A large-scale commodity combination pattern mining method based on multi-objective optimization

Abstract

The application discloses a mining method of large-scale commodity combination mode based on multi-target optimization, designs a mode reduction strategy for mining frequent and efficient item sets of a large-scale transaction database, and proposes a population initialization and population repair strategy as an auxiliary, so that a commodity combination with high sales volume and high profit is mined. The application solves the problem that with the increase of the number of commodities in the commodity transaction database, the mining difficulty exponentially increases, provides multiple commodity combinations for user selection, and thus the accuracy and effectiveness of the commodity combination mode mining in the commodity transaction database are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of optimal pattern set of itemset pattern, in particular to a mining method of large-scale itemset pattern based on multi-objective optimization. BACKGROUND

[0002] The target of itemset mining is to find some novel, difficult to directly observe, useful patterns in a given database, and the frequent high utility itemset mining is one of the research directions, in recent years, it has become an important data mining task. In the traditional pattern mining algorithm, the user is required to specify the minimum support threshold min_sup, the minimum occupancy threshold min_util, and the related importance priority between the two indicators λ. The multi-objective optimization algorithm can solve the problem of frequent high utility itemset mining, which can solve the problem of threshold determination. But these algorithms mostly use 0-1 coding, which makes the size of the decision space and the number of items in a geometric relationship, and with the increase of the dimension of the data set, the decision space grows exponentially, thus causing the dimension disaster problem, and cannot mine high-quality solutions.

[0003] At present, the optimal itemset pattern set recommendation problem in itemset pattern mining contains two types of research methods:

[0004] One type of research method is to weight the support and utility as one target, and then use the traditional method to mine itemsets. This method has a limitation, which requires the user to pre-set necessary parameters, that is, min_sup, min_util and λ. Under the premise that the user lacks prior knowledge in the corresponding field, it is difficult for the user to give a reasonable parameter to run the algorithm, and the setting of the parameter has a great influence on the running time and the running result of the algorithm. If the parameter is set too small, the algorithm will run slowly, which is difficult for the user to bear, and if the parameter is set too large, the algorithm will be over-pruned in the search process, which will cause the optimal solution to be cut off.

[0005] One type of research method is the frequent high utility itemset mining algorithm based on multi-objective optimization, which takes the support and utility value as two targets, and converts the frequent high utility itemset mining problem into a multi-objective optimization problem, and then uses evolutionary algorithm for optimization, finally obtains a set of high-quality solutions. This method can solve the problem of threshold determination, but with the increase of the scale of itemset, the difficulty of existing algorithms to mine high-quality solutions also increases, which cannot meet the needs of users. SUMMARY

[0006] The purpose of this invention is to provide a method for mining large-scale product combination patterns based on multi-objective optimization, which can improve the accuracy and effectiveness of product combination pattern recommendations in large-scale product databases, thereby recommending an optimal set of patterns to users, allowing users to choose the optimal product combination pattern according to their needs.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for mining large-scale commodity combination patterns based on multi-objective optimization includes the following steps:

[0009] S1. Establish a commodity transaction database D, and iterate through the total number of commodity combination patterns in the commodity transaction database D to obtain different types of commodity items {I1, I2, ..., I...} in the commodity transaction database D. j , ..., I n , and the utility value W(T) of each item in each item combination pattern. i ,I j ), where n is the total number of items;

[0010] S2. Optimize the multi-objective optimization problem transformed from all product combination patterns in the product transaction database D using the proposed pattern reduction strategy and a multi-objective optimization method based on fitness calculation, thereby obtaining the optimal set of product combination patterns.

[0011] Preferably, in step S1, D = {T1, T2, ..., T} i ,...,T m} represents the set of all product combination patterns in the product transaction database, T i Let m represent the i-th product combination pattern, and m be the total number of product combination patterns.

[0012] The problem of finding any product combination pattern X is transformed into a multi-objective optimization problem, MaximizeF(X), as shown in equation (1):

[0013] MaximizeF(X)=(supp(X),util(X)) T (1),

[0014] In equation (1), supp(X) represents the support of the product combination pattern X, and we have:

[0015] supp(X)=freq(X) / m (2),

[0016] In equation (2), freq(X) represents the frequency of the product combination pattern X;

[0017] In equation (1), util(X) represents the utility value of the commodity combination pattern X, and we have:

[0018]

[0019] In equation (3), T i I represents any supporting transaction that contains a product combination pattern X. j W(T) is any product contained in X. i ,I j ) represents item I in the product combination pattern X. j In transaction T i The corresponding profit, T k Let W(T) represent any transaction in D. k ) represents transaction T k The profit corresponding to all products in the product.

[0020] Preferably, step S2 comprises the following steps:

[0021] S2.1 Population Coding:

[0022] Based on the number of product item types in all product combination patterns in the product transaction database D, all product items in product combination pattern X are encoded in binary format to obtain an individual product combination pattern X = {x1, x2, ..., x...} i ,...,x n};x i Let x represent the i-th item in the product combination pattern X. i =1 indicates that the i-th item exists in the product combination pattern X. i =0 indicates that the i-th item does not exist in the product combination pattern X; thus, the m product combination patterns are encoded to obtain the encoded product transaction database D′;

[0023] Based on the number n of different types of product items in the product transaction database D, let the code of product item Y be "1" and the code of the remaining n-1 product items be "0", thereby obtaining an individual of the product meta-pattern;

[0024] S2.2 Initialize the population:

[0025] First, define the population size as popSize, and set popSize = 100;

[0026] Next, traverse the encoded commodity transaction database D′, find the individual with the most commodity items in D′, denote its individual length as maxLen, and use it as the upper limit of the number of commodity items in the first generation population, with a lower limit of 1;

[0027] Then, the support and utility values of all individuals in the initialized population are calculated by using formula (2) and formula (3), and the reduced importance of each item in the data set is obtained by summing and averaging the support and utility values of all individuals j ):

[0028]

[0029] In formula (4), supp(I j ) and util(I j ) are the support and utility values of item I j , respectively;

[0030] Finally, a number is randomly taken from [1, maxLen] as the individual length for each individual, and the better k item groups are selected by binary tournament to form the individual;

[0031] S2.3, population evolution:

[0032] First, the iteration number L is initialized to 0;

[0033] Second, the reduced interval radix is defined, and radix is taken as 5. If L%radix≠0, the sorted population with multiple frontiers is selected by using the tournament selection strategy, and a mating pool is obtained as the population of the Lth iteration. Then, the individuals in the population of the Lth iteration are crossed and mutated to generate the population of the (L+1)th iteration. If L%radix=0, population reduction and population repair are performed to obtain the repaired population of the (L+1)th iteration. The above steps are repeated until the maximum iteration number gen is reached.

[0034] Then, the original population and the population of the (L+1)th iteration are sorted by using the non-dominated sorting to obtain the sorted population of the (L+1)th iteration with multiple frontiers;

[0035] Finally, the crowded distance of the sorted population of the (L+1)th iteration with multiple frontiers is calculated according to the Euclidean distance, and the sorted population of the (L+1)th iteration with multiple frontiers is sorted in descending order according to the obtained crowded distance to obtain the population of the (L+1)th iteration with multiple frontiers sorted in descending order according to the crowded distance, i.e., the optimal set of item combination patterns.

[0036] Preferably, in the step S2.3, the population is reduced, and the specific steps are as follows:

[0037] A1, the population is sorted by using the non-dominated sorting algorithm to obtain the sorted population with multiple frontiers. The solutions of the first frontier are regarded as elite solutions, and the solutions of the remaining frontiers are regarded as non-elite solutions. The population reduction is performed on the non-elite solutions.

[0038] A2, determine the reduction order of the commodity item in the next individual by formula (5), the larger the TWU, the higher the priority of the comparison;

[0039]

[0040] Where, W(T i ) is the utility sum of all items in transaction T i ;

[0041] A3, find the highest TWU item I k in the non-elite individual P i , and calculate the correlation Corr(I i , I k ) between I i and other items in P j , wherein the other items are assumed to be I j :

[0042]

[0043] In formula (6), I i and I j are two different commodity items, and I i is assumed to be better than I j ;

[0044] A4, determine whether the item I i needs to be reduced according to the size of Corr(I j , I j ):

[0045] If Corr(I i, , I j ) ≥ 0.5, it means that the probability of {I i , I j} is high, and both of them are retained and used as a whole to judge the correlation with other items, and then return to step A1 for continuous calculation;

[0046] If Corr(I i , I j ) = 0, it means that they are completely irrelevant, and the item I j is reduced from 1 to 0, and then return to step A1 for continuous calculation;

[0047] If 0.5 > Corr(I i , I j ) > 0, the next step is judged;

[0048] A5, calculate the importance of the commodity item in the population DS(P k , Ij The higher the importance, the more important the item is in the dataset;

[0049] A6. Taking into account both the importance of an item in the dataset and the importance of an item in the population, calculate the individual P. k Middle Item I j Attribution-reduction importance II (P) k I j );

[0050] A7. Combine the reduction importance with the current generation number to calculate the reduction reversal probability FP. t (P k ,I j If FP t (P k ,I j If the value is greater than the random value, then item I will be... j Reduced to 0, FP t (P k ,I j The larger the value, the greater the probability of reduction;

[0051] A8. Return to step A1 and continue in sequence until all non-elite individuals are reduced and repaired to obtain the population after the L+1th iteration reduction.

[0052] Preferably, in step S2.3, the population repair refers to providing evolutionary guidance to the individuals to be repaired after reduction. For two randomly selected, non-repeating parent individuals, repair is only performed on the items where the selected states differ. The specific steps and methods are as follows:

[0053] B1. Taking into account the reduced importance of the term across all individuals in the population and its corresponding TWU ranking, calculate the repair importance II of the term. Rep (I j ), used for subsequent individual repairs;

[0054] B2. Randomly select two individuals P1 and P2 without repetition. Then, using a non-dominated ranking method, select the individual with higher fitness as P1. b The other is P w Find a List of items in an individual that have different states of 0 and 1. Calculate the corresponding repair probability by considering each item in the List and the current algebra, and then set P... w The middle section was repaired to P. b The corresponding state in;

[0055] B3. Repeat the above steps until all individuals have been selected, and obtain the population after the L+1th iteration repair.

[0056] Preferably, in step A5, the DS(P) k, I j ) is calculated as:

[0057]

[0058] where DR(P k ) is the number of individuals dominated by individual P k , N is the population size, and P j (I k ) is the status of item I j in individual P t , with value 1 indicating that item I k is selected in individual P j , and 0 otherwise.

[0059] In step A6, the II(P k , I j ) is calculated as:

[0060]

[0061] In step A7, the FP t (P k , I j ) is calculated as:

[0062]

[0063] where II(P k , I j ) is the reduced importance of item I k in individual P j , N is the population size, and TWU k (I Rep ) is the TWU rank of item I j .

[0064] Preferably, in step B1, the II Rep (I j ) is calculated as:

[0065]

[0066] where II(P k , I j ) is the reduced importance of item I k in individual P j , N is the population size, and TWU rank (I j ) is the TWU rank of item I j .

[0067] Preferably, in step B2, the repair probability, RP t (P w , I j ), is calculated as:

[0068] RP t (P w , I j ) = a + (-1)α • II rep (I j )·exp Ψ(b,w,t) (11),

[0069]

[0070] Wherein, the repair item is 1 and the adjustment formula is 0, there is partial difference, controlled by alpha;Alpha = phi w＞b , if P b 0 and P w 1, then 1, otherwise, alpha is 0;II Rep (I j ) is the repair importance of item, t is the current evolution generation, the earlier the generation, the higher the repair importance, the greater the repair probability, the greater the probability of this item being 1.

[0071] The beneficial effects of the application are that:

[0072] The application discloses a mining method of large-scale commodity combination mode based on multi-objective optimization, solves the problem that mining difficulty exponentially increases with the increase of the number of commodities in a commodity transaction database, reduces the search space by reducing irrelevant items, and on this basis, proposes a multi-objective optimization algorithm for mining high-dimensional frequent high-utility item sets based on item set reduction; through the reduction strategy, the repair strategy and the initialization strategy, the accuracy and effectiveness of commodity combination mode recommendation in a large-scale commodity database can be improved, so that an optimal mode set can be recommended to a user, and the user can select the optimal commodity combination mode according to the own demand. The mining method of large-scale commodity combination mode based on multi-objective optimization has the advantages of fast operation speed, high accuracy and good effectiveness. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 : the flow chart of the method of the application;

[0074] Figure 2 : the schematic diagram of population importance calculation of the application;

[0075] Figure 3 : the schematic diagram of population reduction of the application;

[0076] Figure 4 : the schematic diagram of population repair of the application;

[0077] Figure 5 : the schematic diagram of population initialization of the application. DETAILED DESCRIPTION

[0078] The application is further illustrated by the following examples. It should be noted that the examples are only examples and illustrations of the inventive concept, and various modifications or supplements or replacements of the described specific examples by similar ways can be made by those skilled in the art without departing from the inventive concept or beyond the scope defined by the claims, and should be considered to fall within the protection scope of the application.

[0079] As shown in the figure, a mining method of large-scale commodity combination pattern based on multi-objective optimization includes the following steps: Figure 1

[0080] S1, establishing a commodity transaction database D, D={T1, T2,..., Tm} indicates a set of all commodity combination patterns in the commodity transaction database, T i indicates the i-th commodity combination pattern, and m is the total number of commodity combination patterns; traversing the total number of commodity combination patterns m in the commodity transaction database D, obtaining different types of commodity items {I1, I2,..., In} in the commodity transaction database D and the utility value W(T m , In) of the commodity items under each commodity combination pattern, and n is the total number of commodity items; i j n i j

[0081] The mining problem of any one commodity combination pattern X is converted into a multi-objective optimization problem Maximize F(X) as shown in formula (1):

[0082] Maximize F(X)=(supp(X), util(X)) T (1),

[0083] In formula (1), supp(X) represents the support degree of the commodity combination pattern X, and has:

[0084] supp(X)=freq(X) / m (2),

[0085] In formula (2), freq(X) represents the frequency of the commodity combination pattern X;

[0086] In formula (1), util(X) represents the utility value (profit) of the commodity combination pattern X, and has:

[0087]

[0088] In formula (3), T i indicates any one support transaction containing the commodity combination pattern X, and I j ​​​​​​is any one of the items included in X, W(T i , j ) represents the item I j in the combination pattern X i , T k represents the corresponding profit in the transaction T k , W(T k ) represents the profit corresponding to all items in the transaction T

[0089] S2, using the proposed pattern reduction strategy and multi-objective optimization method based on fitness calculation, the multi-objective optimization problem converted from all combination patterns in the commodity transaction database D is optimized, and an optimal set of combination patterns is obtained.

[0090] S2.1, population coding:

[0091] According to the number of item categories of all combination patterns in the commodity transaction database D, all items in the combination pattern X are coded in binary form, and an individual X = {x1, x2,..., xm} of the combination pattern is obtained; xi represents the i-th item in the combination pattern X, xi = 1 indicates that the i-th item exists in the combination pattern X, and xi = 0 indicates that the i-th item does not exist in the combination pattern X; thereby m combination patterns are coded, and the coded commodity transaction database D' is obtained. i n i i i

[0092] According to the number n of different types of items in the commodity transaction database D, the coding of the item Y is "1", and the coding of the remaining n-1 items is "0", thereby obtaining an individual of the item pattern;

[0093] S2.2, initialization of population:

[0094] First, define the population size as popSize, and let popSize = 100;

[0095] Secondly, traverse the coded commodity transaction database D', find an individual with the most items in D', and record its individual length as maxLen, which is used as the upper limit of the number of items in the initial population, and the lower limit is 1;

[0096] Then, the support and utility values of all individuals in the initialization population are calculated using formulas (2) and (3), and the importance of each item in the data set is obtained by summing and averaging SU(I j ):​​​​​

[0097]

[0098] In formula (4), supp(I j ), util(I j ) are support and utility values of item I j respectively;

[0099] Finally, a number is randomly taken from [1, maxLen] as the individual length for each individual, and the better k item groups are selected by binary tournament to form the individual;

[0100] Figure 5 Schematic diagram for initializing population of the application. Figure 5 In the formula, it is assumed that the individual length k=3, three items a, c and g are selected by binary tournament, so that the individual {1, 0, 1, 0, 0, 0, 1, 0, 0} is obtained, which is a part of the initial population, and is adjusted by the subsequent population evolution strategy.

[0101] S2.3, population evolution:

[0102] First, the iteration number L is initialized to 0;

[0103] Secondly, the reduction interval radix is defined, and radix=5 is taken, if L%radix≠0, the strategy of tournament selection is adopted to select the population with multiple frontiers after reordering, to obtain the mating pool as the population of the Lth iteration, and the individuals in the population of the Lth iteration are crossed and mutated to generate the population of the (L+1)th iteration, if L%radix=0, population reduction and population repair are performed first to obtain the repaired population of the (L+1)th iteration, and the above steps are repeated until the maximum iteration number gen is reached.

[0104] Wherein, the specific steps of the above population reduction are:

[0105] A1, the population is sorted by using a non-dominated sorting algorithm to obtain a sorted population with multiple frontiers, the solutions of the first frontier are regarded as elite solutions, and the solutions of the remaining frontiers are regarded as non-elite solutions, and the population reduction part is performed on the non-elite solutions;

[0106] A2, the reduction comparison order of the item in the next individual is determined by using formula (5), the larger the TWU is, the higher the priority of comparison is;

[0107]

[0108] Wherein, W(T i ) is the utility sum of all items in transaction T i .

[0109] A3. Find non-elite individuals P k China's highest TWU item I i Calculate I sequentially i With P k Corr(I) among other items i ,I j ), wherein the other term is assumed to be I. j :

[0110]

[0111] In equation (6), I i I j These are two different product items, assuming I i Superior to I j ;

[0112] A4. According to Corr(I) i ,I j Size judgment item I j Does it need to be reduced?

[0113] If Corr(I) i ,I j )≥0.5, indicating {I i ,I j} is a frequent itemset with a high probability. Keep both items and treat them as a whole to determine their relevance with other items. Then return to step A1 and continue in sequence.

[0114] If Corr(I) i ,I j If ) = 0, it means that the two are completely unrelated. Therefore, term I... j Reduce from 1 to 0, then return to step A1 and continue in sequence;

[0115] If 0.5 > Corr(I) i ,I j If ) > 0, proceed to the next step of judgment;

[0116] A5. Calculate the importance DS(P) of each item in the population. k I j The higher the importance, the more important the item is in the dataset; the DS(P) k I j The formula for calculating ) is:

[0117]

[0118] Among them, DR(P k ) is the number of individuals dominated by an individual, N is the population size, and Pk (I j () represents the status of an item in an individual; a value of 1 indicates that an item is selected in the individual, and a value of 0 indicates that it is not selected.

[0119] A6. Taking into account both the importance of an item in the dataset and the importance of an item in the population, calculate the individual P. k Middle Item I j Attribution-reduction importance II (P) k I j ); the II(P) k I j The formula for calculating ) is:

[0120]

[0121] A7. Combine the reduction importance with the current generation number to calculate the reduction reversal probability FP. t (P k ,I j If FP t (P k ,I j If the value is greater than the random value, then item I will be... j Reduced to 0, FP t (P k ,I j The larger the value of FP, the greater the probability of reduction; t (P k ,I j The formula for calculating ) is:

[0122]

[0123] Among them, II(P) k ,I j ) is an individual P k Middle Item I j The importance of reduction Individual P k The average individual importance of all items in the equation, where γ is the adjustment factor and t is the current generation number;

[0124] A8. Return to step A1 and continue in sequence until all non-elite individuals are reduced and repaired to obtain the population after the L+1th iteration reduction.

[0125] Figure 2 This is a schematic diagram illustrating the population importance calculation method of the present invention. Figure 2 The importance of an item in an individual is calculated using formula (7) on the top side. Then, the information from the bottom left is substituted into formula (4) to calculate the importance of the item in the dataset. Substituting both into formula (8) yields the reduced importance of all items.

[0126] Figure 3 A schematic diagram for the population reduction of the present application. By Figure 3 It can be seen that the priority order of the commodity items in the individual {1, 1, 0, 0, 1, 1, 1, 0, 1} is i, a, g, f, e, b, and the correlation degrees thereof are compared in turn. Commodity item a has a low correlation degree with the highest priority item i but a high importance degree itself, and the calculated flipping probability is less than the random value (assumed to be 0.3), so it is retained as a whole and compared with the subsequent items; commodity item g has a high correlation degree with i, a, and is directly retained; commodity item f has low correlation degree and importance degree, and the flipping probability obtained by substituting into formula (9) is 0.67, which is assumed to be greater than the random number 0.5, so it is reduced; commodity item e is directly retained as g; item b is not related to other items and is directly reduced. Thus, the reduced individual is {1, 0, 0, 0, 1, 0, 1, 0, 1}.

[0127] The above population repair refers to evolution guidance for the reduced individual to be repaired. For two parent individuals randomly and repeatedly selected, only the items with different selected states in the individuals are repaired, and the specific steps and methods are as follows:

[0128] B1, the repair importance II of the item is calculated by comprehensively considering the reduction importance of the item in all individuals in the population and the corresponding TWU ranking, which is used for individual repair in the following; the calculation formula of the II is as follows: Rep j Rep j

[0129]

[0130] wherein II(P k ,I j ) is the reduction importance of item I k in individual P j , N is the population size, and TWU rank (I j ) is the TWU ranking corresponding to item I j .

[0131] B2, two individuals P1 and P2 are randomly and repeatedly selected, the individual with high fitness is selected as P b , and the other is P w , the commodity items with different 0 and 1 states in the individuals are found, the repair probability corresponding to each item in List is calculated by considering the List and the current generation number, and part of the items in P w are repaired to the corresponding states in P b ; the repair probability is specifically as follows:

[0132] RP t (P​​​​w I j )=α+(-1) α ·II rep (I j )·exp Ψ(b,w,t) (11),

[0133]

[0134] Wherein, the repair item is 1 and is 0 when the adjustment formula has partial difference, is controlled by α;α=φ w＞b , if P b 0 and P w 1, then 1, otherwise, α is 0;II Rep (I j ) is the repair importance of item, t is the current evolution generation, the repair importance is higher when the generation is earlier, the repair probability is greater, the probability of this item being 1 is greater.

[0135] B3, repeat the above steps until all individuals are selected, and obtain the population after the L+1th iteration repair.

[0136] Figure 4 It is a schematic diagram of population repair of the application. As can be seen from the figure, Figure 4 The different items in the two individuals are e, f and g, the corresponding repair probability in the poor individual P2={1,0,0,0,1,0,1,0,1} is calculated by formula (11), and then compared with random number, assuming that only item e meets the repair condition, then it is repaired, and finally the individual P2'={1,0,0,0,0,0,1,0,1} is obtained.

[0137] Then, the original population and the L+1th iteration population are sorted by using non-dominated sorting, and the sorted L+1th iteration population with multiple frontiers is obtained;

[0138] Finally, the crowded distance of the sorted L+1th iteration population with multiple frontiers is calculated according to the Euclidean distance, and the sorted L+1th iteration population with multiple frontiers is sorted in descending order according to the obtained crowded distance, and the L+1th iteration population with multiple frontiers sorted in descending order according to the crowded distance is obtained, that is, the optimal set of commodity combination mode set is obtained.

[0139] The application discloses a mining method of large-scale commodity combination mode based on multi-target optimization, solves the problem that mining difficulty exponentially increases with the increase of commodity quantity in a commodity transaction database, reduces the search space by reducing irrelevant terms, and on the basis, proposes a multi-target optimization algorithm for mining high-dimensional frequent high-utility term sets based on term set reduction; through a reduction strategy, a repair strategy and an initialization strategy, the accuracy and effectiveness of commodity combination mode recommendation in a large-scale commodity database can be improved, so that an optimal mode set is recommended for a user, and the user can select an optimal commodity combination mode according to own demand.

[0140] The above is an exemplary description of the application, and obviously, the specific implementation of the application is not limited by the above mode, as long as the method concept and technical scheme of the application are adopted for non-essential improvement or the concept and technical scheme of the application are directly applied to other occasions without improvement, which is within the protection scope of the application.

Claims

1. A method for mining large-scale commodity combination patterns based on multi-objective optimization, characterized in that, Includes the following steps: S1. Establish a commodity transaction database D, and iterate through the total number of commodity combination patterns in the commodity transaction database D to obtain different types of commodity items {I1, I2, ..., I...} in the commodity transaction database D. j , ..., I n ,} and I j In T i The corresponding utility value or profit W(T) i ,I j ), T i This represents a transaction that includes the i-th product combination pattern, i.e., any product combination pattern X, where I... j This represents the number of product items contained in the i-th product combination pattern, i.e., any product combination pattern X, where n is the total number of product items. S2. Optimize the multi-objective optimization problem transformed by all product combination patterns in the product transaction database D using the proposed pattern reduction strategy and the multi-objective optimization method based on fitness calculation, thereby obtaining the optimal set of product combination patterns. In step S1, D = {T1,T2,...,T} i ,...,T m } represents the set of all product combination patterns in the product transaction database, where m is the total number of product combination patterns; The problem of finding any product combination pattern X is transformed into a multi-objective optimization problem, as shown in Equation (1): Maximize F(X): MaximizeF(X)=(supp(X),util(X)) T (1), In equation (1), supp(X) represents the support of the product combination pattern X, and we have: supp(X)=freq(X) / m (2), In equation (2), freq(X) represents the frequency of the product combination pattern X; In equation (1), util(X) represents the utility value of the commodity combination pattern X, and we have: In equation (3), T k Let W(T) represent any transaction in D. k ) represents transaction T k The profit corresponding to all products in the product; The specific steps of step S2 are as follows: S2.1 Population Coding: Based on the number of product item types in all product combination patterns in the product transaction database D, all product items in product combination pattern X are encoded in binary format to obtain an individual product combination pattern X = {x1, x2, ..., x...} i ,...,x n };x i Let x represent the i-th item in the product combination pattern X. i =1 indicates that the i-th item exists in the product combination pattern X. i =0 indicates that the i-th item does not exist in the product combination pattern X; thus, the m product combination patterns are encoded to obtain the encoded product transaction database D′; Based on the number n of different types of product items in the product transaction database D, let the code of product item Y be "1" and the code of the remaining n-1 product items be "0", thereby obtaining an individual of the product meta-pattern; S2.2 Initialize the population: First, define the population size as popSize, and set popSize = 100; Next, traverse the encoded commodity transaction database D′, find the individual with the most commodity items in D′, denote its individual length as maxLen, and use it as the upper limit of the number of commodity items in the first generation population, with a lower limit of 1; Then, the support and utility values ​​of all individuals in the initial population are calculated using equations (2) and (3), and the importance SU(I) of each item in the dataset is obtained by summing and averaging. j ): In equation (4), supp(I j ),util(I j ) are respectively item I j Support and utility value; Finally, for each individual, a number is randomly selected from [1, maxLen] as the individual length, and then the best k items are selected through a binary tournament to form an individual; S2.3 Population Evolution: First, initialize the number of iterations L = 0; Secondly, define the reduction interval radix and take radix = 5. If L%radix ≠ 0, then the tournament selection strategy is used to select from the population with multiple frontiers after re-sorting to obtain the mating pool, which is the population of the Lth iteration. Then, crossover mutation is performed on the individuals in the population of the Lth iteration to generate the population of the L+1th iteration. If L%radix = 0, then first perform population reduction and population repair to obtain the population after the L+1th iteration repair. Repeat the above steps until the maximum number of iterations gen is reached. Then, the original population and the population of the L+1th iteration are sorted using non-dominated sorting to obtain the sorted population of the L+1th iteration with multiple frontiers. Finally, the crowding distance of the population with multiple front edges in the L+1th iteration after sorting is calculated according to Euclidean distance, and the population with multiple front edges in the L+1th iteration after sorting is sorted in descending order according to the obtained crowding distance, so as to obtain the population with multiple front edges in the L+1th iteration after sorting in descending order according to crowding distance, which is the optimal set of commodity combination patterns. In step S2.3, the population reduction is carried out through the following steps: A1. The population is sorted using a non-dominated sorting algorithm to obtain a sorted population with multiple fronts. The solution of the first front is regarded as an elite solution, and the solutions of the other fronts are regarded as non-elite solutions. The population reduction part is performed on the non-elite solutions. A2. Use formula (5) to determine the order of comparison of the items in the individual. The larger the TWU, the higher the priority of the comparison. Among them, W(T) i ) is T i The sum of the utilities of all items in the formula; A3. Find non-elite individuals P k China's highest TWU item I i Calculate I sequentially i With P k Corr(I) among other items i ,I j ), wherein the other term is assumed to be I. j : In equation (6), I i I j These are two different product items, assuming I i Superior to I j ; A4. According to Corr(I) i ,I j Size judgment item I j Does it need to be reduced? If Corr(I) i ,I j )≥0.5, indicating {I i ,I j } is a frequent itemset with a high probability. Keep both items and treat them as a whole to determine their relevance with other items. Then return to step A1 and continue in sequence. If Corr(I) i ,I j If ) = 0, it means that the two are completely unrelated. Therefore, term I... j Reduce from 1 to 0, then return to step A1 and continue in sequence; If 0.5 > Corr(I) i ,I j If )>0, proceed to the next step of judgment; A5. Calculate the importance DS(P) of each item in the population. k ,I j The higher the importance, the more important the item is in the dataset; A6. Taking into account both the importance of an item in the dataset and the importance of an item in the population, calculate the individual P. k Middle Item I j Attribution-reduction importance II (P) k I j ); A7. Combine the reduction importance with the current generation number to calculate the reduction reversal probability FP. t (P k ,I j If FP t (P k ,I j If the value is greater than the random value, then item I will be... j Reduced to 0, FP t (P k ,I j The larger the value, the greater the probability of reduction; A8. Return to step A1 and continue in sequence until all non-elite individuals are reduced and repaired to obtain the population after the L+1th iteration of reduction. In step S2.3, population repair refers to providing evolutionary guidance to the individuals to be repaired after reduction. For two randomly selected, non-repeating parent individuals, repair is only performed on the items where the selected states differ. The specific steps and methods are as follows: B1. Taking into account the reduced importance of the term across all individuals in the population and its corresponding TWU ranking, calculate the repair importance II of the term. Rep (Ij), used for subsequent individual repairs; B2. Randomly select two individuals P1 and P2 without repetition. Then, using a non-dominated ranking method, select the individual with higher fitness as P1. b The other is P w Find a List of items in an individual that have different states of 0 and 1. Calculate the corresponding repair probability by considering each item in the List and the current algebra, and then set P... w The middle section was repaired to P. b The corresponding state in; B3. Repeat the above steps until all individuals have been selected, and obtain the population after the L+1th iteration repair.

2. The method for mining large-scale commodity combination patterns based on multi-objective optimization according to claim 1, characterized in that, In step A5, the DS(P) k I j The formula for calculating ) is: Among them, DR(P k ) is the number of individuals dominated by an individual, N is the population size, and P k (I j () represents the status of an item in an individual; a value of 1 indicates that an item is selected in the individual, and a value of 0 indicates that it is not selected. In step A6, the II(P) k I j The formula for calculating ) is: In step A7, the FP t (P k ,I j The formula for calculating ) is: Among them, II(P) k ,I j ) is an individual P k Middle Item I j The importance of reduction Individual P k The average individual importance of all items in the equation, where γ is the adjustment factor and t is the current generation number.

3. The method for mining large-scale commodity combination patterns based on multi-objective optimization according to claim 2, characterized in that, In step B1, the II Rep (I j The formula for calculating ) is: Among them, II(P) k ,I j ) is an individual P k Middle Item I j The importance of reduction, where N is the population size, TWU rank (I j ) is item I j The corresponding TWU ranking.

4. The method for mining large-scale commodity combination patterns based on multi-objective optimization according to claim 3, characterized in that, In step B2, the repair probability is specifically as follows: RP t (P w ,I j )=α+(-1) α ·II rep (I j )·expΨ(b,w,t) (11), The adjustment formula differs slightly when the item to be repaired is 1 and 0, and this difference is controlled by α; α = φ w>b If P b =0 and P w If α is 1, then it is 1; otherwise, α is 0. Rep (I j ) represents the repair importance of the term, and t is the current evolution generation. The earlier the generation, the higher the repair importance, the greater the repair probability, and the greater the probability that this term is 1.

Images