An evolutionary multi-task based large-scale user recommendation method

Abstract

The application discloses a kind of large-scale user recommendation methods based on evolution multitask, comprising: step 1, obtaining user and article interaction dataset, step 2, by constructing neural network algorithm, the score of the preference of user to article is excavated, and the rating matrix of user to article is obtained;Step 3: by clustering, group users, and similar user of preference interest is as the same category;Step 4, initialization generates multi-task population;Step 5, information migration between individuals in the same user group, information migration between populations in different user groups, and the optimal user solution is selected by environmental selection iteration, and finally the optimal solution is selected as the recommendation list of the user.The application can reduce the time and space consumed in large-scale user recommendation optimization problem, and improve the accuracy of user recommendation result prediction by clustering technology.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of evolutionary computation and data mining, specifically a recommendation method based on evolutionary multi-task optimization. Background Technology

[0002] The purpose of recommender systems is to help users filter useful information from massive amounts of data. Traditional recommender systems only consider the accuracy of recommendations. Besides accuracy, novelty, diversity, and other performance metrics are also important indicators for recommender systems. Therefore, multi-objective recommender systems have become an important research direction in recommender systems. However, as recommender systems need to satisfy multiple metrics, optimizing some metrics will inevitably lead to conflicts with others. Therefore, multi-objective optimization has become an important technical means to solve the problems in recommender systems.

[0003] Existing multi-objective optimization methods for recommender systems include: scalar methods that convert multi-objective metrics into a single-objective problem through weighted summation, or using population-based evolutionary algorithms to optimize multiple objectives simultaneously.

[0004] Scalar methods typically begin by weighted summation of the two objectives, then use a multigradient descent algorithm to combine scalarization with Pareto-efficient SGD, employing KKT conditions to guide the update of the scalar weights. However, this approach can only optimize objectives with gradients.

[0005] When optimizing multi-objective recommendation systems using evolutionary algorithms, a common approach is to run the algorithm separately for each user. However, as an iterative algorithm, this sequential optimization leads to excessively long runtimes when the dataset contains a large number of users. Another approach is to combine the user recommendations into a single chromosome using real-valued solutions, optimizing all users simultaneously in one iteration. However, this method results in excessively long chromosome encodings, making it difficult to optimize all users to the optimal state simultaneously. This can cause the optimization method to easily get trapped in local optima, resulting in recommendations that fail to meet user needs. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a large-scale user recommendation method based on evolutionary multi-task, thereby providing users with recommendation lists that include accuracy, novelty, and diversity, while ensuring the efficiency and effectiveness of the recommendation method.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] The large-scale user recommendation method based on evolutionary multi-task in this invention is characterized by the following steps:

[0009] Step 1: Obtain relevant user and item data:

[0010] Obtain the user set S = {s1, s2, ..., s} u ,…,s |S|}, where |S| represents the number of users, s u Indicates user number u;

[0011] Obtain the item set Q = {q1, q2, ..., q} i ,…,q |Q|}, where |Q| represents the quantity of items, q i This represents the i-th item;

[0012] Obtain the user-item interaction dataset, and use the interaction data of each user and item as a positive sample, and randomly collect items that the user has not interacted with as negative samples;

[0013] Step 2: Obtain the user rating matrix for items through a prediction model, which includes: an encoding layer, a fully connected layer, an attention interaction layer, an interaction output layer, and a prediction layer;

[0014] Step 2.1: The encoding layer for user s with ID u u and the i-th item q i Perform one-hot encoding to obtain user s with ID u. u sparse vector θ u and item q i sparse vectors After mapping the two sparse vectors to E-dimensional vectors respectively, and then obtaining user s with ID u through the mapping of the fully connected layer. u The representation vector p u and the i-th item q i The representation vector o i ;

[0015] Step 2.2, the attention interaction layer on p u With o i Process the data and output the attention vector a. u,i ;

[0016] Step 2.3: The interactive output layer will p u With o i After concatenation, it is then combined with the attention vector a u,i After performing the dot product, we get user s with ID u. u With the i-th item q i The interaction vector f u,i ;

[0017] Step 2.4: The prediction layer processes the interaction vector f u,i After performing multi-layer fully connected processing, the predicted interaction score r is output. u,i;

[0018] Step 2.5: Using the mean squared error of the predicted interaction score as the loss function, the Adam algorithm is used to optimize the prediction model until the maximum number of iterations is reached, thereby obtaining the optimal prediction model and its output interaction score matrix for each user for each item.

[0019] Step 3: Let the number of clusters be K, and top represent the number of candidate items. Select user s with ID u. u The top 10 preferred items are used as the candidate item set (Candicates). u The similarity between users is calculated by counting the number of identical items in each user's candidate item set. Users with similarity are grouped together using a clustering algorithm, thus dividing the users into K groups (|S|) and obtaining the user set U = {U1, U2, ..., U...}. j ,…,U K}, U j Let U represent the user group of the j-th category, and U j ={P j,1 ,P j,2 ,…,P j,m ,…,P j,M};P j,m U j The p-th user in the list;

[0020] Step 4: Population Initialization

[0021] Step 4.1: Define the current iteration count as l, the maximum iteration count as L, and let N be the population size. Using real numbers, group the user group U of the j-th category. j Each user's N recommendation results are encoded into an individual of length T, where each decision variable of the individual represents the index of the recommended item; thus, a population is formed from the N recommendation results of a user. Let the m-th user P... j,m Let the N recommendation results be denoted as the m-th population of the l-th generation, then the user group U of the j-th category... j All the first generation populations are denoted as

[0022] Will U j The m-th user P j,m The nth recommendation result is denoted as the nth individual in the mth population of the lth generation. and This represents the m-th user P in the l-th generation. j,m The index of the t-th recommended item in the n-th recommendation result;

[0023] Step 4.2: Based on the interaction score matrix predicted for each user for each item, start from the m-th user P j,m Candicates j,m T unique item numbers are randomly selected from the list for analysis. Perform initialization;

[0024] Step 4.3: Use equation (1) to obtain the m-th user P of the l-th generation. j,m The nth recommendation result accuracy metrics

[0025]

[0026] In equation (1), This represents the m-th user P. j,m For serial number The rating of the corresponding item;

[0027] Using equation (2), the m-th user P of the l-th generation is obtained. j,m The nth recommendation result Novelty rate index

[0028]

[0029] In equation (2), popular t Indicates serial number The popularity of the corresponding items;

[0030] Using equation (3), the m-th user P of the l-th generation is obtained. j,m The nth recommendation result diversity indicators

[0031]

[0032] In equation (3), label Indicates serial number The corresponding item category label. all Represents the category labels of all items in the user item interaction dataset;

[0033] The l-th generation multi-objective optimization function Maximize is constructed using equation (4).

[0034]

[0035] Step 5: Perform information migration between individuals within the same user group, perform information migration between different user groups, and select the optimal user solution through environmental selection iteration.

[0036] The large-scale user recommendation method based on evolutionary multi-task described in this invention is also characterized in that step 5 includes:

[0037] Step 5.1: Perform information migration between individuals within the same user group:

[0038] Step 5.1.1: Based on equation (4), use the binary tournament selection method from... Select 2×N recommended results to participate in the evolution, and obtain the l-th generation mating pool.

[0039] Step 5.1.2, from the lth generation mating pool Two recommendation results from the lth generation are selected and denoted as follows: and Then, a crossover operation is performed to obtain two l-th generation crossover recommendation results. in, express The index of the t-th recommended item. express The index of the t-th recommended item;

[0040] Step 5.1.3, with probability m P right Perform mutation operation:

[0041] Randomly select a number r′ from {1,2,3…,T}, and then... Candicates (set of candidate items for the corresponding user) j,1 Randomly select one with The index of the r′-th recommended item By replacing different item numbers, we can obtain... Recommendation results of the first generation variant

[0042] Step 5.1.4: For the first generation mating pool... After performing crossover and mutation operations on all recommended results according to steps 5.1.2-5.1.3, the mutation recommendation results of all populations in generation l are obtained. Will and User group U merged into category j jThe l-th generation M new population is obtained, and the fitness value of each recommendation result in the l-th generation M new population is measured by Equation (4). Thus, environmental selection is carried out on the l-th generation M new population through non-dominated ranking and crowding distance, and the best N recommendation results are retained as the m-th population of the l+1 generation.

[0043] Step 5.2: Perform inter-population information migration between different user groups:

[0044] If the m-th user of the l-th generation is P j,m If more than half of the recommendation results remain unchanged across multiple generations, then calculate the user group U for the j-th category. j The similarity between user groups and other group categories is used to select the user group with the highest similarity for pairing with the m-th user P in the l-th generation. j,m Crossover and mutation operations are performed among all the recommended results to obtain the m-th population of the (l+1)th generation, following the process in step 5.1.4.

[0045] Step 5.3: After assigning l+1 to l, determine whether l reaches L. If not, return to step 4.3 and execute sequentially; otherwise, randomly select an individual from the m-th population of the L-th generation as the user group U of the j-th category. j The m-th user P j,m One of the recommended results.

[0046] Step 5.1.2 includes:

[0047] Step a, Judgment and If they belong to the same user, proceed to step b; otherwise, proceed to step c.

[0048] Step b, for and With probability c P Perform cross operations:

[0049] Randomly select a number r from {1,2,3…,T}, and... The first r position and By swapping the first r positions, we obtain two cross-recommendation results for the l-th generation. and in, express The index of the r-th recommended item. express The index of the r-th recommended item;

[0050] Step c, will and After merging, a new generation l recommendation result is obtained. Using equation (1) as the fitness value, a binary bidding method is used to select the corresponding indices of T items from the new generation l recommendation result and form two generation l cross recommendation results accordingly. by As The generated cross-recommendation results are as follows: As The resulting cross-recommendation results.

[0051] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing any of the large-scale user recommendation methods, and the processor is configured to execute the program stored in the memory.

[0052] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of any of the large-scale user recommendation methods.

[0053] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0054] 1. This invention is the first to apply evolutionary multitasking to the optimization of multi-objective recommender systems. Traditional optimization methods can only generate optimized recommendation results for one user at a time, without considering the correlation between different tasks. This invention can generate optimal recommendation results for multiple users simultaneously in a single optimization process, greatly reducing the computational load of the recommender algorithm.

[0055] 2. To avoid the negative transfer that may result from simultaneous optimization, this invention uses a clustering algorithm to group multiple users with similar interests in the data into a user group in the optimization method. At the same time, a new operator is designed to speed up the convergence of the algorithm, thereby improving the speed of recommendation.

[0056] 3. This invention reduces computation, improves recommendation efficiency, and ensures recommendation accuracy by designing population optimization schemes, genetic operators, and information transfer strategies for different populations. Attached Figure Description

[0057] Figure 1 This is a flowchart of the method of the present invention;

[0058] Figure 2 This is an example diagram of user clustering in this invention;

[0059] Figure 3 This is an example diagram of population initialization for the present invention;

[0060] Figure 4 This is an example diagram illustrating the generation of the next generation population according to the present invention. Detailed Implementation

[0061] In this embodiment, a large-scale user recommendation method based on evolutionary multi-task is described, such as... Figure 1 As shown, the procedure is as follows:

[0062] Step 1: Obtain relevant user and item data:

[0063] Obtain the user set S = {s1, s2, ..., s} u ,…,s |S|}, where |S| represents the number of users, s u Indicates user number u;

[0064] Obtain the item set Q = {q1, q2, ..., q} i ,…,q |Q|}, where |Q| represents the quantity of items, q i This represents the i-th item;

[0065] Obtain the user-item interaction dataset, and use the interaction data of each user and item as a positive sample, and randomly collect items that the user has not interacted with as negative samples;

[0066] Step 2: Obtain the user rating matrix for items through the prediction model. The prediction model includes: encoding layer, fully connected layer, attention interaction layer, interaction output layer, and prediction layer.

[0067] Step 2.1: The encoding layer processes user s with ID u. u and the i-th item q i Perform one-hot encoding to obtain user s with ID u. u sparse vector θ u and item q i sparse vectors After mapping the two sparse vectors to E-dimensional vectors respectively, and then through the mapping of a fully connected layer, user s with ID u is obtained. u The representation vector p u and the i-th item q i The representation vector o i ;

[0068] Step 2.2, Attention Interaction Layer on p u With o i Process the data and output the attention vector a. u,i ;

[0069] Step 2.3, the interactive output layer will p u With o i After concatenation, it is then combined with the attention vector a u,i After performing the dot product, we get user s with ID u. u With the i-th item q iThe interaction vector f u,i ;

[0070] Step 2.4: The prediction layer processes the interaction vector f u,i After performing multi-layer fully connected processing, the predicted interaction score r is output. u,i ;

[0071] Step 2.5: Using the mean squared error of the predicted interaction score as the loss function, the Adam algorithm is used to optimize the prediction model until the maximum number of iterations is reached, thereby obtaining the optimal prediction model and its output interaction score matrix for each user for each item.

[0072] Step 3: Let the number of clusters be K, and top represent the number of candidate items. Select user s with ID u. u The top 10 preferred items are used as the candidate item set (Candicates). u The similarity between users is calculated by counting the number of identical items in each user's candidate item set. Users with similarity are grouped into one category using a clustering algorithm, thus |S| divides the users into K categories. In this embodiment, K = |S| / 10, resulting in a user set U = {U1, U2, ..., U...}. j ,…,U K}, U j Let U represent the user group of the j-th category, and U j ={P j,1 ,P j,2 ,…,P j,m ,…,P j,M};P j,m U j The p-th user in the list; such as Figure 2 For example, suppose there are candidate item sets for 4 users s1, s2, s3, and s4. By counting the number of identical items among different users, the similarity between s1 and s2 is 7, and the similarity between s3 and s4 is 7. Therefore, the final clustering result is that s1 and s2 belong to class U1, and s3 and s4 belong to class U2.

[0073] Step 4: Population Initialization

[0074] Step 4.1: Define the current iteration count as l and the maximum iteration count as L. In this embodiment, L = 100. Let N be the number of individuals in the population. In this embodiment, N = 10. Use real numbers to represent the user group U of the j-th category. j Each user's N recommendation results are encoded into an individual of length T. In this embodiment, T = 10, and each decision variable of the individual represents the sequence number of the recommended item; thus, a population is formed by the N recommendation results of a user. Let the m-th user P... j,mLet the N recommendation results be denoted as the m-th population of the l-th generation, then the user group U of the j-th category... j All the first generation populations are denoted as

[0075] Will U j The m-th user P j,m The nth recommendation result is denoted as the nth individual in the mth population of the lth generation. and This represents the m-th user P in the l-th generation. j,m The index of the t-th recommended item in the n-th recommendation result. The population individuals are composed of the recommendation results of all users in this class, and are encoded as a matrix. Figure 3 The example shown is a population of individuals with 2 users and a recommendation length of 10.

[0076] Step 4.2: Based on the interaction score matrix predicted for each user for each item, start from the m-th user P j,m Candicates j,m T unique item numbers are randomly selected from the list for analysis. Perform initialization;

[0077] Step 4.3: Use equation (1) to obtain the m-th user P of the l-th generation. j,m The nth recommendation result accuracy metrics

[0078]

[0079] In equation (1), This represents the m-th user P. j,m right Item rating;

[0080] Using equation (2), the m-th user P of the l-th generation is obtained. j,m The nth recommendation result Novelty rate index

[0081]

[0082] In equation (2), popular t express The popularity of an item;

[0083] Using equation (3), the m-th user P of the l-th generation is obtained. j,m The nth recommendation result diversity indicators

[0084]

[0085] In equation (3), label express Item category label all Represents the category labels of all items in the user item interaction dataset;

[0086] The l-th generation multi-objective optimization function Maximize is constructed using equation (4).

[0087]

[0088] Step 5: Perform information migration between individuals within the same user group, perform information migration between different user groups, and select the optimal user solution through environmental selection iteration.

[0089] Step 5.1: Perform information migration between individuals within the same user group:

[0090] Step 5.1.1: Based on equation (4), use the binary tournament selection method from... Select 2×N recommended results to participate in the evolution, and obtain the l-th generation mating pool.

[0091] Step 5.1.2, from the lth generation mating pool Two recommendation results from the lth generation are selected and denoted as follows: and And perform cross operations, express The index of the t-th recommended item. express The index of the t-th recommended item:

[0092] Step a, Judgment and If they belong to the same user, proceed to step b; otherwise, proceed to step c.

[0093] Step b, for and With probability c P Perform cross operations:

[0094] Randomly select a number r from {1,2,3…,T}, and... The first r position and By swapping the first r positions, we obtain two cross-recommendation results for the l-th generation. and in, express The index of the r-th recommended item. express The index of the r-th recommended item;

[0095] Step c, will and After merging, a new generation l recommendation result is obtained. Using equation (1) as the fitness value, a binary bidding method is used to select the corresponding indices of T items from the new generation l recommendation result and form two generation l cross recommendation results accordingly. by As The generated cross-recommendation results are as follows: As The resulting cross-recommendation results;

[0096] Step 5.1.3, with probability m P right Perform mutation operation:

[0097] Randomly select a number r′ from {1,2,3…,T}, and then... Candicates (set of candidate items for the corresponding user) j,1 Randomly select one with The index of the r′-th recommended item By replacing different item numbers, we can obtain... Recommendation results of the first generation variant

[0098] Figure 4 An example is given to illustrate the specific operation process of the crossover and mutation operators. The user group in the example includes two users, denoted as s1 and s2, and X1 and X2 are two recommendation results for s1 and s2. The length of each user's recommendation result is 10, and the probability is c. P ,m PThe values ​​are 0.5 and 0.5 respectively. First, a crossover operation is performed. Assuming the generated random number is 0.3, then a crossover operation is performed on X1 and X2. For X1, a number is randomly selected from {1,2,...,10}, let's say it's 3. Then, crossover (1,23,15,9,5,12,4,18,22,14) and (18,2,16,8,20,24,4,25,17,30) to get (1,23,15,8,20,24,4,25,17,30) and (18,2,16,9,5,12,4,18,22,14). Similarly, suppose X2 randomly selects a number, say 4. The numbers (10, 2, 7, 6, 5, 3, 11, 13, 17, 19) and (10, 2, 15, 6, 21, 12, 11, 27, 28, 29) are then crossed to obtain (10, 2, 7, 6, 21, 12, 11, 27, 28, 29) and (10, 2, 15, 6, 5, 3, 11, 13, 17, 19), resulting in two children. Next, the mutation operator is executed for each child. Assuming the generated random number is 0.4, then for Perform a mutation operation, randomly selecting a number from {1, 2, ..., 10}, let's say it's 4. Then, select the candidate item set (Candicates) from user s1. s1 Choose a new item 6 to replace the original item 8. Similarly, perform a similar operation on s2, and finally... After obtaining the mutation operator These are two new recommendation results.

[0099] Step 5.1.4: For the first generation mating pool... After performing crossover and mutation operations on all recommended results according to steps 5.1.2-5.1.3, the mutation recommendation results of all populations in generation l are obtained. Will and User group U merged into category j j The l-th generation M new population is obtained, and the fitness value of each recommendation result in the l-th generation M new population is measured by Equation (4). Thus, environmental selection is carried out on the l-th generation M new population through non-dominated ranking and crowding distance, and the best N recommendation results are retained as the m-th population of the l+1 generation.

[0100] Step 5.2: Perform inter-population information migration between different user groups:

[0101] If the m-th user of the l-th generation is P j,m If more than half of the recommendation results remain unchanged across multiple generations, then calculate the user group U for the j-th category. jThe similarity between user groups and other group categories is used to select the user group with the highest similarity for pairing with the m-th user P in the l-th generation. j,m Crossover and mutation operations are performed among all the recommended results to obtain the m-th population of the (l+1)th generation, following the process in step 5.1.4.

[0102] Step 5.3: After assigning l+1 to l, determine whether l reaches L. If not, return to step 4.3 and execute sequentially; otherwise, randomly select an individual from the m-th population of the L-th generation as the user group U of the j-th category. j The m-th user P j,m One of the recommended results.

[0103] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0104] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A large-scale user recommendation method based on evolutionary multi-task, characterized in that, The procedure is as follows: Step 1: Obtain relevant user and item data: Get user set ,in, Indicates the number of users. express User number; Acquire Item Set ,in, Indicates the quantity of items. Indicates the first One item; Obtain the user-item interaction dataset, and use the interaction data of each user and item as a positive sample, and randomly collect items that the user has not interacted with as negative samples; Step 2: Obtain the user rating matrix for items through a prediction model, which includes: an encoding layer, a fully connected layer, an attention interaction layer, an interaction output layer, and a prediction layer; Step 2.1, the coding layer pair User No. and the i-th item Perform one-hot encoding to obtain User No. sparse vectors and items sparse vectors Then the two sparse vectors are mapped to... After the dimensional vector is obtained, it is then mapped through the fully connected layer to get... User No. representation vector and the i-th item representation vector ; Step 2.2, the attention interaction layer... and Process the data and output the attention vector. ; Step 2.3, the interactive output layer will and After concatenation, then with the attention vector After performing the dot product, we get User No. With the i-th item Interaction vectors ; Step 2.4: The prediction layer processes the interaction vector. After performing multi-layer fully connected processing, the predicted interaction score is output. ; Step 2.5: Using the mean squared error of the predicted interaction score as the loss function, the Adam algorithm is used to optimize the prediction model until the maximum number of iterations is reached, thereby obtaining the optimal prediction model and its output interaction score matrix for each user for each item. Step 3: Let the number of clusters be... , Indicates the number of candidate items, selected. User No. The former A set of preferred items as candidate items The similarity between users is calculated by counting the number of identical items in each user's candidate item set and using this count as the similarity score. Users with similar scores are grouped together using a clustering algorithm, thus... Users are divided into Group categories and obtain user sets. , Indicates the first User groups of group type, and ; express The first in One user; Step 4: Population Initialization Step 4.1: Define the current iteration number as... The maximum number of iterations is ,make The population size is represented by the real number system. User groups of group categories Each user Each recommendation result is encoded into a single string of length [length missing]. Each individual is a decision variable representing the sequence number of the recommended item; thus, a user's decision is determined by... The n recommendation results form a population, let the nth recommendation result be a population. individual users of The nth recommendation result is denoted as the i-th recommendation result. The generation The first population, then the second... User groups of group categories All of the first Generation population is recorded as ; Will The Middle individual users The The nth recommendation result is denoted as the i-th recommendation result. The generation The first in each population individual ,and ; Indicates the first The generation individual users The The index of the t-th recommended item in the recommended results; Step 4.2: Based on the interaction score matrix predicted for each user for each item, starting from the... individual users Candidate item set Random selection The serial number of each unique item is used to... Perform initialization; Step 4.3: Use equation (1) to obtain the first... The generation individual users The Recommended results accuracy metrics : (1) In equation (1), Indicates the first individual users For serial number The rating of the corresponding item; Using equation (2) to obtain the first The generation individual users The Recommended results Novelty rate index : (2) In equation (2), Indicates serial number The popularity of the corresponding items; Using equation (3) to obtain the first The generation individual users The Recommended results diversity indicators : (3) In equation (3), Indicates serial number The corresponding item category label, Represents the category labels of all items in the user item interaction dataset; Construct the first using equation (4) Alternate multi-objective optimization function : (4) Step 5: Perform information migration between individuals within the same user group, perform information migration between different user groups, and select the optimal user solution through environmental selection iteration. Step 5.1: Perform information migration between individuals within the same user group: Step 5.1.1: Based on equation (4), use the binary bidding method to select from... Select The recommended result participates in the evolution, resulting in the... Mating pool ; Step 5.1.2, from the first Mating pool Selected from Let the two recommendation results be denoted as follows: and And perform a crossover operation to obtain two... Cross-recommendation results , ;in, express The index of the t-th recommended item. express The index of the t-th recommended item; Step 5.1.3, using probability right Perform mutation operation: Random from Choose one number ,from The set of candidate items corresponding to the user Randomly select one with The Middle The serial number of each recommended item By replacing different item numbers, we can obtain... The Recommendation results of generational mutations ; Step 5.1.4, for the first Mating pool After performing crossover and mutation operations on all recommended results according to steps 5.1.2-5.1.3, we obtain the [number]th [result]. Recommended variation results for all populations ,Will and merged into the first User groups of group categories The generation A new population, and the first population is measured by equation (4). generation The fitness value of each recommendation result in the new population is used to evaluate the fitness of the first recommendation result by non-dominated ranking and crowding distance. generation A new population undergoes environmental selection, retaining the optimal [species / species]. The recommendation result is used as the first recommendation result. The generation Individual populations; Step 5.2: Perform inter-population information migration between different user groups: If the first The generation individual users If more than half of the recommended results remain unchanged across multiple generations, then the calculation of the [number]th generation... User groups of group categories The similarity between user groups and other group categories is used to select the user group with the highest similarity for comparison with the first category. The generation individual users Crossover and mutation operations are performed among all the recommended results to obtain the result according to step 5.1.

4. The generation Individual populations; Step 5.3, Assign to Then, make a judgment Has it been achieved? If the condition is not met, return to step 4.3 and execute sequentially; otherwise, start from generation L. One individual is randomly selected from each population as the first... User groups of group categories The Middle individual users One of the recommended results.

2. The large-scale user recommendation method based on evolutionary multi-task as described in claim 1, characterized in that, Step 5.1.2 includes: Step a, Judgment and If they belong to the same user, proceed to step b; otherwise, proceed to step c. Step b, for and With probability Perform cross operations: Random from Choose one number ,Will The former Position and The former By swapping the bits, we obtain the th... Two cross-recommendation results and ;in, express The index of the r-th recommended item. express The index of the r-th recommended item; Step c, will and After merging, we get a first The new recommendation results are presented in a format that... For fitness values, a binary bidding method is used from the first... In the new recommendation results, select the serial numbers corresponding to T items and form two corresponding numbers. Cross-recommendation results , ,by As The generated cross-recommendation results are as follows: As The resulting cross-recommendation results.

3. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the large-scale user recommendation methods of claims 1-2, the processor being configured to execute the programs stored in the memory.

4. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, performs the steps of any of the large-scale user recommendation methods described in claims 1-2.

Images