A content recommendation method, electronic device, and storage medium

Abstract

The application discloses a content recommendation method, an electronic device and a storage medium, wherein the content recommendation method comprises the following steps: acquiring a content knowledge graph, an associated object of a target object and at least two pieces of to-be-recommended content, wherein the to-be-recommended content comprises historical browsing content of each object on different terminal devices, and each object comprises the target object and the associated object; constructing an object-content knowledge graph related to the target object according to the content knowledge graph, the associated object of the target object and the at least two pieces of to-be-recommended content; determining a target vector corresponding to the target object according to a feature vector corresponding to each feature node in the object-content knowledge graph for each to-be-recommended content; determining a confidence degree of each to-be-recommended content according to a preset content vector of each to-be-recommended content and the target vector corresponding to each to-be-recommended content; and recommending the to-be-recommended content whose confidence degree meets a recommendation condition to the target object.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data recommendation, and in particular to a content recommendation method, an electronic device and a storage medium. BACKGROUND

[0002] In a content recommendation scenario, a content knowledge graph is constructed for a media library. After an object clicks on a certain content or watches a certain content on a current terminal device, a recommendation system supplements content information by using the relationship between content nodes in the content knowledge graph, so as to recommend other related contents to the object based on the historical content watched on the current terminal device and the relationship between the content nodes to expand the content that the object can watch. However, at least the amount of data obtained is single when related contents are recommended to the object based on the historical browsing records of the object, which leads to inaccurate recommendation. SUMMARY

[0003] Therefore, the embodiments of the present application provide a content recommendation method, an electronic device and a storage medium.

[0004] The technical solution of the present application is implemented as follows:

[0005] The embodiments of the present application provide a content recommendation method, which comprises the following steps:

[0006] obtaining a content knowledge graph, associated objects of a target object and at least two to-be-recommended contents, wherein each entity node in the content knowledge graph corresponds to a representation vector, the to-be-recommended contents include historical browsing contents of each object on different terminal devices, and the each object includes the target object and the associated objects;

[0007] constructing an object-content knowledge graph related to the target object according to the content knowledge graph, the associated objects of the target object and the at least two to-be-recommended contents;

[0008] determining a target vector corresponding to the target object according to a feature vector corresponding to each feature node in the object-content knowledge graph for each to-be-recommended content, wherein the feature node includes an object node and the entity node, and the feature vector includes the representation vector;

[0009] determining a confidence degree of each to-be-recommended content according to a preset content vector of each to-be-recommended content and a target vector corresponding to each to-be-recommended content;

[0010] recommending the to-be-recommended content whose confidence degree meets a recommendation condition to the target object.

[0011] The embodiments of the present application provide a content recommendation device, which comprises the following steps:

[0012] an acquisition module configured to acquire a content knowledge graph, an associated object of a target object, and at least two pieces of to-be-recommended content, wherein each entity node in the content knowledge graph corresponds to a representation vector, and the to-be-recommended content includes historical browsing content of each object at different terminal devices, and the each object includes the target object and the associated object;

[0013] a processing module configured to construct an object-content knowledge graph related to the target object according to the content knowledge graph, the associated object of the target object, and the at least two pieces of to-be-recommended content;

[0014] a determination module configured to determine, for each piece of to-be-recommended content, a target vector corresponding to the target object according to a feature vector corresponding to each feature node in the object-content knowledge graph, wherein the feature node includes an object node and the entity node, and the feature vector includes the representation vector;

[0015] The determination module is further configured to determine a confidence degree of each piece of to-be-recommended content according to a preset content vector of each piece of to-be-recommended content and a target vector corresponding to each piece of to-be-recommended content.

[0016] The processing module is further configured to recommend, to the target object, to-be-recommended content that satisfies a recommendation condition.

[0017] An embodiment of the present application provides an electronic device, which includes a processor, a memory, and a communication bus.

[0018] The communication bus is configured to realize communication connection between the processor and the memory.

[0019] The processor is configured to execute an information processing program stored in the memory, so as to realize the content recommendation method.

[0020] An embodiment of the present application provides a storage medium, which stores one or more programs, and the one or more programs can be executed by one or more processors to realize the content recommendation method.

[0021] The content recommendation method, electronic device, and storage medium provided in this application acquire a content knowledge graph, associated objects of a target object, and at least two pieces of content to be recommended. Each entity node in the content knowledge graph corresponds to a representation vector. The content to be recommended includes the historical browsing content of each object on different terminal devices. Each object includes the target object and associated objects. Based on the content knowledge graph, associated objects of the target object, and at least two pieces of content to be recommended, an object-content knowledge graph related to the target object is constructed. For each piece of content to be recommended, the feature vector corresponding to each feature node in the object-content knowledge graph is used to determine the target object's corresponding... The target vector, where feature nodes include object nodes and entity nodes, and feature vectors include representation vectors; based on the preset content vector of each content to be recommended and the target vector corresponding to each content to be recommended, the confidence level of each content to be recommended is determined; content to be recommended that meets the recommendation criteria is recommended to the target object; thus, the historical browsing content of the target object and related objects on different mobile terminal devices, such as large-screen TV terminal devices and small-screen mobile terminal devices, is obtained, expanding the number of content to be recommended to the target object, while improving the accuracy of content recommendation on the terminal devices used by the target object, such as large-screen TV terminal devices or small-screen mobile terminal devices. Attached Figure Description

[0022] Figure 1 A flowchart illustrating a content recommendation method provided in an embodiment of this application;

[0023] Figure 2 A schematic diagram of a content knowledge graph provided in an embodiment of this application;

[0024] Figure 3 A schematic diagram of an object-content knowledge graph provided in an embodiment of this application;

[0025] Figure 4 A flowchart illustrating another content recommendation method provided in this application embodiment;

[0026] Figure 5 A flowchart illustrating yet another content recommendation method provided in an embodiment of this application;

[0027] Figure 6 A flowchart illustrating a content recommendation method provided in another embodiment of this application;

[0028] Figure 7 This is a structural block diagram of the target knowledge graph provided in the embodiments of this application;

[0029] Figure 8 A flowchart illustrating another content recommendation method provided in another embodiment of this application;

[0030] Figure 9 A schematic diagram of the water ripple set of the k-th hop content related to the target object and the c-th hop object preference propagation provided in the embodiments of this application;

[0031] Figure 10 A flowchart illustrating yet another content recommendation method provided in another embodiment of this application;

[0032] Figure 11 A flowchart illustrating a content recommendation method provided in yet another embodiment of this application;

[0033] Figure 12 This is a schematic diagram of the structure of a content recommendation device provided in an embodiment of this application;

[0034] Figure 13 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0035] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0036] It should be understood that the phrases "embodiments of this application" or "foreign embodiments" throughout the specification mean that a specific feature, structure, or characteristic related to an embodiment is included in at least one embodiment of this application. Therefore, "embodiments of this application" or "in the foreign embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. In the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0037] Unless otherwise specified, any step in the embodiments of this application performed by the electronic device may be executed by the processor of the electronic device. It is also worth noting that the embodiments of this application do not limit the order in which the electronic device performs the following steps. Furthermore, the methods used to process data in different embodiments may be the same or different methods. It should also be noted that any step in the embodiments of this application can be executed independently by the electronic device; that is, when the electronic device performs any step in the following embodiments, it may not depend on the execution of other steps.

[0038] To better understand the content recommendation method provided in the embodiments of this application, the content recommendation methods in related technologies will be described first.

[0039] In related technologies, traditional recommendation systems primarily help users discover content of interest from massive amounts of information by mining the binary relationship between users and items, thus satisfying their personalized needs. With the promotion of mobile internet and social media applications, as well as the emergence of social marketing (such as group buying and group purchasing), social recommendation systems have gradually developed. Social recommendation, based on existing traditional recommendation systems, incorporates social attribute information related to users or items (such as friend relationships, trust relationships, and item tags) as important influencing factors into the recommendation model to improve the overall performance of the recommendation system. From this perspective, social recommendation is a recommendation method that uses the social attribute information of users or items as additional input to traditional recommendation systems.

[0040] Meanwhile, with the increasingly widespread application of knowledge graphs in content recommendation, supplementing content information based on the relationships between content nodes in the content graph can, to some extent, solve the problem of data sparsity and improve the accuracy of recommendation algorithms. For example, knowledge graph-based recommendation algorithms (Multi-task Learning for KG enhanced Recommendation, MKR) alternately embed knowledge graphs and recommendation models into the training process to achieve the transfer of content information. RippleNet uses the node clicked by the user as the source point and expands outwards like ripples on water, absorbing the multi-hop information of the source point and expanding this expanded information to the user's range of interests.

[0041] However, recommendation systems supplement content information by utilizing the relationships between content nodes in a content knowledge graph. Based on the currently viewed content and the relationships between content nodes, they recommend other related content to the user to expand the range of content the user might view. Because recommending other related content to a user solely based on their personal viewing or browsing history results in a limited data source, leading to inaccurate recommendations.

[0042] To address the problems existing in related technologies, this application provides a content recommendation method, which is applied to a content recommendation device or electronic device. The content recommendation method provided in this embodiment can be implemented by a computer program, which, when executed, completes each step of the content recommendation method provided in this embodiment. In some embodiments, the computer program can be executed by a processor in an electronic device. Figure 1 This is a flowchart illustrating a content recommendation method provided in an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0043] Step 101: Obtain the content knowledge graph, the related objects of the target object, and at least two pieces of content to be recommended.

[0044] In the content knowledge graph, each entity node corresponds to a representation vector. The content to be recommended includes the historical browsing content of each object on different terminal devices. Each object includes the target object and related objects.

[0045] In this embodiment, the content knowledge graph can be a graph obtained by fusing and aligning content data from various media databases. The content knowledge graph includes entity nodes and the connections between them, and each entity node corresponds to a representation vector, which uniquely represents the corresponding entity node. For example, referring to… Figure 2 As shown, Figure 2 What is shown is a content knowledge graph.

[0046] Here, the media database includes, but is not limited to, the content databases of various video applications (APPs) installed on different terminal devices. Different terminal devices include mobile terminal devices and fixed terminal devices. Mobile terminal devices can be portable devices with a screen size smaller than a certain threshold, including but not limited to mobile phones and tablets. Fixed terminal devices include fixed devices with a screen size greater than or equal to a certain threshold, including but not limited to television terminal devices. Various video APPs include, but are not limited to, AA Video APP, BB Video APP, CC Video APP, DD Video APP, and EE Video APP. This unifies and integrates the media databases of different applications on different mobile terminal devices, such as large-screen television terminal devices and small-screen mobile terminal devices, ensuring the completeness of the content to be recommended.

[0047] Here, a knowledge graph, also known as a knowledge base in some contexts, is a knowledge system formed by structuring human knowledge in the real world. In a knowledge graph, vast amounts of knowledge, such as information from open databases and encyclopedias, are typically expressed as relational datasets. Within these datasets, basic facts are abstracted as entities, while rules, logic, reasoning, and other relational information are abstracted as relationships between entities. If entities are mapped to nodes and relationships to edges, this knowledge can be further presented in the form of a graph. Specifically, one representation method for knowledge graphs is the triple representation method, i.e., (h, r, t), where h and t are the head entity and tail entity, respectively, and r describes the relationship between the head entity h and the tail entity t.

[0048] In this embodiment of the application, the associated objects of the target object can be associated objects and associated objects that have a close relationship with the target object. For example, the associated objects of the target object can be family associated objects that have a close family relationship with the target object, and the associated objects of the target object can also be family associated objects that have a close work relationship with the target object.

[0049] In this embodiment, the content to be recommended can be the historical browsing content that all objects have viewed or watched on different terminal devices within a preset time period. The content to be recommended can also be the historical browsing content that the target object and related objects have viewed or watched on different terminal devices within a preset time period. In this way, by obtaining the historical browsing content of the target object and related objects on different mobile terminal devices such as large-screen TV terminal devices and small-screen mobile terminal devices, the number of content to be recommended to the target object is expanded, and the accuracy of content recommendation on the large-screen TV terminal device or small-screen mobile terminal device used by the target object is improved.

[0050] Here, we define an object collection using all objects and all content as an example. and content collection Based on the historical browsing content of each object on different terminal devices, an interaction matrix between the object and the content is constructed. This interaction matrix can be represented by formula (1).

[0051]

[0052] Where u represents the object and v represents the content. This represents the object's interaction history regarding viewed content, with values ​​of 0 and 1. A value of 0 indicates that the object has not viewed the content; a value of 1 indicates that the object has viewed the content. In this case, the content viewed by the object can be considered as potential content for recommendation, allowing for the selection of some content from this pool to recommend to the target object.

[0053] Step 102: Based on the content knowledge graph, the related objects of the target object, and at least two pieces of content to be recommended, construct an object-content knowledge graph related to the target object.

[0054] In this embodiment, after the electronic device obtains a content knowledge graph, associated objects of the target object, and at least two pieces of content to be recommended, it can construct a content knowledge graph of the content to be recommended that the target object has viewed, i.e., a target object-content knowledge graph, based on the content knowledge graph and the content to be recommended that the target object has viewed. Simultaneously, based on the content knowledge graph and the content to be recommended that each associated object has viewed, it constructs a content knowledge graph of the content to be recommended that the associated objects have viewed, i.e., an associated object-content knowledge graph. The target object-content knowledge graph and the associated object-content knowledge graph are then fused to obtain an object-content knowledge graph related to the target object. For example, refer to... Figure 3 As shown, Figure 3 This shows an object-content knowledge graph related to the target object, where... Figure 3 In the diagram, light gray nodes represent object nodes, and dark gray nodes represent entity nodes.

[0055] Step 103: For each piece of content to be recommended, determine the target vector corresponding to the target object based on the feature vectors corresponding to each feature node in the object-content knowledge graph.

[0056] Among them, feature nodes include object nodes and entity nodes. Object nodes include nodes corresponding to the target object and associated objects, and feature vectors include representation vectors.

[0057] In this embodiment of the application, the object-content knowledge graph includes multiple feature nodes and the relationships between each feature node.

[0058] In this embodiment of the application, for each piece of content to be recommended, in the object-content knowledge graph, the feature node corresponding to the target object is taken as the center node and the feature node corresponding to the content to be recommended is taken as the target node, and at least one link path between the center node and the target node is determined; then, based on the feature vectors of all feature nodes on the at least one link path, the target vector corresponding to the target object of each piece of content to be recommended is determined.

[0059] Step 104: Determine the predicted value of each content to be recommended based on the preset content vector of each content to be recommended and the target vector corresponding to each content to be recommended.

[0060] In this embodiment, the preset content vector of the content to be recommended is predetermined. Confidence level is used to indicate the degree of interest the target audience has in the content to be recommended.

[0061] In this embodiment of the application, the predicted value of each content to be recommended can be determined by the following formula (2).

[0062]

[0063] in, For each piece of content to be recommended, the predicted value is... Let v be the target vector for each piece of content to be recommended, and let v be the preset content vector for each piece of content to be recommended.

[0064] Step 105: Recommend the content to be recommended that meets the recommendation criteria to the target audience.

[0065] In this embodiment, the confidence level meeting the recommendation criteria includes: content to be recommended with a confidence level greater than or equal to a recommendation threshold, which can be a pre-set value, such as 0.7, 0.85, etc.; the confidence levels of each content to be recommended and the target object are sorted in descending order, and the content to be recommended with the highest confidence level in the permutation combination, or the content to be recommended with the top N confidence levels in the permutation combination, where N is a positive integer. The content to be recommended is then pushed to the terminal device corresponding to the target object, such as a TV terminal device or a mobile terminal device, thereby providing users with content they may be interested in.

[0066] The content recommendation method provided in this application obtains a content knowledge graph, associated objects of a target object, and at least two pieces of content to be recommended. Each entity node in the content knowledge graph corresponds to a representation vector. The content to be recommended includes the historical browsing content of each object on different terminal devices. Each object includes the target object and associated objects. Based on the content knowledge graph, associated objects of the target object, and at least two pieces of content to be recommended, an object-content knowledge graph related to the target object is constructed. For each piece of content to be recommended, a target vector corresponding to the target object is determined based on the feature vectors corresponding to each feature node in the object-content knowledge graph. Feature nodes include object nodes and entity nodes, and feature vectors include representation vectors. The confidence level of each piece of content to be recommended is determined based on its preset content vector and the target vector. Content with confidence levels that meet the recommendation criteria is recommended to the target object. This method obtains the historical browsing content of the target object and associated objects on different mobile terminal devices, such as large-screen TV terminals and small-screen mobile terminals, expanding the number of content recommendations for the target object and improving the accuracy of content recommendations on the target object's terminal devices (large-screen TV terminals or small-screen mobile terminals).

[0067] Based on the foregoing embodiments, this application also provides a content recommendation method, applied to electronic devices, such as... Figure 4 The diagram shown is a flowchart illustrating another content recommendation method provided in this application embodiment. Figure 1 and Figure 4 As shown, the method includes the following steps:

[0068] Step 201: Obtain the raw communication data of the target object.

[0069] In this embodiment of the application, the original communication data may be the communication data obtained when the target object communicates with each of the multiple different first objects within a preset time period, and the original communication data of the target object includes multiple.

[0070] Here, the preset time period can be the most recent month, the most recent two months, or the most recent quarter.

[0071] In this embodiment of the application, before obtaining the original communication data of the target object, all objects owned by the operator can be filtered according to preset filtering rules to obtain the photographed object. Here, the photographed object includes the target object.

[0072] Here, the filtering rules include at least one of the following: filtering non-real-name objects, including but not limited to objects corresponding to numbers not registered with an ID card number, objects corresponding to numbers registered with a company name, and objects corresponding to numbers not maintained in the operator's database; filtering abnormal status objects, including but not limited to objects with statuses such as "suspended," "cancelled," "inactive," and "unknown"; filtering Subscriber Identity Module (SIM) objects and Machine-to-machine (M2M) objects, i.e., objects corresponding to numbers with 0 calls and greater than 0 data usage in the last 3 months within a preset time period; filtering test objects, i.e., objects corresponding to numbers with 0 pass counts, greater than 0 SMS messages, and 0 data usage in the last 3 months within a preset time period; filtering objects that have been roaming outside the province for a long time, i.e., objects corresponding to numbers whose calls in the last 3 months have all been roaming calls within a preset time period; and filtering low-consumption protection numbers, i.e., objects corresponding to numbers whose average revenue per user (arpu) in the last 6 months is less than or equal to 5 yuan within a preset time period.

[0073] Step 202: Extract features from the original communication data to obtain communication features.

[0074] The communication features include the peer number that connects to the local number corresponding to the target object, and the peer number corresponds to the first object.

[0075] In this embodiment of the application, the communication features include multiple features, including but not limited to whether the local number, the remote number, and the object's permanent base station are consistent. As shown in Table 1, Table 1 shows the communication features extracted from the original communication data.

[0076]

[0077] Table 1

[0078] Step 203: Input the communication features into the regression model corresponding to the feature conditions satisfied by the communication features, and obtain the correlation degree between the target object and the first object output by the regression model, thereby obtaining multiple correlation degrees between the target object and each of the multiple first objects.

[0079] In this embodiment, the feature conditions satisfied by the communication features are used to classify the communication features so that the communication features can be input into the corresponding trained regression model.

[0080] In this embodiment, the communication characteristics satisfy the following conditions: the communication characteristics satisfy the home characteristic condition and / or the communication characteristics satisfy the work characteristic condition. The home characteristic condition includes high communication frequency, strong stability, nighttime / weekend / holiday location, and high base station overlap. The work characteristic condition includes high communication frequency, strong stability, daytime / weekday location, and high base station overlap. Here, the regression model includes a home regression model and a work regression model. If the communication characteristics satisfy the home characteristic condition, the communication characteristics are input into the pre-trained home regression model corresponding to the home characteristic condition; if the communication characteristics satisfy the work characteristic condition, the communication characteristics are input into the pre-trained work regression model corresponding to the work characteristic condition.

[0081] In this embodiment, a regression model is used to determine the degree of association between the target object and the first object based on communication characteristics. The degree of association characterizes the closeness between the target object and the first object, and is also called the probability value of the binary family relationship pair between the target object and the first object n. Here, the regression model can be expressed by the following formula (3),

[0082]

[0083] in, Characterizes the degree of association between the target object and the first object n. This represents the k-th communication feature in the communication features. Characterizing communication features The corresponding regression coefficients.

[0084] In this embodiment, the electronic device extracts features from each piece of raw communication data of the target object to obtain communication features corresponding to each first object, thereby obtaining communication features of all first objects communicating with the target object. For all communication features, each communication feature is input into a regression model corresponding to the feature conditions satisfied by that communication feature to obtain the correlation degree between the target object and each first object output by the regression model, thereby obtaining multiple correlation degrees between the target object and multiple first objects.

[0085] Here, taking the family regression model as an example, we will explain the process of obtaining the regression coefficients corresponding to the communication characteristics in the family regression model.

[0086] Obtain the sample phone numbers of all sample objects. Based on these sample phone numbers, filter out the family network objects and combine them with the call information table of the objects associated with the photo object table to generate a call data table for the sample objects. Divide the call data table into positive sample call data and negative sample call data according to the following rules.

[0087] Here, the rules for filtering positive sample call data are as follows: family relationship pairs with multiple phone numbers are removed by using the ID card information of the target sample object, the number of people in the family network is controlled to be 2-3, and the identity document (ID) of the family network is consistent. The family relationship pairs of the filtered sample objects are set to 1, thereby obtaining positive sample objects with close relationships with the target sample object. Here, the family relationship pair is also called the number relationship pair.

[0088] Here, the rules for filtering negative sample call data are as follows: remove the same network relationship pairs with the same ID card number of the target sample object, the family network IDs that are inconsistent, and the fact that one party has a family network while the other party does not. Set the family relationship pairs of the filtered sample objects to 0, thereby obtaining negative sample objects that have no close relationship with the target sample object.

[0089] Furthermore, feature extraction is performed on the positive sample call data between the target sample object and the positive sample object, and on the negative sample call data between the target sample object and the negative sample object, to obtain positive sample call features and negative sample call features; and the sample correlation degree of family relationship pairs between the target sample object and the sample objects is obtained, i.e. ,and The positive sample call features, negative sample call features, and sample correlation are input into the log-likelihood function to obtain multiple nonlinear log-likelihood functions. The Newton-Raphson algorithm is then used to iteratively solve these multiple nonlinear log-likelihood functions. When the Newton-Raphson algorithm converges, the regression coefficients corresponding to each sample call feature in the above family regression model are obtained.

[0090] Here, the log-likelihood function can be expressed by the following formula (4),

[0091]

[0092] in, This represents the log-likelihood between the target sample object and the sample object n, where N represents the total number of positive and negative sample objects. This represents the sample correlation degree between the target sample object and sample object n. Indicates positive and negative sample objects.

[0093] Here, formula (4) is transformed to obtain the transformed log-likelihood function. The positive sample call features, negative sample call features, and sample correlation are then input into the transformed log-likelihood function to obtain the regression coefficients corresponding to each sample call feature in the above family regression model. The transformed log-likelihood function can be expressed by the following formula (5).

[0094]

[0095] Step 204: Filter multiple correlations and determine the target social network of the target object based on the filtered correlations.

[0096] The target social network includes associated objects.

[0097] In this embodiment of the application, the filtering process for multiple correlation degrees can be understood as filtering out correlation degrees that are greater than the correlation degree threshold from multiple correlation degrees. The correlation degree threshold can be a pre-set value, such as 0.5, 0.6, etc.

[0098] In this embodiment of the application, the target social network is a network composed of the target object and related objects with close relationships.

[0099] In one scenario, the electronic device inputs communication features into a family regression model corresponding to the family feature conditions satisfied by the communication features. After obtaining the correlation degree between the target object and the first object output by the family regression model, it filters out the correlation degrees that are greater than the correlation degree threshold from multiple correlation degrees. Based on the filtered correlation degrees, it determines the family social network of the target object. The target social network includes the target family social network, and the network density between all objects included in the target family social network satisfies the first density condition.

[0100] In another scenario, the electronic device inputs communication features into the working regression model corresponding to the working feature conditions satisfied by the communication features. After obtaining the correlation degree between the target object and the first object output by the working regression model, it filters out the correlation degrees that are greater than the correlation degree threshold from multiple correlation degrees. Based on the filtered correlation degrees, it determines the working social network of the target object. The target social network includes the target working social network, and the network density between all objects included in the target working social network satisfies the second density condition.

[0101] In this embodiment of the application, step 204, which determines the social network of the target object based on the filtered relevance, is combined with... Figure 5To elaborate further,

[0102] Step 241: Initialize network nodes, treating each network object as a network node.

[0103] The network object includes the target object and the first object corresponding to the filtered relevance.

[0104] Step 242: Based on the filtered correlation degree, calculate the modularity corresponding to each network node when it is assigned to the community of its neighboring network nodes, and determine the community corresponding to the largest modularity among multiple modularities as the community to which the network node is assigned.

[0105] Step 243: Aggregate the divided communities into a single point, reconstruct the network based on the generated community structure, and continue dividing the next network node until the structure in the network no longer changes, thus obtaining the social network of the target object.

[0106] In this embodiment, the number of objects in the target object's social network increases significantly over time, while object recommendation requires a certain degree of real-time performance. To quickly obtain the social network composed of objects closely related to the target object, the Fast Unfolding graph model algorithm can be used. This algorithm uses modularity to characterize the tightness of discovered communities. First, each network object is initialized as a network node, including the target object and the first object corresponding to the filtered association degree. Second, based on the filtered association degree, the modularity corresponding to assigning each network node to the community of its adjacent network nodes is calculated, and the community corresponding to the highest modularity among multiple modularities is determined as the community assigned to the network node. Then, the assigned communities are aggregated into a single point, and the network is reconstructed based on the generated community structure. Finally, the assignment of the next network node is continued, repeating the above two steps until the network structure no longer changes, ultimately obtaining the target object's social network.

[0107] Here, modularity refers to the proportion of edges connecting vertices within a community structure in a network, minus the expected proportion of edges connecting any two nodes within the same community structure. Modularity can characterize the quality of such partitioning; the higher the modularity, the better the social network partitioning. The formula for calculating modularity can be expressed by formula (6).

[0108]

[0109] Where m represents the sum of the correlation degrees of all network nodes in the network. The degree of association between network node i and network node j is the degree of association of the binary family relationship pair. This represents the sum of the associativity of the edges connected to network node i as a vertex. This indicates the home or work circle to which network node i is assigned. This function is used to determine whether network node i and network node j are in the same home circle or the same work circle. If they are, it returns 1; otherwise, it returns 0.

[0110] For example, in determining the family social network of a target individual, the Fast Unfolding algorithm infers that the key person among family members is generally the head of household. The relationship between the target individual and the key family member can be determined by combining the basic information of the target individual and the information differences between individuals, using the following rules:

[0111] Rule 1: If the target is a user of this website and is of a different gender from the key family member, and the absolute value of the age difference is less than 10 years, then the target is determined to be the spouse of the key family member.

[0112] Rule 2: If the target is a user of this website and is more than 20 years older than the key family member, then the target is determined to be the parents of the key family member;

[0113] Rule 3: If the target is a user of this website and is more than 20 years younger than the key family member, then the target is determined to be the child of the key family member.

[0114] Rule 4: If the target is a user of this website and is of the same gender as the key family member, and the absolute age difference is within 6 years; or, if the target is of a different gender than the key family member, but the first 6 digits of their ID card are the same and the absolute age difference is less than 6 years, then the target is determined to be a peer of the key family member.

[0115] Rule 5: The target object is an object from another network.

[0116] It should be noted that by extracting features from the original communication data of the target object, the communication features when communicating with each first object are obtained. Based on the communication features, the correlation between the target object and the first object is determined. Then, based on the correlation, the fast-unfolding graph model algorithm is used to expand into small family social circles and / or work social circles, and finally a family close relationship group and / or work close relationship group of the target object is generated.

[0117] Based on the foregoing embodiments, this application also provides a content recommendation method, applied to electronic devices, such as... Figure 6 The diagram shown is a flowchart illustrating another content recommendation method provided in this application embodiment. Figure 1 and Figure 6 As shown, the method includes the following steps:

[0118] Step 301: Obtain the original content data from the media database and perform data deduplication on the original content data to obtain the content database.

[0119] In this embodiment, since the media database includes content databases from various video applications installed on different terminal devices, data deduplication can be understood as data alignment and standardization of the original content data. Here, because different media databases may contain duplicate content with consistent content, data alignment is required for the original content data. For example, the media databases corresponding to AA Video APP and DD Video APP each include the content "WBSYS". Data alignment is needed for multiple "WBSYS" entries obtained from different media databases to obtain a content database containing only one "WBSYS". Of course, different media databases may use different identifiers to represent the same content, so data standardization is required for the content represented by different identifiers. For example, the media database corresponding to AA Video APP includes the content "TJ", and the media database corresponding to DD Video APP includes the content "RZJTZTJ". Since "RZJTZTJ" and "TJ" both represent the same content, data standardization is required for the content represented by "RZJTZTJ" and "TJ".

[0120] Step 302: Extract knowledge from the target fields in the content database to obtain the initial knowledge graph.

[0121] The initial knowledge graph includes entities, relationships between entities, and entity attribute information.

[0122] In this embodiment, the target field refers to a valid field of the content included in the content database. Here, after obtaining the content database, the target fields of the content included in the content database, such as a movie and TV series database, are extracted.

[0123] In this embodiment, knowledge extraction includes entity extraction, relation extraction, and attribute extraction. Entity extraction involves extracting content based on a target field. Entity extraction, also known as named entity recognition, refers to automatically identifying named entities from a dataset; for example, extracting entities such as "movie," "actor," "director," and "tag" from a content database. Relationship extraction involves extracting the relationships between entities; for example, extracting the movie-director relationship, the movie-actor relationship, and the director-actor relationship. Attribute extraction involves collecting attribute information of specific entities from different information sources, including but not limited to the main content and behind-the-scenes footage.

[0124] Of course, after extracting entities from the target fields in the content database, the similarity between entities can be calculated. This involves converting entities into vectors using word vectors (e.g., word2vector), calculating the similarity between the corresponding vectors, and determining whether entities overlap based on their similarity scores. This process then performs entity disambiguation. Further, after entity disambiguation, entity tables, such as movie tables or actor tables, are built based on the remaining entities.

[0125] Of course, after extracting the relationships between entities from the target fields in the content database, you can also build relationship tables, such as movie-director relationship tables, movie-actor relationship tables, and director-actor relationship tables.

[0126] In this embodiment, the electronic device extracts entities, relations, and attributes from the target fields in the content database, constructs a neo4j graph database, and generates the constructed knowledge graph triples, thereby initializing the knowledge graph.

[0127] Step 303: Perform knowledge reasoning on the initial knowledge graph to obtain the target knowledge graph.

[0128] Here, knowledge reasoning can be understood as reasoning about different series of the same content. For example, the initial knowledge graph includes the content "SDYJQ," and the current content database stores series 1 to series 8 of "SDYJQ," namely "SDYJQ1" to "SDYJQ8." When "SDYJQ9" appears, reasoning can be performed on the series of "SDYJQ" stored in the content database, adding "SDYJQ9" to the initial knowledge graph to obtain the target knowledge graph.

[0129] Step 304: Perform unsupervised learning on each entity node in the target knowledge graph to obtain a content knowledge graph that represents each entity node through representation vectors.

[0130] In this embodiment, the target knowledge graph includes multiple entity nodes and the relationships between them. Unsupervised learning is performed on each entity node in the target knowledge graph, that is, the embedding vector of the entity node is learned in an unsupervised manner. A representation vector is generated for each entity and relationship in the target knowledge graph, while maintaining the original structure or semantic information of the graph. Unsupervised feature learning algorithms include, but are not limited to, deep graph infomax (DGI), skip-gram, one-hot encoding, Deepwalk, node2vev, etc. The final representation vector of the entity node can represent the relationship between the entity node and its neighboring entity nodes. For example, this embodiment uses the Random Walk method in DeepWalk to perform unsupervised learning on each entity node in the target knowledge graph. Random Walk is a depth-first traversal algorithm that allows repeated visits to visited nodes. Given the current starting node, nodes are randomly sampled from its neighbors as the next visited node, and this process is repeated until the length of the visited sequence meets a preset condition. Here, the process of constructing the target knowledge graph is described in the flowchart. Figure 7 As shown.

[0131] Based on the foregoing embodiments, this application also provides a content recommendation method, applied to electronic devices, such as... Figure 8 The diagram shown is a flowchart illustrating another content recommendation method provided in this application embodiment. Figure 1 and Figure 8 As shown, the method includes the following steps:

[0132] Step 401: Based on the object-content knowledge graph, construct a ripple set of the k-th hop content related to the target object and the preference propagation of the c-th hop object.

[0133] Where 0≤k≤K, K is the total number of entities that can be jumped, and 0≤c≤C, C is the total number of objects that can be jumped.

[0134] Here, we are referring to the knowledge graph triples constructed in the object-content knowledge graph. The c-th hop of the target object u's intimate relationship with the target object. ,here Here, we no longer distinguish between family-based and work-based close relationships between objects. The tail entity associated with the k-th, c-th hop of the target object u can be represented by formula (7):

[0135]

[0136] (7)

[0138] in, This represents the set of content viewed by the target object u in the history. h represents the head entity in the knowledge graph, r represents the relation, and t represents the tail entity. k represents the number of skipped entities, K represents the total number of skippable entities, c represents the number of skipped objects, and C represents the total number of skippable objects. It should be noted that the total number of skippable entities K and the total number of skippable objects C are predefined.

[0139] Here, the set of water ripples for the c-th jump of the target object can be represented by formula (8).

[0140]

[0141] (8)

[0143] in, This represents the k,c-th hop triplet relationship of the target object u's historical access content entities in the knowledge graph and social graph.

[0144] In the embodiments of this application, reference is made to Figure 9 As shown, Figure 9 This shows the set of ripples representing the propagation of the k-th hop content related to the target object and the c-th hop object preference. Among them, Figure 9 The left-hand diagram in the image represents an object-content knowledge graph. Figure 9 The dashed box in the middle of the image represents the set of ripples representing the propagation of the k=1 hop content related to the target object and the c=1 hop object preference. Figure 9 The right-hand diagram represents the set of ripples representing the propagation of the k=2 hop content related to the target object and the c=1 hop object preference.

[0145] Step 402: For each triple in the ripple set, based on the relationship between the head entity, the head entity and the tail entity, and the feature vector of the tail entity, determine the similarity between each content to be recommended and the head entity.

[0146] Here, for the constructed ripple set of content related to the target object at the k-th hop and the preference propagation of the target object at the c-th hop, for each triple in the ripple set, the feature vector corresponding to the head entity, the feature vector corresponding to the relationship between the head entity and the tail entity, and the feature vector of the tail entity are multiplied together to obtain the similarity between each content to be recommended and the head entity.

[0147] Here, the similarity is normalized using the softmax function, and the similarity can be calculated using formula (9).

[0148]

[0149] in, Content to be recommended and entity exist Similarity in vector space , They are relationships and head node eigenvectors.

[0150] Step 403: Based on similarity, perform a weighted summation of the feature vectors corresponding to the tail entities in the ripple set to obtain the first sub-target vector corresponding to each content to be recommended in the ripple set.

[0151] In this embodiment, the electronic device, for each triple in the ripple set, determines the similarity between each content to be recommended and the head entity based on the relationship between the head entity, the head entity, and the tail entity, and the feature vector of the tail entity, and then sorts the ripple set... Similarity corresponding to the middle tail The first sub-target vector corresponding to each content to be recommended in the ripple set is obtained by weighted summation using the following formula (10). .

[0152]

[0153] in, Tail feature node The corresponding feature vector, the first sub-target vector The relevant representation of the target object's historical viewing content in the knowledge graph at the k-th hop content and the c-th hop object.

[0154] Step 404: Add the first sub-target vectors corresponding to the (K+C) sets of water ripples related to the target object to obtain the target vector corresponding to each target object of the content to be recommended.

[0155] In this embodiment, the electronic device constructs (K+C) sets of ripples related to the target object based on the object-content knowledge graph. That is, the representations obtained through multiple interest diffusions in the RippleNet algorithm are added together to finally obtain the target vector corresponding to each target object of the content to be recommended. Here, the target vector corresponding to each target object of the content to be recommended can be represented by the following formula (11).

[0156]

[0157] Where u is the target vector corresponding to each target object of the content to be recommended. This is the first sub-target vector of the target object's historical viewed content within the knowledge graph at the k-th hop content and c-th hop object. Thus, by using a ripple set to obtain the target vector for each piece of content to be recommended, the accuracy of content recommendations for the target object's terminal devices—whether large-screen TVs or small-screen mobile devices—is improved.

[0158] Based on the foregoing embodiments, this application also provides a content recommendation method, applied to electronic devices, such as... Figure 10 The diagram shown is a flowchart illustrating another content recommendation method provided in this application embodiment. Figure 1 and Figure 10 As shown, the method includes the following steps:

[0159] Step 501: For each piece of content to be recommended, take the feature node that matches the target object as the center node and the feature node that matches each piece of content to be recommended as the target node, and construct a knowledge graph triple in the object-content knowledge graph.

[0160] Step 502: For a knowledge graph triple, determine at least one link path from the central node to the target node.

[0161] Step 503: Perform weighted aggregation on the feature vectors corresponding to all feature nodes in each of the at least one link path to obtain the target vector.

[0162] In this embodiment, for each piece of content to be recommended, in the object-content knowledge graph, with the feature node corresponding to the target object as the central node and the feature node corresponding to the content to be recommended as the target node, for the knowledge graph triples constructed in the object-content knowledge graph, at least one link path is determined between the central node and the target node. The feature vectors corresponding to all feature nodes on each of the at least one link path are weighted and aggregated to obtain the target vector corresponding to the target object of each piece of content to be recommended.

[0163] For example, the video content watched by the target object within a preset time period includes GDYFDR and HZGG, and the video content watched by the target object's associated objects within the preset time period includes WBSYS. Here, taking GDYFDR as an example of the content to be recommended, the feature node corresponding to the target object is determined as the central node, and the feature node corresponding to GDYFDR is determined as the target node, thereby determining at least one link path between the central node and the target node. Here, for each knowledge graph triple in the object-content knowledge graph, since the target object has watched GDYFDR, with "target object" as the head entity and "content watched by the target object" as the relation, the determined tail entity includes the content "GDYFDR", so the first link path is the central node (target object) - target node (GDYFDR); further, the feature vectors corresponding to all feature nodes on the first link path are weighted and aggregated to obtain the third sub-target vector corresponding to each target object of the content to be recommended.

[0164] Furthermore, since the target object's associated objects have viewed WBSYS, using "target object" as the head entity and "target object's close objects" as the relation, the determined tail entity includes "associated objects"; using "associated objects" as the head entity and "content viewed by associated objects" as the relation, the determined tail entity includes the content "WBSYS"; using "WBSYS" as the head entity and "title-actor" as the relation, the determined tail entity includes the actor "XZ"; furthermore, using "XZ" as the head entity and "actor-TV series / movies acted in" as the relation, the determined tail entity includes... The tail entity includes the TV series "GDYFDR"; thus, the second link path is obtained as: central node (target object) - intermediate node (related object) - intermediate node (WBSYS) - intermediate node - (XZ) - target node (GDYFDR); further, the feature vectors corresponding to all feature nodes on the second link path are weighted and aggregated to obtain the fourth sub-target vector corresponding to each target object of the content to be recommended. The third sub-target vector and the fourth sub-target vector are added together to obtain the target vector corresponding to each target object of the content to be recommended.

[0165] In this embodiment of the application, step 503, which involves weighted aggregation of the feature vectors corresponding to all feature nodes on each of the at least one link path to obtain the target vector, is combined with... Figure 9 To explain,

[0166] Step 531: Based on the relationship between two adjacent feature nodes on each link path, determine the weight of the latter feature node in the two adjacent feature nodes.

[0167] In this embodiment, if two adjacent feature nodes are object nodes, the correlation between the two adjacent object nodes is determined as the similarity between the two adjacent feature nodes, that is, the correlation between the two adjacent object nodes is used as the weight of the latter object. Here, the correlation between the two adjacent object nodes can be determined by the above formula (9), and this application does not impose any specific restrictions on this.

[0168] Step 532: Based on the weight of the next feature node in each link path, weight the feature vector corresponding to the next feature to obtain the second sub-target vector corresponding to each link path.

[0169] Step 533: Determine the sum of the second sub-target vectors corresponding to all link paths, which is the target vector.

[0170] In this embodiment, based on the weight corresponding to the next feature node in each link path, the feature vector corresponding to the next feature is weighted to obtain a weighted feature vector for each next feature node. The weighted feature vectors of all next feature nodes in each link path are then summed to obtain the second sub-target vector for each link path. Finally, the sum of the second sub-target vectors for all link paths is determined as the target vector for each content to be recommended. Thus, by determining the target vector for each content to be recommended based on the link path, the accuracy of content recommendation for the terminal devices used by the target device (large-screen TV or small-screen mobile device) is improved.

[0171] Based on the foregoing embodiments, this application also provides a content recommendation device 600, which can be applied to... Figure 1 , Figures 4 to 6 , Figure 8 , Figures 10 to 11 In a corresponding embodiment of the content recommendation method, referring to Figure 12 As shown, the content recommendation device 600 includes:

[0172] The acquisition module 601 is used to acquire a content knowledge graph, associated objects of the target object, and at least two pieces of content to be recommended. Each entity node in the content knowledge graph corresponds to a representation vector. The content to be recommended includes the historical browsing content of each object on different terminal devices. Each object includes the target object and associated objects.

[0173] Processing module 602 is used to construct an object-content knowledge graph related to the target object based on the content knowledge graph, the associated objects of the target object, and at least two pieces of content to be recommended;

[0174] The determination module 603 is used to determine the target vector corresponding to the target object for each content to be recommended, based on the feature vectors corresponding to each feature node in the object-content knowledge graph. The feature nodes include object nodes and entity nodes, and the feature vectors include representation vectors.

[0175] The determination module 603 is also used to determine the confidence level of each content to be recommended based on the preset content vector of each content to be recommended and the target vector corresponding to each content to be recommended;

[0176] The processing module 602 is also used to recommend content to the target object that meets the recommendation criteria.

[0177] In some embodiments, the acquisition module 601 is further configured to acquire the original communication data of the target object; the processing module 602 is further configured to extract features from the original communication data to obtain communication features, wherein the communication features include the peer number that makes a call connection with the local number corresponding to the target object, and the peer number corresponds to the first object; input the communication features into the regression model corresponding to the feature conditions satisfied by the communication features to obtain the correlation degree between the target object and the first object output by the regression model, thereby obtaining multiple correlation degrees between the target object and multiple first objects; filter the multiple correlation degrees, and determine the target social network of the target object based on the filtered correlation degrees, wherein the target social network includes associated objects.

[0178] In some embodiments, the communication characteristics satisfying the characteristic conditions include: the communication characteristics satisfying the home characteristic conditions, and / or the communication characteristics satisfying the work characteristic conditions; wherein, the home characteristic conditions include the communication characteristics satisfying high communication frequency, nighttime weekend and holiday location, and high base station overlap, and the work characteristic conditions include the communication characteristics satisfying high communication frequency, daytime workday location, and high base station overlap.

[0179] In some embodiments, the processing module 602 is further configured to initialize network nodes with each network object as a network node, wherein the network object includes the target object and the first object corresponding to the filtered association degree; based on the filtered association degree, calculate the modularity corresponding to the community where each network node is assigned to its neighboring network nodes, and determine the community corresponding to the largest modularity among multiple modularities as the community assigned to the network node; aggregate the assigned communities into a point, reconstruct the network according to the generated community structure, and continue the assignment of the next network node until the structure in the network no longer changes, thereby obtaining the target social network of the target object.

[0180] In some embodiments, the acquisition module 601 is further configured to acquire original content data from the media database and perform data deduplication on the original content data to obtain a content database; the processing module 602 is further configured to extract knowledge from the target fields in the content database to obtain an initial knowledge graph, wherein the initial knowledge graph includes entities, relationships between entities, and attribute information of entities; perform knowledge reasoning on the initial knowledge graph to obtain a target knowledge graph; and perform unsupervised learning on each entity node in the target knowledge graph to obtain a content knowledge graph that represents each entity node through representation vectors.

[0181] In some embodiments, the processing module 602 is further configured to construct a ripple set of k-th hop content related to the target object and the preference propagation of the c-th hop object based on the object-content knowledge graph, where 0≤k≤K, K is the total number of hopable entities, 0≤c≤C, and C is the total number of hopable objects; the determining module 603 is further configured to determine the similarity between each content to be recommended and the head entity for each triple in the ripple set based on the relationship between the head entity, the head entity and the tail entity, and the feature vector of the tail entity; the processing module 602 is further configured to perform a weighted summation of the feature vectors corresponding to the tail entities in the ripple set based on the similarity to obtain the first sub-target vector corresponding to each content to be recommended in the ripple set; and to sum the first sub-target vectors corresponding to the (K+C) ripple sets related to the target object to obtain the target vector corresponding to the target object for each content to be recommended.

[0182] In some embodiments, the processing module 602 is further configured to, for each content to be recommended, construct a knowledge graph triplet in the object-content knowledge graph by taking the feature node matching the target object as the center node and the feature node matching each content to be recommended as the target node; the determining module 603 is further configured to, for the knowledge graph triplet, determine at least one link path from the center node to the target node; the processing module 602 is further configured to, weighted aggregate the feature vectors corresponding to all feature nodes on each link path in the at least one link path, to obtain the target vector.

[0183] In some embodiments, the determining module 603 is further configured to determine the weight of the next feature node among two adjacent feature nodes based on the relationship between them; the processing module 602 is further configured to weight the feature vector corresponding to the next feature node in each link path based on the weight of the next feature node in each link path to obtain the second sub-target vector corresponding to each link path; the determining module 603 is further configured to determine the sum of the second sub-target vectors corresponding to all link paths as the target vector.

[0184] It should be noted that the specific implementation process of the steps performed by each module in this embodiment can be referred to Figure 1 , Figures 4 to 6 , Figure 8 , Figures 10 to 11 The implementation process of the content recommendation method provided in the corresponding embodiment will not be described in detail here.

[0185] Based on the foregoing embodiments, embodiments of this application also provide an electronic device 7, which can be applied to... Figure 1 , Figures 4 to 6 , Figure 8 , Figures 10 to 11 In a corresponding embodiment of the content recommendation method, referring to Figure 13 As shown, the electronic device 7 may include: a processor 701, a memory 702, and a communication bus 703, wherein:

[0186] The communication bus 703 is used to realize the communication connection between the processor 701 and the memory 702.

[0187] The processor 701 executes a program for a processing method stored in the memory 702 to implement the reference. Figure 1 , Figures 4 to 6 , Figure 8 , Figures 10 to 11 The corresponding implementation provides a content recommendation method.

[0188] Based on the foregoing embodiments, embodiments of this application provide a computer-readable storage medium storing one or more programs, which can be executed by one or more processors to achieve the following: Figure 1 , Figures 4 to 6 , Figure 8 , Figures 10 to 11 The corresponding implementation provides a content recommendation method.

[0189] It should be noted that the aforementioned computer-readable storage media can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc-read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0190] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0191] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0192] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0193] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0194] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0195] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0196] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A content recommendation method, characterized in that, The method includes: The system acquires a content knowledge graph, associated objects of the target object, and at least two pieces of content to be recommended. Each entity node in the content knowledge graph corresponds to a representation vector. The content to be recommended includes the historical browsing content of each object on different terminal devices. Each object includes the target object and the associated objects. Based on the content knowledge graph, the associated objects of the target object, and the at least two pieces of content to be recommended, construct an object-content knowledge graph related to the target object; For each piece of content to be recommended, a target vector corresponding to the target object is determined based on the feature vectors corresponding to each feature node in the object-content knowledge graph, wherein the feature nodes include object nodes and entity nodes, and the feature vectors include the representation vectors; The confidence level of each content to be recommended is determined based on the preset content vector of each content to be recommended and the target vector corresponding to each content to be recommended. Content whose confidence level meets the recommendation criteria is recommended to the target object; The acquisition of the associated object of the target object includes: Obtain the original communication data of the target object; The original communication data is subjected to feature extraction to obtain communication features, wherein the communication features include the peer number that makes a call connection with the local number corresponding to the target object, and the peer number corresponds to the first object; The communication features are input into the regression model corresponding to the feature conditions satisfied by the communication features to obtain the correlation degree between the target object and the first object output by the regression model, thereby obtaining multiple correlation degrees between the target object and multiple first objects. The multiple correlations are filtered, and based on the filtered correlations, the target social network of the target object is determined, wherein the target social network includes the associated object; The communication characteristics satisfying the characteristic conditions include: the communication characteristics satisfying the home characteristic conditions, and / or the communication characteristics satisfying the work characteristic conditions; wherein, the home characteristic conditions include the communication characteristics satisfying high communication frequency, nighttime, weekend and holiday location, and high base station overlap, and the work characteristic conditions include the communication characteristics satisfying high communication frequency, daytime workday location, and high base station overlap.

2. The method according to claim 1, characterized in that, The step of determining the target social network of the target object based on the filtered relevance includes: Each network object is treated as a network node, and the network nodes are initialized, wherein the network objects include the target object and the first object corresponding to the filtered relevance; Based on the filtered correlation, the modularity corresponding to each network node being assigned to the community of its neighboring network nodes is calculated, and the community corresponding to the largest modularity among multiple modularities is determined as the community to which the network node is assigned. The divided communities are aggregated into a single point, and the network is reconstructed based on the generated community structure. This process continues until the structure of the network no longer changes, thus obtaining the target social network of the target object.

3. The method according to claim 1, characterized in that, The acquired content knowledge graph includes: Obtain the original content data from the media database and perform data deduplication on the original content data to obtain the content database; Knowledge is extracted from the target fields in the content database to obtain an initial knowledge graph, wherein the initial knowledge graph includes entities, relationships between entities, and attribute information of the entities; The initial knowledge graph is used to perform knowledge reasoning to obtain the target knowledge graph; Unsupervised learning is performed on each entity node in the target knowledge graph to obtain the content knowledge graph that represents each entity node through the representation vector.

4. The method according to claim 1, characterized in that, For each piece of content to be recommended, determining the target vector corresponding to the target object based on the feature vectors corresponding to each feature node in the object-content knowledge graph includes: Based on the object-content knowledge graph, a ripple set of content related to the k-th hop and the preference propagation of the c-th hop object is constructed, where 0≤k≤K, K is the total number of hopable entities, and 0≤c≤C is the total number of hopable objects. For each triple in the ripple set, the similarity between each content to be recommended and the head entity is determined based on the relationship between the head entity, the head entity and the tail entity, and the feature vector of the tail entity. Based on the similarity, the feature vectors corresponding to the tail entities in the ripple set are weighted and summed to obtain the first sub-target vector corresponding to each content to be recommended in the ripple set. The first sub-target vectors corresponding to the K+C sets of water ripples related to the target object are added together to obtain the target vector corresponding to the target object for each content to be recommended.

5. The method according to claim 1, characterized in that, For each piece of content to be recommended, determining the target vector corresponding to the target object based on the feature vectors corresponding to each feature node in the object-content knowledge graph includes: For each piece of content to be recommended, the feature node that matches the target object is taken as the center node, and the feature node that matches each piece of content to be recommended is taken as the target node, thus constructing a knowledge graph triplet in the object-content knowledge graph; For the knowledge graph triple, determine at least one link path from the central node to the target node; The target vector is obtained by weighted aggregation of the feature vectors corresponding to all feature nodes on each of the at least one link path.

6. The method according to claim 5, characterized in that, The step of weighted aggregation of the feature vectors corresponding to all feature nodes on each of the at least one link path to obtain the target vector includes: Based on the relationship between two adjacent feature nodes on each link path, determine the weight of the next feature node in the two adjacent feature nodes. Based on the weight corresponding to the next feature node in each link path, the feature vector corresponding to the next feature is weighted to obtain the second sub-target vector corresponding to each link path. The sum of the second sub-target vectors corresponding to all link paths is determined as the target vector.

7. An electronic device, characterized in that, The electronic device includes: a processor, a memory, and a communication bus; The communication bus is used to realize the communication connection between the processor and the memory; The processor is used to execute information processing programs stored in the memory to implement the content recommendation method as described in any one of claims 1 to 6.

8. A storage medium, characterized in that, The storage medium stores one or more programs, which can be executed by one or more processors to implement the content recommendation method as described in any one of claims 1 to 6.

Images