A hybrid recommendation method, device, electronic device and storage medium

Abstract

The present invention discloses a hybrid recommendation method, device, electronic device and storage medium, which relates to the field of data processing, including: obtaining the behavior data of old users purchasing goods on the Internet and calculating the preference value of the old users for the interacted goods, the heat value of the goods and the similarity between different old users, calculating the preference value of the old users for the non-interacted goods according to the similarity between different old users based on the collaborative filtering idea, and optimizing the preference information of the old users and the goods on this basis; inputting the preference information of the old users and the goods into an implicit feature vector prediction model to obtain the implicit feature vector matrix of the old users and the implicit feature vector matrix of the goods; fixing the implicit feature vector matrix of the goods, optimizing the implicit feature vector matrix of the old users by using a genetic algorithm, and obtaining the optimal implicit feature vector matrix of the old users; generating a recommendation list according to the optimal implicit feature vector of the old users, and solving the problems of data sparsity and cold start.

Description

Technical Field

[0001] The present invention relates to the field of data processing, and in particular to a hybrid recommendation method, device, electronic equipment and storage medium. Background Art

[0002] In the tide of the digital age, e-commerce has become an indispensable part of the business field. In the field of e-commerce recommendation, the collaborative filtering (CF) model is known for its wide popularity, far-reaching influence and wide application. Although collaborative filtering has achieved great success in the field of recommendation systems, its limitations have begun to emerge as the scale of data continues to increase.

[0003] First, collaborative filtering faces the problem of data sparsity, that is, due to incomplete interaction information between users and products, the accuracy and comprehensiveness of recommendation results are affected. Second, the user cold start problem also plagues collaborative filtering models because there is a lack of sufficient historical data for new users to make recommendations. Summary of the invention

[0004] The purpose of this application is to propose a hybrid recommendation method, device, electronic device and storage medium to address the above-mentioned technical problems.

[0005] In a first aspect, the present invention provides a hybrid recommendation method, comprising the following steps:

[0006] Obtain the behavior data of old users who purchase goods online and process the interaction information between old users and goods. Calculate the preference value of old users for the goods they interact with, the popularity value of goods and the similarity between different old users based on the interaction information between old users and goods. Calculate the preference value of old users for goods they do not interact with based on the similarity between different old users based on the collaborative filtering idea, and optimize the preference information between old users and goods based on the preference value of old users for goods they interact with.

[0007] Input the preference information of old users and products into the implicit feature vector prediction model to obtain the implicit feature vector matrix of old users and the implicit feature vector matrix of products. The implicit feature vector matrix of old users is the set of implicit feature vectors of all old users, and the implicit feature vector matrix of products is the set of implicit feature vectors of all products.

[0008] The implicit feature vector matrix of the fixed product is used to optimize the implicit feature vector matrix of the old users by using the genetic algorithm to obtain the optimal implicit feature vector matrix of the old users;

[0009] A recommendation list of evolved old users is generated according to the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, and the heat value of the product.

[0010] Preferably, the interaction information is represented in the following format:

[0011] (User, Item, Act, Timestamp);

[0012] Among them, User represents the user ID, Item represents the product ID, Act represents the quantitative value of the user's interaction behavior on the product, and Timestamp represents the timestamp of the user's interaction behavior on the product.

[0013] As a preferred method, the preference value of the old user for the interacted product, the popularity value of the product and the similarity between different old users are calculated based on the interaction information between the old user and the product. The preference value of the old user for the non-interacted product is calculated based on the similarity between different old users based on the collaborative filtering idea. The preference information of the old user and the product is optimized based on the preference value of the old user for the interacted product. Specifically, it includes:

[0014] The corresponding preference value is calculated based on the interaction information between each old user and the product. The calculation formula is as follows:

[0015] ;

[0016] in, Indicates the preference value corresponding to each interactive information of old user u for the interactive product i; and They respectively represent the latest and earliest timestamps when old user u interacts with the interactive product i; Represents the quantitative value of the interaction behavior of old user u on the interactive product i;

[0017] The comprehensive preference value of old users for the interacted products is calculated based on the preference value corresponding to the transaction information of the interacted products, as shown in the following formula:

[0018] ;

[0019] in, Represents the comprehensive preference value of old user u for the interactive product i;

[0020] The similarity between different old users is calculated based on the comprehensive preference value of the old users for the interacted products, the set of products interacted by the old users, and the popularity value of the products. The calculation formula is as follows:

[0021] ;

[0022] ;

[0023] in, represents the set of all old users, represents the set of old users who have interacted with product g. It means to find the number of elements in a set. Indicates the heat value of product g; and They represent the set of commodities interacted by old user u and the set of commodities interacted by old user v respectively; express and The intersection of represents the comprehensive preference value of old user u for the interactive product g, represents the comprehensive preference value of old user v for the interactive product g, Represents the similarity between old user u and old user v;

[0024] The following formula is used to calculate the preference information of old user u and product j:

[0025] ;

[0026] in, represents the comprehensive preference value of old user u for product j, It means to get the list of nearest neighbor old users of old user u by sorting them from large to small similarity. Represents traversal For each old user d in the table, the set of items interacted by old user u is The collection of products obtained after the update, , represents the set of products interacted by old user d, represents the similarity between old user u and old user d, represents the comprehensive preference value of old user d for product j, It represents the comprehensive preference value of old user u for product j. max() means taking the maximum value of the elements in the brackets.

[0027] As a preferred method, the implicit feature vector matrix of the fixed product is optimized by using a genetic algorithm to obtain the optimal implicit feature vector matrix of the old user, which specifically includes:

[0028] S31: Based on the implicit feature vector of the old user u in the implicit feature vector matrix P of the old user , the implicit feature vector of each product l in the implicit feature vector matrix Q of the product Generate a recommendation list for old users before evolution based on the popularity value of the product , as shown below:

[0029] ;

[0030] Among them, the implicit feature vector matrix of the product output by the implicit feature vector prediction model is expressed as , m represents the total number of commodities, and the latent feature vector matrix of old users output by the latent feature vector prediction model is expressed as , n represents the total number of old users, Represents the implicit feature vector of product l and the implicit feature vector of old user u The inner product of , k represents the heat penalty factor, and its value is [0, 1]. Indicates the popularity value of product l, Indicates that the list is sorted from largest to smallest. Before taking the list elements, I represents the set of all commodities;

[0031] Define fitness as shown below:

[0032] ;

[0033] in, Represents the implicit feature vector of old user u The fitness of Indicates the actual purchase list of old user u. Indicates the recommendation list of old users before evolution The position of product l calculated from the sorting position 1, Indicates the recommendation list of old users before evolution Product l is the product that the old user actually purchases;

[0034] S32: Randomly initialize t candidate implicit feature vectors of old user u and compare them with the implicit feature vector of old user u. The implicit feature vector population of old user u with the number of individuals t+1 is obtained by mixing:

[0035] ;

[0036] in, represents the implicit feature vector population of old user u, represents the candidate latent feature vector of the t-th old user u, that is, the t-th individual in the candidate latent feature vector population of the old user u;

[0037] Set the number of population evolutions N, and evolve the implicit feature vector population of old user u in parallel based on the distributed framework , the evolution rules are as follows:

[0038] 1) Create an initially empty population template ;

[0039] 2) Calculation The fitness of each individual in the population is calculated, and the individuals are sorted from large to small according to their fitness, and the top half of the individuals are selected and put into the population template. ;

[0040] in, Represents the implicit feature vector population of old user u The fitness of each individual in

[0041] 3) According to the principle that the greater the individual's fitness, the greater the probability of being selected, calculate the implicit feature vector population of old user u The probability of each individual being selected , the calculation formula is as follows:

[0042] ;

[0043] 4) Based on the implicit feature vector population of old users u The probability of each individual being selected From the latent feature vector population of old user u Select multiple individuals for crossover evolution, and select the implicit feature vector population of old user u based on the preset probability μ Select multiple individuals for mutation evolution, and put the individuals obtained by cross evolution and mutation evolution into the population template obtained in 2). ;

[0044] 5) Calculate the population template obtained in 4) The fitness of each individual in , and select the top t with the highest fitness as the candidate implicit feature vector population of old user u in the next round of evolution;

[0045] 6) Repeat 2) to 5) until the population converges or reaches the number of population evolutions N. Before each round of evolution, the population template Clear and get the population template from the last evolution The individual corresponding to the maximum value of the fitness of all individuals in is taken as the implicit feature vector of the optimal old user u;

[0046] 7) Repeat 1) to 6) for each old user's implicit feature vector in the old user's implicit feature vector matrix P to obtain the optimal old user's implicit feature vector matrix.

[0047] Preferably, generating a recommendation list of evolved old users according to the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, and the heat value of the products specifically includes:

[0048] The following formula is used to generate the recommendation list of the evolved old user u: :

[0049] ;

[0050] in, Represents the implicit feature vector of product l and the optimal latent feature vector of old user u The inner product of , k represents the heat penalty factor, and its value is [0, 1]. Indicates the popularity value of product l, Indicates that the list is sorted from largest to smallest. Before taking the list elements, and I represents the set of all commodities.

[0051] As a preferred embodiment, the method further includes: obtaining the behavior data of the new user purchasing the product on the network and processing the interaction information between the new user and the product, and performing a cold start based on the interaction information between the new user and the product. The specific steps are as follows:

[0052] S51: Randomly initialize new users The hidden feature vector of the new user The initial hidden eigenvector of , the implicit feature vector of each product in the implicit feature vector matrix of the product and the heat value of the product are fuzzy recommended to generate the initial fuzzy recommendation list for the new user , as shown below:

[0053] ;

[0054] in, Represents the implicit feature vector of product l and new users The initial hidden eigenvector of The inner product of , k represents the heat penalty factor, and its value is [0, 1]. Indicates the popularity value of product l, Indicates that the list is sorted from largest to smallest. Before taking the list elements;

[0055] Taking the initial fuzzy recommendation list as the current fuzzy recommendation list;

[0056] According to the interaction information between the new user and the product, it is determined whether the new user has any interaction behavior with the products in the current fuzzy recommendation list. If so, the current fuzzy recommendation list is used as the new user in the 0th iteration. fuzzy recommendation list; otherwise, select the one behind the current fuzzy recommendation list in the collection of all products in a gradient descent manner according to the heat value. The next fuzzy recommendation list is generated for each item, and the next fuzzy recommendation list is used as the current fuzzy recommendation list and this step is repeated until it is determined that the new user has an interactive behavior with the items in the current fuzzy recommendation list and the current fuzzy recommendation list is used as the new user in the 0th iteration. Create a fuzzy recommendation list that is initially empty Used to save historical interaction information of new users;

[0057] S52: According to the new user New users in the round iteration Fuzzy recommendation list The interactive behavior of the products in the generated new user's interactive information on the products , is a positive integer, based on the new user's interaction information on the product , the user's implicit feature vector matrix P and the product's implicit feature vector matrix Q calculate the new user's new implicit feature vector , as shown below:

[0058] ;

[0059] in, Indicates that based on new users Interactive information about products The first latent eigenvector calculated by the latent eigenvector matrix P of the user; Indicates based on new user Interactive information about products The second latent eigenvector calculated from the latent eigenvector matrix Q of the product; and are the weight of the first latent feature vector and the weight of the second latent feature vector, and the sum of the two is 1. and The calculation formula is as follows:

[0060] ;

[0061] ;

[0062] in, Indicates that among all old users The set of users who interact with the products. represents the implicit feature vector representation of the old user h in the old user implicit feature vector matrix, represents the implicit feature vector of the known product c in the implicit feature vector matrix of the known product, Indicates new user Interactive information about products A collection of products in Indicates new user The similarity between the old user h and Indicates new user The comprehensive preference value for the interacted product c;

[0063] S53: Based on the new implicit feature vector of the new user Create a new user New Recommended List ;

[0064] S54: Determine a new recommendation list for the new user based on the interaction information between the new user and the product Is there any interaction between the products in the list? If so, the new recommendation list for the new user will be As in New users in the round iteration Fuzzy recommendation list Otherwise, select the product that is located at the new user in the collection of all products in a gradient descending manner according to the heat value. New Recommended List Behind The next recommendation list is generated for each product, and the next recommendation list is used as a new recommendation list and step S54 is repeated until a new recommendation list for the new user is determined. There is interaction between the products in the list and the new recommendation list for new users As in New users in the round iteration Fuzzy recommendation list ;

[0065] S55: Based on new users Interactive information about products , confirm that it will be New users in the round iteration The hidden eigenvector of Or the new latent feature vector of the new user Updated to , the calculation formula is as follows:

[0066] ;

[0067] ;

[0068] ;

[0069] in, Indicates that according to New users in the round iteration The hidden eigenvector of Generated recommendation list, Indicates new user The comprehensive preference value for the interactive product a, Indicates new user The comprehensive preference value for the interacted product b;

[0070] S56: Repeat steps S52-S55 until Convergence or reaching the maximum number of iterations M, the new user obtained in the last iteration The latent feature vector of The hidden eigenvector of , and the new user obtained from the last iteration The recommendation list generated by the latent feature vector of Recommended list .

[0071] In a second aspect, the present invention provides a hybrid recommendation device, comprising:

[0072] The data processing module is configured to obtain the behavior data of old users purchasing goods on the Internet and process the interaction information between the old users and the goods, calculate the preference value of the old users for the interacted goods, the popularity value of the goods and the similarity between different old users according to the interaction information between the old users and the goods, calculate the preference value of the old users for the goods that have not been interacted with according to the similarity between different old users based on the collaborative filtering idea, and optimize the preference information between the old users and the goods based on the preference value of the old users for the interacted goods;

[0073] A prediction module is configured to input the preference information of old users and commodities into a latent feature vector prediction model to obtain an latent feature vector matrix of old users and an latent feature vector matrix of commodities, wherein the latent feature vector matrix of old users is a set of latent feature vectors of all old users, and the latent feature vector matrix of commodities is a set of latent feature vectors of all commodities;

[0074] The optimization module is configured to fix the implicit feature vector matrix of the commodity, and uses a genetic algorithm to optimize the implicit feature vector matrix of the old users to obtain the optimal implicit feature vector matrix of the old users;

[0075] The list generation module is configured to generate a recommendation list of evolved old users based on the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, and the heat value of the product.

[0076] In a third aspect, the present invention provides an electronic device comprising one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner in the first aspect.

[0077] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any implementation manner in the first aspect.

[0078] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in any implementation manner in the first aspect.

[0079] Compared with the prior art, the present invention has the following beneficial effects:

[0080] (1) The hybrid recommendation method proposed in the present invention is based on the collaborative filtering idea and calculates the preference value of old users for non-interacted commodities according to the similarity between different old users, thereby enriching the interaction data between users and commodities. In terms of processing data sparsity, local information and global information are comprehensively considered, and local prediction technology is used to improve the accuracy of global recommendation. Genetic algorithm is used to optimize the implicit feature vectors of old users, and the optimal implicit feature vectors of old users are used to generate a recommendation list, so the recommendation accuracy is higher.

[0081] (2) The hybrid recommendation method proposed in the present invention makes reasonable use of the implicit feature vectors of old users and products in dealing with the cold start of new users, and adopts iterative update technology to quickly obtain the implicit feature vectors of new users.

[0082] (3) The hybrid recommendation method proposed in this invention is of great significance for improving the accuracy of the recommendation system and meeting user needs, and can provide enterprises with a competitive advantage in the highly competitive market. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments 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.

[0084] Figure 1 A schematic flow chart of a hybrid recommendation method according to an embodiment of the present application;

[0085] Figure 2A schematic diagram of a hybrid recommendation device according to an embodiment of the present application;

[0086] Figure 3 A schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0088] Figure 1 A hybrid recommendation method provided by an embodiment of the present application is shown, comprising the following steps:

[0089] S1, obtain the behavioral data of old users purchasing goods on the Internet and process the interaction information between old users and goods. According to the interaction information between old users and goods, calculate the preference value of old users for interacted goods, the popularity value of goods and the similarity between different old users. Based on the collaborative filtering idea, calculate the preference value of old users for non-interacted goods according to the similarity between different old users, and optimize the preference information of old users and goods based on the preference value of old users for interacted goods.

[0090] In a specific embodiment, the interaction information is represented in the following format:

[0091] (User, Item, Act, Timestamp);

[0092] Among them, User represents the user ID, Item represents the product ID, Act represents the quantitative value of the user's interactive behavior on the product, and Timestamp represents the timestamp of the user's interactive behavior on the product.

[0093] Specifically, the behavior data of old users are collected from the website log of the e-commerce platform. The behavior data records the interaction data of old users on a certain product. The interaction data includes interaction behavior, interaction frequency and interaction time. The behavior data recorded in the website log is processed to obtain each piece of interaction information. The interaction information contains the quantitative value of the user's interaction behavior on the product. The interaction behavior includes click behavior, collection behavior and purchase behavior. In one embodiment, the quantitative value corresponding to the click behavior is 1, the quantitative value corresponding to the collection behavior is 2, and the quantitative value corresponding to the purchase behavior is 3. In other words, the quantitative value of the click behavior that least reflects the user's preference for the product is set to 1, followed by the quantitative value of the collection behavior is 2, and the quantitative value of the purchase behavior that best reflects the user's preference for the product is 3. In other embodiments, quantitative values ​​of different sizes can also be set according to specific scenarios.

[0094] In a specific embodiment, the preference value of the old user for the interacted product, the popularity value of the product and the similarity between different old users are calculated based on the interaction information between the old user and the product. The preference value of the old user for the non-interacted product is calculated based on the similarity between different old users based on the collaborative filtering idea. The preference information of the old user and the product is optimized based on the preference value of the old user for the interacted product. Specifically, it includes:

[0095] The corresponding preference value is calculated based on the interaction information between each old user and the product. The calculation formula is as follows:

[0096] ;

[0097] in, Indicates the preference value corresponding to each interactive information of old user u for the interactive product i; and They respectively represent the latest and earliest timestamps when old user u interacts with the interactive product i; Represents the quantitative value of the interaction behavior of old user u on the interactive product i;

[0098] The comprehensive preference value of old users for the interacted products is calculated based on the preference value corresponding to the transaction information of the interacted products, as shown in the following formula:

[0099] ;

[0100] in, Represents the comprehensive preference value of old user u for the interactive product i;

[0101] The similarity between different old users is calculated based on the comprehensive preference value of the old users for the interacted products, the set of products interacted by the old users, and the popularity value of the products. The calculation formula is as follows:

[0102] ;

[0103] ;

[0104] in, represents the set of all old users, represents the set of old users who have interacted with product g. It means to find the number of elements in a set. Indicates the heat value of product g; and They represent the set of commodities interacted by old user u and the set of commodities interacted by old user v respectively; express and The intersection of represents the comprehensive preference value of old user u for the interactive product g, represents the comprehensive preference value of old user v for the interactive product g, Represents the similarity between old user u and old user v;

[0105] The following formula is used to calculate the preference information of old user u and product j:

[0106] ;

[0107] in, represents the comprehensive preference value of old user u for product j, It means to get the list of nearest neighbor old users of old user u by sorting them from large to small similarity. Represents traversal For each old user d in the table, the set of items interacted by old user u is The collection of products obtained after the update, , represents the set of products interacted by old user d, represents the similarity between old user u and old user d, represents the comprehensive preference value of old user d for product j, It represents the comprehensive preference value of old user u for product j. max() means taking the maximum value of the elements in the brackets.

[0108] Specifically, an old user has multiple interactions with a certain product, so there are also multiple interaction information between the old user and a certain product; the old user's real-time preference value for the product is estimated through implicit feedback data, and each interaction information between the old user and a certain product can calculate the corresponding preference value. Therefore, the old user has multiple preference values ​​for a certain product. The multiple preference values ​​are summed to calculate the comprehensive preference value of the old user and a certain product. The similarity of two different old users is calculated based on the comprehensive preference value of two different old users for a certain product, the heat value of a certain product, and the set of products interacted by two different old users. There are two core ideas: first, the more common products two different old users interact with, the higher the similarity between the two; second, among the common products interacted by two different old users, the lower the heat value of the product can reflect the difference between the two different old users. According to the similarity between the old user u and other old users in the set of all users, the nearest neighbor list of the old user u is determined. The number of old users in the nearest neighbor list is set according to the actual situation and is not limited here. Adopting the user-based collaborative filtering idea, the comprehensive preference value of users for certain non-interacted products is calculated based on the similarity values ​​between different old users. In other words, the comprehensive preference value of old user u for non-interacted products is predicted based on the nearest neighbor list of old user u, and the comprehensive preference value of old user u for interacted products and the comprehensive preference value of old user u for non-interacted products are combined to obtain the preference information of old user u and products. Therefore, the interaction information between users and products is enriched and data sparsity is alleviated.

[0109] In the calculation formula of the preference information of old user u and product, and each time a All old users v in the to update.

[0110] S2, input the preference information of old users and products into the implicit feature vector prediction model to obtain the implicit feature vector matrix of old users and the implicit feature vector matrix of products. The implicit feature vector matrix of old users is the set of implicit feature vectors of all old users, and the implicit feature vector matrix of products is the set of implicit feature vectors of all products.

[0111] Specifically, the input of the latent feature vector prediction model is the preference information of old users and commodities, and the output is the latent feature vector of old users and the latent feature vector of commodities, and the latent feature vector matrix of old users and the latent feature vector matrix of commodities are formed. In one embodiment, the ALS model of the distributed computing framework of the Apache Spark platform is trained and optimized to obtain the latent feature vector prediction model; in other embodiments, other existing network structures can also be used, which is not limited here.

[0112] S3, the implicit feature vector matrix of the fixed product, uses the genetic algorithm to optimize the implicit feature vector matrix of the old users, and obtains the optimal implicit feature vector matrix of the old users.

[0113] In a specific embodiment, step S3 specifically includes:

[0114] S31: Based on the implicit feature vector of the old user u in the implicit feature vector matrix P of the old user , the implicit feature vector of each product l in the implicit feature vector matrix Q of the product Generate a recommendation list for old users before evolution based on the popularity value of the product , as shown below:

[0115] ;

[0116] Among them, the implicit feature vector matrix of the product output by the implicit feature vector prediction model is expressed as , m represents the total number of commodities, and the latent feature vector matrix of old users output by the latent feature vector prediction model is expressed as , n represents the total number of old users, Represents the implicit feature vector of product l and the implicit feature vector of old user u The inner product of , k represents the heat penalty factor, and its value is [0, 1]. Indicates the popularity value of product l, Indicates that the list is sorted from largest to smallest. Before taking the list elements, I represents the set of all commodities;

[0117] Define fitness as shown below:

[0118] ;

[0119] in, Represents the implicit feature vector of old user u The fitness of Indicates the actual purchase list of old user u. Indicates the recommendation list of old users before evolution The position of product l calculated from the sorting position 1, Indicates the recommendation list of old users before evolution Product l is the product that the old user actually purchases;

[0120] S32: Randomly initialize t candidate implicit feature vectors of old user u and compare them with the implicit feature vector of old user u. The implicit feature vector population of old user u with the number of individuals t+1 is obtained by mixing:

[0121] ;

[0122] in, represents the implicit feature vector population of old user u, represents the candidate latent feature vector of the t-th old user u, that is, the t-th individual in the candidate latent feature vector population of the old user u;

[0123] Set the number of population evolutions N, and evolve the implicit feature vector population of old user u in parallel based on the distributed framework , the evolution rules are as follows:

[0124] 1) Create an initially empty population template ;

[0125] 2) Calculation The fitness of each individual in the population is calculated, and the individuals are sorted from large to small according to their fitness, and the top half of the individuals are selected and put into the population template. ;

[0126] in, Represents the implicit feature vector population of old user u The fitness of each individual in

[0127] 3) According to the principle that the greater the individual's fitness, the greater the probability of being selected, calculate the implicit feature vector population of old user u The probability of each individual being selected , the calculation formula is as follows:

[0128] ;

[0129] 4) Based on the implicit feature vector population of old users u The probability of each individual being selected From the latent feature vector population of old user u Select multiple individuals for crossover evolution, and select the implicit feature vector population of old user u based on the preset probability μ Select multiple individuals for mutation evolution, and put the individuals obtained by cross evolution and mutation evolution into the population template obtained in 2). ;

[0130] 5) Calculate the population template obtained in 4) The fitness of each individual in , and select the top t with the highest fitness as the candidate implicit feature vector population of old user u in the next round of evolution;

[0131] 6) Repeat 2) to 5) until the population converges or reaches the number of population evolutions N. Before each round of evolution, the population template Clear and get the population template from the last evolution The individual corresponding to the maximum value of the fitness of all individuals in is taken as the implicit feature vector of the optimal old user u;

[0132] 7) Repeat 1) to 6) for each old user's implicit feature vector in the old user's implicit feature vector matrix P to obtain the optimal old user's implicit feature vector matrix.

[0133] Specifically, for each old user, the old user's implicit feature vector output by the implicit feature vector prediction model is mixed with a plurality of randomly generated t candidate implicit feature vectors of old users u to form a population, thereby accelerating the convergence speed of the genetic algorithm. The implicit feature vector of the old user is optimized by the genetic algorithm to obtain the optimal implicit feature vector of the old user.

[0134] S4, generating a recommendation list of evolved old users according to the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, and the heat value of the product.

[0135] In a specific embodiment, step S4 specifically includes:

[0136] The following formula is used to generate the recommendation list of the evolved old user u: :

[0137] ;

[0138] in, Represents the implicit feature vector of product l and the optimal latent feature vector of old user u The inner product of , k represents the heat penalty factor, and its value is [0, 1]. Indicates the popularity value of product l, Indicates that the list is sorted from largest to smallest. Before taking the list elements, and I represents the set of all commodities.

[0139] Specifically, based on the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, a recommendation list can be generated through a recommendation formula.

[0140] In a specific embodiment, it also includes: obtaining the behavior data of the new user purchasing goods on the network and processing to obtain the interaction information between the new user and the goods, and performing a cold start based on the interaction information between the new user and the goods. The specific steps are as follows:

[0141] S51: Randomly initialize new users The hidden feature vector of the new user The initial hidden eigenvector of , the implicit feature vector of each product in the implicit feature vector matrix of the product and the heat value of the product are fuzzy recommended to generate the initial fuzzy recommendation list for the new user , as shown below:

[0142] ;

[0143] in, Represents the implicit feature vector of product l and new users The initial hidden eigenvector of The inner product of , k represents the heat penalty factor, and its value is [0, 1]. Indicates the popularity value of product l, Indicates that the list is sorted from largest to smallest. Before taking the list elements;

[0144] Taking the initial fuzzy recommendation list as the current fuzzy recommendation list;

[0145] According to the interaction information between the new user and the product, it is determined whether the new user has any interaction behavior with the products in the current fuzzy recommendation list. If so, the current fuzzy recommendation list is used as the new user in the 0th iteration. fuzzy recommendation list; otherwise, select the one behind the current fuzzy recommendation list in the collection of all products in a gradient descent manner according to the heat value. The next fuzzy recommendation list is generated for each item, and the next fuzzy recommendation list is used as the current fuzzy recommendation list and this step is repeated until it is determined that the new user has an interactive behavior with the items in the current fuzzy recommendation list and the current fuzzy recommendation list is used as the new user in the 0th iteration. Create a fuzzy recommendation list that is initially empty Used to save historical interaction information of new users;

[0146] S52: According to the new user New users in the round iteration Fuzzy recommendation list The interactive behavior of the products in the generated new user's interactive information on the products , is a positive integer, based on the new user's interaction information on the product , the user's implicit feature vector matrix P and the product's implicit feature vector matrix Q calculate the new user's new implicit feature vector , as shown below:

[0147] ;

[0148] in, Indicates based on new user Interactive information about products The first latent eigenvector calculated by the latent eigenvector matrix P of the user; Indicates based on new user Interactive information about products The second latent eigenvector calculated from the latent eigenvector matrix Q of the product; and are the weight of the first latent feature vector and the weight of the second latent feature vector, and the sum of the two is 1. and The calculation formula is as follows:

[0149] ;

[0150] ;

[0151] in, Indicates that among all old users The set of users who interact with the products. represents the implicit feature vector representation of the old user h in the old user implicit feature vector matrix, represents the implicit feature vector of the known product c in the implicit feature vector matrix of the known product, Indicates new user Interactive information about products A collection of products in Indicates new user The similarity between the old user h and Indicates new user The comprehensive preference value for the interacted product c;

[0152] S53: Based on the new implicit feature vector of the new user Create a new user New Recommended List ;

[0153] S54: Determine a new recommendation list for the new user based on the interaction information between the new user and the product Is there any interaction between the products in the list? If so, the new recommendation list for the new user will be As in New users in the round iteration Fuzzy recommendation list Otherwise, select the product that is located at the new user in the collection of all products in a gradient descending manner according to the heat value. New Recommended List Behind The next recommendation list is generated for each product, and the next recommendation list is used as a new recommendation list and step S54 is repeated until a new recommendation list for the new user is determined. There is interaction between the products in the list and the new recommendation list for new users As in New users in the round iteration Fuzzy recommendation list ;

[0154] S55: Based on new users Interactive information about products , confirm that it will be New users in the round iteration The hidden eigenvector of Or the new latent feature vector of the new user Updated to , the calculation formula is as follows:

[0155] ;

[0156] ;

[0157] ;

[0158] in, Indicates that according to New users in the round iteration The hidden eigenvector of Generated recommendation list, Indicates new user The comprehensive preference value for the interactive product a, Indicates new user The comprehensive preference value for the interacted product b;

[0159] S56: Repeat steps S52-S55 until Convergence or reaching the maximum number of iterations M, the new user obtained in the last iteration The latent feature vector of The hidden eigenvector of , and the new user obtained from the last iteration The recommendation list generated by the latent feature vector of Recommended list .

[0160] Specifically, old users have behavioral data from the beginning, while new users do not have behavioral data at the beginning. Behavioral data is obtained during the recommendation process, and the implicit feature vector of the new user is updated while the behavioral data is obtained. For new users, since they do not have enough behavioral data, the above steps S1 and S2 cannot be used to obtain the corresponding implicit feature vector, which leads to a cold start problem. For all old users, the implicit feature vectors of all the optimal old users constitute the optimal implicit feature vector matrix of old users. An initial fuzzy recommendation list is generated for new users based on the heat value of the product. If the user does not interact with the products in this initial recommendation list, the fuzzy recommendation will continue in a gradient descending manner according to the heat value until the user interacts with the products in the recommendation list.

[0161] The principle of selecting products by descending heat value gradient is as follows: Assume that the heat values ​​of products are ranked from large to small as a1, b1, c1, d1, e1, f1, and the initial recommendation list is a1, b1. If the new user has no interaction with the products in the initial recommendation list, the next recommendation list is generated in descending heat value: c1, d1, until the new user interacts with the generated recommendation list. Whether an interaction occurs can be judged based on the interaction information obtained by processing the behavior data of the new user. The data format of the interaction information between the old user and the product is (User, Item, Act, Timestamp). If there is an interaction, the product ID in the interaction information between the old user and the product can be matched in the recommendation list.

[0162] Further references Figure 2 As an implementation of the methods shown in the above figures, the present application provides an embodiment of a hybrid recommendation device. Figure 1 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0163] The embodiment of the present application provides a hybrid recommendation device, including:

[0164] The data processing module 1 is configured to obtain the behavior data of the old user purchasing goods on the Internet and process the interaction information between the old user and the goods, calculate the preference value of the old user for the interacted goods, the popularity value of the goods and the similarity between different old users according to the interaction information between the old user and the goods, calculate the preference value of the old user for the goods that have not been interacted with according to the similarity between different old users based on the collaborative filtering idea, and optimize the preference information between the old user and the goods based on the preference value of the old user for the interacted goods;

[0165] Prediction module 2 is configured to input the preference information of old users and commodities into the implicit feature vector prediction model to obtain the implicit feature vector matrix of old users and the implicit feature vector matrix of commodities, wherein the implicit feature vector matrix of old users is the set of implicit feature vectors of all old users, and the implicit feature vector matrix of commodities is the set of implicit feature vectors of all commodities;

[0166] The optimization module 3 is configured to fix the implicit feature vector matrix of the commodity, and optimizes the implicit feature vector matrix of the old user by using a genetic algorithm to obtain the optimal implicit feature vector matrix of the old user;

[0167] The list generation module 4 is configured to generate a recommendation list of evolved old users based on the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, and the heat value of the product.

[0168] Figure 3 Schematic diagram of the hardware structure of the electronic device provided by the embodiment of the present invention. Figure 3 As shown, the electronic device of this embodiment includes: a processor 301 and a memory 302; wherein the memory 302 is used to store computer-executable instructions; the processor 301 is used to execute the computer-executable instructions stored in the memory to implement the various steps performed by the electronic device in the above embodiment. For details, please refer to the relevant description in the above method embodiment.

[0169] Optionally, the memory 302 may be independent or integrated with the processor 301 .

[0170] When the memory 302 is independently provided, the electronic device further includes a bus 303 for connecting the memory 302 and the processor 301 .

[0171] The embodiment of the present invention further provides a computer storage medium, in which computer execution instructions are stored. When the processor 301 executes the computer execution instructions, the above method is implemented.

[0172] The embodiment of the present invention further provides a computer program product, including a computer program. When the computer program is executed by the processor 301, the above method is implemented.

[0173] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules is only a logical function division, and there may be other division methods in actual implementation, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0174] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to implement the solution of this embodiment.

[0175] In addition, each functional module in each embodiment of the present invention may be integrated into one processing unit, or each module may exist physically separately, or two or more modules may be integrated into one unit. The above-mentioned module-composed unit may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0176] The above-mentioned integrated module implemented in the form of a software function module can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium, including a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor 301 to perform some steps of the methods of various embodiments of the present application.

[0177] It should be understood that the processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), etc. A general-purpose processor may be a microprocessor or the processor 301 may be any conventional processor 301, etc. The steps of the method disclosed in the invention may be directly embodied in the execution of the hardware processor 301, or may be executed by a combination of hardware and software modules in the processor 301.

[0178] The memory 302 may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk, or an optical disk.

[0179] The bus 303 may be an Industry Standard Architecture (ISA), a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus 303 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, the bus 303 in the drawings of the present application is not limited to only one bus 303 or one type of bus 303.

[0180] The above storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The storage medium can be any available medium that can be accessed by a general or special purpose computer.

[0181] An exemplary storage medium is coupled to the processor 301, so that the processor 301 can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be a component of the processor 301. The processor 301 and the storage medium can be located in an application specific integrated circuit (Application Specific Integrated Circuits, referred to as ASIC). Of course, the processor 301 and the storage medium can also exist as discrete components in an electronic device or a main control device.

[0182] Those skilled in the art can understand that all or part of the steps of implementing the above-mentioned method embodiments can be completed by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, the steps of the above-mentioned method embodiments are executed; and the aforementioned storage medium includes: ROM, RAM, disk or optical disk and other media that can store program codes.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A hybrid recommendation method, characterized in that: The following steps are involved: Obtain the behavior data of old users purchasing goods on the Internet and process them to obtain the interaction information between the old users and the goods. According to the interaction information between the old users and the goods, calculate the preference value of the old users for the interacted goods, the popularity value of the goods and the similarity between different old users. Based on the collaborative filtering idea, calculate the preference value of the old users for the goods that have not been interacted with according to the similarity between different old users, and optimize the preference information between the old users and the goods based on the preference value of the old users for the interacted goods. Inputting the preference information of the old users and commodities into the implicit feature vector prediction model to obtain the implicit feature vector matrix of the old users and the implicit feature vector matrix of the commodities, wherein the implicit feature vector matrix of the old users is the set of implicit feature vectors of all the old users, and the implicit feature vector matrix of the commodities is the set of implicit feature vectors of all the commodities; The implicit feature vector matrix of the product is fixed, and the implicit feature vector matrix of the old user is optimized by using a genetic algorithm to obtain the optimal implicit feature vector matrix of the old user, which specifically includes: S31: according to the implicit feature vector p of the old user u in the implicit feature vector matrix P of the old user u , the implicit feature vector q of each product l in the implicit feature vector matrix Q of the product l Generate the recommendation list r of old users before evolution with the popularity value of the product u , as shown below: The implicit feature vector matrix of the product output by the implicit feature vector prediction model is represented as Q = {q1, q2, ..., q m }, m represents the total number of commodities, and the implicit feature vector matrix of the old users output by the implicit feature vector prediction model is expressed as P = {p1, p2, ..., p n }, n represents the total number of old users, (q l , p u ) represents the implicit feature vector q of product l l and the implicit feature vector p of old user u u The inner product of k represents the heat penalty factor, which takes a value of [0, 1], pop l represents the popularity value of product l, Sort_ represents the list is sorted from largest to smallest, take(RecN) represents taking the first RecN elements of the list, and I represents the set of all products; Define fitness as shown below: Among them, Fitness(p u ) represents the implicit feature vector p of old user u u The fitness of u represents the actual purchase list of old user u, position(r u , l) represents the recommendation list r of the old user before evolution u The position of product l calculated from the sorting position 1, rel l Represents the recommendation list r of old users before evolution u Product l is the first product that regular users will actually purchase; S32: Randomly initialize t candidate implicit feature vectors of old user u and compare them with the implicit feature vector p of old user u. u The implicit feature vector population of old user u with the number of individuals t+1 is obtained by mixing: Population u =(p u ,p u1 ,p u2 ,...,p ut ),u∈U; Among them, Population u represents the implicit feature vector population of old user u, p ut represents the candidate latent feature vector of the t-th old user u, that is, the t-th individual in the candidate latent feature vector population of the old user u; Set the number of population evolutions N, and evolve the implicit feature vector population of old user u in parallel based on the distributed framework. u , the evolution rules are as follows: 1) Create an initially empty population template NewPop; 2) Calculate Population u The fitness of each individual in the population is calculated, and the individuals are sorted from large to small according to their fitness. The top half of the individuals are selected and put into the population template NewPop. Among them, Fitness(p) represents the implicit feature vector population Population of old user u u The fitness of each individual in 3) According to the principle that the greater the individual's fitness, the greater the probability of being selected, calculate the implicit feature vector population Population of old user u u The probability that each individual is selected is p , the calculation formula is as follows: 4) Population based on the implicit feature vector population of the old user u u The probability of each individual being selected is p Population of latent feature vectors from the old user u u Select multiple individuals for crossover evolution, and select Population from the implicit feature vector population of the old user u based on the preset probability μ u Select multiple individuals for mutation evolution, and put the individuals obtained by cross evolution and mutation evolution into the population template NewPop obtained in 2); 5) Calculate the fitness of each individual in the population template NewPop obtained in 4), and select the top t with the highest fitness as the candidate implicit feature vector population of the old user u in the next round of evolution; 6) Repeat 2) to 5) until the population converges or reaches the number of population evolutions N. Before each round of evolution, the population template NewPop is cleared, and the individual corresponding to the maximum fitness of all individuals in the population template NewPop obtained in the last evolution is taken as the implicit feature vector of the optimal old user u; 7) Repeat 1) to 6) for each implicit feature vector of the old user in the implicit feature vector matrix P of the old users to obtain the optimal implicit feature vector matrix of the old users; A recommendation list for evolved old users is generated according to the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, and the heat value of the product.

2. The hybrid recommendation method according to claim 1, characterized in that: The interaction information is expressed in the following format: (User, Item, Act, Timestamp); Among them, User represents the user ID, Item represents the product ID, Act represents the quantitative value of the user's interactive behavior on the product, and Timestamp represents the timestamp of the user's interactive behavior on the product.

3. The hybrid recommendation method according to claim 2, characterized in that: The preference value of the old user for the interacted product, the popularity value of the product and the similarity between different old users are calculated based on the interaction information between the old user and the product. The preference value of the old user for the non-interacted product is calculated based on the similarity between different old users based on the collaborative filtering idea. The preference information of the old user and the product is optimized based on the preference value of the old user for the interacted product. Specifically, it includes: The corresponding preference value is calculated based on each piece of interaction information between the old user and the product. The calculation formula is as follows: Among them, s ui Indicates the preference value corresponding to each interactive information of old user u for the interactive product i; and Respectively represent the latest and earliest timestamps when old user u interacts with the interactive product i; V ui Represents the quantitative value of the interaction behavior of old user u on the interactive product i; The comprehensive preference value of the old user for the interacted product is calculated according to the preference value corresponding to the transaction information of the interacted product by the old user, as shown in the following formula: Score ui =∑s ui ; Among them, Score ui Represents the comprehensive preference value of old user u for the interactive product i; The similarity between different old users is calculated based on the comprehensive preference value of the old users for the interacted products, the set of products interacted by the old users, and the popularity value of the products. The calculation formula is as follows: Among them, U represents the set of all old users, U g represents the set of old users who have interacted with product g. Indicates the number of elements in a collection, pop g Indicates the heat value of product g; I u and I v I represents the set of commodities interacted by old user u and the set of commodities interacted by old user v respectively; uv Indicates I u and I v The intersection of Score ug Score represents the comprehensive preference value of old user u for the interactive product g. vg represents the comprehensive preference value of old user v for the interacted product g, Sim(u, v) represents the similarity between old user u and old user v; The following formula is used to calculate the preference information of old user u and product j: Among them, Predict uj represents the comprehensive preference value of old user u for product j, N sort Indicates the list of nearest neighbor old users of old user u, sorted by similarity from large to small, I′ u Indicates traversal N sort For each old user d in the set I of products interacted by old user u u The set of products obtained after the update, I′ u =I u ∪I d , I d represents the set of products interacted by old user d, Sim(u, d) represents the similarity between old user u and old user d, Score dj Score represents the comprehensive preference value of old user d for product j. uj It represents the comprehensive preference value of old user u for product j, and max() means taking the maximum value of the elements in the brackets.

4. The hybrid recommendation method according to claim 1, characterized in that: Generating a recommendation list of evolved old users according to the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old user, and the heat value of the product, specifically includes: The following formula is used to generate the recommendation list r′ of the evolved old user u u : Among them, (q l , p′ u ) represents the implicit feature vector q of product l l and the optimal implicit feature vector p′ of old user u u The inner product of k represents the heat penalty factor, which takes a value of [0, 1], pop l Represents the popularity value of product l, Sort_ indicates that the list is sorted from large to small, take(RecN) means taking the first RecN elements of the list, and I represents the set of all products.

5. The hybrid recommendation method according to claim 3, characterized in that: Also includes: The behavior data of new users purchasing goods on the Internet is obtained and the interaction information between the new users and the goods is obtained by processing. A cold start is performed based on the interaction information between the new users and the goods. The specific steps are as follows: S51: Randomly initialize the implicit feature vector of the new user u′, according to the initial implicit feature vector of the new user u′ The implicit feature vector of each product in the implicit feature vector matrix of the product and the heat value of the product are fuzzily recommended to generate an initial fuzzy recommendation list for the new user As shown below: in, The implicit feature vector q representing the product l l and the initial latent feature vector of the new user u′ The inner product of k represents the heat penalty factor, which takes a value of [0, 1], pop l represents the popularity value of product l, Sort_ represents the sorting of the list from largest to smallest, take(RecN′) represents taking the first RecN′ elements of the list; Using the initial fuzzy recommendation list as the current fuzzy recommendation list; According to the interaction information between the new user and the product, it is determined whether the new user has an interactive behavior with the products in the current fuzzy recommendation list. If so, the current fuzzy recommendation list is used as the fuzzy recommendation list of the new user u′ in the 0th round of iteration; otherwise, the RecN′ products behind the current fuzzy recommendation list are selected from the set of all products in a gradient descending manner according to the heat value to generate the next fuzzy recommendation list, and the next fuzzy recommendation list is used as the current fuzzy recommendation list and this step is repeated until it is determined that the new user has an interactive behavior with the products in the current fuzzy recommendation list and the current fuzzy recommendation list is used as the fuzzy recommendation list of the new user u′ in the 0th round of iteration, and an initially empty list H is created. u′ Used to store the historical interaction information of the new user; S52: Fuzzy recommendation list of new user u′ in the iter-1th iteration based on the new user The interactive behavior of the products in the generated new user's interactive information H on the products u′ , iter is a positive integer, based on the interactive information H of the new user on the product u′ , the implicit feature vector matrix P of the old user and the implicit feature vector matrix Q of the product to calculate the new implicit feature vector p of the new user new , as shown below: p new =w user ·UserP+w item ·ItemP; Where UserP represents the interaction information H based on the new user u′ for the product u′ The first latent feature vector is calculated by the latent feature vector matrix P of the old user; ItemP represents the interactive information H based on the new user u′ for the product u′ The second latent eigenvector calculated from the latent eigenvector matrix Q of the product; w user and w item are the weights of the first latent feature vector and the second latent feature vector respectively, and the sum of the two is 1. The calculation formulas of UserP and ItemP are as follows: Among them, Neighbor means the number of old users who are closest to H. u′ The set of users who have interactive behaviors with the products, p h represents the implicit feature vector representation of old user h in the implicit feature vector matrix of old users, q c represents the implicit feature vector of the known product c in the implicit feature vector matrix of the known product, H u′ (item) represents the interaction information H of the new user u′ on the item u′ The collection of products in, Sim(u′, h) represents the similarity between new user u′ and old user h, Score u′c represents the comprehensive preference value of the new user u′ for the interacted product c; S53: Based on the new implicit feature vector p of the new user new Generate a new recommendation list r for new user u′ new ; S54: Determine the new recommendation list r of the new user for the new user based on the interaction information between the new user and the product new Is there any interaction behavior among the products in the list? If so, the new recommendation list r of the new user is new As the fuzzy recommendation list for the new user u′ in the iter iteration Otherwise, select the new recommendation list r for the new user u′ from the set of all products in a gradient descending manner according to the heat value. new The next recommendation list is generated for the following RecN″ items, and the next recommendation list is used as a new recommendation list and step S54 is repeated until the new user determines the new recommendation list r for the new user. new There is interaction between the products in the list and the new recommendation list r of the new user new As the fuzzy recommendation list for the new user u′ in the iter iteration S55: Based on the interactive information H of the new user u′ on the product u′ , determine the implicit feature vector of the new user u′ in the iter-1th iteration Or the new latent feature vector p of the new user new Updated to The calculation formula is as follows: Among them, r (iter-1) Represents the implicit feature vector of the new user u′ in the iter-1th iteration Generated recommendation list, Score u′a Score represents the comprehensive preference value of new user u′ for the interactive product a. u′b represents the comprehensive preference value of the new user u′ for the interacted product b; S56: Repeat steps S52-S55 until When the maximum number of iterations M is reached, the implicit feature vector of the new user u′ obtained in the last iteration is used as the implicit feature vector p of the new user u′. u′ , and the recommendation list generated according to the implicit feature vector of the new user u′ obtained in the last iteration is used as the recommendation list r of the new user u′ u′ .

6. A hybrid recommendation device, characterized in that: include: The data processing module is configured to obtain the behavior data of the old user purchasing goods on the network and process the interaction information between the old user and the goods, calculate the preference value of the old user for the interacted goods, the popularity value of the goods and the similarity between different old users according to the interaction information between the old user and the goods, calculate the preference value of the old user for the goods that have not been interacted with according to the similarity between different old users based on the collaborative filtering idea, and optimize the preference information between the old user and the goods based on the preference value of the old user for the interacted goods; A prediction module is configured to input the preference information of the old users and the products into a latent feature vector prediction model to obtain an latent feature vector matrix of the old users and an latent feature vector matrix of the products, wherein the latent feature vector matrix of the old users is a set of latent feature vectors of all the old users, and the latent feature vector matrix of the products is a set of latent feature vectors of all the products; The optimization module is configured to fix the implicit feature vector matrix of the product, optimize the implicit feature vector matrix of the old user by using a genetic algorithm, and obtain the optimal implicit feature vector matrix of the old user, specifically including: S31: according to the implicit feature vector p of the old user u in the implicit feature vector matrix P of the old user u , the implicit feature vector q of each product l in the implicit feature vector matrix Q of the product l Generate the recommendation list r of old users before evolution with the popularity value of the product u , as shown below: The implicit feature vector matrix of the product output by the implicit feature vector prediction model is represented as Q = {q1, q2, ..., q m }, m represents the total number of commodities, and the implicit feature vector matrix of the old users output by the implicit feature vector prediction model is expressed as P = {p1, p2, ..., p n }, n represents the total number of old users, (q l , p u ) represents the implicit feature vector q of product l l and the implicit feature vector p of old user u u The inner product of k represents the heat penalty factor, which takes a value of [0, 1], pop l represents the popularity value of product l, Sort_ represents the list is sorted from largest to smallest, take(RecN) represents taking the first RecN elements of the list, and I represents the set of all products; Define fitness as shown below: Among them, Fitness(p u ) represents the implicit feature vector p of old user u u The fitness of u represents the actual purchase list of old user u, position(r u , l) represents the recommendation list r of the old user before evolution u The position of product l calculated from the sorting position 1, rel l Represents the recommendation list r of old users before evolution u Product l is the first product that regular users will actually purchase; S32: Randomly initialize t candidate implicit feature vectors of old user u and compare them with the implicit feature vector p of old user u. u The implicit feature vector population of old user u with the number of individuals t+1 is obtained by mixing: Population u =(p u ,p u1 ,p u2 ,...,p ut ),u∈U; Among them, Population u represents the implicit feature vector population of old user u, p ut represents the candidate latent feature vector of the t-th old user u, that is, the t-th individual in the candidate latent feature vector population of the old user u; Set the number of population evolutions N, and evolve the implicit feature vector population of old user u in parallel based on the distributed framework. u , the evolution rules are as follows: 1) Create an initially empty population template NewPop; 2) Calculate Population u The fitness of each individual in the population is calculated, and the individuals are sorted from large to small according to their fitness. The top half of the individuals are selected and put into the population template NewPop. Among them, Fitness(p) represents the implicit feature vector population Population of old user u u The fitness of each individual in 3) According to the principle that the greater the individual's fitness, the greater the probability of being selected, calculate the implicit feature vector population Population of old user u u The probability that each individual is selected is p , the calculation formula is as follows: 4) Population based on the implicit feature vector population of the old user u u The probability of each individual being selected is p Population of latent feature vectors from the old user u u Select multiple individuals for crossover evolution, and select Population from the implicit feature vector population of the old user u based on the preset probability μ u Select multiple individuals for mutation evolution, and put the individuals obtained by cross evolution and mutation evolution into the population template NewPop obtained in 2); 5) Calculate the fitness of each individual in the population template NewPop obtained in 4), and select the top t with the highest fitness as the candidate implicit feature vector population of the old user u in the next round of evolution; 6) Repeat 2) to 5) until the population converges or reaches the number of population evolutions N. Before each round of evolution, the population template NewPop is cleared, and the individual corresponding to the maximum fitness of all individuals in the population template NewPop obtained in the last evolution is taken as the implicit feature vector of the optimal old user u; 7) Repeat 1) to 6) for each implicit feature vector of the old user in the implicit feature vector matrix P of the old users to obtain the optimal implicit feature vector matrix of the old users; The list generation module is configured to generate a recommendation list for evolved old users based on the implicit feature vector of each product in the implicit feature vector matrix of the products, the implicit feature vector of each optimal old user in the implicit feature vector matrix of the optimal old users, and the heat value of the product.

7. An electronic device comprising: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images