Exercise recommendation method and device, electronic equipment and computer readable storage medium

Abstract

Embodiments of the present application provide a sports recommendation method and device, electronic equipment and computer readable storage medium, and relate to the technical field of computers. The method comprises: obtaining user personalized information of a target user, and obtaining at least one candidate interested item corresponding to the target user; taking the user personalized information and a sports interest label as respective label information, and for each candidate interested item, determining the relevance between each label information and the candidate interested item; and based on the relevance corresponding to each candidate interested item, determining a target sports item recommended to the target user from the candidate interested items. The embodiments of the present application fully consider user personalized information in the process of recommending sports items, and thus the recommended sports items are more suitable for the user's own situation, and personalized sports recommendation is achieved.

Description

Technical Field

[0001] This application relates to the field of exercise recommendation technology, and more specifically, to an exercise recommendation method, apparatus, electronic device, and computer-readable storage medium. Background Technology

[0002] With the national strategy of promoting proactive health, sports medicine is receiving increasing attention and importance, especially for people with chronic diseases and those in a sub-healthy state. Regular and scientific exercise helps maintain health and improve quality of life. Scientific research has confirmed that exercise has a positive effect on people with chronic diseases such as diabetes, hypertension, hyperlipidemia, and obesity, helping to improve glucose metabolism, regulate blood pressure, improve blood lipids, and control weight.

[0003] However, most current exercise programs are either manually created or recommended to users based on similar user programs. These methods are not very effective and provide a poor user experience, and therefore require improvement. Summary of the Invention

[0004] The purpose of this application is to address at least one of the aforementioned technical deficiencies, particularly the technical deficiency that exercise recommendations do not take into account individual characteristics.

[0005] According to some embodiments of this application, a method for recommending exercise is provided, the method comprising:

[0006] Obtain personalized information of the target user, which includes the target user's personal information and sports interest tags, including the types of sports and / or sports that the target user is interested in.

[0007] Obtain at least one candidate item of interest for the target user;

[0008] User personal information and sports interest tags are treated as a type of tag information. For each candidate item of interest, the association between each type of tag information and that candidate item of interest is determined.

[0009] Based on the correlation between the two types of tag information corresponding to each candidate item of interest, the target sports items recommended to the target user are determined from each candidate item of interest.

[0010] According to some embodiments of this application, a method for recommending exercise is provided, the method comprising:

[0011] Obtain the sports interest tags of the target users, which include the types of sports and / or sports that the target users are interested in;

[0012] Based on sports interest tags and a pre-configured sports knowledge graph, at least one candidate sport of interest for the target user is determined from multiple candidate sports included in the sports knowledge graph. The sports knowledge graph contains the association between each candidate sport and its corresponding sports type.

[0013] Based on each candidate item of interest, a target sport is selected to recommend to the target user.

[0014] According to some other embodiments of this application, an exercise recommendation device is provided, the device comprising:

[0015] The first acquisition module is used to acquire the target user's personalized information, which includes the target user's personal information and sports interest tags. The sports interest tags include the types of sports and / or sports that the target user is interested in.

[0016] The second acquisition module is used to acquire at least one candidate item of interest for the target user.

[0017] The correlation determination module is used to treat user personal information and sports interest tags as a type of tag information, and for each candidate interest item, determine the correlation between each type of tag information and that candidate interest item;

[0018] The target sport determination module is used to determine the target sport to recommend to the target user based on the correlation between the two types of tag information corresponding to each candidate sport of interest.

[0019] According to other embodiments of this application, a sports recommendation device is provided, the device comprising:

[0020] The tag acquisition module is used to acquire sports interest tags that the target user is interested in, including sports types and / or sports events that the target user is interested in;

[0021] The candidate interest item determination module is used to determine at least one candidate interest item of the target user from multiple candidate sports items contained in the sports knowledge graph based on sports interest tags and a pre-configured sports knowledge graph. The sports knowledge graph contains the association relationship between each candidate sports item and its corresponding sports type.

[0022] The target sport determination module is used to determine the target sport to recommend to the target user based on each candidate sport of interest.

[0023] According to other embodiments of this application, an electronic device is provided, the electronic device comprising:

[0024] One or more processors;

[0025] Memory;

[0026] One or more applications, wherein the applications are stored in memory and configured to be executed by one or more processors, the applications being configured to: execute the motion recommendation method of the first or second aspect of this application.

[0027] Some embodiments of this application provide a computing device, including: a processor, a memory, a communication interface, and a communication bus, wherein the processor, the memory, and the communication interface communicate with each other through the communication bus;

[0028] The memory is used to store at least one executable instruction that causes the processor to perform an operation corresponding to the motion recommendation method of the first or second aspect of this application.

[0029] According to another aspect of this application, a computer-readable storage medium is provided, characterized in that when a computer program is executed by a processor, it implements the motion recommendation method of the first or second aspect of this application.

[0030] Some embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the motion recommendation method of the first or second aspect of this application.

[0031] According to one aspect of this application, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations of the first or second aspect described above.

[0032] The beneficial effects of the technical solution provided in this application are:

[0033] In one optional implementation of this application, when determining the target sport to recommend to the target user, the user's personal information and sports interest tags are obtained. For a candidate sport of interest, the correlation between the user's personal information and sports interest tags and the sport is determined. Based on the correlation between the user's personal information and sports interest tags and the candidate sport of interest, the target sport to recommend to the target user is determined. Thus, in the process of recommending sports, the user's personal attributes and sports interests are fully considered, and the recommended sports are more in line with the user's own situation, realizing personalized sports recommendations.

[0034] In another optional implementation scheme provided in this application, at least one candidate interest item for the target user is determined from multiple candidate sports items included in the sports knowledge graph based on sports interest tags and a pre-configured sports knowledge graph. Based on each candidate interest item, a target sports item recommended to the target user is determined. This realizes the expansion of sports interest tags to obtain associated candidate interest items based on the user's own sports interest tags and the correlation between each candidate sports item and sports type in the sports knowledge graph, so as to recommend a richer range of target sports items to the user that are more suitable for the user's own situation. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0036] Figure 1 One of the flowcharts of an exercise recommendation method provided in this application embodiment;

[0037] Figure 2 This is an architecture diagram of the matching degree calculation model in an embodiment of this application;

[0038] Figure 3 A schematic diagram of a scenario for determining the optimal path corresponding to a motion interest tag in an embodiment of this application;

[0039] Figure 4 A second schematic flowchart illustrating an exercise recommendation method provided in an embodiment of this application;

[0040] Figure 5 This is one of the structural schematic diagrams of an exercise recommendation device provided in an embodiment of this application;

[0041] Figure 6 This is a second schematic diagram of the structure of a sports recommendation device provided in an embodiment of this application;

[0042] Figure 7 This is a schematic diagram of the architecture of a sports recommendation system provided in an embodiment of this application;

[0043] Figure 8 This is a schematic diagram of the structure of an electronic device for motion recommendation provided in an embodiment of this application. Detailed Implementation

[0044] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain this application, and should not be construed as limiting this application.

[0045] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0046] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0047] This application provides a method for recommending exercise. The executing entity of this method can be any electronic device, such as an application with exercise recommendation functionality. The method can be executed by the application's server or a terminal device; optionally, the method can be executed by the server. Figure 1 The diagram shown is a flowchart illustrating an exercise recommendation method provided in an embodiment of this application. The method may include:

[0048] Step S101: Obtain the target user's personalized information.

[0049] The target user can be any user. For example, it could be any registered user of the aforementioned application. User personalization information can include the target user's personal information and sports interest tags. The personal information can include the target user's personal attributes and health status information, such as gender, age, height, weight, and current and past medical history. Sports interest tags can be information about sports that the target user is interested in, such as the types and / or events of sports that the target user is interested in. Specifically, sports types can include, for example, ball sports, water sports, gymnastics, etc., and sports events can include, for example, badminton, swimming, aerobics, etc.

[0050] In addition, the aforementioned personalized user information can be obtained by receiving personalized user information input by the target user through the terminal device, or it can be obtained through a personalized user information database. The personalized user information database can be constructed by user information obtained from user questionnaires, or it can be constructed by analyzing user historical behavior data.

[0051] S102: Obtain at least one candidate item of interest for the target user.

[0052] Specifically, the candidate sports interests of the target user can be used to characterize sports that the user may be interested in, or sports that the user may be interested in. Optionally, the candidate sports interests can be multiple candidate sports (various sports identified through statistical methods or other methods), or at least one item after filtering multiple candidate sports. For example, multiple candidate sports can be pre-filtered based on the user's personal information and at least one of the sports interest tags to determine multiple candidate sports that the user may be interested in.

[0053] Optionally, one can obtain information on sports interests from multiple sample users, statistically determine the number of users interested in each sport, and calculate the probability of interest for each sport. Based on these probabilities and the users' sports interest tags, candidate sports interests for the target user can be predicted. Alternatively, one can analyze the target user's historical sports information to predict candidate sports interests. Furthermore, one can also predict candidate sports interests for the target user using a pre-constructed knowledge graph containing the relationships between sports interests and sports types among multiple sample users.

[0054] S103: Treat user personal information and sports interest tags as separate tags. For each candidate item of interest, determine the correlation between each tag and the candidate item of interest.

[0055] S104: Based on the correlation between the two types of tag information corresponding to each candidate item of interest, determine the target sports item to recommend to the target user from each candidate item of interest.

[0056] Specifically, because users' personal information and interest tags differ, the sports they might be interested in also vary. For example, users of different genders and ages may have different interests in different sports. Similarly, users may have different sports interest tags, meaning they are interested in different sports or types of sports, thus potentially leading to different sports they might be interested in. Therefore, user personal information and sports interest tags can each be treated as a type of tag information. For each candidate sport of interest, the correlation between each tag information and that candidate sport of interest can be determined. This correlation can be used to characterize the relationship between each tag information and the candidate sport of interest, such as the degree of matching between each tag information and the candidate sport of interest.

[0057] Furthermore, the correlation between each of the aforementioned label information and the candidate item of interest can be calculated using a neural network learning model. Alternatively, the correlation between the features corresponding to each label information and the features corresponding to the candidate item of interest can be calculated directly.

[0058] Optionally, among the candidate items of interest, those that meet the preset correlation criteria can be selected as target sports items to recommend to the target user. For example, when the correlation is the matching degree between the tag information and the candidate items of interest, the candidate items of interest with a matching degree greater than a preset threshold can be selected as target sports items to recommend to the target user. In addition, a preset number of candidate items of interest can be selected as target sports items to recommend to the target user, either in descending order of matching degree or in ascending order of matching degree. For example, when sorted in descending order of matching degree, the top 5 candidate items of interest with the highest matching degree can be selected as target sports items to recommend to the target user.

[0059] This application embodiment determines the target sports to recommend to the target user by determining the correlation between the user's personal information and sports interest tags and the candidate sports of interest. Thus, in the process of recommending sports, the user's personal attributes and sports interests are fully considered, and the recommended sports are more in line with the user's own situation, thereby realizing personalized sports recommendations.

[0060] In one embodiment of this application, step S103, determining the association between each tag information and the candidate item of interest, may include:

[0061] Obtain the first tag feature for each tag information, and the project feature for each candidate item of interest;

[0062] For each candidate item of interest, the relevance of each tag information is determined based on the item's characteristics and the first tag feature of each tag information.

[0063] Specifically, user personal information and sports interest tags are treated as a type of tag information, and the first tag feature of each tag information is obtained. The first tag feature of each tag information can be the initial feature vector corresponding to each tag information.

[0064] For user personal information, it can first be encoded. For example, one-hot encoding can be used to obtain the binary code corresponding to the user's personal information. For instance, if gender information in the user's personal information is categorical, with only two possible values ​​(male and female), then one-hot encoding would represent male as [0,1] and female as [1,0]. If the user's personal information includes continuous information, such as age, this continuous information can be discretized, for example, by dividing it into intervals of 10-20, 20-30, and so on, thus converting the numerical information into categorical information before one-hot encoding. This encoding can then be further transformed to obtain the initial feature vector corresponding to the user's personal information. For example, embedding can be used to convert the one-hot encoding into a low-dimensional dense vector. The dimension of the initial feature vector corresponding to the user's personal information can be set to a preset value, such as 100 dimensions.

[0065] For each sport or sport type in the sports interest tags, the initial feature vector corresponding to the sports interest tag can be obtained by querying a pre-trained model. For example, the initial feature vector of the sports interest tag can be obtained by querying a pre-trained model based on a sports knowledge graph. For example, the dimension of the initial feature vector corresponding to the sports interest tag can be set to a predetermined value. For example, the dimension of the initial feature vector corresponding to the sports interest tag and the dimension of the initial feature vector corresponding to the user's personal information can be set to the same preset value.

[0066] Among them, the pre-trained model based on knowledge graphs can use the TransE (Translating Embedding) algorithm (or improved algorithms of TransE, such as TransH, TransR, and TransD) to map the entities and relations contained in the triples of the knowledge graph to low-dimensional dense vectors, thereby obtaining the embedding vector representation of each entity and relation in the knowledge graph, that is, the feature vectors corresponding to the entity and relation respectively. The loss function of TransE can be calculated as follows:

[0067]

[0068] Where the hyperparameter γ is a positive number, S is a triple in the knowledge graph, h and t are entity vectors in the triple, r is a relation vector, and d is the distance between h+r and t.

[0069] S′ is constructed by replacing either the head or tail entity in S with only one of them replaced. That is, S′ satisfies:

[0070] S′ (h,r,t) ={(h',r,t)|h'∈E}∪{(h,r,t')|t'∈E}

[0071] Where E is the set of all triples in the knowledge graph.

[0072] Furthermore, the loss function L can be minimized using stochastic gradient descent, thereby obtaining the embedding vector for each entity and relation.

[0073] Furthermore, the project features of each candidate item of interest can be obtained from the initial feature vector corresponding to that item, or by querying a pre-trained model based on a motion knowledge graph. For example, the dimension of the initial feature vector corresponding to each candidate item of interest can be set to a predetermined value. For instance, the dimension of the initial feature vector corresponding to each candidate item of interest can be set to the same predetermined value as the dimension of the initial feature vector corresponding to the user's personal information and / or the initial feature vector corresponding to the motion interest tag. Thus, for each candidate item of interest, the correlation can be determined based on the project features of that item and the first tag feature of each tag information, for example, by determining the correlation through a neural network model.

[0074] Furthermore, step S104 may specifically include:

[0075] For each candidate item of interest, the first label feature of each label information is weighted according to the correlation between each label information and the item, so as to obtain the second label feature corresponding to each label information of the item.

[0076] For each candidate item of interest, the matching degree between the target user and the item is obtained based on the second tag feature of the two tag information and the item feature of the item.

[0077] Based on the matching degree of each candidate interest item, the target sports item recommended to the target user is determined from each candidate interest item.

[0078] Specifically, in this application embodiment, user personal information and sports interest tags can be used as a type of tag information. For each candidate item of interest, a second tag feature corresponding to each tag information of that item can be used to characterize the tag feature that takes into account the correlation between each tag information and the item.

[0079] It should be noted that, for each candidate item of interest, when weighting the first label feature of each label information based on its relevance to the item, if each label information includes multiple pieces of information, for each piece of information, the first sub-label feature of that piece of information can be weighted based on its relevance to the item to obtain the corresponding second sub-label feature. A specific implementation scheme can be found in the following embodiment. Furthermore, if the label information includes only one piece of information, its weight value can be 1 when weighting the first label feature corresponding to that label information.

[0080] Then, based on the second tag feature corresponding to each tag information of the project, and the project feature, the matching degree corresponding to the project can be calculated through the matching degree calculation model, for example, the second sub-model in the matching degree calculation model used to calculate the matching degree.

[0081] Furthermore, from among the candidate items of interest, those with a matching degree greater than a preset threshold can be selected as target sports items to recommend to the target user. Alternatively, a preset number of candidate items of interest can be selected as target sports items to recommend to the target user, arranged in descending order of matching degree. For example, when a user logs into the corresponding sports recommendation app, target sports items can be pushed to the user via floating windows, in-app messages, or other means. In addition, when changes are detected in the user's personal information, such as changes in the user's health status or changes in the sports interest tags, the redefined target sports items can be pushed to the user via SMS or other prompts. For example, recommendations can also be made to the user via voice broadcast.

[0082] Therefore, this embodiment of the application determines the second tag feature corresponding to each tag information for each project based on the first tag feature of each tag information and the correlation between each tag information and the project; for each candidate project of interest, the matching degree between the target user and the project is obtained based on the second tag features of the two tag information and the project features of the project, and the target sports project recommended to the target user is determined from each candidate project of interest. Thus, by determining the matching degree corresponding to the project based on the correlation between the user's personal information and sports interest tags and the candidate projects of interest, and recommending sports projects based on the matching degree corresponding to each project, the personal attributes of the user and the matching degree between the user's sports interests and the sports projects to be recommended are fully considered, and the recommended sports projects are more in line with the user's own situation, thus realizing personalized sports recommendations.

[0083] In another embodiment of this application, for each type of tag information, the tag information includes at least one piece of information, and the first tag feature of each type of tag information includes the first sub-tag feature corresponding to each piece of information;

[0084] For each candidate item of interest, the relevance of each tag information is determined based on the item's characteristics and the first tag feature of each tag information. This may include:

[0085] For each piece of information in each tag, the correlation between the information and the candidate sport of interest is determined based on the first sub-tag feature corresponding to the information and the project feature of the project.

[0086] For each candidate item of interest, the first label feature of each label information is weighted based on its relevance to the item, resulting in a second label feature corresponding to each label information for that item. This second label feature may include:

[0087] For each candidate item of interest, the relevance of each item in each tag information to that item is used to weight the first sub-tag feature corresponding to each item of information to obtain the second sub-tag feature corresponding to that item of information.

[0088] For each type of label information, the second sub-label features of each item contained in the label information are fused to obtain the second label feature of the label information.

[0089] Specifically, in this embodiment, for each candidate item of interest, determining the correlation between each tag information and the candidate item of interest can be achieved through the first sub-model in the matching degree calculation model, and determining the matching degree between the target user and the item can be achieved through the second sub-model in the matching degree calculation model, combined with... Figure 2 This is an architecture diagram of the matching degree calculation model in an embodiment of this application. Figure 2 In this model, the matching degree calculation model can include a first sub-model and a second sub-model. The first sub-model can be implemented using an attention unit; for example, it can include the Attention module shown in the diagram. The second sub-model can be implemented using a classifier for calculating the matching degree; for example, it can include the MLP (Multilayer Perceptron) module and the sigmoid module shown in the diagram. Alternatively, both the first and second sub-models can be implemented using neural networks, such as CNN, RNN, GRU, LSTM, etc. As shown in the diagram, user personal information includes n items, which can be represented as x1, x2, x3…x n Where n represents the number of items in the user's personal information, and the first sub-label feature corresponding to each item in the user's personal information can be represented as u1, u2, u3...u n The sports interest tag includes m items of information, which can be represented as ID1, ID2, ID3...ID m 'm' represents the number of information items included in the motion interest label, and the first sub-label feature corresponding to each information item in the motion interest label can be represented as v1, v2, v3...v m Furthermore, multiple candidate items of interest can be represented as id1, id2, id3…id t t represents the number of candidate items of interest, and the item features corresponding to each candidate item of interest can be represented as c1, c2, c3…c t The user's personal information may include gender, age, height, weight, and past medical history. Sports interest tags may include information on sports types the user is interested in, such as ball sports, water sports, dance, and gymnastics, or sports events the user is interested in, such as badminton, swimming, and resistance bands. For user personal information [x1, x2, x3…x], ... n The first sub-label features corresponding to each piece of information contained in the document can be obtained by performing one-hot encoding on each piece of information to obtain the corresponding binary code. This binary code can then be converted into a low-dimensional dense vector of a predetermined dimension, which is the first sub-label feature corresponding to that piece of information. For example, one-hot encoding can be converted into a low-dimensional dense vector using embedding. For example, for user information x... i Its corresponding first sub-label feature u i It can be calculated in the following ways:

[0090] u i =W i x i

[0091] Among them, W i For feature xi The corresponding embedding matrix, W i It can be obtained through training a pre-trained model.

[0092] For sports interest tags [v1, v2, v3…v] m The first sub-label features corresponding to each piece of information contained herein can be obtained by querying a pre-trained model based on a motion knowledge graph for each piece of information.

[0093] For each item in each of the two types of tag information mentioned above—user personal information and sports interest tags—the correlation between that item and the candidate sport of interest is determined based on the first sub-tag feature corresponding to that item and the project feature of that project. This can include:

[0094] The first sub-label feature corresponding to this information is multiplied with the item feature of this item to obtain the first associated feature;

[0095] Input the first sub-label feature corresponding to this information and the project feature of this project into the weighted network model to obtain the second associated feature;

[0096] Based on the first and second association features, determine the association between this item and this information.

[0097] Specifically, based on the ID of the kth candidate project of interest among the candidate projects of interest. k And the ID of the i-th sports interest tag contained in the sports interest tags. i For example, candidate project IDs k With sports interest tag ID i First associated feature v ik The Hadamard product can be used to calculate the product, and the specific calculation method is as follows:

[0098]

[0099] Among them, v i Sports interest tag ID i The corresponding first sub-label feature, c k For candidate project IDs k Corresponding project characteristics.

[0100] Because the aforementioned first association feature will v i and c k The first associated feature is calculated by multiplying the feature values ​​at the corresponding positions. Therefore, it can be understood that the first associated feature is used to characterize the static feature relationship between each piece of information in each type of label information and the item.

[0101] In addition, candidate project IDs k With sports interest tag ID i The second associated feature v' ik It can be calculated using a weighted network model, and the calculation principle is as follows:

[0102] v' ik =v i w1c k

[0103] Where W1 is the weight matrix.

[0104] Since the weight matrix W can be obtained through continuous training and adjustment, it is understandable that the second association feature is used to characterize the dynamic feature relationship between each piece of information in each label and the item.

[0105] Furthermore, based on the first and second association features, the association features corresponding to the item and the information can be determined. For example, the average of the first and second association features can be used to obtain the candidate item of interest ID. k With sports interest tag ID i Corresponding related features Right now,

[0106]

[0107] Furthermore, in some embodiments, when calculating the association features between candidate items of interest and motion interest tags, for example, using the candidate item of interest ID... k With sports interest tag ID i For example, its corresponding association features You can also select the candidate project IDs mentioned above. k With sports interest tag ID i First associated feature v ik To perform calculations, that is, Or simply by the ID of the candidate project of interest k With sports interest tag ID i The second associated feature v' ik To perform calculations, that is,

[0108] Furthermore, it can be based on sports interest tag IDs i Based on the corresponding association features, determine the ID of the candidate item of interest. k With sports interest tag ID i correlation α ik For example, the correlation α between the numerical range of 0 and 1 can be obtained through normalization. ik This can eliminate The impact of singular data. α ik The specific calculation method is as follows:

[0109]

[0110] Furthermore, using sports interest tag IDs i For example, for candidate interested project IDs k When determining the second sub-label feature corresponding to a sports interest tag, the relevance of each piece of information in the sports interest tag can be used as the weight of each piece of information. The first sub-label features corresponding to each piece of information are then weighted to obtain the second sub-label feature corresponding to each piece of information. For example, for the sports interest tag ID... i Its corresponding second sub-label feature can be α ik v i Then, the second sub-label features corresponding to each piece of information in the sports interest tags are fused (e.g., summed) to obtain the second label feature of the sports interest tag. For example, the second sub-label features corresponding to each sports interest tag can be summed to obtain the second label feature V of the sports interest tag, that is:

[0111]

[0112] Where m is the number of tags contained in the motion interest tag, and V is the second tag feature corresponding to the motion interest tag of the k-th candidate interest item.

[0113] In summary, the second label feature V of the sports interest label can be obtained through the above calculation process. In addition, the second label feature U corresponding to the user's personal information can be determined by a similar calculation method.

[0114] For example, using the ID of the kth candidate project of interest among the candidate projects of interest. k And the j-th item of user personal information x contained in the user's personal information. j For example, candidate project IDs k With user personal information x j The first associated feature u jk The Hadamard product can be used to calculate the product, and the specific calculation method is as follows:

[0115]

[0116] Among them, u j For user personal information x j The corresponding first sub-label feature, c k For candidate project IDs kCorresponding project characteristics.

[0117] Candidate Interested Item ID k With user personal information x j The second association feature u' jk It can be calculated using a weighted network model, and the calculation principle is as follows:

[0118] u' jk =u j w1c k

[0119] Where W1 is the weight matrix.

[0120] The first and second association features mentioned above are added together and averaged to obtain the candidate item of interest ID. k With user personal information x j Corresponding related features Right now,

[0121]

[0122] Then, based on the aforementioned user personal information x j Corresponding related features Determine the ID of the candidate project of interest k With user personal information x j The correlation b jk For example, the correlation b of the numerical range between 0 and 1 can be obtained through normalization. jk b jk The specific calculation method is as follows:

[0123]

[0124] Furthermore, the second sub-label features corresponding to each user's personal information are summed to obtain the second label feature U of the user's personal information, namely:

[0125]

[0126] Where n is the number of tags contained in the sports interest tag, and U is the second tag feature corresponding to the sports interest tag of the kth candidate interest item.

[0127] In addition to the above-described methods for calculating the second tag feature of tag information (user personal information or sports interest tags), in some embodiments, the second tag feature of each type of tag information can also be calculated in the following ways:

[0128] For candidate project IDs k Taking the calculation of the second label feature of motion interest labels as an example, we can analyze the IDs of candidate interested items.k Project characteristics c k Transform to obtain candidate item IDs of interest k The corresponding query vector is calculated as follows:

[0129] Query=w q c k

[0130] In addition, for sports interest tag ID i The corresponding first sub-label feature v i Transform to obtain motion interest tag ID i Corresponding Key i (Key) and Value i (value),

[0131] Key i =w k v i

[0132] Value i =w v v i

[0133] Then, based on the candidate interested project ID k The corresponding query vector and motion interest tag ID i Corresponding Key i It is possible to determine the ID of the candidate project of interest. k With sports interest tag ID i Corresponding related features The specific calculation method is as follows:

[0134]

[0135] Furthermore, normalization can be used to obtain candidate item IDs with values ​​ranging from 0 to 1. k With sports interest tag ID i correlation α ik The specific calculation method is as follows:

[0136]

[0137] Furthermore, α can be... ik As the corresponding Value i The weight of the value i We perform weighted analysis to obtain the second sub-label feature α corresponding to each piece of information. ik Value iThen, the second sub-label features corresponding to each piece of information in the sports interest label are fused (e.g., summed) to obtain the second label feature V of the sports interest label. The specific calculation method is as follows:

[0138]

[0139] Among them, w q w k w v Query and Key respectively i Value i The corresponding weight matrix.

[0140] Therefore, in this embodiment, for each type of tag information, the corresponding second tag feature is a user feature vector that takes into account the correlation weight between the various information contained in the tag information and the candidate items of interest. Thus, when determining the matching degree of the candidate items of interest based on the second tag feature, the influence of the various information contained in the tag information is fully considered. Therefore, the target items recommended to the target user are more accurate.

[0141] In some embodiments of this application, for each candidate item of interest, the matching degree between the target user and the item is obtained based on the second tag feature of the two tag information and the item feature of the item, including:

[0142] For each candidate item of interest, the second label feature corresponding to the two label information and the item feature of the item are concatenated. Based on the concatenated feature, the matching degree of the item is obtained.

[0143] For example, the second tag feature corresponding to the two types of tag information and the project feature of the project are concatenated. Based on the concatenated feature, the matching degree corresponding to the project can be obtained through the second sub-model in the matching degree calculation model of this application embodiment.

[0144] like Figure 2 As shown, candidate interested item IDs k For example, after obtaining the second label feature V corresponding to the sports interest label for the project, and the second label feature U corresponding to the user's personal information for the project, the candidate interested project IDs can be... k The corresponding project features, as well as the second label features V and U, are concatenated, and the candidate project IDs are obtained based on the concatenated vector. k The corresponding matching degree.

[0145] Specifically, based on the candidate project ID kFor example, by increasing the number of channels, i.e., the contact method, the second label feature V corresponding to the sports interest label of the project, the second label feature U corresponding to the user's personal information of the project, and the project feature c of the candidate interested projects can be used. k Concatenate the vectors to obtain the concatenated vector. Right now,:

[0146]

[0147] Furthermore, in some embodiments, the second label feature V corresponding to the above-mentioned motion interest label, the second label feature U corresponding to the user's personal information, and the item feature c of the candidate interest item can also be added together by adding the feature maps, i.e., by the add method. k The vector is then concatenated to obtain the concatenated vector.

[0148] Furthermore, the above concatenated vectors can be... The input is fed into an MLP, and then a sigmoid transformation is performed to obtain the candidate item of interest ID. k Corresponding matching degree The specific calculation method is as follows:

[0149]

[0150] Where f is a multilayer perceptron.

[0151] in, It is a ReLU activation function. w, w0, b1, and b0 are the training parameters of the second sub-model.

[0152] Therefore, this embodiment determines the matching degree of a project based on the second tag features and the project features of the candidate projects of interest, so that the matching degree of each project takes into account the association between each tag information and the candidate projects of interest.

[0153] In some embodiments of this application, determining the relevance and identifying the target sport to recommend to the target user based on the relevance corresponding to each candidate sport of interest is achieved through a matching degree calculation model, which is trained in the following manner:

[0154] Obtain a training sample set. Each training sample in the training sample set includes: the sample user's personalized information and the sample user's real label for each candidate sport. The real label indicates whether the candidate sport is the sample user's target sport.

[0155] Each training sample is input into the initial neural network model to obtain the prediction result corresponding to each training sample. For each training sample, the prediction result includes the predicted matching degree of the sample user to each candidate item.

[0156] The training loss value is determined based on the true label and predicted matching degree of each training sample for each candidate item.

[0157] If the preset training termination condition is met, the training ends, and the neural network model at the end of training is used as the matching degree calculation model. If the training termination condition is not met, the model parameters of the neural network model are adjusted, and the adjusted model is trained again based on each training sample.

[0158] Specifically, the matching degree calculation model in this application embodiment may include a first sub-model for calculating the correlation between each type of tag information (user personal information or sports interest tag) and candidate items of interest, wherein the first sub-model may be an attention module, and a second sub-model for calculating the matching degree between the target user and the item based on the correlation between each type of tag information and the candidate items of interest.

[0159] During training, gradient descent can be used to continuously calculate the loss function of the matching degree based on the backpropagation principle, and the model parameters in the model can be adjusted according to the loss function of the matching degree. The loss function L of the matching degree can be calculated in the following way:

[0160]

[0161] Where H represents the number of sample users, and t represents the number of candidate items. y represents the matching degree of the candidate items calculated by the model. hk y represents the true match of candidate projects. hk ∈{0,1}.

[0162] Furthermore, the model parameters can be the weight matrices w1 and w2 in the first sub-model of the above embodiments. q w k w v And the weight matrices w and w0, and the biases b1 and b0 in the second sub-model.

[0163] During the training process described above, the number of training samples can be selected according to the actual situation. For example, in this embodiment, the number of training samples can be 32. Furthermore, 1% of neurons can be randomly removed during training to prevent overfitting and thus obtain better training results. In addition, the F1 score can be used as the evaluation metric for the model, and the training termination condition can include convergence of the training loss value, the training loss value being less than a preset threshold, or reaching a certain number of iterations, such as 100 iterations.

[0164] This application embodiment calculates the correlation between each piece of information in the user's personal information and the candidate items of interest, as well as the correlation between each piece of information in the sports interest tags and the candidate items of interest, through a matching degree calculation model. This fully considers the correlation between each piece of information in the user's personal information and the sports interest tags and the candidate items of interest, making the recommended sports more suitable for the user's various situations and more accurate.

[0165] In another embodiment of this application, step S102: obtaining at least one candidate item of interest corresponding to the target user may include:

[0166] Based on sports interest tags and a pre-configured sports knowledge graph, at least one candidate sport of interest for the target user is determined from multiple candidate sports included in the sports knowledge graph. The sports knowledge graph contains the association between each candidate sport and its corresponding sports type.

[0167] Specifically, the sports knowledge graph includes multiple nodes and multiple edges. Each node corresponds to a candidate sport or a sport type, and the edges are the connections between the candidate sport and the sport type to which the candidate sport belongs.

[0168] For each motion interest tag, the weight of each edge in the motion knowledge graph is determined by probabilistic walk. Based on the weight of each edge in the motion knowledge graph and the set path length, the optimal path with the node corresponding to the motion interest tag as the starting point is determined. The optimal path is the path with the largest sum of weights of the edges contained in the candidate paths with the node corresponding to the motion interest tag as the starting point.

[0169] The candidate sports items contained in the optimal path corresponding to each sports interest tag are identified as the candidate interest items of the target user.

[0170] Specifically, in this embodiment, the optimal path, starting from the node corresponding to the motion interest tag, can be used to represent the path with the highest degree of association with the node corresponding to the motion interest tag; that is, the optimal path in which each node has the highest degree of association with the node corresponding to the motion interest tag. For each edge, the weight represents the degree of association between the nodes connected by that edge.

[0171] In this embodiment, during the determination of the optimal path, the node corresponding to the motion interest tag can be used as the starting point. Among the nodes connected to the node corresponding to the motion interest tag, the node with the highest correlation is selected as the next node. Furthermore, for each determined node, the node with the highest correlation among its connected nodes can be selected as the next node, following the same method. In this process, the correlation between nodes can be calculated using the probabilistic walk strategy described in detail below, and the correlation between nodes can be used as the weight of the edge connecting two nodes. Therefore, the determined optimal path is the path with the largest sum of edge weights among the candidate paths starting from the node corresponding to the motion interest tag.

[0172] Combination Figure 3 This is a schematic diagram illustrating the scenario of determining the optimal path corresponding to the sports interest tag in this application embodiment. For ease of description, the diagram exemplarily shows a sports knowledge graph containing nodes a, b, c1, c2, and c3. In this sports knowledge graph, sports items and their respective sports types are connected by edges. Node b is aerobic exercise, i.e., a sports type node. The nodes a, c1, c2, and c3 connected to it are the sports items included in the aerobic exercise (sports type) corresponding to node b, namely, aerobics (node ​​a), aerobics (node ​​c1), cycling (node ​​c2), and running (node ​​c3), respectively. The aerobics corresponding to node a can also be the sports type of the aerobics corresponding to node c1, so node a can be connected to node c1.

[0173] Taking the determination of the optimal path corresponding to node a as an example, firstly, taking node a as the starting point, when determining the next node of node a, we can choose the node with the highest degree of association with node a among the nodes connected to node a. The degree of association between the two nodes, i.e., the weight of the edge connecting the two nodes, can be determined through a probabilistic walk strategy. For example, when selecting the next node of the starting node, the weight P of each edge... ax It can be calculated in the following ways:

[0174]

[0175] Where E(a) is the set of adjacent nodes of node a, and da E(a) contains the number of nodes.

[0176] like Figure 3 As shown, nodes connected to node a are node b and node c1. Using the calculation method described above, the weight values ​​of edges ab and ac1 connected to a are both 1 / 2. In this case, since the weight values ​​of each edge are equal, a node can be randomly selected as the next node of node a. Furthermore, for nodes b and c1, the next node can be selected for each node separately. Then, based on the paths determined by nodes b and c1 respectively, the path with the largest sum of weight values ​​among the paths determined by nodes b and c1 is selected as the optimal path.

[0177] Taking node b as the next node after node a as an example, since node b is not the starting node, the edges connected to node b (such as...) Figure 3 The weight P of edges bc1, bc2, bc3 shown in the figure by It can be calculated in the following ways:

[0178]

[0179] Where E(b) is the set of nodes adjacent to node b, and d b It is the number of nodes contained in E(b).

[0180]

[0181] Where d(a,y) is the number of edges contained in the shortest path between node y and the previous node (node ​​a) of node b, and d(a,y)∈{0,1,2}.

[0182] Furthermore, in the probabilistic walk strategy, p is the return probability, which is the probability that p controls node b to repeatedly visit the previous node. In this embodiment, it can be characterized as the probability that the next node is determined to be node a when selecting the next node of b (hence the name p as the return probability). If p is large, the probability of visiting the previous node in the next walk is smaller, and vice versa.

[0183] In the probabilistic walk strategy, q represents the input-output probability, which controls the strategy type of node b's next random move (hence the term q as input-output probability). That is, in this embodiment, it can be characterized as the strategy type for selecting node b's next node. If q is relatively large, vertex b's next move tends to visit the nearest neighbor of the vertex in the previous step, i.e., Breadth First Search (BFS); if q is relatively small, vertex b's next move tends to visit the vertex farther from the vertex in the previous step, i.e., Depth First Search (DFS).

[0184] Experiments show that p = 0.25 and q = 0.25 are the most favorable values ​​for selecting the optimal path.

[0185] Therefore, the weight values ​​of each edge connected to node b can be obtained through the above calculations. For example, when p = 0.25 and q = 0.25, the weight P of the edge between b and c2 can be calculated. bc2 For example, since the shortest path between node a (the node preceding node b) and node c2 contains 2 edges, therefore, α pq (a,y)=1 / q=1 / 0.25=4, and node b has 4 adjacent nodes. Therefore,

[0186] Furthermore, for the edge bc1 between b and c1, the shortest path between node a (the node preceding node b) and node c1 contains 1 edge. Therefore, α pq Given (a,y)=1, the weight of the edge between b and c1 can be calculated. For the edge bc3 between b and c3, the shortest path between node a (the node preceding node b) and node c3 contains 2 edges. Therefore, α pq (a,y)=1 / q=1 / 0.25=4, therefore, the weight of the edge between b and c3 can be calculated.

[0187] Therefore, we can obtain the weights of each edge connected to node b. We can choose the node connected to the edge with the largest weight value as the next node of node b. Since there are two nodes corresponding to the edge with the largest weight value, namely c2 and c3, we can randomly select one node as the next node of node b.

[0188] Similarly, by iteratively executing the above method, a preset number of nodes are selected, and the path formed by these selected nodes is taken as the optimal path. For example, the preset number of nodes could be 5, meaning the optimal path could include 5 nodes. If a selected node has no next node (i.e., it is the last node in its path), and the number of selected nodes has not yet reached the preset number, the path formed by the selected nodes can be taken as the optimal path. For example, if the preset number of nodes is 5, but the second node determined during the process of determining the optimal path is the end point of a path (i.e., it has no next node), then the selection process can stop, and the path formed by the two determined nodes can be taken as the optimal path. In this embodiment, each motion interest tag can correspond to one optimal path.

[0189] Therefore, it can be understood that in this embodiment, the optimal path corresponding to each motion interest tag is related to the values ​​of p (return probability) and q (input-output probability), as well as the graph connection form formed by each node and edge in the knowledge graph.

[0190] Furthermore, after determining the optimal path corresponding to each sports interest tag, candidate sports items corresponding to the nodes in each optimal path can be selected, duplicate candidate sports items can be removed, and the remaining candidate sports items can be used as the target user's candidate interest items.

[0191] This embodiment determines the optimal path of a sports interest tag in a sports knowledge graph, and then determines candidate items of interest. Since the knowledge graph contains multiple candidate sports items and sports types, as well as the degree of association between them, the candidate sports item with the highest degree of association with the sports interest tag is determined as the candidate item of interest based on the degree of association between the sports interest tag and each of the candidate sports items and sports types associated with it, thus realizing the expansion of the candidate items of interest.

[0192] In another embodiment of this application, step S101, obtaining the sports interest tags that the target user is interested in, includes:

[0193] Obtain the target users' real sports interest tags;

[0194] Based on actual sports interest tags and prior knowledge of sports, potential sports interest tags are identified. The sports interest tags that the target users are interested in include both actual and potential sports interest tags.

[0195] Among them, the prior knowledge of motion includes the first interest probability corresponding to each target information and the second interest probability corresponding to any two target information, where the target information is a candidate sport or sport type;

[0196] Prior knowledge of motion is obtained in the following ways:

[0197] Acquire data from multiple sample users, each sample user data including the actual motion interest label of a sample user;

[0198] For each target information, the probability of first interest is determined based on the proportion of sample users who are interested in that target information in all sample user data.

[0199] For any two target information pieces, the probability of the second interest is determined based on the proportion of sample users who are simultaneously interested in both target information pieces in all sample user data.

[0200] Specifically, the target user's real sports interest tags are the sports projects and / or sports types that the user is actually interested in. These real sports interest tags can be obtained by receiving sports interest information input by the target user through the terminal device, or by obtaining user information databases. The user information databases can be constructed by obtaining user information through questionnaires.

[0201] In addition, potential sports interest tags can be predicted to be the sports interests that users may be interested in. For example, potential sports interest tags can be determined based on real sports interest tags and prior knowledge of sports.

[0202] Specifically, prior knowledge of motion can be obtained through statistical analysis of a large amount of sample user data. Among them, the first interest probability corresponding to the target information can be used to describe the probability that the user is interested in the target information, and the second interest probability corresponding to any two target information can be used to describe the probability that the user is interested in any two target information at the same time.

[0203] The determination of the first probability of interest for each target information in the prior knowledge of motion, and the second probability of interest for any two target information, is illustrated through the following example:

[0204] For ease of description, sample user data from 5 sample users are selected for illustration, as shown in the following table. For example, the table contains 5 sample users, and the actual sports interest tags of the above 5 sample users involve a total of 3 sports interest tags, namely C1, C2, and C3. In the table, "1" indicates that the sports interest tag is the actual sports interest tag of the corresponding user, that is, the user is interested in the sports interest tag, and "0" indicates that the sports interest tag is not the actual sports interest tag of the corresponding user, that is, the user is not interested in the sports interest tag.

[0205] User 1 1 0 1 User 2 0 1 1 User 3 1 1 1 User 4 0 1 1 User 5 1 0 0

[0206] Table 1

[0207] For example, as shown in Table 1, User 1 is interested in C1 and C3, User 2 is interested in C2 and C3, and User 3 is interested in C1, C2, and C3, etc. Therefore, taking the first probability of interest corresponding to C1 as an example, based on the above sample user data, we can obtain that the users interested in C1 are User 1, User 3, and User 5. Therefore, the first probability of interest corresponding to C1 is the number of users interested in it / the number of sample users, that is, P(C1) = 3 / 5. Furthermore, taking the second probability of interest corresponding to C1 and C3 as an example, the users interested in both C1 and C3 are User 1 and User 3. The second probability of interest corresponding to C1 and C3 is the number of users interested in both C1 and C3 / the number of sample users, that is, P(C1C3) = 2 / 5.

[0208] Therefore, when determining potential sports interest tags based on real sports interest tags and prior sports knowledge, a pre-determined probability calculation model can be used to calculate the probability that a user is interested in other target information in prior sports knowledge when they are interested in one or more real sports interest tags. Candidate sports items or sports types corresponding to target information with an interest probability greater than a preset threshold are determined as potential sports interest tags.

[0209] The following example uses a user's actual exercise interest tags C1 and C2. Using a probability calculation model, the probability P(C3|C1C2) that the user is interested in C3 is calculated. The formula is as follows:

[0210]

[0211] Where, P(C1|C3)=P(C1C3) / P(C1), P(C2|C3)=P(C2C3) / P(C2).

[0212] P(C1) is the first interest probability corresponding to C1, and P(C1C3) is the second interest probability corresponding to C1 and C3.

[0213] It should be noted that in the above example, the sports interest labels C1, C2, and C3 refer to sports events or sports types, respectively. For example, prior knowledge of sports can include the first probability of interest corresponding to a sports event and the second probability of interest corresponding to any two sports events. Thus, based on the actual sports interest events and prior knowledge of sports, potential sports interest events can be determined. Similarly, potential sports interest types can be determined.

[0214] In this embodiment, potential sports interest tags are determined based on real sports interest tags and prior sports knowledge. Since the prior sports knowledge includes statistical patterns from multiple sample user data, the system can predict users' potential sports interest tags based on big data statistical results.

[0215] This application provides a method for recommending exercise. The method can be executed by any electronic device, such as an application with exercise recommendation functionality. The method can be executed by the application's server or a terminal device. Figure 4 As shown, this is a second flowchart illustrating an exercise recommendation method provided in an embodiment of this application. The method may include:

[0216] S401: Obtain sports interest tags that the target user is interested in.

[0217] Specifically, sports interest tags can be sports information that the target user is interested in, including the type of sports and / or the sports activities that the target user is interested in. For example, it can include the type of sports and / or the sports activities that the target user is interested in. Specifically, the type of sports can include, for example, ball sports, water sports, gymnastics, etc., and the sports activities can include, for example, badminton, swimming, aerobics, etc.

[0218] In addition, the aforementioned sports interest tags can be obtained by receiving personalized user information input by the target user through the terminal device, or by obtaining a personalized user information database. The personalized user information database can be constructed by obtaining user information through questionnaires or by analyzing user historical behavior data.

[0219] S402: Based on the sports interest tags and the pre-configured sports knowledge graph, determine at least one candidate sport of interest for the target user from multiple candidate sports included in the sports knowledge graph, wherein the sports knowledge graph contains the association between each candidate sports and its corresponding sports type.

[0220] Specifically, a sports knowledge graph can be constructed using multiple sample user data sets. Each sample user data set includes the sports interest tags of the sample user. A neural network model can be used to determine the relationship between the sports items in the sports interest tags of each sample user and the sports types to which they belong.

[0221] S403: Based on each candidate item of interest, determine the target sport to recommend to the target user.

[0222] This application embodiment determines at least one candidate sport of interest for a target user from multiple candidate sport items included in the sport knowledge graph based on sport interest tags and a pre-configured sport knowledge graph. Based on each candidate sport of interest, it determines the target sport to recommend to the target user. This realizes the expansion of sport interest tags to obtain associated candidate sport of interest based on the user's own sport interest tags and the relationship between each candidate sport item and sport type in the sport knowledge graph, so as to recommend a richer range of target sport items that are more suitable for the user's own situation.

[0223] In another embodiment of this application, the knowledge graph includes multiple nodes and multiple edges, wherein each node corresponds to a candidate sport or a sport type, and the edge is the connection between the candidate sport and the sport type to which the candidate sport belongs.

[0224] Step S402 may specifically include:

[0225] For each motion interest tag, the weight of each edge in the motion knowledge graph is determined by probabilistic walk. Based on the weight of each edge in the motion knowledge graph and the set path length, the optimal path with the node corresponding to the motion interest tag as the starting point is determined. The optimal path is the path with the largest sum of weights of the edges among the candidate paths with the node corresponding to the motion interest tag as the starting point.

[0226] Based on the candidate sports items contained in the optimal path corresponding to each sports interest tag, the candidate interest items of the target user are determined.

[0227] Specifically, in this embodiment, the optimal path, starting from the node corresponding to the motion interest tag, can be used to represent the path with the highest degree of association with the node corresponding to the motion interest tag; that is, each node in the optimal path has the highest degree of association with the node corresponding to the motion interest tag. The weight represents the degree of association between the two nodes connected by the edge.

[0228] In this embodiment, during the determination of the optimal path, the node corresponding to the motion interest tag can be used as the starting point. Among the nodes connected to the node corresponding to the motion interest tag, the node with the highest correlation is selected as the next node. Furthermore, for each determined node, the node with the highest correlation among its connected nodes can be selected as the next node, following the same method. In this process, the correlation between nodes can be calculated using the probabilistic walk strategy described in detail below, and the correlation between nodes can be used as the weight of the edge connecting two nodes. Therefore, the determined optimal path is the path with the largest sum of edge weights among the candidate paths starting from the node corresponding to the motion interest tag.

[0229] Combination Figure 3 This is a schematic diagram illustrating the scenario of determining the optimal path corresponding to the sports interest tag in this application embodiment. For ease of description, the diagram exemplarily shows a sports knowledge graph containing nodes a, b, c1, c2, and c3. In this sports knowledge graph, sports items and their respective sports types are connected by edges. Node b is aerobic exercise, i.e., a sports type node. The nodes a, c1, c2, and c3 connected to it are the sports items included in the aerobic exercise (sports type) corresponding to node b, namely, aerobics (node ​​a), aerobics (node ​​c1), cycling (node ​​c2), and running (node ​​c3), respectively. The aerobics corresponding to node a can also be the sports type of the aerobics corresponding to node c1, so node a can be connected to node c1.

[0230] Taking the determination of the optimal path corresponding to node a as an example, firstly, taking node a as the starting point, when determining the next node of node a, we can choose the node with the highest degree of association with node a among the nodes connected to node a. The degree of association between the two nodes, i.e., the weight of the edge connecting the two nodes, can be determined through a probabilistic walk strategy. For example, when selecting the next node of the starting node, the weight P of each edge... ax It can be calculated in the following ways:

[0231]

[0232] Where E(a) is the set of adjacent nodes of node a, and d a E(a) contains the number of nodes.

[0233] As shown in Figure 3, nodes b and c1 are connected to node a. Using the calculation method described above, the weights of the edges ab and ac1 connected to a are both 1 / 2. In this case, since the weights of each edge are equal, a node can be randomly selected as the next node of node a. Furthermore, for nodes b and c1, the next node can be selected for each node separately. Then, based on the paths determined by nodes b and c1 respectively, the path with the largest sum of weights among the paths determined by nodes b and c1 is selected as the optimal path.

[0234] Taking node b as the next node after node a as an example, since node b is not the starting node, the edges connected to node b (such as...) Figure 3 The weight P of edges bc1, bc2, bc3 shown in the figure by It can be calculated in the following ways:

[0235]

[0236] Where E(b) is the set of nodes adjacent to node b, and d b It is the number of nodes contained in E(b).

[0237]

[0238] Where d(a,y) is the number of edges contained in the shortest path between node y and the previous node (node ​​a) of node b, and d(a,y)∈{0,1,2}.

[0239] Furthermore, in the probabilistic walk strategy, p is the return probability, which is the probability that p controls node b to repeatedly visit the previous node. In this embodiment, it can be characterized as the probability that the next node is determined to be node a when selecting the next node of b (hence the name p as the return probability). If p is large, the probability of visiting the previous node in the next walk is smaller, and vice versa.

[0240] In the probabilistic walk strategy, q represents the input-output probability, which controls the strategy type of node b's next random move (hence the term q as input-output probability). That is, in this embodiment, it can be characterized as the strategy type for selecting node b's next node. If q is relatively large, vertex b's next move tends to visit the nearest neighbor of the vertex in the previous step, i.e., Breadth First Search (BFS); if q is relatively small, vertex b's next move tends to visit the vertex farther from the vertex in the previous step, i.e., Depth First Search (DFS).

[0241] Experiments show that p = 0.25 and q = 0.25 are the most favorable values ​​for selecting the optimal path.

[0242] Therefore, the weight values ​​of each edge connected to node b can be obtained through the above calculations. For example, when p = 0.25 and q = 0.25, the weight P of the edge between b and c2 can be calculated. bc2 For example, since the shortest path between node a (the node preceding node b) and node c2 contains 2 edges, therefore, α pq (a,y)=1 / q=1 / 0.25=4, and node b has 4 adjacent nodes. Therefore,

[0243] Furthermore, for the edge bc1 between b and c1, the shortest path between node a (the node preceding node b) and node c1 contains 1 edge. Therefore, α pq Given (a,y)=1, the weight of the edge between b and c1 can be calculated. For the edge bc3 between b and c3, the shortest path between node a (the node preceding node b) and node c3 contains 2 edges. Therefore, α pq (a,y)=1 / q=1 / 0.25=4, therefore, the weight of the edge between b and c3 can be calculated.

[0244] Therefore, we can obtain the weights of each edge connected to node b. We can choose the node connected to the edge with the largest weight value as the next node of node b. Since there are two nodes corresponding to the edge with the largest weight value, namely c2 and c3, we can randomly select one node as the next node of node b.

[0245] Similarly, by iteratively executing the above method, a preset number of nodes are selected, and the path formed by these selected nodes is taken as the optimal path. For example, the preset number of nodes could be 5, meaning the optimal path could include 5 nodes. If a selected node has no next node (i.e., it is the last node in its path), and the number of selected nodes has not yet reached the preset number, the path formed by the selected nodes can be taken as the optimal path. For example, if the preset number of nodes is 5, but the second node determined during the process of determining the optimal path is the end point of a path (i.e., it has no next node), then the selection process can stop, and the path formed by the two determined nodes can be taken as the optimal path. In this embodiment, each motion interest tag can correspond to one optimal path.

[0246] Therefore, it can be understood that in this embodiment, the optimal path corresponding to each motion interest tag is related to the values ​​of p (return probability) and q (input-output probability), as well as the graph connection form formed by each node and edge in the knowledge graph.

[0247] Furthermore, after determining the optimal path corresponding to each sports interest tag, candidate sports items corresponding to the nodes in each optimal path can be selected, duplicate candidate sports items can be removed, and the remaining candidate sports items can be used as the target user's candidate interest items.

[0248] This embodiment determines the optimal path of a sports interest tag in a sports knowledge graph, and then determines candidate items of interest. Since the knowledge graph contains multiple candidate sports items and sports types, as well as the degree of association between them, the candidate sports item with the highest degree of association with the sports interest tag is determined as the candidate item of interest based on the degree of association between the sports interest tag and each of the candidate sports items and sports types associated with it, thus realizing the expansion of the candidate items of interest.

[0249] In another embodiment of this application, step S401 may specifically include:

[0250] Obtain the target users' real sports interest tags;

[0251] Based on actual sports interest tags and prior knowledge of sports, potential sports interest tags are identified. The sports interest tags that the target users are interested in include both actual and potential sports interest tags.

[0252] Among them, the prior knowledge of motion includes the first interest probability corresponding to each target information and the second interest probability corresponding to any two target information, where the target information is a candidate sport or sport type;

[0253] Prior knowledge of motion is obtained in the following ways:

[0254] Acquire data from multiple sample users, each sample user data including the actual motion interest label of a sample user;

[0255] For each target information, the probability of first interest is determined based on the proportion of sample users who are interested in that target information in all sample user data.

[0256] For any two target information pieces, the probability of the second interest is determined based on the proportion of sample users who are simultaneously interested in both target information pieces in all sample user data.

[0257] Specifically, the target user's real sports interest tags are the sports projects and / or sports types that the user is actually interested in. These real sports interest tags can be obtained by receiving sports interest information input by the target user through the terminal device, or by obtaining user information databases. The user information databases can be constructed by obtaining user information through questionnaires.

[0258] In addition, potential sports interest tags can be predicted to be the sports interests that users may be interested in. For example, potential sports interest tags can be determined based on real sports interest tags and prior knowledge of sports.

[0259] Specifically, prior knowledge of motion can be obtained through statistical analysis of a large amount of sample user data. Among them, the first interest probability corresponding to the target information can be used to describe the probability that the user is interested in the target information, and the second interest probability corresponding to any two target information can be used to describe the probability that the user is interested in any two target information at the same time.

[0260] The determination of the first probability of interest for each target information in the prior knowledge of motion, and the second probability of interest for any two target information, is illustrated through the following example:

[0261] For ease of description, sample user data from 5 sample users are selected for illustration, as shown in the following table. For example, the table contains 5 sample users, and the actual sports interest tags of the above 5 sample users involve a total of 3 sports interest tags, namely C1, C2, and C3. In the table, "1" indicates that the sports interest tag is the actual sports interest tag of the corresponding user, that is, the user is interested in the sports interest tag, and "0" indicates that the sports interest tag is not the actual sports interest tag of the corresponding user, that is, the user is not interested in the sports interest tag.

[0262] User 1 1 0 1 User 2 0 1 1 User 3 1 1 1 User 4 0 1 1 User 5 1 0 0

[0263] Table 1

[0264] For example, as shown in Table 1, User 1 is interested in C1 and C3, User 2 is interested in C2 and C3, and User 3 is interested in C1, C2, and C3, etc. Therefore, taking the first probability of interest corresponding to C1 as an example, based on the above sample user data, we can obtain that the users interested in C1 are User 1, User 3, and User 5. Therefore, the first probability of interest corresponding to C1 is the number of users interested in it / the number of sample users, that is, P(C1) = 3 / 5. Furthermore, taking the second probability of interest corresponding to C1 and C3 as an example, the users interested in both C1 and C3 are User 1 and User 3. The second probability of interest corresponding to C1 and C3 is the number of users interested in both C1 and C3 / the number of sample users, that is, P(C1C3) = 2 / 5.

[0265] Therefore, when determining potential sports interest tags based on real sports interest tags and prior sports knowledge, a pre-determined probability calculation model can be used to calculate the probability that a user is interested in other target information in prior sports knowledge when they are interested in one or more real sports interest tags. Candidate sports items or sports types corresponding to target information with an interest probability greater than a preset threshold are determined as potential sports interest tags.

[0266] The following example uses a user's actual exercise interest tags C1 and C2. Using a probability calculation model, the probability P(C3|C1C2) that the user is interested in C3 is calculated. The formula is as follows:

[0267]

[0268] Where, P(C1|C3)=P(C1C3) / P(C1), P(C2|C3)=P(C2C3) / P(C2).

[0269] P(C1) is the first interest probability corresponding to C1, and P(C1C3) is the second interest probability corresponding to C1 and C3.

[0270] It should be noted that in the above example, the sports interest labels C1, C2, and C3 refer to sports events or sports types, respectively. For example, prior knowledge of sports can include the first probability of interest corresponding to a sports event and the second probability of interest corresponding to any two sports events. Thus, based on the actual sports interest events and prior knowledge of sports, potential sports interest events can be determined. Similarly, potential sports interest types can be determined.

[0271] In this embodiment, potential sports interest tags are determined based on real sports interest tags and prior sports knowledge. Since the prior sports knowledge includes statistical patterns from multiple sample user data, it is possible to predict users' potential sports interest tags based on big data statistical results.

[0272] In another embodiment of this application, the method may further include:

[0273] Obtain the target user's personal information.

[0274] Based on the candidate items of interest, target sports items are determined to be recommended to the target user, including:

[0275] For each candidate item of interest, determine the association between each type of tag information and that candidate item of interest, wherein the types of tag information are at least one of user personal information and sports interest tags;

[0276] Based on the relevance of each candidate item of interest, target sports items are determined from each candidate item of interest and recommended to the target user.

[0277] User personal information may include the target user's personal attributes and physical health information, such as gender, age, height, weight, and physical health information such as current medical history and past medical history.

[0278] In addition, the aforementioned user personal information can be obtained by receiving user information input from target users through terminal devices, or it can be obtained through a user information database. The user information database can be constructed by user information obtained through questionnaires or by analyzing user historical behavior data.

[0279] Determine the association between each tag and the candidate item of interest, including:

[0280] Obtain the first tag feature for each tag information, and the project feature for each candidate item of interest;

[0281] For each candidate item of interest, the relevance of each tag information is determined based on the item characteristics of the item and the first tag characteristics of each tag information.

[0282] Based on the correlation between the two types of tag information corresponding to each candidate item of interest, target sports items recommended to the target user are determined from each candidate item of interest, including:

[0283] For each candidate item of interest, the first label feature of each label information is weighted according to the correlation between each label information and the item, so as to obtain the second label feature corresponding to each label information of the item.

[0284] For each candidate item of interest, the matching degree between the target user and the item is obtained based on the second tag feature of the two tag information and the item feature of the item.

[0285] Based on the matching degree of each candidate interest item, the target sports item recommended to the target user is determined from each candidate interest item.

[0286] For each of the candidate items of interest, the correlation between each type of label information and the candidate item of interest is determined by the first sub-model.

[0287] For each of the candidate items of interest, the matching degree between the target user and the item is determined through a second sub-model.

[0288] Furthermore, determining the relevance and identifying the target sport to recommend to the target user based on the relevance corresponding to each candidate item of interest is achieved through a matching degree calculation model. The matching degree calculation model includes the first sub-model and the second sub-model, and is trained in the following way:

[0289] Obtain a training sample set. Each training sample in the training sample set includes: the sample user's personalized information and the sample user's real label for each candidate sport. The real label indicates whether the candidate sport is the sample user's target sport.

[0290] Each training sample is input into the initial neural network model to obtain the prediction result corresponding to each training sample. For each training sample, the prediction result includes the predicted matching degree of the sample user to each candidate item.

[0291] The training loss value is determined based on the true label and predicted matching degree of each training sample for each candidate item.

[0292] If the preset training termination condition is met, the training ends, and the neural network model at the end of training is used as the matching degree calculation model. If the training termination condition is not met, the model parameters of the neural network model are adjusted, and the adjusted model is trained again based on each training sample.

[0293] Specifically, the correlation between each tag information and the candidate item of interest can be used to characterize the relationship between each tag information and the candidate item of interest, such as the matching relationship between each tag information and the candidate item of interest.

[0294] In this embodiment, the correlation between each type of tag information and the candidate item of interest is determined. Based on the correlation corresponding to each candidate item of interest, the specific implementation method for determining the target sports item to be recommended to the target user from each candidate item of interest is the same as the specific implementation method in the above embodiment for calculating the matching degree of the candidate sports item through the matching degree calculation model and determining the target sports item to be recommended to the target user based on the matching degree of each candidate sports item. Furthermore, the training method of the matching degree calculation model involved is also the same, and will not be repeated here.

[0295] This application's embodiments determine the correlation between each type of tag information and the candidate item of interest; based on the correlation corresponding to each candidate item of interest, a target sport to be recommended to the target user is determined from among the candidate items of interest. Thus, by determining the target sport to be recommended to the target user based on the correlation between the user's personal information and sport interest tags and the candidate items of interest, the recommendation process fully considers the user's personal attributes and sport interests, resulting in recommended sport items that are more suitable for the user's own situation, achieving personalized sport recommendations.

[0296] Figure 1 and Figure 4 The exercise recommendation method embodiments shown can be applied individually or in combination. To better illustrate the exercise recommendation method of this application, a specific embodiment will be described below.

[0297] In practical scenarios, the system can first receive personalized user information input by the target user via a terminal device or obtain personalized user information from a user personalization database. This personalized information can include the user's personal information and sports interest tags. Here, sports interest tags refer to the target user's actual sports interest tags. Based on these actual sports interest tags and combined with the prior sports knowledge involved in the above embodiments, the potential sports interest tags of the target user are predicted. Then, the actual and potential sports interest tags are used as the target user's interest entity set. Based on all sports interest tags in this interest entity set and the sports knowledge graph in the above embodiments, the optimal path corresponding to each sports interest tag in the interest entity set is determined in the sports knowledge graph. The candidate sports items involved in the determined set of optimal paths are then used as the user's candidate interest items. Furthermore, for each candidate interest item, the matching degree calculation model involved in the above embodiments is used to calculate the matching degree between the candidate interest item and the target user. Then, according to the matching degree from largest to smallest, a preset number of candidate interest items are selected as target sports items to recommend to the target user.

[0298] This application provides an exercise recommendation device, such as... Figure 5 As shown, the exercise recommendation device 50 may include: a first acquisition module 501, a second acquisition module 502, a correlation determination module 503, and a target exercise determination module 504, wherein,

[0299] The first acquisition module 501 is used to acquire the user's personalized information, which includes the user's personal information and sports interest tags, including the sports types and / or sports projects that the target user is interested in; the second acquisition module 502 is used to acquire at least one candidate project of interest corresponding to the target user.

[0300] Relevance determination module 503,

[0301] This is used to treat user personal information and sports interest tags as separate tags, and for each candidate item of interest, to determine the association between each tag and that candidate item of interest.

[0302] The target sport determination module 504 is used to determine the target sport to recommend to the target user from among the candidate sport of interest based on the correlation between the two types of tag information corresponding to each candidate sport of interest. In another embodiment of this application, the correlation determination module 503 may include:

[0303] The feature acquisition unit is used to acquire the first label feature for each type of label information, as well as the project feature for each candidate item of interest;

[0304] The correlation determination unit is used to determine the correlation of each tag information for each candidate item of interest based on the item characteristics of the item and the first tag characteristics of each tag information.

[0305] The target sport determination module 504 may include:

[0306] The feature determination unit is used to, for each candidate item of interest, weight the first label feature of each label information based on the correlation between each label information and the item, and obtain the second label feature corresponding to each label information of the item.

[0307] The matching degree determination unit is used to determine the matching degree between the target user and the project for each candidate item of interest, based on the second tag feature of the two tag information and the project feature of the project.

[0308] The sports activity determination unit is used to determine the target sports activity to recommend to the target user based on the matching degree of each candidate interest activity.

[0309] In another embodiment of this application,

[0310] For each type of label information, the label information includes at least one piece of information, and the first label feature of each type of label information includes the first sub-label feature corresponding to each piece of information;

[0311] The correlation determination unit is specifically used to determine the correlation between each piece of information in each type of label information and the item feature of the item based on the first sub-label feature corresponding to the information and the item feature of the item.

[0312] The feature determination unit is specifically used for,

[0313] For each candidate item of interest, the relevance of each item in each tag information to the item is used, and the first sub-tag feature corresponding to each item of information is weighted to obtain the second sub-tag feature corresponding to that item of information.

[0314] For each type of tag information, the second sub-tag features of all the information contained in that tag information are fused to obtain the second tag feature of that tag information. In another embodiment of this application, the association determination unit is further configured to,

[0315] The first associated feature is obtained by multiplying the first sub-label feature corresponding to this information with the feature value at the corresponding position in the project feature of this item.

[0316] Input the first sub-label feature corresponding to this information and the project feature of this project into the weighted network model to obtain the second associated feature;

[0317] Based on the first and second association features, determine the association between this item and this information.

[0318] In another embodiment of this application, the matching degree determination unit is specifically used for,

[0319] For each candidate item of interest, the second label feature corresponding to the two label information and the item feature of the item are concatenated. Based on the concatenated feature, the matching degree of the item is obtained.

[0320] In another embodiment of this application, for each of the candidate items of interest, determining the correlation between each type of tag information and the candidate item of interest is achieved through a first sub-model.

[0321] In another embodiment of this application, for each of the candidate items of interest, the matching degree between the target user and the item is determined by a second sub-model.

[0322] In another embodiment of this application, determining the relevance and identifying the target sports item to recommend to the target user based on the relevance corresponding to each candidate item of interest is achieved through a matching degree calculation model. The matching degree calculation model includes the first sub-model and the second sub-model, and the matching degree calculation model is trained in the following manner:

[0323] Obtain a training sample set. Each training sample in the training sample set includes: the sample user's personalized information and the sample user's real label for each candidate sport. The real label indicates whether the candidate sport is the sample user's target sport.

[0324] Each training sample is input into the initial neural network model to obtain the prediction result corresponding to each training sample. For each training sample, the prediction result includes the predicted matching degree of the sample user to each candidate item.

[0325] The training loss value is determined based on the true label and predicted matching degree of each training sample for each candidate item.

[0326] If the training loss value meets the preset training termination condition, the training ends, and the neural network model at the end of training is used as the matching degree calculation model. If the training termination condition is not met, the model parameters of the neural network model are adjusted, and the adjusted model is trained again based on each training sample.

[0327] In another embodiment of this application, the second acquisition module is specifically used to determine at least one candidate interest item of the target user from multiple candidate sports items contained in the sports knowledge graph based on sports interest tags and a pre-configured sports knowledge graph, wherein the sports knowledge graph contains the association between each candidate sports item and its sports type among multiple candidate sports items. In another embodiment of this application, the first acquisition module is specifically used to: acquire the target user's real sports interest tags.

[0328] Based on real sports interest tags and prior sports knowledge, we determine the potential sports interest tags of the target users. The sports interest tags that the target users are interested in include real sports interest tags and potential sports interest tags.

[0329] Among them, the prior knowledge of motion includes the first interest probability corresponding to each target information and the second interest probability corresponding to any two target information, where the target information is a candidate sport or sport type, and the two target information are information of the same type;

[0330] Prior knowledge of motion is obtained in the following ways:

[0331] Acquire data from multiple sample users, each sample user data including the actual motion interest label of a sample user;

[0332] For each target information, the first probability of interest for that target information is determined based on the proportion of sample users who are interested in that target information in all sample user data.

[0333] For any two target information pieces, the second interest probability corresponding to the two target information pieces is determined based on the proportion of sample users who are simultaneously interested in both target information pieces in all sample user data.

[0334] This application embodiment obtains the personalized information of the target user, including the target user's personal information and sports interest tags, and obtains at least one candidate interest item corresponding to the target user. The user's personal information and sports interest tags are each treated as a type of tag information. For each candidate interest item, the correlation between each type of tag information and that candidate interest item is determined. Based on the correlation corresponding to each candidate interest item, a target sports item to be recommended to the target user is determined from among the candidate interest items. Therefore, by determining the target sports item to be recommended to the target user based on the correlation between the user's personal information and sports interest tags and the candidate interest items, the recommendation process fully considers the user's personal attributes and sports interests, resulting in recommended sports items that are more suitable for the user's own situation, thus achieving personalized sports recommendations.

[0335] This application provides an exercise recommendation device, such as... Figure 6 As shown, the exercise recommendation device 60 may include: a tag acquisition module 601, a candidate interest item determination module 602, and a target exercise item determination module 603, wherein,

[0336] The tag acquisition module 601 is used to acquire sports interest tags that the target user is interested in, wherein the sports interest tags include the types of sports and / or sports that the target user is interested in;

[0337] The candidate interest item determination module 602 is used to determine at least one candidate interest item of the target user from multiple candidate sports items contained in the sports knowledge graph based on sports interest tags and a pre-configured sports knowledge graph. The sports knowledge graph contains the association relationship between each candidate sports item and its corresponding sports type among the multiple candidate sports items.

[0338] Target sport determination module 603, used for

[0339] Based on each candidate item of interest, a target sport is selected to recommend to the target user.

[0340] In another embodiment of this application, the knowledge graph includes multiple nodes, multiple edges, and a weight corresponding to each edge. Each node corresponds to a candidate sport or a sport type, and the edge is the connection between the candidate sport and the sport type to which the candidate sport belongs. For each edge, the weight represents the degree of association between the nodes connected by the edge.

[0341] The candidate interest item determination module is specifically used for,

[0342] For each sports interest tag, the optimal path is determined based on the weight of each edge in the sports knowledge graph and the set path length, with the node corresponding to the sports interest tag as the starting point. The optimal path is the path with the largest sum of edge weights among the candidate paths with the node corresponding to the sports interest tag as the starting point.

[0343] Based on the candidate sports items contained in the optimal path corresponding to each sports interest tag, the candidate interest items of the target user are determined.

[0344] In another embodiment of this application, the tag acquisition module is specifically used to acquire the target user's real motion interest tags;

[0345] Based on actual sports interest tags and prior knowledge of sports, potential sports interest tags are identified. The sports interest tags that the target users are interested in include both actual and potential sports interest tags.

[0346] Among them, the prior knowledge of the sport includes the first interest probability corresponding to each target information and the second interest probability corresponding to any two target information, where the target information is a candidate sport or sport type;

[0347] Prior knowledge of motion is obtained in the following ways:

[0348] Acquire data from multiple sample users, each sample user data including the actual motion interest label of a sample user;

[0349] For each target information, the probability of first interest is determined based on the proportion of sample users who are interested in that target information in all sample user data.

[0350] For any two target information pieces, the probability of the second interest is determined based on the proportion of sample users who are simultaneously interested in both target information pieces in all sample user data.

[0351] In another embodiment of this application, the device further includes a personal information acquisition module for acquiring user personal information of the target user.

[0352] The target sport determination module is specifically used for,

[0353] For each candidate item of interest, determine the association between each type of tag information and that candidate item of interest, wherein the types of tag information are at least one of user personal information and sports interest tags;

[0354] Based on the relevance of each candidate item of interest, the target sport to be recommended to the target user is determined from among the candidate items of interest. In another embodiment of this application, determining the relevance and identifying the target sport to be recommended to the target user from among the candidate items of interest is achieved through a matching degree calculation model, which is trained in the following way:

[0355] Obtain a training sample set. Each training sample in the training sample set includes: the sample user's personalized information and the sample user's real label for each candidate sport. The real label indicates whether the candidate sport is the sample user's target sport.

[0356] Each training sample is input into the initial neural network model to obtain the prediction result corresponding to each training sample. For each training sample, the prediction result includes the predicted matching degree of the sample user to each candidate item.

[0357] The training loss value is determined based on the true label and prediction result of each training sample corresponding to the candidate item.

[0358] If the training loss value meets the preset training termination condition, the training ends, and the neural network model at the end of training is used as the matching degree calculation model. If the training termination condition is not met, the model parameters of the neural network model are adjusted, and the adjusted model is trained again based on each training sample.

[0359] This application embodiment determines at least one candidate sport of interest for a target user from multiple candidate sport items included in the sport knowledge graph based on sport interest tags and a pre-configured sport knowledge graph. Based on each candidate sport of interest, it determines the target sport to recommend to the target user. This realizes the expansion of sport interest tags to obtain associated candidate sport of interest based on the user's own sport interest tags and the relationship between each candidate sport item and sport type in the sport knowledge graph, so as to recommend a richer range of target sport items that are more suitable for the user's own situation.

[0360] Based on the same inventive concept, this application provides an exercise recommendation system, the architecture of which is shown in the following diagram. Figure 7 As shown, it includes: an input unit, a processor, and a display unit.

[0361] The input unit is used to receive personalized information from the target user.

[0362] The processor is electrically connected to the input unit and is used to determine the target sports program to be recommended to the target user using any of the sports recommendation processing methods in this application.

[0363] The display unit is electrically connected to the processor and is used to display the target sports recommended to the target user.

[0364] Optionally, the exercise recommendation system in this application embodiment is specifically a separate terminal device, which can be an electronic device with strong computing power, such as a desktop computer, a laptop computer, or a 2-in-1 computer.

[0365] Optionally, the exercise recommendation system in this application embodiment includes a cloud device and a terminal device connected by communication. The cloud device may be an electronic device with strong computing power, such as a single server, server cluster, or distributed server, and has a processor for executing steps S101 to S104 in the above-described exercise recommendation method, as well as the expansion processing of each step in steps S101 to S104. The terminal device may be an electronic device with weaker computing power, such as a smartphone or tablet computer, and has an input unit, a processor, and a display unit for executing step S101 in each method for determining the target exercise to be recommended to the target user based on the target user's personalized information, as well as the expansion processing of that step.

[0366] This application provides an electronic device, comprising: a memory and a processor; at least one program stored in the memory, which, when executed by the processor, can achieve the following compared to the prior art: This application provides an electronic device that acquires personalized user information of a target user, including the target user's personal information and sports interest tags, and acquires at least one candidate interest item corresponding to the target user; it uses the user's personal information and sports interest tags as tag information, and for each candidate interest item, determines the correlation between each tag information and the candidate interest item; based on the correlation corresponding to each candidate interest item, it determines the target sports item to be recommended to the target user from each candidate interest item. Thus, by determining the target sports item to be recommended to the target user based on the correlation between the user's personal information and sports interest tags and the candidate interest items, the recommendation process fully considers the user's personal attributes and sports interests, resulting in a more suitable recommended sports item for the user, achieving personalized sports recommendation. An optional embodiment provides an electronic device, such as... Figure 8 As shown, Figure 8 The illustrated electronic device 4000 includes a processor 4001 and a memory 4003. The processor 4001 and the memory 4003 are connected, for example, via a bus 4002. Optionally, the electronic device 4000 may further include a transceiver 4004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 4004 is not limited to one type, and the structure of the electronic device 4000 does not constitute a limitation on the embodiments of this application.

[0367] Processor 4001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. Processor 4001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0368] Bus 4002 may include a pathway for transmitting information between the aforementioned components. Bus 4002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 4002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 8 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0369] The memory 4003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0370] The memory 4003 stores application code (computer program) that executes the solution of this application, and its execution is controlled by the processor 4001. The processor 4001 executes the application code stored in the memory 4003 to implement the content shown in the foregoing method embodiments.

[0371] Electronic devices include, but are not limited to: mobile phones, laptops, multimedia players, desktop computers, etc.

[0372] This application provides a computer-readable storage medium storing a computer program that, when run on a computer, enables the computer to execute the corresponding content in the aforementioned method embodiments.

[0373] This application embodiment obtains the personalized information of the target user, including the target user's personal information and sports interest tags, and obtains at least one candidate interest item corresponding to the target user. The user's personal information and sports interest tags are each treated as a type of tag information. For each candidate interest item, the correlation between each type of tag information and that candidate interest item is determined. Based on the correlation corresponding to each candidate interest item, a target sports item to be recommended to the target user is determined from among the candidate interest items. Therefore, by determining the target sports item to be recommended to the target user based on the correlation between the user's personal information and sports interest tags and the candidate interest items, the recommendation process fully considers the user's personal attributes and sports interests, resulting in recommended sports items that are more suitable for the user's own situation, thus achieving personalized sports recommendations.

[0374] The modules / units described in some embodiments of this disclosure can be implemented by at least one of software, hardware, and firmware.

[0375] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0376] The above are only some embodiments of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for recommending exercise, characterized in that, include: Obtain the target user's personalized information, which includes the target user's personal information and sports interest tags, and the sports interest tags include the types of sports and / or sports that the target user is interested in; Obtain at least one candidate item of interest corresponding to the target user; The user's personal information and the sports interest tags are each treated as a type of tag information. For each candidate item of interest, the association between each type of tag information and the candidate item of interest is determined. Based on the correlation between the two types of tag information corresponding to each candidate interest item, a target sports item recommended to the target user is determined from each candidate interest item; Determining the association between each of the tag information and the candidate item of interest includes: Obtain the first tag feature for each type of tag information, and the project feature for each candidate item of interest; For each candidate item of interest, the relevance of each tag information is determined based on the item characteristics of that item and the first tag characteristic of each tag information. The step of determining the target sport to recommend to the target user from each of the candidate interest items based on the correlation between the two types of tag information corresponding to each candidate interest item includes: For each of the candidate items of interest, the first tag features of each tag information are weighted according to the correlation between each tag information and the item, so as to obtain the second tag features corresponding to each tag information of the item. For each of the candidate items of interest, the matching degree between the target user and the item is obtained based on the second tag feature of the two types of tag information and the item feature of the item; Based on the matching degree of each of the candidate items of interest, a target sport is determined from each of the candidate items of interest to be recommended to the target user.

2. The exercise recommendation method according to claim 1, characterized in that, For each type of tag information, the tag information includes at least one piece of information, and the first tag feature of each type of tag information includes the first sub-tag feature corresponding to each piece of information; For each candidate item of interest, the relevance of each tag information is determined based on the item characteristics of that item and the first tag feature of each tag information, including: For each item in each type of label information, the correlation between the item and the candidate sport of interest is determined based on the first sub-label feature corresponding to the item and the project feature of the project. For each candidate item of interest, the first tag feature of each tag information is weighted based on its correlation with the item to obtain a second tag feature corresponding to each tag information of the item, including: For each of the candidate items of interest, the relevance of each item in each of the label information to the item is used to weight the first sub-label feature corresponding to each item to obtain the second sub-label feature corresponding to that item. For each type of tag information, the second sub-tag features of each item contained in the tag information are fused to obtain the second tag feature of the tag information.

3. The exercise recommendation method according to claim 2, characterized in that, For each item in each type of label information, the correlation between that item and the candidate sport of interest is determined based on the first sub-label feature corresponding to that item and the project feature of that item, including: The first sub-label feature corresponding to this information is multiplied with the item feature of this item to obtain the first associated feature; Input the first sub-label feature corresponding to this information and the project feature of this project into the weighted network model to obtain the second associated feature; Based on the first association feature and the second association feature, the association between the project and the information is determined.

4. The exercise recommendation method according to claim 1, characterized in that, For each candidate item of interest, the matching degree between the target user and the item is obtained based on the second tag feature of the two types of tag information and the item feature of the item, including: For each candidate item of interest, the second label feature corresponding to the two types of label information and the item feature of the item are concatenated, and the matching degree corresponding to the item is obtained based on the concatenated feature.

5. The exercise recommendation method according to any one of claims 1 to 4, characterized in that, For each of the candidate items of interest, the correlation between each type of label information and the candidate item of interest is determined by the first sub-model.

6. The exercise recommendation method according to claim 5, characterized in that, For each of the candidate items of interest, the matching degree between the target user and the item is determined through a second sub-model.

7. The exercise recommendation method according to claim 6, characterized in that, Determining the correlation and, based on the correlation corresponding to each candidate item of interest, determining the target sports item to recommend to the target user from the candidate items of interest is achieved through a matching degree calculation model. The matching degree calculation model includes a first sub-model and a second sub-model, and is trained in the following manner: Obtain a training sample set, wherein each training sample in the training sample set includes: the sample user's personalized information and the sample user's real label for each candidate sport, wherein the real label indicates whether the candidate sport is the sample user's target sport. Each training sample is input into an initial neural network model to obtain a prediction result corresponding to each training sample. For each training sample, the prediction result includes the predicted matching degree of the sample user to each candidate item. The training loss value is determined based on the true label and predicted matching degree of each candidate item corresponding to each training sample; If the preset training termination condition is met, the training ends, and the neural network model at the end of the training is used as the matching degree calculation model. If the training termination condition is not met, the model parameters of the neural network model are adjusted, and the adjusted model is trained again based on each training sample.

8. The exercise recommendation method according to any one of claims 1 to 4, characterized in that, The step of obtaining at least one candidate item of interest corresponding to the target user includes: Based on the sports interest tags and the pre-configured sports knowledge graph, at least one candidate sport of interest for the target user is determined from multiple candidate sports included in the sports knowledge graph, wherein the sports knowledge graph contains the association between each candidate sports and its corresponding sports type.

9. The exercise recommendation method according to any one of claims 1 to 4, characterized in that, Obtain sports interest tags that the target users are interested in, including: Obtain the target user's actual sports interest tags; Based on the actual sports interest tags and prior sports knowledge, the potential sports interest tags of the target user are determined. The sports interest tags that the target user is interested in include the actual sports interest tags and the potential sports interest tags. The prior knowledge of the motion includes a first probability of interest corresponding to each target information and a second probability of interest corresponding to any two target information, wherein the target information is a candidate sport or sport type, and the two target information are information of the same type; The prior knowledge of motion is obtained through the following methods: Acquire multiple sample user data, wherein the sample user data includes the actual motion interest tag of a sample user; For each target information, the first probability of interest for that target information is determined based on the proportion of sample users who are interested in that target information in all sample user data. For any two target information pieces, the second interest probability corresponding to the two target information pieces is determined based on the proportion of sample users who are simultaneously interested in both target information pieces in all sample user data.

10. A method for recommending exercise, characterized in that, include: Obtain the sports interest tags of the target user, wherein the sports interest tags include the types of sports and / or sports that the target user is interested in; Based on the sports interest tags and the pre-configured sports knowledge graph, at least one candidate interest item of the target user is determined from multiple candidate sports items included in the sports knowledge graph, wherein the sports knowledge graph contains the association relationship between each candidate sports item and its corresponding sports type among the multiple candidate sports items; Based on each of the candidate items of interest, a target sport to be recommended to the target user is determined from each of the candidate items of interest; The method further includes: Obtain the target user's personal information. The step of determining the target sport to recommend to the target user based on each of the candidate items of interest includes: For each of the candidate items of interest, the correlation between each type of tag information and the candidate item of interest is determined, wherein the types of tag information are the user's personal information and the sports interest tags; Based on the correlation corresponding to each of the candidate items of interest, a target sport to be recommended to the target user is determined from each of the candidate items of interest; The step of determining the association between each type of tag information and the candidate item of interest includes: Obtain the first tag feature for each type of tag information, and the project feature for each candidate item of interest; For each candidate item of interest, the relevance of each tag information is determined based on the item characteristics of that item and the first tag characteristic of each tag information. The step of determining the target sport to recommend to the target user from each of the candidate interest items based on the correlation between the two types of tag information corresponding to each candidate interest item includes: For each of the candidate items of interest, the first tag features of each tag information are weighted according to the correlation between each tag information and the item, so as to obtain the second tag features corresponding to each tag information of the item. For each of the candidate items of interest, the matching degree between the target user and the item is obtained based on the second tag feature of the two types of tag information and the item feature of the item; Based on the matching degree of each of the candidate items of interest, a target sport is determined from each of the candidate items of interest to be recommended to the target user.

11. The exercise recommendation method according to claim 10, characterized in that, The knowledge graph includes multiple nodes and multiple edges, wherein each node corresponds to a candidate sport or a sport type, and the edge is the connection between the candidate sport and the sport type to which the candidate sport belongs; The step of determining at least one candidate interest item for the target user from multiple candidate sports items included in the sports knowledge graph, based on the sports interest tags and a pre-configured sports knowledge graph, includes: For each of the aforementioned motion interest tags, the weights of each edge in the motion knowledge graph are determined by probabilistic walks. Based on the weights of each edge in the motion knowledge graph and the set path length, the optimal path with the node corresponding to the motion interest tag as the starting point is determined. The optimal path is the path with the largest sum of edge weights among the candidate paths with the node corresponding to the motion interest tag as the starting point. Based on the candidate sports items contained in the optimal path corresponding to each of the aforementioned sports interest tags, the candidate sports items of interest for the target user are determined.

12. The exercise recommendation method according to claim 10, characterized in that, The process of obtaining the sports interest tags that the target user is interested in includes: Obtain the target user's actual sports interest tags; Based on the actual sports interest tags and prior sports knowledge, potential sports interest tags are determined. The sports interest tags that the target user is interested in include both the actual sports interest tags and the potential sports interest tags. The prior knowledge of the sport includes a first probability of interest for each target information and a second probability of interest for any two target information, wherein the target information is the candidate sport or sport type. The prior knowledge of motion is obtained through the following methods: Acquire multiple sample user data, wherein the sample user data includes the actual motion interest tag of a sample user; For each piece of target information, the first probability of interest is determined based on the proportion of sample users who are interested in the target information in all sample user data; For any two target information items, the second interest probability is determined based on the proportion of sample users who are simultaneously interested in both target information items in all sample user data.

13. A sports recommendation device, characterized in that, include: The first acquisition module is used to acquire the user's personalized information of the target user. The user's personalized information includes the target user's personal information and sports interest tags. The sports interest tags include the sports types and / or sports projects that the target user is interested in. The second acquisition module is used to acquire at least one candidate item of interest corresponding to the target user; The correlation determination module is used to treat the user's personal information and the sports interest tags as a type of tag information, and for each candidate interest item, determine the correlation between each type of tag information and the candidate interest item; The target sport determination module is used to determine the target sport to be recommended to the target user from each of the candidate interest items based on the correlation between the two types of tag information corresponding to each candidate interest item; The correlation determination module is specifically used to obtain the first tag feature of each type of tag information and the project feature of each candidate item of interest; for each candidate item of interest, the correlation corresponding to each type of tag information is determined based on the project feature of the project and the first tag feature of each type of tag information. The target sport determination module is specifically used to, for each candidate sport of interest, weight the first tag feature of each tag information according to the correlation between each tag information and the sport to obtain the second tag feature corresponding to each tag information of the sport; for each candidate sport of interest, obtain the matching degree between the target user and the sport based on the second tag features of the two tag information and the sport feature of the sport; and determine the target sport to be recommended to the target user from each candidate sport of interest based on the matching degree corresponding to each candidate sport of interest.

14. A sports recommendation device, characterized in that, include: The tag acquisition module is used to acquire sports interest tags that the target user is interested in, wherein the sports interest tags include sports types and / or sports events that the target user is interested in; The candidate interest item determination module is used to determine at least one candidate interest item of the target user from multiple candidate sports items included in the sports knowledge graph based on the sports interest tags and a pre-configured sports knowledge graph, wherein the sports knowledge graph contains the association relationship between each candidate sports item and its corresponding sports type among the multiple candidate sports items; The target sport determination module is used to determine, based on each of the candidate interest items, a target sport to recommend to the target user. The target sport determination module is specifically used to determine the correlation between each type of tag information and each candidate sport of interest for each candidate sport of interest, wherein the types of tag information are user personal information and sport interest tags; based on the correlation corresponding to each candidate sport of interest, the module determines the target sport to be recommended to the target user from each candidate sport of interest; The step of determining the association between each type of tag information and the candidate item of interest includes: Obtain the first tag feature for each type of tag information, and the project feature for each candidate item of interest; For each candidate item of interest, the relevance of each tag information is determined based on the item characteristics of that item and the first tag characteristic of each tag information. The step of determining the target sport to recommend to the target user from each of the candidate interest items based on the correlation between the two types of tag information corresponding to each candidate interest item includes: For each of the candidate items of interest, the first tag features of each tag information are weighted according to the correlation between each tag information and the item, so as to obtain the second tag features corresponding to each tag information of the item. For each of the candidate items of interest, the matching degree between the target user and the item is obtained based on the second tag feature of the two types of tag information and the item feature of the item; Based on the matching degree of each of the candidate items of interest, a target sport is determined from each of the candidate items of interest to be recommended to the target user.

15. An electronic device, characterized in that, The electronic device includes: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to: perform the exercise recommendation method according to any one of claims 1 to 9, or perform the exercise recommendation method according to any one of claims 10 to 12.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the exercise recommendation method according to any one of claims 1 to 9, or the exercise recommendation method according to any one of claims 10 to 12.

Images