Multi-objective evolutionary algorithm-based frequent and efficient co-location pattern mining method

Abstract

According to the method, the idea and method of a multi-objective evolutionary algorithm are combined, frequent and efficient co-location modes are mined from a spatial data set, and the co-location modes are excellent in the aspects of the frequency and the utility value. The co-location mode has two traditional measurement methods, namely the frequency and the utility value, in practical application, a user possibly needs to consider the frequency and the utility value at the same time to make a comprehensive decision, but the existing method is difficult to effectively consider the frequency and the utility value. The mining problem is converted into a multi-objective optimization problem by taking the frequency and the utility value as two objectives. According to the method, potential interesting modes possibly lost in single measurement mining are avoided, meanwhile, the step that a user sets a frequency threshold value or a utility value threshold value is omitted, and the optimal co-location mode can be obtained only through one-time mining. In order to improve the mining speed, the concept of a group utility value is introduced, and a second-order auxiliary tool is provided by combining the downward closing characteristic of the group utility value and the frequency. The tool helps to reduce the search space, thereby accelerating the mining process.

Description

Technical Field

[0001] The present invention relates to the technical fields of spatial data mining, co-location pattern mining and multi-objective evolutionary algorithm, and in particular to a frequent and highly effective co-location pattern mining method based on a multi-objective evolutionary algorithm. Background Art

[0002] Co-location pattern mining is an important branch of data mining. Co-location patterns can be used in many spatial data application scenarios, and the mining results can be used to make corresponding decisions for users. A co-location pattern represents the spatial characteristics of a subset, and the characteristics of the pattern have adjacent geographical relationships in spatial distribution. For example, when city builders do urban planning, they can use co-location patterns to plan the distribution of various types of urban POIs such as restaurants, shops, supermarkets, and hospitals in advance, and design the combination of different urban facility POIs, so as to better construct a more reasonable urban distribution.

[0003] There are two traditional measurement methods for co-location patterns, namely frequency and utility value, and these two measurement methods focus on different mining tasks. Frequency refers to the frequency of co-location patterns, and utility value refers to the sum of the values ​​of the patterns. However, for urban planners, they may need to consider both the frequency and utility value of urban POIs, but the existing methods cannot take these two measurement methods into account at the same time. At the same time, frequency and participation are conflicting to a certain extent. When a corresponding single threshold is set, some patterns that users may be interested in will be lost, such as patterns with high utility but low frequency, or patterns with high frequency but low utility value, but these patterns also have corresponding reference value and significance in reality.

[0004] In order to avoid the mining task from missing patterns with reference value and significance, this paper combines the ideas and methods of multi-objective optimization into co-location pattern mining. Using multi-objective evolutionary algorithm mining not only allows the mining task to mine co-location patterns from the perspective of utility value and frequency at the same time, but also saves the process of users setting thresholds, so that users do not need to set thresholds multiple times to mine co-location patterns of interest. The co-location patterns mined by the multi-objective evolutionary algorithm meet the needs of users to examine multiple aspects of information, and are also conducive to users making more comprehensive decisions in actual application scenarios. Summary of the invention

[0005] The purpose of the present invention is to provide a frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm, aiming to mine co-location patterns from both high-utility and frequency perspectives while avoiding missing some patterns that may be of interest to users.

[0006] To achieve the above object, the present invention provides a frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm, comprising the following steps:

[0007] (1) Obtain the cluster and total utility value of the spatial data set;

[0008] (2) Initialize the second-order auxiliary tools based on the information of the spatial dataset;

[0009] (3) Genetically encode the cluster and transform it into an initial population using an initialization strategy;

[0010] (4) Calculate the participation and utility value of each individual in the population, then perform non-dominated sorting on the population, and select individuals distributed on the Pareto frontier and with better performance as the next generation of parents. The parents generate a new next generation population through the improved two-point mutation method, and update the second-order auxiliary tool through the population individuals. Repeat step (4) until the maximum number of evolutions is reached;

[0011] (5) compile the last generation of population individuals into the form of homotopic patterns;

[0012] A spatial dataset refers to a collection of spatial instances, which includes spatial coordinate information, its features, and instance number information. In the process of calculating clusters, the proximity relationship between all spatial instances is calculated, and finally the collection of generated clusters is extracted. The generated clusters can be used for subsequent participation and utility value calculations. The total utility value of a spatial dataset refers to the sum of the utility values ​​represented by all spatial instances.

[0013] The second-order auxiliary tool refers to the flag used to mark the second-order homonymous pattern. When one of the second-order homonymous patterns is not dominated by the existing homonymous patterns in terms of participation and utility value, its flag is 1, otherwise its flag is 0. First, the feature set in the spatial data set is obtained, and then all the second-order homonymous patterns are generated through permutations and combinations, and the flags of all second-order homonymous patterns are initially set to 1.

[0014] Gene encoding is a binary encoding form that converts the homology pattern into a spatial feature set. The feature of the homology pattern corresponds to the position of the spatial feature set and needs to be set to 1, otherwise it is set to 0. The length of each gene encoding is the length of the spatial feature set. The form of gene encoding is conducive to the multi-objective evolutionary algorithm to perform mutation and selection steps. The initialization strategy refers to the method of generating the initialization population. First, the groups are sorted according to the length, and the groups with the longest length are selected to be converted into gene encoding as the initialization population. The number of selections is the size of the population.

[0015] The participation and utility value of each population individual are calculated based on the group instance, and then the population individuals are sorted by non-domination. According to the non-dominated sorting results, individuals distributed on the Pareto front and with better performance are selected. The number of individuals selected is the size of the population. These selected individuals serve as the parents of the next generation for evolutionary mutation. The obtained parents are used to generate new population individuals by using the improved two-point mutation method with the help of the second-order auxiliary tool as the next generation population. The new population needs to update the second-order auxiliary tool to determine whether the new population individuals dominate the second-order homotopic pattern, so as to reduce the possibility of generating dominated individuals. Then repeat step (4) until the maximum number of evolutions set by the user is reached.

[0016] The last generation of population individuals is the optimal set of isotopic patterns produced during the entire mining process. When the evolution is completed, the genetic coding form needs to be converted into the isotopic pattern form.

[0017] Among them, we proposed a second-order auxiliary tool to speed up the mining process. The second-order auxiliary tool records the status of all second-order syntopic patterns in the spatial data set. When the status code of one of the second-order syntopic patterns is 0, it means that the second-order syntopic pattern is dominated by other syntopic patterns in terms of participation and utility value. According to the downward closure characteristics of participation and group utility value, it is necessary to prune the second-order syntopic pattern and its superset syntopic pattern. This is reflected in the improved two-point mutation method later. The gene code corresponding to the second-order syntopic pattern needs to be set to 0, that is, its superset syntopic pattern is not generated, so as to reduce the mining search space and speed up the mining process.

[0018] The present invention provides a frequent and highly effective co-location pattern mining method based on a multi-objective evolutionary algorithm to generate all clusters from a spatial data set. According to the number of features in the spatial data set, the second-order auxiliary tool is initialized, and the participation and cluster utility value of each second-order co-location pattern are calculated. In the framework of the evolutionary algorithm, an initialization population is first generated through an initialization strategy, and the participation and utility value of each individual are calculated, and then a non-dominated sorting is performed. In the cyclic evolution process, excellent individuals are selected from the non-dominated sorting results of the previous generation population, and then an improved mutation method is used to generate new offspring, and the second-order auxiliary tools need to be updated at the same time. Finally, after reaching the maximum number of evolutions, the solution set on the Pareto frontier is the optimal co-location pattern. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0020] Figure 1 It is a flow chart of a frequent and highly effective co-location pattern mining method based on a multi-objective evolutionary algorithm of the present invention.

[0021] Figure 2 This is a pseudo code introduction to the overall framework of the multi-objective evolutionary algorithm of the present invention.

[0022] Figure 3 This is an introduction to the improved two-point mutation pseudo code of the present invention.

[0023] Figure 4 It is an improved two-point mutation flow chart of the frequent and highly effective isotope pattern mining method based on a multi-objective evolutionary algorithm of the present invention. DETAILED DESCRIPTION

[0024] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0025] See also Figure 1 The present invention proposes a frequent and highly effective co-location pattern mining method based on a multi-objective evolutionary algorithm, comprising the following steps:

[0026] S1: Traverse the data set to obtain all groups and total utility values;

[0027] S2: Initialize the second-order auxiliary tool based on the information of the spatial dataset;

[0028] S3: Generate an initialization population using an initialization strategy and convert it into a genetic encoding form;

[0029] S4: Evaluate the current population individuals and perform non-dominated sorting. In the sorting results, select the better population individuals as the parent generation. The parent generation generates a new population through the improved two-point mutation method and updates the second-order auxiliary tool until the maximum number of evolutions is reached;

[0030] S5: Convert the individuals of the last first generation distributed on the Pareto frontier into the form of homotopic patterns;

[0031] When traversing the data set, it is necessary to calculate the proximity relationship of each spatial instance, then obtain the cluster according to the definition of the cluster, and finally accumulate the utility value of each instance to obtain the total utility value.

[0032] According to the number of features in the spatial data set, all second-order co-location patterns are obtained by permutation and combination, and the state of each second-order co-location pattern is recorded using a second-order auxiliary tool, which is initially set to 0.

[0033] The obtained clusters are first sorted according to their length, and then the initialization strategy is used to generate the initialization population, which is converted into genetic coding according to the characteristics of each instance of the cluster.

[0034] Calculate the participation and utility value of each individual in the population, then perform non-dominated sorting, select individuals with excellent performance in both participation and utility value as the next generation parents, and then use the improved two-point mutation method to generate new population individuals. At the same time, it is necessary to update the status code of the second-order auxiliary tool in time, and repeat this step until the maximum number of evolutions is reached.

[0035] The input data required by this algorithm are the spatial dataset, the location features in the dataset, the minimum threshold and the maximum number of evolutions defined by the user.

[0036] Furthermore, the present invention is further described in detail with reference to specific embodiments and pseudo codes:

[0037] See also Figure 2 ,The specific steps of the algorithm are as follows:

[0038] Algorithm 1 shows the overall pseudo code for mining frequent and high-utility co-location patterns.

[0039] (Step 1-2) First, based on the spatial data set provided by the user and the minimum distance threshold set by the user, all clusters are generated and the number of features of the spatial data set is counted.

[0040] (Step 3-4) Then, according to the number of features in the spatial data set, all second-order homonymous patterns are generated by permutations and combinations, and the participation and utility value of each second-order homonymous pattern are calculated. The second-order auxiliary tool is initialized, and the status code of each second-order homonymous pattern is set to 0.

[0041] (Step 5-6) Calculate the total utility value of the spatial data set for calculating the utility rate, perform gene encoding on all clusters, and generate an initialization population using the initialization strategy.

[0042] (Step 7-8) Next, the participation and utility value of each initialized population individual are calculated, and a non-dominated sort is performed based on the calculation results.

[0043] (Step 11) Using the binary tournament method, select individuals with better performance from the non-dominated sorting results of the previous generation as the parents of the next generation.

[0044] (Step 12) In Algorithm 2 ( Figure 3 ) is described in detail.

[0045] (Steps 13-15) According to the elite retention mechanism, the offspring generated by mutation are merged with the previous parent generation, and the participation and utility values ​​are calculated, followed by non-dominated sorting.

[0046] (Step 16) Based on the participation and utility values ​​of the generated offspring, the status code of the second-order auxiliary tool is updated to determine whether there is a second-order homonymous pattern dominated by individuals in the new population. The calculation formula for participation is as follows.

[0047]

[0048] The calculation formula of the utility value is as follows.

[0049]

[0050] (Step 19) Select the solution set distributed on the Pareto front from the final population offspring, and convert it into a common pattern, that is, the final frequent and high-utility common pattern.

[0051] See also Figure 3 ,The specific steps of the algorithm are as follows:

[0052] And Algorithm 2( Figure 3 ) is the algorithm 1( Figure 2 ) is a detailed description of step 12 in the improved two-point mutation pseudocode introduction. The parent generation generates new population individuals through the improved two-point mutation method.

[0053] (Step 1-2) Obtain non-dominated second-order co-patterns and dominated second-order co-patterns from the second-order auxiliary tool respectively.

[0054] (Step 3) Perform mutation processing on each selected parent generation.

[0055] (Step 4-10) Next, a co-pattern is randomly selected from the non-dominated second-order co-patterns, and the encoding is changed according to the genetic encoding of the selected co-pattern.

[0056] (Steps 11-15) It is necessary to traverse each dominated second-order homology pattern, repair the selected parent individual gene, and set the corresponding gene code to 0.

[0057] (Step 16-22) Determine whether the gene length of the individual after the mutation exceeds 1. If it does not exceed 1, it is necessary to select an individual in the non-dominated second-order homology pattern for gene change.

[0058] (Steps 23-24) Finally, each individual that has undergone genetic mutation serves as a population individual of the next generation.

[0059] See also Figure 4, combined with examples to illustrate the specific steps of the improved two-point mutation flow chart:

[0060] Step 1: Genetically encode the homology pattern {A,E,G} that needs to be mutated into the form of {1,0,0,0,1,0,1}.

[0061] Step 2: Determine the status of all second-order parity patterns recorded by the current second-order auxiliary tool. The two parity patterns {A, C} and {E, G} are dominated, and the status code is 1.

[0062] Step 3: Perform an improved two-point mutation on the isotope pattern {A, E, G}, and the final result is the isotope pattern {A, B}.

[0063] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm, characterized in that: The following steps are involved: (1) Obtain the cluster and total utility value of the spatial data set; (2) Initialize the second-order auxiliary tools based on the information of the spatial dataset; (3) Genetically encode the cluster and transform it into an initial population using an initialization strategy; (4) Calculate the participation and utility value of each individual in the population, then perform non-dominated sorting on the population, and select individuals distributed on the Pareto frontier and with better performance as the next generation of parents. The parents generate a new next generation population through the improved two-point mutation method, and update the second-order auxiliary tool through the population individuals. Repeat step (4) until the maximum number of evolutions is reached; (5) Compile the last generation of population individuals into the form of isotopic patterns.

2. The frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm as claimed in claim 1, characterized in that: The spatial dataset described in step (1) refers to a collection of spatial instances, which includes spatial coordinate information, its features, and instance number information. In the process of calculating clusters, the proximity relationship between all spatial instances is calculated, and finally the collection of generated clusters is extracted. The generated clusters can be used for subsequent participation and utility value calculations. The total utility value of the spatial dataset refers to the sum of the utility values ​​represented by all spatial instances.

3. The frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm as claimed in claim 1, characterized in that: The second-order auxiliary tool described in step (2) refers to a flag used to mark the second-order co-location pattern. When one of the second-order co-location patterns is not dominated by the existing co-location patterns in terms of participation and utility value, its flag is 1, otherwise its flag is 0. First, the feature set in the spatial data set is obtained, and then all the second-order co-location patterns are generated by permutations and combinations, and the flags of all the second-order co-location patterns are initially set to 1.

4. The frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm according to claim 1, characterized in that: The gene encoding described in step (3) is a binary encoding form of converting the homology pattern into a spatial feature set. The feature of the homology pattern corresponds to the position of the spatial feature set and needs to be set to 1, otherwise it is set to 0. The length of each gene encoding is the length of the spatial feature set. The form of gene encoding is conducive to the multi-objective evolutionary algorithm to perform mutation and selection steps. The initialization strategy refers to the method of generating an initialization population. First, the groups are sorted according to the length, and the groups with the longest length are selected to be converted into gene encoding as the initialization population. The number of selections is the size of the population.

5. The frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm according to claim 1, characterized in that: In step (4), the participation and utility value of each population individual are calculated according to the group, and then the population individuals are sorted in a non-dominated manner. According to the non-dominated sorting results, select individuals that are distributed on the Pareto front and have better performance. The number of individuals selected is the size of the population. These selected individuals serve as the parents of the next generation for evolutionary mutation. The obtained parents are used to generate new population individuals using the two-point mutation method improved by the second-order auxiliary tool as the next generation population. The new population needs to update the second-order auxiliary tool to determine whether the new population individuals dominate the second-order homotopic pattern, so as to reduce the possibility of generating dominated individuals. Then repeat step (4) until the maximum number of evolutions set by the user is reached.

6. The frequent and high-utility co-location pattern mining method based on a multi-objective evolutionary algorithm according to claim 1, characterized in that: The last generation of population individuals in step (5) is the optimal isotope pattern set generated in the entire mining process. When the evolution is completed, the gene coding form needs to be converted into the isotope pattern form.

Images