-occurring event behaviors within the time window, and η is the distance decay parameter.
CN202110009698.3A 2021-01-05 2021-01-05 Dynamic space-time event prediction method and system Active CN112633607B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110009698.3A CN112633607B (en) 2021-01-05 2021-01-05 Dynamic space-time event prediction method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110009698.3A CN112633607B (en) 2021-01-05 2021-01-05 Dynamic space-time event prediction method and system

Publications (2)

Publication Number Publication Date
CN112633607A CN112633607A (en) 2021-04-09
CN112633607B true CN112633607B (en) 2023-06-30

Family

ID=75291431

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110009698.3A Active CN112633607B (en) 2021-01-05 2021-01-05 Dynamic space-time event prediction method and system

Country Status (1)

Country Link
CN (1) CN112633607B (en)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113837472B (en) * 2021-09-26 2024-03-12 杭州海康威视系统技术有限公司 Method and equipment for predicting event executives

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111681059A (en) * 2020-08-14 2020-09-18 支付宝(杭州)信息技术有限公司 Training method and device of behavior prediction model
CN111798241A (en) * 2020-05-18 2020-10-20 北京三快在线科技有限公司 Transaction data processing method and device, electronic equipment and readable storage medium

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10423983B2 (en) * 2014-09-16 2019-09-24 Snap Inc. Determining targeting information based on a predictive targeting model

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111798241A (en) * 2020-05-18 2020-10-20 北京三快在线科技有限公司 Transaction data processing method and device, electronic equipment and readable storage medium
CN111681059A (en) * 2020-08-14 2020-09-18 支付宝(杭州)信息技术有限公司 Training method and device of behavior prediction model

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Big data analysis of web map service log;Chen Di et al.;Journal of chinese computer systems;第36卷(第1期);全文 *
事件库构建技术综述;薛聪等;信息安全学报;第4卷(第2期);83-106 *

Also Published As

Publication number Publication date
CN112633607A (en) 2021-04-09

Similar Documents

Publication Publication Date Title
CN113313947B (en) Road condition evaluation method of short-term traffic prediction graph convolution network
CN111079931A (en) State space probabilistic multi-time-series prediction method based on graph neural network
CN114299723B (en) Traffic flow prediction method
Esquivel et al. Spatio-temporal prediction of Baltimore crime events using CLSTM neural networks
CN113762595B (en) Traffic time prediction model training method, traffic time prediction method and equipment
CN114419878B (en) Method, electronic device and storage medium for predicting urban road network global traffic state
CN112766597A (en) Bus passenger flow prediction method and system
CN113780665B (en) Private car stay position prediction method and system based on enhanced recurrent neural network
CN114648097A (en) Elevator trapping feature analysis and time series prediction model construction method based on deep learning, obtained model and prediction method
CN116258573A (en) Agricultural product supply chain credit risk identification and evolution prediction method
CN117148197A (en) Lithium ion battery life prediction method based on integrated transducer model
CN112633607B (en) Dynamic space-time event prediction method and system
CN115641720A (en) Traffic prediction method and system based on space-time fusion graph neural network
CN115114128A (en) Satellite health state evaluation system and evaluation method
Liang et al. Crime prediction with missing data via spatiotemporal regularized tensor decomposition
CN114238765A (en) Block chain-based position attention recommendation method
CN112201348B (en) Knowledge-aware-based multi-center clinical data set adaptation device
Leke et al. Proposition of a theoretical model for missing data imputation using deep learning and evolutionary algorithms
CN116523001A (en) Method, device and computer equipment for constructing weak line identification model of power grid
CN114399901B (en) Method and equipment for controlling traffic system
Chien et al. Bayesian multi-temporal-difference learning
Zhang et al. Granger causal inference for interpretable traffic prediction
CN114358186A (en) Data processing method and device and computer readable storage medium
CN115457081A (en) Hierarchical fusion prediction method based on graph neural network
CN115333957A (en) Service flow prediction method and system based on user behaviors and enterprise service characteristics

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