Recommendation model training, information push method, device, equipment and medium

By constructing a recommendation model based on a relationship network, assigning weights to different relationship levels and optimizing those weights, the problem of low accuracy in information push in existing technologies is solved, enabling more targeted information push and improving user reception rates.

CN113158047BActive Publication Date: 2024-11-29PING AN TECH (SHENZHEN) CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202110434646.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-04-22
Publication Date
2024-11-29
Estimated Expiration
2041-04-22

AI Technical Summary

Technical Problem

Existing prediction models have low accuracy in information push, resulting in non-targeted information pushes that may not be of interest to users, leading to low user reception rates.

Method used

By constructing a recommendation model based on relationship networks, the social relationship networks of sample users are obtained, different weights are assigned to different relationship levels, sample user vectors are generated using neighbor user vectors, and the relationship level weights are updated and iterated through a preset parameter optimization algorithm until the convergence condition is met, thus forming a social recommendation model.

Benefits of technology

It improved the accuracy of information push notifications, ensuring that the pushed information better matches user preferences and increasing the user reception rate.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113158047B_ABST
    Figure CN113158047B_ABST
Patent Text Reader

Abstract

The present invention discloses a recommendation model training, information push method, device, equipment and medium based on a relationship network. The method inputs the neighbor user vector associated with each neighbor user in the social relationship network of the sample user into the initial recommendation model containing the relationship level weight, so as to generate a sample user vector corresponding to the sample user according to each neighbor user vector; according to the sample user vector, determine the recommendation prediction value corresponding to the sample user, and obtain the recommendation true value corresponding to the sample user; according to the recommendation prediction value and the recommendation true value, determine the prediction loss value of the initial recommendation model; when the prediction loss value does not reach the preset convergence condition, update and iterate the weights of each relationship level through a preset parameter optimization algorithm, until the prediction loss value reaches the preset convergence condition, the converged initial recommendation model is recorded as a socialized recommendation model. The present invention improves the prediction accuracy of the initial recommendation model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of prediction model technology, and in particular to a recommendation model training and information push method, device, equipment and medium based on a relationship network. Background Art

[0002] With the development of science and technology, more and more applications will push various types of information to each user, such as recommended information such as newly released movies, new welfare event information, etc., so that users can quickly access the content of new products or new activities.

[0003] At present, information push or product push for users is carried out by collecting user information and then predicting it based on the user information through a prediction model, so as to push information or products related to the user information to the user. However, this method has the following shortcomings: after the prediction model predicts the information related to the user information, in order to improve the efficiency of information push, it will choose to push all information to the user, which will lead to the phenomenon of not being able to push targeted information, resulting in users not checking the information they are not interested in at all, which leads to a low prediction accuracy of the model. Summary of the invention

[0004] The embodiments of the present invention provide a recommendation model training and information push method, device, equipment and medium based on a relationship network to solve the problem of low model prediction accuracy.

[0005] A recommendation model training method based on a relational network, comprising:

[0006] Acquire a social relationship network of a sample user; the social relationship network includes a plurality of relationship levels associated with the sample user, each relationship level includes at least one neighbor user, and each neighbor user is associated with a neighbor user vector;

[0007] Inputting each of the neighbor user vectors into an initial recommendation model including initial parameters to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors; the initial parameters including a relationship level weight associated with each of the relationship levels;

[0008] Determine, according to the sample user vector, a recommendation prediction value corresponding to the sample user, and obtain a recommendation true value corresponding to the sample user;

[0009] Determining a prediction loss value of the initial recommendation model according to the recommendation prediction value and the recommendation true value;

[0010] When the predicted loss value does not reach the preset convergence condition, the weights of each relationship level are updated and iterated through a preset parameter optimization algorithm until the predicted loss value reaches the preset convergence condition, and the initial recommendation model after convergence is recorded as a social recommendation model.

[0011] An information push method, comprising:

[0012] receiving an information push instruction sent by a user to be pushed; the information push instruction includes a user relationship network associated with the user to be pushed; the user relationship network includes a plurality of user levels associated with the user to be pushed, each of the user levels includes at least one relationship user, and each of the relationship users is associated with an information push vector;

[0013] Inputting each of the information push vectors into a social recommendation model to determine the user preference information of the user to be pushed according to each of the information push vectors; the social recommendation model is obtained according to the above-mentioned recommendation model training method based on the relationship network;

[0014] Obtaining an information list; the information list includes at least one information to be pushed;

[0015] According to the user preference information, the information to be pushed is sorted to obtain an information push sequence, and the information push sequence is pushed to the user to be pushed.

[0016] A recommendation model training device based on a relational network, comprising:

[0017] A social relationship network acquisition module, used to acquire a social relationship network of a sample user; the social relationship network includes a plurality of relationship levels associated with the sample user, each relationship level includes at least one neighbor user, and each neighbor user is associated with a neighbor user vector;

[0018] A sample user vector generation module, configured to input each of the neighbor user vectors into an initial recommendation model including initial parameters, so as to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors; the initial parameters including a relationship level weight associated with each of the relationship levels;

[0019] A movie preference prediction module, used to determine the recommendation prediction value corresponding to the sample user according to the sample user vector, and obtain the recommendation true value corresponding to the sample user;

[0020] A prediction loss value determination module, used to determine the prediction loss value of the initial recommendation model according to the recommended prediction value and the recommended true value;

[0021] The model convergence determination module is used to update and iterate the weights of each relationship level through a preset parameter optimization algorithm when the predicted loss value does not reach the preset convergence condition, until the predicted loss value reaches the preset convergence condition, and record the initial recommendation model after convergence as a social recommendation model.

[0022] An information push device, characterized by comprising:

[0023] An information push instruction receiving module is used to receive an information push instruction sent by a user to be pushed; the information push instruction includes a user relationship network associated with the user to be pushed; the user relationship network includes multiple user levels associated with the user to be pushed, each user level includes at least one relationship user, and each relationship user is associated with an information push vector;

[0024] A user preference information determination module, used to input each of the information push vectors into a social recommendation model to determine the user preference information of the user to be pushed according to each of the information push vectors; the social recommendation model is obtained according to the above-mentioned recommendation model training method based on the relationship network;

[0025] An information list acquisition module is used to acquire an information list; the information list contains at least one information to be pushed;

[0026] The information push module is used to sort the information to be pushed according to the user preference information, obtain an information push sequence, and push the information push sequence to the user to be pushed.

[0027] A computer device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned recommendation model training method based on a relationship network when executing the computer program, or implements the above-mentioned information push method when executing the computer program.

[0028] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned recommendation model training method based on a relationship network, or implements the above-mentioned information push method when the computer program is executed by a processor.

[0029] The above-mentioned recommendation model training, information push method, device, equipment and medium based on relationship network, the method obtains the social relationship network of sample users; the social relationship network includes multiple relationship levels associated with the sample users, one relationship level includes at least one neighbor user, and one neighbor user is associated with a neighbor user vector; each neighbor user vector is input into an initial recommendation model including initial parameters to generate a sample user vector corresponding to the sample user according to each neighbor user vector; the initial parameters include relationship level weights associated with each relationship level; according to the sample user vector, the recommendation prediction value corresponding to the sample user is determined, and the recommendation true value corresponding to the sample user is obtained; according to the recommendation prediction value and the recommendation true value, the prediction loss value of the initial recommendation model is determined; when the prediction loss value does not reach the preset convergence condition, the weights of each relationship level are updated and iterated through a preset parameter optimization algorithm until the prediction loss value reaches the preset convergence condition, and the initial recommendation model after convergence is recorded as a socialized recommendation model.

[0030] The recommendation model training method based on the relationship network of the present invention introduces a social relationship network and assigns different relationship level weights to neighbor users of different relationship levels in the social relationship network. That is, when generating a sample user vector corresponding to a sample user through the neighbor user vector of the neighbor user, the information provided by each relationship level (that is, the neighbor user vector) has different effects on the sample user, and the information provided by the neighbor user of the relationship level closer to the sample user is more valuable (the relationship level weight is higher). In this way, the accuracy of the initial recommendation model can be improved by determining the above relationship level weights. Furthermore, when the prediction accuracy of the initial recommendation model does not meet the requirements, the relationship level weights are updated and iterated through the preset parameter optimization algorithm, so that the relationship level weights are adapted to each relationship level, thereby improving the prediction accuracy of the initial recommendation model.

[0031] The information push method of the present invention obtains a social recommendation model trained by the above-mentioned recommendation model training method based on the relationship network, combines the user relationship network of the user to be pushed, and then determines the user preference information of the user to be pushed according to the information push vector of each relationship user in the user relationship network, that is, uses the information push vector of the relationship user to predict the movie preference of the user to be pushed, thereby improving the accuracy of information push. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0033] Figure 1 It is a schematic diagram of an application environment of a recommendation model training method or an information push method based on a relationship network in an embodiment of the present invention;

[0034] Figure 2 is a flow chart of a recommendation model training method based on a relationship network in one embodiment of the present invention;

[0035] Figure 3 is a flow chart of a recommendation model training method based on a relationship network in one embodiment of the present invention;

[0036] Figure 4 is a flow chart of an information push method in one embodiment of the present invention;

[0037] Figure 5 It is a principle block diagram of a recommendation model training device based on a relational network in one embodiment of the present invention;

[0038] Figure 6 It is a principle block diagram of a recommendation model training device based on a relational network in one embodiment of the present invention;

[0039] Figure 7 is a principle block diagram of an information push device in one embodiment of the present invention;

[0040] Figure 8 is a schematic diagram of a computer device in one embodiment of the present invention. DETAILED DESCRIPTION

[0041] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of 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.

[0042] The recommendation model training method based on the relationship network provided in the embodiment of the present invention can be applied as follows: Figure 1 Specifically, the recommendation model training method based on the relationship network is applied in the recommendation model training system based on the relationship network, and the recommendation model training system based on the relationship network includes: Figure 1The client and server shown in the figure communicate with each other through the network to solve the problem of low model prediction accuracy. The client, also known as the user end, refers to the program corresponding to the server and providing local services to customers. The client can be installed on but not limited to various personal computers, laptops, smart phones, tablets and portable wearable devices. The server can be implemented as an independent server or a server cluster consisting of multiple servers.

[0043] In one embodiment, if Figure 2 As shown in the figure, a recommendation model training method based on relationship network is provided, and the method is applied in Figure 1 The server in the example is used as an example to illustrate the process, including the following steps:

[0044] S10: Acquire a social relationship network of the sample user; the social relationship network includes a plurality of relationship levels associated with the sample user, one relationship level includes at least one neighbor user, and one neighbor user is associated with a neighbor user vector.

[0045] It can be understood that the social relationship network is constructed and generated by neighbor users who have direct relationships (such as relatives, friends, colleagues, etc.) or indirect relationships (such as friends of relatives, relatives of colleagues, etc.) with the sample user. The social relationship network contains multiple relationship levels. For example, the neighbor users who have a direct relationship with the sample user can be set in the first relationship level, and other neighbor users who have a direct relationship with the neighbor users in the first relationship level can be set in the second relationship level, etc. The essence of the neighbor user vector is a triple, which contains the relationship portrait vector, product portrait vector and product portrait score corresponding to the neighbor user.

[0046] In one embodiment, before obtaining the social relationship network of the sample user, the following steps are included:

[0047] Neighbor users associated with the sample user are obtained from the user attribute library, and each of the neighbor users is inserted into a relationship level determination sequence.

[0048] It can be understood that the user attribute library can be a social platform used by the sample users. The relationship level determination sequence is used to store neighbor users whose relationship levels have been determined, and each neighbor user can be inserted into the relationship level determination sequence by random insertion.

[0049] Neighbor users who have a direct relationship with the sample user are recorded as a first relationship level, and neighbor users who have a direct relationship with neighbor users at the first relationship level are recorded as a second relationship level.

[0050] It can be understood that the directly related persons refer to neighbor users who have a direct association with the sample user, such as, for example, the sample user's relatives, friends, colleagues, etc., and then the neighbor users who have a direct relationship with the sample user are recorded as the first relationship level. Further, the neighbor users who are associated with the neighbor users of the first relationship level are recorded as the second relationship level, such as, for example, the parents of the sample user's spouse, the friends of the sample user's friends, etc. It should be noted that in this embodiment, only the first relationship level and the second relationship level are proposed as examples, that is, there may also be neighbor users who have a direct relationship with the neighbor users of the second relationship level, which are recorded as the third relationship level, which will not be repeated here.

[0051] After the relationship levels of all the neighbor users in the relationship level determination sequence are determined, it is determined that the construction of the social relationship network of the sample user is completed.

[0052] It can be understood that after the relationship levels of all neighbor users in the relationship level determination sequence are determined, a social relationship network centered on the sample user is formed, which includes the sample user, neighbor users associated with the sample user, and the relationship levels corresponding to each neighbor user.

[0053] In one embodiment, after the relationship levels of all the neighbor users in the relationship level determination sequence are determined, the method further includes:

[0054] From the social relationship network, neighbor users at each relationship level are clustered to determine a relationship portrait vector corresponding to each neighbor user.

[0055] It can be understood that each neighbor user has a corresponding user label, which represents the individual information of the neighbor user, such as gender, age, occupation, etc., and then the neighbor users are classified according to their individual information, such as by age group, gender, occupation, etc., and then the neighbor users at each relationship level are clustered, and the relationship portrait vector corresponding to each neighbor user is obtained.

[0056] The product information associated with all the neighboring users is obtained, and the product information is clustered to determine the product portrait vector corresponding to each of the neighboring users.

[0057] For example, in the movie recommendation scenario, product information can be obtained based on the neighbor user's movie ticket purchase records on various online movie ticketing platforms, or offline movie ticket purchase records, etc.; in the music recommendation scenario, product information can be obtained based on the neighbor user's music playback records on various online music platforms, or music album purchase records, etc., and then after obtaining the product information associated with all neighbor users, the neighbor user's product information is clustered, and the product portrait vector corresponding to the neighbor user is determined. Furthermore, if a neighbor user may have watched or browsed many movies, and a movie may belong to multiple movie categories, such as a movie may have both science fiction and horror movie category labels, then according to each neighbor user's product information, the specific product category contained in the neighbor user's product information is determined, and then the product portrait vector corresponding to each neighbor user is generated.

[0058] The product ratings of the neighboring users corresponding to the product information associated with the neighboring users are obtained, and the product image ratings of the neighboring users for the product image vectors are determined according to the product ratings.

[0059] It can be understood that the product rating refers to the personal rating of the neighbor user's product; for example, in the movie recommendation scenario, the product rating can be the neighbor user's rating of the movies he has watched, and then after the product information is clustered and the product portrait vector corresponding to each neighbor user is determined, since the product portrait vector contains all product categories watched by the corresponding neighbor user, the product ratings corresponding to the products belonging to the same product category are weighted averaged to determine the product portrait score of each neighbor user for the product portrait vector. Exemplarily, in a movie recommendation scenario, suppose a neighbor user has watched a total of 10 movies, and these 10 movies include three movie categories: science fiction, action, and romance. Suppose one movie corresponds to one movie category, and among these 10 movies, 6 belong to the science fiction category, 3 belong to the action category, and 1 belongs to the romance category. Then, the score of the science fiction category in the product portrait vector of the neighbor user is the average of the product scores of the 6 movies, the score of the action category in the product portrait vector of the neighbor user is the average of the product scores of the 3 movies, and the score of the romance category in the product portrait vector of the neighbor user is the product score of 1 movie. Then, the product portrait score of each neighbor user for each product portrait vector is determined according to the scores of these three categories.

[0060] A neighbor user vector associated with each neighbor user is generated based on the relationship portrait vector, the product portrait vector and the product portrait score corresponding to each neighbor user.

[0061] After determining the relationship portrait vector, product portrait vector and product portrait score corresponding to each neighbor user, a neighbor user vector associated with each neighbor user is generated based on the relationship portrait vector, product portrait vector and product portrait score.

[0062] S20: Inputting each of the neighbor user vectors into an initial recommendation model including initial parameters to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors; the initial parameters include a relationship level weight associated with each of the relationship levels.

[0063] It can be understood that the initial recommendation model can be an XGBoost model, or a model built based on a graph neural network. The sample user vector is obtained by aggregating the neighbor user vectors of all neighbor users in the social relationship network. The relationship level weight refers to the weight value pre-assigned to each relationship level in the social relationship network. Since the relationship between each relationship level and the sample user is different, the first relationship level has a direct relationship with the sample user, so it has a greater influence on the sample user, and thus the first relationship level can be assigned a higher relationship level weight. As the relationship level increases, the relationship with the sample user becomes smaller and smaller, and thus the outermost relationship level in the social relationship network is assigned the lowest relationship level weight.

[0064] In one embodiment, if Figure 3 As shown, in step S20, that is, inputting each of the neighbor user vectors into an initial recommendation model including initial parameters to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors, including:

[0065] S201: Starting from the first relationship level of the social relationship network, each neighbor user vector in the first relationship level is input into the initial recommendation model in sequence, and each neighbor user vector in the first relationship level is encrypted and aggregated by the average aggregation function in the initial recommendation model to obtain an aggregation vector corresponding to the first relationship level.

[0066] Optionally, the average aggregation function may adopt an average aggregation function based on homomorphic encryption to ensure information security when information is transmitted between sample users and neighbor users, improve the security of information transmission, and protect user privacy.

[0067] In a specific implementation, step S201 includes:

[0068] A key and a secret key sent from the sample user are received, and the key is sent to all neighbor users in the first relationship level, so that each neighbor user encrypts the neighbor user vector associated with it according to the key and then outputs it.

[0069] It can be understood that the secret key and the secret key are a pair of encryption keys pre-generated by the sample user, the secret key is used to enable the neighbor user to encrypt the neighbor user vector, and the secret key is used to decrypt.

[0070] Specifically, after each neighbor user vector in the first relationship level is input into the initial recommendation model in turn, the key and secret key sent by the sample user are received, and the key is sent to all neighbor users in the first relationship level, so that all neighbor users in the first relationship level encrypt the neighbor user vector associated with them according to the key, and output the encrypted neighbor user vector.

[0071] Receive the encrypted neighbor user vectors output by each neighbor user in the first relationship level, and send the encrypted neighbor user vectors to the sample user, so that the sample user aggregates the encrypted neighbor user vectors through a homomorphic encryption algorithm to obtain and output an encrypted aggregate vector.

[0072] Specifically, after each of the neighbor users encrypts the neighbor user vector associated with it according to the key and outputs it, the encrypted neighbor user vector output by each neighbor user in the first relationship level is received, and the encrypted neighbor user vector is sent to the sample user, so that the sample user aggregates the encrypted neighbor user vectors through the homomorphic encryption algorithm to obtain an encrypted aggregate vector and output it. It can be understood that since the sample user also has the key and the secret key, the sample user can aggregate the encrypted neighbor user vectors; in addition, after receiving the encrypted neighbor user vector output by each neighbor user in the first relationship level, the encrypted neighbor user vectors can be directly aggregated through the homomorphic encryption algorithm to obtain an encrypted aggregate vector.

[0073] The encrypted aggregation vector output by the sample user is received, the encrypted aggregation vector is decrypted by using the key, and the aggregation vector is determined according to the decrypted encrypted aggregation vector.

[0074] Specifically, after the sample user is allowed to aggregate the encrypted neighbor user vectors through a homomorphic encryption algorithm to obtain and output an encrypted aggregate vector, the encrypted aggregate vector output by the sample user is received, the encrypted aggregate vector is decrypted through a key, and the aggregate vector is determined based on the decrypted encrypted aggregate vector and the total number of neighbor users in the first relationship level.

[0075] S202: After all the relationship levels in the social relationship network have determined their corresponding aggregation vectors, the sample user vector is determined according to the aggregation vector corresponding to each relationship level and the relationship level weight.

[0076] It can be understood that in this embodiment, after each aggregation vector corresponding to a relationship level is determined, the sample user vector is updated. Exemplarily, before the aggregation vector corresponding to the first relationship level is determined, the sample user vector corresponding to the sample user can be set to a null value; after the aggregation vector corresponding to the first relationship level is determined, the aggregation vector is updated according to the aggregation vector of the first relationship level and the relationship level weight of the first relationship level; after the aggregation vector corresponding to the second relationship level is determined, the aggregation vector of the second relationship level and the relationship level weight of the second relationship level are added to the aggregation vector at this time through the concat function (the aggregation vector at this time is also updated according to the aggregation vector of the first relationship level and the relationship level weight of the first relationship level), and then after all relationship levels in the social relationship network have determined the corresponding aggregation vectors, that is, after the final aggregation vector is updated according to the aggregation vectors corresponding to all relationship levels and the relationship weights, the final aggregation vector is normalized to obtain the sample user vector.

[0077] S30: Determine a recommendation prediction value corresponding to the sample user according to the sample user vector, and obtain a recommendation true value corresponding to the sample user.

[0078] The recommendation prediction value refers to the prediction obtained by the initial recommendation model based on the neighbor user vector of the neighbor user. For example, in the movie recommendation scenario, the recommendation prediction value can represent the predicted sample user's preference score for various movie categories. The recommendation true value refers to the sample user's true score for various commodities. The recommendation true value can be determined based on the sample user's commodity purchase record, etc.; for example, in the movie recommendation scenario, the recommendation true value can represent the sample user's true preference score for various movie categories.

[0079] Specifically, after each of the neighbor user vectors is input into the initial recommendation model containing initial parameters to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors, the sample user vector is connected to the fully connected layer to predict the sample user's movie preference according to the sample user vector through the fully connected layer to obtain a predicted score corresponding to each movie category, that is, a recommended prediction value.

[0080] S40: Determine a prediction loss value of the initial recommendation model according to the recommendation prediction value and the recommendation true value.

[0081] Specifically, after determining the recommendation prediction value corresponding to the sample user according to the sample user vector and obtaining the recommendation true value corresponding to the sample user, based on the recommendation prediction value and the recommendation true value, for example, in a movie recommendation scenario, according to the same movie category, determine the gap between the predicted score and the actual score of the sample user for this type of movie, and then determine the prediction loss value of the initial recommendation model.

[0082] S50: When the predicted loss value does not reach the preset convergence condition, the weights of each relationship level are updated and iterated through a preset parameter optimization algorithm until the predicted loss value reaches the preset convergence condition, and the initial recommendation model after convergence is recorded as a social recommendation model.

[0083] It can be understood that the convergence condition can be the condition that the predicted loss value is less than the set threshold, that is, when the predicted loss value is less than the set threshold (or the relationship level weights corresponding to each relationship level converge), the training is stopped; the convergence condition can also be the condition that the predicted loss value is very small after 10,000 calculations and will not decrease anymore, that is, when the predicted loss value is very small after 10,000 calculations and will not decrease, the training is stopped, and the initial recommendation model after convergence is recorded as a social recommendation model.

[0084] Furthermore, after determining the predicted loss value of the initial recommendation model based on the recommended predicted value and the recommended true value, when the predicted loss value does not reach the preset convergence condition, the weights of each relationship level are updated and iterated through the preset parameter optimization algorithm according to the predicted loss value, and the neighbor user vectors corresponding to the sample user are input into the initial recommendation model containing the initial parameters, so that when the predicted loss value corresponding to the sample user reaches the preset convergence condition, the social relationship network of another sample user is selected for training, and the above steps S10 to S40 are performed to obtain the predicted loss value corresponding to the sample user, and when the predicted loss value does not reach the preset convergence condition, the weights of each relationship level are updated and iterated again through the preset parameter optimization algorithm according to the predicted loss value, so that the predicted loss value corresponding to the sample user reaches the preset convergence condition.

[0085] In this way, after the initial recommendation model is trained through the social relationship networks of different sample users, the results output by the initial recommendation model can continue to approach the accurate results, making the recognition accuracy higher and higher, until the predicted loss values ​​corresponding to each sample user reach the preset convergence conditions, and the initial recommendation model after convergence is recorded as a socialized recommendation model.

[0086] Furthermore, in this embodiment, the preset parameter optimization algorithm can optimize the relationship level weights based on the differential privacy SGD method, so that a more suitable relationship level weight is assigned to each relationship level, thereby improving the prediction accuracy of the initial recommendation model.

[0087] In this embodiment, by introducing a social relationship network and assigning different relationship level weights to neighbor users of different relationship levels in the social relationship network, that is, when generating a sample user vector corresponding to a sample user through the neighbor user vector of the neighbor user, the information provided by each relationship level (that is, the neighbor user vector) has different effects on the sample user, and the information provided by the neighbor user of the relationship level closer to the sample user is more valuable (the relationship level weight is higher). In this way, by determining the above relationship level weights, the accuracy of the initial recommendation model can be improved. Furthermore, when the prediction accuracy of the initial recommendation model does not meet the requirements, the relationship level weight is updated and iterated through the preset parameter optimization algorithm, so that the relationship level weight is adapted to each relationship level, thereby improving the prediction accuracy of the initial recommendation model.

[0088] In one embodiment, if Figure 4 As shown, a method for pushing information is proposed, comprising:

[0089] S60: receiving an information push instruction sent by a user to be pushed; the information push instruction includes a user relationship network associated with the user to be pushed; the user relationship network includes multiple user levels associated with the user to be pushed, each user level includes at least one relationship user, and each relationship user is associated with an information push vector.

[0090] It can be understood that the user to be pushed is a user waiting to be sent the information to be pushed. The information push instruction can be automatically generated after the user to be pushed uploads the user relationship network, or it can be sent by the user to be pushed through a mobile terminal or other device. The user relationship network is constructed and generated by each relationship user who has a direct relationship (such as relatives, friends, colleagues, etc.) or an indirect relationship (such as friends of relatives, relatives of colleagues, etc.) with the user to be pushed. The user relationship network contains multiple user levels. Exemplarily, the relationship users who have a direct relationship with the user to be pushed can be set in the first user level, and other relationship users who have a direct relationship with the relationship users in the first user level can be set in the second user level, etc. The essence of the information push vector is a triple, which contains the user portrait vector corresponding to the relationship user, the user product portrait vector and the user product rating.

[0091] S70: Input each of the information push vectors into a social recommendation model to determine the user preference information of the user to be pushed according to each of the information push vectors; the social recommendation model is obtained according to the above-mentioned recommendation model training method based on the relationship network.

[0092] Specifically, after receiving the information push instruction sent by the user to be pushed, each information push vector in the user relationship network is input into the social recommendation model, and the social recommendation model obtained by the recommendation model training method based on the relationship network in the above embodiment is used to predict the movie preference of the user to be pushed according to each social recommendation model, so as to obtain the user preference information corresponding to the user to be pushed.

[0093] S80: Obtain an information list; the information list includes at least one information to be pushed.

[0094] Exemplarily, in a movie recommendation scenario, the information list may be a movie information list, which includes at least one currently showing movie (that is, the above-mentioned information to be pushed), that is, movies for which tickets are currently being sold on major online platforms or offline platforms; a currently showing movie may correspond to one or more movie category tags.

[0095] S90: Sort the information to be pushed according to the user preference information to obtain an information push sequence, and push the information push sequence to the user to be pushed.

[0096] Specifically, after obtaining the information list, the information to be pushed is sorted according to the user preference information, that is, according to the user preference information, the interest level of the user to be pushed in the information to be pushed is determined; illustratively, the user preference information records the scores corresponding to each type of information (the score represents the interest level of the user to be pushed in different categories of information, and the higher the score, the more interested the user to be pushed is), and then the information to be pushed is sorted according to the scores corresponding to each type of information in the information to be pushed and in the user preference information (the score can be sorted from high to low, or from low to high), an information push sequence is generated, and then the information push sequence is pushed to the user to be pushed. Exemplarily, the information push sequence can be pushed to the user to be pushed so that the user to be pushed selects the information to be pushed according to the information push sequence, or the information to be pushed with the highest score in the information push sequence is pushed to the user to be pushed.

[0097] It should be understood that the order of execution of the steps in the above embodiment does not necessarily mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of the present invention.

[0098] In one embodiment, a recommendation model training device based on a relational network is provided, and the recommendation model training device based on a relational network corresponds one-to-one to the recommendation model training method based on a relational network in the above embodiment. Figure 5 As shown, the device includes a social relationship network acquisition module 10, a sample user vector generation module 20, a movie preference prediction module 30, a prediction loss value determination module 40 and a model convergence determination module 50. The functional modules are described in detail as follows:

[0099] A social relationship network acquisition module 10 is used to acquire a social relationship network of a sample user; the social relationship network includes a plurality of relationship levels associated with the sample user, each relationship level includes at least one neighbor user, and each neighbor user is associated with a neighbor user vector;

[0100] A sample user vector generation module 20, configured to input each of the neighbor user vectors into an initial recommendation model including initial parameters, so as to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors; the initial parameters including a relationship level weight associated with each of the relationship levels;

[0101] The movie preference prediction module 30 is used to determine the recommendation prediction value corresponding to the sample user according to the sample user vector, and obtain the recommendation true value corresponding to the sample user;

[0102] A predicted loss value determination module 40, configured to determine a predicted loss value of the initial recommendation model according to the recommended predicted value and the recommended true value;

[0103] The model convergence determination module 50 is used to update and iterate the weights of each relationship level through a preset parameter optimization algorithm when the predicted loss value does not reach the preset convergence condition, until the predicted loss value reaches the preset convergence condition, and record the initial recommendation model after convergence as a social recommendation model.

[0104] Preferably, the recommendation model training device based on the relationship network also includes:

[0105] A neighbor user acquisition module, used to acquire neighbor users associated with the sample user in a user attribute library, and insert each of the neighbor users into a relationship level determination sequence;

[0106] A relationship level determination module, configured to record neighbor users having a direct relationship with the sample user as a first relationship level, and record neighbor users having a direct relationship with neighbor users of the first relationship level as a second relationship level;

[0107] The social relationship network construction module is used to determine that the construction of the social relationship network of the sample user is completed after the relationship levels of all the neighbor users in the relationship level determination sequence are determined.

[0108] Preferably, the recommendation model training device based on the relationship network also includes:

[0109] A crowd clustering module, used to cluster neighbor users at each relationship level in the social relationship network, and determine a relationship portrait vector corresponding to each neighbor user;

[0110] A product clustering module, used to obtain product information associated with all the neighboring users, and cluster the product information to determine a product portrait vector corresponding to each of the neighboring users;

[0111] A product rating determination module is used to obtain the product ratings of each of the neighboring users for the product information associated with the neighboring users, and determine the product image ratings of each of the neighboring users for each of the product image vectors based on the product ratings;

[0112] The neighbor user vector generation module is used to generate a neighbor user vector associated with each neighbor user based on the relationship portrait vector, the product portrait vector and the product portrait score.

[0113] Preferably, if Figure 6 As shown, the sample user vector generation module 20 includes:

[0114] The encryption aggregation unit 201 is used to input each neighbor user vector in the first relationship level of the social relationship network into the initial recommendation model in sequence, and perform encryption aggregation processing on each neighbor user vector in the first relationship level through the average aggregation function in the initial recommendation model to obtain an aggregation vector corresponding to the first relationship level;

[0115] The sample user vector generating unit 202 is configured to determine the sample user vector according to the aggregation vector corresponding to each relationship level and the relationship level weight after all the relationship levels in the social relationship network have determined the corresponding aggregation vectors.

[0116] Preferably, the encryption aggregation unit 201 includes:

[0117] a key sending subunit, configured to enable the sample user to send a key to all neighbor users in the first relationship level, so that each neighbor user encrypts a neighbor user vector associated with it by using the key and then sends the encrypted vector to the sample user;

[0118] An encryption aggregation subunit, configured to enable the sample user to aggregate neighbor user vectors of all the neighbor users in the first relationship level through a homomorphic encryption algorithm to obtain an encrypted aggregation vector;

[0119] The vector decryption subunit is used to enable the sample user to decrypt the encrypted aggregate vector using a key corresponding to the key, and determine the aggregate vector according to the decrypted encrypted aggregate vector.

[0120] For the specific limitations of the recommendation model training device based on the relational network, please refer to the limitations of the recommendation model training method based on the relational network above, which will not be repeated here. Each module in the above-mentioned recommendation model training device based on the relational network can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0121] In one embodiment, if Figure 7 As shown, an information push device is proposed, comprising:

[0122] The information push instruction receiving module 60 is used to receive an information push instruction sent by a user to be pushed; the information push instruction includes a user relationship network associated with the user to be pushed; the user relationship network includes multiple user levels associated with the user to be pushed, each user level includes at least one relationship user, and each relationship user is associated with an information push vector;

[0123] A user preference information determination module 70 is used to input each of the information push vectors into a social recommendation model to determine the user preference information of the user to be pushed according to each of the information push vectors; the social recommendation model is obtained according to the above-mentioned recommendation model training method based on the relationship network;

[0124] The information list acquisition module 80 is used to acquire an information list; the information list contains at least one information to be pushed;

[0125] The information push module 90 is used to sort the information to be pushed according to the user preference information, obtain an information push sequence, and push the information push sequence to the user to be pushed.

[0126] For the specific definition of the information push device, please refer to the definition of the information push method above, which will not be repeated here. Each module in the above-mentioned information push device can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0127] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 8 As shown. The computer device includes a processor, a memory, a network interface and a database connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the data used by the recommendation model training method based on the relationship network or the information push method in the above-mentioned embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a recommendation model training method based on a relationship network or an information push method is implemented.

[0128] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the recommendation model training method based on the relationship network in the above-mentioned embodiment is implemented, or when the processor executes the computer program, the information push method in the above-mentioned embodiment is implemented.

[0129] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the recommendation model training method based on the relationship network in the above embodiment is implemented, or when the computer program is executed by a processor, the information push method in the above embodiment is implemented.

[0130] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0131] Those skilled in the art can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0132] The embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. Such modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included in the protection scope of the present invention.

Claims

1. A recommendation model training method based on a relational network, characterized in that: include: Obtain the social relationship network of sample users; The social relationship network includes a plurality of relationship levels associated with the sample user, each of the relationship levels includes at least one neighbor user, and each of the neighbor users is associated with a neighbor user vector; Inputting each of the neighbor user vectors into an initial recommendation model including initial parameters to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors; the initial parameters including a relationship level weight associated with each of the relationship levels; The first relationship level has a direct relationship with the sample user, and the first relationship level is given the highest relationship level weight. As the relationship level increases, the relationship with the sample user becomes smaller and smaller, and the outermost relationship level in the social relationship network is given the lowest relationship level weight; Determine, according to the sample user vector, a recommendation prediction value corresponding to the sample user, and obtain a recommendation true value corresponding to the sample user; Determining a prediction loss value of the initial recommendation model according to the recommendation prediction value and the recommendation true value; When the predicted loss value does not reach the preset convergence condition, the weights of each relationship level are updated and iterated by a preset parameter optimization algorithm until the predicted loss value reaches the preset convergence condition, and the initial recommendation model after convergence is recorded as a social recommendation model; Before obtaining the social relationship network of the sample user, the following steps are included: Obtaining neighbor users associated with the sample user in a user attribute library, and inserting each of the neighbor users into a relationship level determination sequence; Recording neighbor users who have a direct relationship with the sample user as a first relationship level, and recording neighbor users who have a direct relationship with neighbor users at the first relationship level as a second relationship level; After the relationship levels of all the neighbor users in the relationship level determination sequence are determined, it is determined that the construction of the social relationship network of the sample user is completed; After the relationship levels of all the neighbor users in the relationship level determination sequence are determined, the method includes: From the social relationship network, clustering neighbor users at each relationship level is performed to determine a relationship portrait vector corresponding to each neighbor user; Acquire product information associated with all the neighbor users, perform product clustering on the product information, and determine a product portrait vector corresponding to each of the neighbor users; Obtaining the product ratings of the neighbor users for the product information associated with the neighbor users, and determining the product image ratings of the neighbor users for the product image vectors according to the product ratings; A neighbor user vector associated with each neighbor user is generated based on the relationship portrait vector, the product portrait vector and the product portrait score corresponding to each neighbor user.

2. The recommendation model training method based on relationship network according to claim 1, characterized in that: The step of inputting each of the neighbor user vectors into an initial recommendation model including initial parameters to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors includes: Starting from the first relationship level of the social relationship network, each neighbor user vector in the first relationship level is sequentially input into the initial recommendation model, and each neighbor user vector in the first relationship level is encrypted and aggregated by the average aggregation function in the initial recommendation model to obtain an aggregate vector corresponding to the first relationship level; After all the relationship levels in the social relationship network have their corresponding aggregation vectors determined, the sample user vector is determined according to the aggregation vector corresponding to each relationship level and the relationship level weight.

3. The recommendation model training method based on relationship network according to claim 2, characterized in that: The step of performing encrypted aggregation processing on each of the neighbor user vectors in the first relationship level by using the average aggregation function in the initial recommendation model to obtain an aggregation vector corresponding to the first relationship level includes: receiving a key and a secret key sent from the sample user, and sending the key to all neighbor users in the first relationship level, so that each neighbor user encrypts the neighbor user vector associated with it according to the key and then outputs it; receiving the encrypted neighbor user vectors output by each neighbor user in the first relationship level, and sending the encrypted neighbor user vectors to the sample user, so that the sample user aggregates the encrypted neighbor user vectors through a homomorphic encryption algorithm to obtain and output an encrypted aggregate vector; The encrypted aggregation vector output by the sample user is received, the encrypted aggregation vector is decrypted by using the key, and the aggregation vector is determined according to the decrypted encrypted aggregation vector.

4. An information push method, characterized in that: include: receiving an information push instruction sent by a user to be pushed; the information push instruction includes a user relationship network associated with the user to be pushed; the user relationship network includes a plurality of user levels associated with the user to be pushed, each user level includes at least one relationship user, and each relationship user is associated with an information push vector; Input each of the information push vectors into a social recommendation model to determine the user preference information of the user to be pushed according to each of the information push vectors; the social recommendation model is obtained according to the recommendation model training method based on the relationship network as claimed in any one of claims 1 to 3; Obtaining an information list; the information list includes at least one information to be pushed; According to the user preference information, the information to be pushed is sorted to obtain an information push sequence, and the information push sequence is pushed to the user to be pushed.

5. A recommendation model training device based on a relational network, characterized in that: The recommendation model training device based on the relationship network is used to execute the recommendation model training method based on the relationship network as claimed in any one of claims 1 to 3, comprising: A social relationship network acquisition module, used to acquire a social relationship network of a sample user; the social relationship network includes a plurality of relationship levels associated with the sample user, each relationship level includes at least one neighbor user, and each neighbor user is associated with a neighbor user vector; A sample user vector generation module, configured to input each of the neighbor user vectors into an initial recommendation model including initial parameters, so as to generate a sample user vector corresponding to the sample user according to each of the neighbor user vectors; the initial parameters including a relationship level weight associated with each of the relationship levels; A movie preference prediction module, used to determine the recommendation prediction value corresponding to the sample user according to the sample user vector, and obtain the recommendation true value corresponding to the sample user; A prediction loss value determination module, used to determine the prediction loss value of the initial recommendation model according to the recommended prediction value and the recommended true value; The model convergence determination module is used to update and iterate the weights of each relationship level through a preset parameter optimization algorithm when the predicted loss value does not reach the preset convergence condition, until the predicted loss value reaches the preset convergence condition, and record the initial recommendation model after convergence as a social recommendation model.

6. An information push device, characterized in that: include: An information push instruction receiving module is used to receive an information push instruction sent by a user to be pushed; the information push instruction includes a user relationship network associated with the user to be pushed; the user relationship network includes multiple user levels associated with the user to be pushed, each user level includes at least one relationship user, and each relationship user is associated with an information push vector; A user preference information determination module, used to input each of the information push vectors into a social recommendation model to determine the user preference information of the user to be pushed according to each of the information push vectors; the social recommendation model is obtained according to the recommendation model training method based on the relationship network as claimed in any one of claims 1 to 3; An information list acquisition module is used to acquire an information list; the information list contains at least one information to be pushed; The information push module is used to sort the information to be pushed according to the user preference information, obtain an information push sequence, and push the information push sequence to the user to be pushed.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the recommendation model training method based on the relationship network as described in any one of claims 1 to 3, or when the processor executes the computer program, it implements the information push method as described in claim 4.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, it implements the recommendation model training method based on the relationship network as described in any one of claims 1 to 3, or when the computer program is executed by a processor, it implements the information push method as described in claim 4.

Citation Information

Patent Citations

  • Method for obtaining commodity recommendation sequence and commodity recommendation

    CN109064285A

  • A recommendation method based on a user community and a scoring combined community

    CN109918562A