Information recommendation method and device based on knowledge graph

Abstract

The present application discloses a method and device for information recommendation based on a knowledge graph, which relates to the field of data processing technology; the method includes: based on the target user's behavioral data, performing preference analysis on the target user to generate a user portrait of the target user; the user portrait includes the target user's positive preferences and negative preferences; based on the user portrait, performing path analysis in a preset product knowledge graph to generate at least one candidate path from the user node to the recommended product node; scoring each candidate path to obtain a path score for each candidate path; based on the path score of each candidate path, selecting a target path from the at least one candidate path; and using the recommended product corresponding to the target path as the target product to be recommended. The technical solution provided by the present application can improve the accuracy of preference content prediction results, thereby improving the accuracy of information recommendation results.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and specifically to an information recommendation method and device based on a knowledge graph. Background Art

[0002] With the rapid development of the internet and the explosive growth of data, the amount of information users face daily is increasing exponentially. Filtering content of interest from this massive amount of information has become a key challenge. Information recommendation systems aim to provide users with personalized content recommendations by analyzing their interests, behavior patterns, and context. Existing technologies often use collaborative filtering algorithms or content-based recommendation algorithms to predict content that users may be interested in based on user behavior data, such as clicks, purchases, and ratings. However, relying solely on positive user behavior data can lead to problems such as data sparsity, cold starts, and overfitting, resulting in inaccurate prediction and recommendation results.

[0003] Therefore, how to overcome the above-mentioned technical problems and defects becomes a key issue that needs to be solved. Summary of the Invention

[0004] The purpose of this application is to provide an information recommendation method and device based on knowledge graph to solve the problems raised in the above background technology, improve the accuracy of preference content prediction results, and thus improve the accuracy of information recommendation results.

[0005] The technical solution of the embodiment of the present application is implemented as follows:

[0006] This embodiment of the present application provides an information recommendation method based on a knowledge graph, the method comprising:

[0007] Based on the target user's behavioral data, a preference analysis is performed on the target user to generate a user profile of the target user; the user profile includes the target user's positive preferences and negative preferences;

[0008] Based on the user profile, path analysis is performed in a preset product knowledge graph to generate at least one candidate path from the user node to the recommended product node; the candidate path includes a first candidate path and a second candidate path, the first candidate path is determined based on the positive preference, and the second candidate path is determined based on the negative preference;

[0009] Score each candidate path to obtain a path score for each candidate path;

[0010] selecting a target path from the at least one candidate path based on the path score of each candidate path;

[0011] The recommended product corresponding to the target path is used as the target product to be recommended.

[0012] In the above solution, the behavior data includes user sessions; performing preference analysis on the target user based on the behavior data of the target user to generate a user profile of the target user includes:

[0013] Determining, based on the session data, recommended keywords presented to the target user in each interaction and reply content of the target user in response to each recommended keyword;

[0014] Determining the target user's response type to each recommended keyword based on the reply content; the response type includes a clear response, an ambiguous response, a clear rejection, and ending the conversation;

[0015] When the response type is a clear response or an ambiguous response, the corresponding recommended keyword is used as a positive preference;

[0016] When the response type is explicit rejection and session termination, the corresponding recommended keyword is used as a negative preference.

[0017] In the above solution, scoring each candidate path to obtain a path score for each candidate path includes:

[0018] Score each candidate path to obtain an initial score for each candidate path;

[0019] Based on the response type corresponding to each path, the initial score of each candidate path is updated to obtain the path score of each candidate path.

[0020] In the above solution, the behavior data includes browsing data; performing preference analysis on the target user based on the behavior data of the target user to generate a user profile of the target user includes:

[0021] Filtering data of target users browsing preset recommended web pages from the browsing data to obtain target browsing data; different recommended web pages are associated with different recommended products;

[0022] Determining, based on the target browsing data, the type of browsing behavior of the target user during each browsing of a preset recommended web page; the type of browsing behavior includes interested browsing and uninterested browsing;

[0023] When the browsing behavior type is interested browsing, the theme of the corresponding recommended web page is used as a positive preference;

[0024] When the type of the browsing behavior is browsing without interest, the theme of the corresponding recommended web page is regarded as a negative preference.

[0025] In the above solution, based on the user profile, path analysis is performed in a preset product knowledge graph to generate at least one candidate path from the user node to the recommended product node, including:

[0026] Based on the user profile, performing path analysis in a preset product knowledge graph to generate at least one first candidate path and at least one second candidate path from the user node to the recommended product node;

[0027] Determining the target user's preference for each of at least one preset recommended product based on the target user's family relationship information;

[0028] determining a target product from the at least one recommended product based on the preference level;

[0029] Based on the target product, performing path analysis in a preset product knowledge graph to generate at least one third candidate path from the application node to the recommended product node;

[0030] At least one candidate path from the user node to the recommended product node is obtained based on the first candidate path, the second candidate path, and the third candidate path.

[0031] In the above solution, determining the target user's preference for each of at least one preset recommended product based on the target user's family relationship information includes:

[0032] Based on the family relationship information of the target user, determining an intent group that matches the target user from at least one intent group associated with the recommended product, and obtaining at least one target group for the target user;

[0033] Determining the target user's attention to each target group based on the target user's consumption data;

[0034] Based on the at least one target group and the intended group associated with each of the at least one preset recommended product, establishing an association relationship between the target user, the target group, and the product; the association relationship includes a degree of association between each product and each target group;

[0035] Based on the target user's attention to each target group and the relevance of each product to each target group, the preference degree of the target user for each recommended product in the at least one recommended product is determined.

[0036] In the above solution, determining the target user's attention to each target group based on the target user's consumption data includes:

[0037] Determining the target user's basic attention to each target group based on the target user's consumption data;

[0038] Based on the public opinion analysis within the first period, determine the correlation score between each target group and public opinion;

[0039] The corresponding basic attention is weighted using the association score of each target group to obtain the attention of the target user to each target group.

[0040] In the above solution, the analysis of public opinion within the first time period is used to determine the correlation score between each target group and public opinion, including:

[0041] Based on the target user's behavior browsing data, second public opinion data browsed by the target user is filtered from the first public opinion data; the first public opinion data includes data of hot public opinions within a first time period; the second public opinion data includes at least one event;

[0042] Performing semantic analysis on the second public opinion data to obtain information about the subject person of each event;

[0043] Based on the subject person information of all events, determine the relevance between each target group of the target users and each event;

[0044] Based on the relevance of each target group to each event and the popularity of each event, the correlation score between each target group and public opinion is determined; the correlation score formula is expressed as:

[0045]

[0046] Among them, S g represents the correlation score between the target group g and public opinion, B g,i represents the relevance of the target group g to the event i in the first time period, β i represents the heat coefficient of event i, β0 is the correction coefficient, G is the target group collection, This is the second collection of public opinion data.

[0047] In the above solution, the method further includes: configuring a product knowledge graph; the configuring the product knowledge graph includes:

[0048] Obtain product labels of all preset recommended products and generate a first triple set including a label-recommended product relationship; the product label includes the type of the recommended product;

[0049] Based on the applicable populations of all preset recommended products, a second triplet set including recommended products and intended groups is generated;

[0050] Based on historical behavior data of at least one user including the target user, generating a third triplet set including user-tag relationships, a fourth triplet set including user-recommended keyword relationships, and a fifth triplet set including user-recommended webpage relationships;

[0051] A product knowledge graph is generated based on the first triple set, the second triple set, the third triple set, the fourth triple set, and the fifth triple set.

[0052] The present application also provides an information recommendation device based on a knowledge graph, the device comprising:

[0053] A data analysis unit, configured to perform a preference analysis on the target user based on the target user's behavior data, and generate a user profile of the target user; the user profile includes the target user's positive preferences and negative preferences;

[0054] a path analysis unit configured to perform path analysis in a preset product knowledge graph based on the user profile to generate at least one candidate path from a user node to a recommended product node; the candidate path comprising a first candidate path and a second candidate path, the first candidate path being determined based on the positive preference and the second candidate path being determined based on the negative preference; scoring each candidate path to obtain a path score for each candidate path; and selecting a target path from the at least one candidate path based on the path score for each candidate path;

[0055] The processing unit is configured to use the recommended product corresponding to the target path as the target product to be recommended.

[0056] The information recommendation method and device based on the knowledge graph provided in the embodiment of the present application can improve the comprehensiveness and accuracy of the user portrait by introducing negative preferences and integrating positive preferences and negative preferences to construct user portraits, so as to improve the accuracy of the recommendation results when recommending information based on the user portrait, and achieve the effect of optimizing the user experience; further, since it no longer relies solely on limited positive feedback data, it can effectively alleviate the problems of data sparsity and cold start in the recommendation of new users or new items, thereby improving the accuracy and reliability of the information recommendation results; further, the first candidate path determined based on the positive preference and the second candidate path determined based on the negative preference are selected and ranked by the scoring machine. The optimal target path is selected by the system. Since different aspects of user interests are taken into consideration, the deviation that may be caused by single-dimensional recommendation is reduced. Therefore, the matching degree between recommended content and actual user needs is improved, the push of irrelevant or unpopular content is reduced, and the user experience is optimized. Furthermore, each candidate path is quantitatively scored, and the final recommended path is determined by screening the best recommended path based on the scoring results. Since the scoring process comprehensively considers multiple factors such as path length and association strength, the objectivity and rationality of the evaluation can be guaranteed. Moreover, since the recommendation decision is based on data analysis rather than simple rule matching, the accuracy and reliability of the recommendation results can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of a flow chart of an information recommendation method based on a knowledge graph provided in an embodiment of the present application;

[0058] Figure 2 This is a flow chart of S101 in the information recommendation method based on knowledge graph in an embodiment of the present application;

[0059] Figure 3 This is a flow chart of S102 in the information recommendation method based on knowledge graph in an embodiment of the present application;

[0060] Figure 4 A schematic diagram of the structure of an information recommendation device based on a knowledge graph provided in an embodiment of the present application. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0062] In recent years, knowledge graphs can not only capture the complex relationships between entities, but also provide rich semantic information and contextual support for recommendations. With the development of knowledge graph technology, more and more fields are applying knowledge graph structured knowledge to enhance the performance of recommendation systems.

[0063] Traditional recommendation algorithms mainly rely on users' positive behavior data for prediction, which leads to some limitations in the prediction results. First, many new users or new items have data sparsity due to the lack of sufficient interaction records, resulting in poor recommendation results. Second, for newly joined users or newly launched products, due to the small number of samples, it is difficult to quickly establish an effective recommendation model, resulting in the problem of cold start. Third, over-reliance on historical preferences may cause the recommendation results to be too narrow, ignoring the user's potential interests, resulting in overfitting of the prediction results.

[0064] Based on this, in various embodiments of the present application, by introducing negative preferences and integrating positive and negative preferences to construct user portraits, the comprehensiveness and accuracy of user portraits can be improved, so that when information is recommended based on user portraits, the accuracy of recommendation results can be improved, and the effect of optimizing user experience can be achieved; further, since it no longer relies solely on limited positive feedback data, the problems of data sparsity and cold start can be effectively alleviated in the recommendation of new users or new items, thereby improving the accuracy and reliability of information recommendation results; further, the first candidate path determined based on the positive preference and the second candidate path determined based on the negative preference, and the optimal one is selected through the scoring mechanism. The target path takes into account different aspects of user interests and reduces the deviation that may be caused by single-dimensional recommendation. Therefore, it can improve the matching degree between recommended content and user's actual needs, reduce the push of irrelevant or unpopular content, and optimize user experience. Furthermore, by quantitatively scoring each candidate path and screening out the best recommended path based on the scoring results, the final recommended path is determined. Since the scoring process comprehensively considers multiple factors, such as path length, association strength, etc., it can ensure the objectivity and rationality of the evaluation. Moreover, since the recommendation decision is based on data analysis rather than simple rule matching, it can improve the accuracy and reliability of the recommendation results.

[0065] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0066] The present application provides an information recommendation method based on knowledge graph, which is applied to electronic devices, specifically personal computers, industrial computers, mobile terminals and other intelligent devices, such as Figure 1 As shown, the method may include S101 to S105. S101 to S105 are described in detail below in conjunction with specific embodiments.

[0067] S101: Based on the target user's behavior data, perform a preference analysis on the target user to generate a user profile of the target user; the user profile includes the target user's positive preferences and negative preferences.

[0068] In actual application, you can first collect the target user's behavioral data from multiple channels, and then process the above behavioral data through data analysis tools to preliminarily identify the target user's interests and dislikes.

[0069] In actual application, user preferences can be analyzed based on the content of user conversations, including call recordings and online conversations between sales customer service and users.

[0070] Based on this, in one embodiment, the behavioral data includes user sessions.

[0071] Based on this, in one embodiment, if Figure 2 As shown, performing preference analysis on the target user based on the target user's behavior data to generate a user profile of the target user, i.e., S101, may include:

[0072] S201: Determine, based on the session data, recommended keywords presented to a target user in each interaction and the target user's reply content in response to each recommended keyword;

[0073] S202: Determine the target user's response type to each recommended keyword based on the reply content; the response type includes clear response, ambiguous response, clear rejection, and end of conversation;

[0074] S203: When the response type is a clear response or an ambiguous response, the corresponding recommended keyword is used as a positive preference;

[0075] S204: When the response type is explicit rejection and session termination, the corresponding recommended keyword is used as a negative preference.

[0076] In actual application, for S201, when the customer service and the target user complete a reply, it can be regarded as an interaction. For example, when the customer service asks "There is product A that is suitable for you. Can I introduce it to you now?", and the target user replies "Yes", a query and reply are completed, that is, an interaction is completed; here, the recommended keyword can be content related to the information topic to be recommended, such as "product A" in the above interaction.

[0077] In actual application, for S202, when the user does not clearly reply to the customer service question, that is, the user replies to the inquiry but the reply content is not directed to the customer service inquiry, it can be regarded as the user's low interest, but not strong resistance. At this time, it can be judged that the user's response type is a vague response; at the same time, when the user clearly replies to the customer service question and it is a positive reply, it can be regarded as a clear response. Correspondingly, when the user's content is a clear reply to refuse, it can be regarded as a clear refusal.

[0078] For example, when customer service asks, "Funds are mainly divided into money market funds, bond funds, hybrid funds, and stock funds. Which one do you prefer?", the recommended keywords are money market funds, bond funds, hybrid funds, and stock funds. If the user replies "money market funds," the user response type is a clear response. Since the user gave a clear response, the platform can add "money market funds" to the user's positive preferences. If the user replies, "I don't know much about bond funds," although the user's response is not very clear, the platform can infer from the context that the user is more interested in "bond funds." Therefore, the platform can cautiously add "hybrid funds" as a potential positive preference to Xiao Li's positive preference list. Correspondingly, if the user replies with negative terms such as "I'm not interested in any of them" or "I don't have time now," the user response type is a clear rejection, and all the above recommended keywords are added to the user's negative preference list. If the user directly ends the conversation after hearing the question, for example, by hanging up the phone or leaving the online chat interface, it can be determined that the user has a strong resistance to the keyword, and the user response type is end of conversation. In this case, all the above recommended keywords are added to the user's negative preference list.

[0079] In actual application, user preferences can also be determined based on user browsing behavior.

[0080] Based on this, in one embodiment, the behavior data includes browsing data; performing preference analysis on the target user based on the behavior data of the target user to generate a user profile of the target user may include:

[0081] Filtering data of target users browsing preset recommended web pages from the browsing data to obtain target browsing data; different recommended web pages are associated with different recommended products;

[0082] Determining, based on the target browsing data, the type of browsing behavior of the target user during each browsing of a preset recommended web page; the type of browsing behavior includes interested browsing and uninterested browsing;

[0083] When the browsing behavior type is interested browsing, the theme of the corresponding recommended web page is used as a positive preference;

[0084] When the type of the browsing behavior is browsing without interest, the theme of the corresponding recommended web page is regarded as a negative preference.

[0085] In actual application, conditions for different browsing behavior types can be pre-configured; specifically, the types of browsing behaviors may include actively searching to enter a web page, actively clicking on a third-party link to enter a web page, and entering a web page after being redirected by a third-party application; when actively searching to enter a web page or actively clicking on a third-party link to enter a web page, the type is interested browsing; when entering a web page after being automatically redirected by a third-party application, if the browsing time is greater than a preset judgment threshold, such as 5s, or there is interactive behavior, such as sliding the web page up and down or partially zooming in on the web page, the type is interested browsing, otherwise it is uninterested browsing.

[0086] S102: Based on the user portrait, path analysis is performed in a preset product knowledge graph to generate at least one candidate path from the user node to the recommended product node; the candidate path includes a first candidate path and a second candidate path, the first candidate path is determined based on the positive preference, and the second candidate path is determined based on the negative preference.

[0087] In actual application, recommended products can be understood as preset products to be recommended, such as insurance products, financial products, etc.; recommended products can also be called recommended information.

[0088] In one embodiment, if Figure 3 As shown, the process of performing path analysis in a preset product knowledge graph based on the user profile to generate at least one candidate path from the user node to the recommended product node, i.e., S102, may include:

[0089] S301: Based on the user profile, perform path analysis in a preset product knowledge graph to generate at least one first candidate path and at least one second candidate path from the user node to the recommended product node;

[0090] S302: Determine the target user's preference for each of at least one preset recommended product based on the target user's family relationship information;

[0091] S303: Determine a target product from the at least one recommended product based on the preference level;

[0092] S304: Based on the target product, perform path analysis in a preset product knowledge graph to generate at least one third candidate path from the application node to the recommended product node;

[0093] S305: Based on the first candidate path, the second candidate path, and the third candidate path, obtain at least one candidate path from the user node to the recommended product node.

[0094] In actual application, based on the user portrait, path analysis is performed in the preset product knowledge graph to generate at least one first candidate path and at least one second candidate path from the user node to the recommended product node. Based on the positive preference, path analysis can be performed in the preset product knowledge graph to obtain at least one first candidate path from the user node to the recommended product node; based on the negative preference, path analysis can be performed in the preset product knowledge graph to obtain at least one second candidate path from the user node to the recommended product node.

[0095] Here, due to the introduction of the user's family relationship and the combination of the needs and characteristics of family members, the user's real needs and potential interests can be identified more accurately, the relevance and applicability of the recommended content are improved, and thus the matching degree between the recommended products and the user is improved.

[0096] In one embodiment, determining the target user's preference for each of at least one preset recommended product based on the target user's family relationship information, i.e., S302, may include:

[0097] Based on the family relationship information of the target user, determining an intent group that matches the target user from at least one intent group associated with the recommended product, and obtaining at least one target group for the target user;

[0098] Determining the target user's attention to each target group based on the target user's consumption data;

[0099] Based on the at least one target group and the intended group associated with each of the at least one preset recommended product, establishing an association relationship between the target user, the target group, and the product; the association relationship includes a degree of association between each product and each target group;

[0100] Based on the target user's attention to each target group and the relevance of each product to each target group, the preference degree of the target user for each recommended product in the at least one recommended product is determined.

[0101] In actual application, the intention group associated with the recommended product can be understood as the applicable group of the recommended product. For example, health insurance, education fund and other products are suitable for children under 18 years old, so the intended group of the product is the children group; the target group of the target user can be understood as the intention group to which each family member of the target user belongs. For example, if user A has a 6-year-old child among his family members, then user A's target group includes the children group.

[0102] It should be noted that the intent groups and the intent groups associated with each product can be pre-configured according to the product content and application scenarios, and this application does not limit this.

[0103] In one embodiment, determining the target user's attention to each target group based on the target user's consumption data may include:

[0104] Determining the target user's basic attention to each target group based on the target user's consumption data;

[0105] Based on the public opinion analysis within the first period, determine the correlation score between each target group and public opinion;

[0106] The corresponding basic attention is weighted using the association score of each target group to obtain the attention of the target user to each target group.

[0107] In actual application, when determining the target user's basic attention to each target group based on the target user's consumption data, the target user's basic attention to each family member can be calculated based on the target user's consumption amount and consumption frequency for each family member, that is, the target user's basic attention to the target group to which each family member belongs, wherein the higher the consumption amount, the higher the basic attention, and the higher the consumption frequency, the higher the basic attention; illustratively, the target user's basic attention to each target group can be calculated using the following formula:

[0108]

[0109] Among them, A i represents the target user’s basic attention to family member i, that is, the basic attention to the target group to which family member i belongs, C i represents the target user's spending amount for family member i, F i represents the target user’s consumption frequency of family member i, and n represents the number of family members.

[0110] In one embodiment, determining the correlation score between each target group and public opinion based on the public opinion analysis within the first time period may include:

[0111] Based on the target user's behavior browsing data, second public opinion data browsed by the target user is filtered from the first public opinion data; the first public opinion data includes data of hot public opinions within a first time period; the second public opinion data includes at least one event;

[0112] Performing semantic analysis on the second public opinion data to obtain information about the subject person of each event;

[0113] Based on the subject person information of all events, determine the relevance between each target group of the target users and each event;

[0114] Based on the relevance of each target group to each event and the popularity of each event, the correlation score between each target group and public opinion is determined; the correlation score formula is expressed as:

[0115]

[0116] Among them, S g represents the correlation score between the target group g and public opinion, B g,i represents the relevance of the target group g to the event i in the first time period, β i represents the heat coefficient of event i, β0 is the correction coefficient, G is the target group collection, This is the second collection of public opinion data.

[0117] In actual application, the first duration can be configured according to the actual application scenario and the update of hot public opinion. For example, the first duration can be within one month or within one week. The specific configuration is not limited in the embodiment of this application.

[0118] In actual application, the correlation relationship between different event types and each intention group can be pre-built. When determining the correlation between each target group of the target user and each event based on the subject character information of all events, the event type can be determined according to the subject character information of each event, and then the correlation between the event and each target group can be determined according to the time type.

[0119] In actual application, event heat can be obtained from a third-party public opinion monitoring and data analysis platform; after obtaining the time heat, the event heat of all events can be normalized to obtain the heat coefficient of each event, and then the correlation score between each target group and public opinion can be calculated based on the relevance of each target group to each event and the heat coefficient of each event.

[0120] In actual application, after determining the correlation score between each target group and public opinion, the correlation score can be normalized first, and then the normalized correlation score can be used to weight the corresponding basic attention to obtain the attention of the target user to each target group.

[0121] In actual application, the user's willingness to purchase is affected by the external public opinion environment. For example, from October to December, the virus is prevalent. For users with children and elderly people in their family, they pay more attention to such behaviors, so they are more willing to purchase medical insurance for children and the elderly. The embodiment of the present application determines the user's attention to each target group by introducing the user's attention to public opinion and the correlation between public opinion and each target group. In addition to traditional consumption behavior data, it can accurately capture the user's real-time interests and changes in purchasing intentions, adding a new dimension to the user portrait, thereby improving the completeness and accuracy of the user portrait, and further improving the prediction accuracy of the recommendation model.

[0122] In actual application, after determining the target user's attention to each target group, the association relationship between the target user, target group and product can be established based on the at least one target group and the intention group associated with each recommended product in the preset at least one recommended product; in this process, the recommended products corresponding to the target group can be determined based on which intention groups each recommended product is suitable for, thereby establishing the association relationship between the target user, target group and product.

[0123] In actual application, after determining the target user's attention to each target group and establishing the association relationship between the target user, target group and product, the target user's preference for each of the at least one recommended product can be determined based on the target user's attention to each target group and the association degree of each product with each target group. In this process, for each recommended product, the target group associated with the recommended product can be first determined, and then the product of the target user's attention to the target group and the association degree between the product and the target group can be calculated to obtain the preference component for the target group. Then, the preference components of all target groups can be added together to obtain the target user's preference for the recommended product. For example, the target user's preference for the recommended product can be expressed as follows:

[0124]

[0125] Among them, P k Indicates the target user’s preference for the recommended product k, A i ′ represents the target user’s attention to family member i, that is, the attention to the target group to which family member i belongs, R ki It represents the correlation between the recommended product k and the target group (i.e., the target group to which family member i belongs).

[0126] In actual application, the correlation between each product and the target group can also be called the adaptability of each product to the target group; it should be pointed out that the correlation can be pre-configured according to the product content and application scenario, and this application does not limit this; for example, the correlation between children's medical insurance and children's group, young parents group, and family pillar group is 80%, 70% and 20% respectively, and the correlation between family health insurance and family pillar group, couple joint protection group, and child growth support group is 90%, 80% and 50% respectively.

[0127] In actual application, for S303 , the N recommended products with the highest preference levels may be used as target products, where N is an integer greater than 0.

[0128] S103: Score each candidate path to obtain a path score for each candidate path.

[0129] In actual application, considering that the response types corresponding to the recommended keywords in positive preference and negative preference are different, the user's preference strength presented is also different. Therefore, when determining the path score, the path score can be weighted according to the user's preference strength for the recommended keywords in the path.

[0130] Based on this, in one embodiment, scoring each candidate path to obtain a path score for each candidate path includes:

[0131] Score each candidate path to obtain an initial score for each candidate path;

[0132] Based on the response type corresponding to each path, the initial score of each candidate path is updated to obtain the path score of each candidate path.

[0133] In one embodiment, updating the initial score of each candidate path based on the response type corresponding to each path to obtain the path score of each candidate path may include:

[0134] Based on the response type corresponding to each path, the initial score of each candidate path is weighted; among them, the assignment of a clear response is greater than the assignment of an ambiguous response, and the assignment of ending the session is greater than the assignment of a clear rejection.

[0135] In practical applications, different response types can be quantitatively assigned first, and then the assignment results can be used as weights to update the initial scores.

[0136] Here, in the process of calculating the path score, the user's preference for the recommended products corresponding to the path is introduced to calculate the final score of the path, that is, the credibility of the path score result is introduced. Since the path score takes into account the credibility of the path score result while considering the optimal path, it can improve the accuracy and reliability of the path score result, thereby improving the accuracy of the recommendation result.

[0137] S104: Selecting a target path from the at least one candidate path based on the path score of each candidate path.

[0138] In practical applications, the top M paths with the highest path scores can be used as target paths, where M is an integer greater than 0.

[0139] S105: Taking the recommended product corresponding to the target path as the target product to be recommended.

[0140] In actual application, the product knowledge graph can be pre-configured according to the products to be recommended.

[0141] Based on this, in one embodiment, the method further includes: configuring a product knowledge graph.

[0142] In one embodiment, the configuration product knowledge graph may include:

[0143] Obtain product labels of all preset recommended products and generate a first triple set including a label-recommended product relationship; the product label includes the type of the recommended product;

[0144] Based on the applicable populations of all preset recommended products, a second triplet set including recommended products and intended groups is generated;

[0145] Based on historical behavior data of at least one user including the target user, generating a third triplet set including user-tag relationships, a fourth triplet set including user-recommended keyword relationships, and a fifth triplet set including user-recommended webpage relationships;

[0146] A product knowledge graph is generated based on the first triple set, the second triple set, the third triple set, the fourth triple set, and the fifth triple set.

[0147] In summary, the information recommendation method based on knowledge graph provided by the embodiment of the present application, by introducing negative preferences, realizes the construction of user portraits by integrating positive preferences and negative preferences, which can improve the comprehensiveness and accuracy of user portraits, thereby improving the accuracy of recommendation results when recommending information based on user portraits, and achieving the effect of optimizing user experience; further, since it no longer relies solely on limited positive feedback data, it can effectively alleviate the problems of data sparsity and cold start in the recommendation of new users or new items, thereby improving the accuracy and reliability of information recommendation results; further, the first candidate path determined based on the positive preference and the second candidate path determined based on the negative preference, and through the evaluation The optimal target path is selected by the sub-mechanism. Since different aspects of user interests are taken into consideration, the deviation that may be caused by single-dimensional recommendation is reduced. Therefore, the matching degree between recommended content and actual user needs is improved, the push of irrelevant or unpopular content is reduced, and the user experience is optimized. Furthermore, each candidate path is quantitatively scored, and the final recommended path is determined by screening the best recommended path based on the scoring results. Since the scoring process comprehensively considers multiple factors such as path length and association strength, the objectivity and rationality of the evaluation can be guaranteed. Moreover, since the recommendation decision is based on data analysis rather than simple rule matching, the accuracy and reliability of the recommendation results can be improved.

[0148] In order to implement the information recommendation method based on knowledge graph of the present application, the embodiment of the present application also provides an information recommendation device based on knowledge graph, which is set on an electronic device, such as Figure 4 As shown, the device may include:

[0149] The data analysis unit 401 is configured to perform a preference analysis on the target user based on the target user's behavior data and generate a user profile of the target user; the user profile includes the target user's positive preferences and negative preferences;

[0150] A path analysis unit 402 is configured to perform path analysis in a preset product knowledge graph based on the user profile to generate at least one candidate path from the user node to the recommended product node; the candidate path includes a first candidate path and a second candidate path, the first candidate path being determined based on the positive preference and the second candidate path being determined based on the negative preference; score each candidate path to obtain a path score for each candidate path; and select a target path from the at least one candidate path based on the path score of each candidate path;

[0151] The processing unit 403 is configured to use the recommended product corresponding to the target path as the target product to be recommended.

[0152] In one embodiment, the behavior data includes user sessions; the data analysis unit 401 may be configured to:

[0153] Determining, based on the session data, recommended keywords presented to the target user in each interaction and reply content of the target user in response to each recommended keyword;

[0154] Determining the target user's response type to each recommended keyword based on the reply content; the response type includes a clear response, an ambiguous response, a clear rejection, and ending the conversation;

[0155] When the response type is a clear response or an ambiguous response, the corresponding recommended keyword is used as a positive preference;

[0156] When the response type is explicit rejection and session termination, the corresponding recommended keyword is used as a negative preference.

[0157] In one embodiment, the path analysis unit 402 may be configured to:

[0158] Score each candidate path to obtain an initial score for each candidate path;

[0159] Based on the response type corresponding to each path, the initial score of each candidate path is updated to obtain the path score of each candidate path.

[0160] In one embodiment, the behavior data includes browsing data; the data analysis unit 401 may be configured to:

[0161] Filtering data of target users browsing preset recommended web pages from the browsing data to obtain target browsing data; different recommended web pages are associated with different recommended products;

[0162] Determining, based on the target browsing data, the type of browsing behavior of the target user during each browsing of a preset recommended web page; the type of browsing behavior includes interested browsing and uninterested browsing;

[0163] When the browsing behavior type is interested browsing, the theme of the corresponding recommended web page is used as a positive preference;

[0164] When the type of the browsing behavior is browsing without interest, the theme of the corresponding recommended web page is regarded as a negative preference.

[0165] In one embodiment, the path analysis unit 402 may be configured to:

[0166] Based on the user profile, performing path analysis in a preset product knowledge graph to generate at least one first candidate path and at least one second candidate path from the user node to the recommended product node;

[0167] Determining the target user's preference for each of at least one preset recommended product based on the target user's family relationship information;

[0168] determining a target product from the at least one recommended product based on the preference level;

[0169] Based on the target product, performing path analysis in a preset product knowledge graph to generate at least one third candidate path from the application node to the recommended product node;

[0170] At least one candidate path from the user node to the recommended product node is obtained based on the first candidate path, the second candidate path, and the third candidate path.

[0171] In one embodiment, the path analysis unit 402 is specifically configured to:

[0172] Based on the family relationship information of the target user, determining an intent group that matches the target user from at least one intent group associated with the recommended product, and obtaining at least one target group for the target user;

[0173] Determining the target user's attention to each target group based on the target user's consumption data;

[0174] Establishing an association relationship between the target user, the target group, and the product based on the at least one target group and the intended group associated with each recommended product in the at least one preset recommended product;

[0175] Based on the target user's attention to each target group and the relevance of each product to each target group, the preference degree of the target user for each recommended product in the at least one recommended product is determined.

[0176] In one embodiment, the path analysis unit 402 is specifically configured to:

[0177] Determining the target user's basic attention to each target group based on the target user's consumption data;

[0178] Based on the public opinion analysis within the first period, determine the correlation score between each target group and public opinion;

[0179] The corresponding basic attention is weighted using the association score of each target group to obtain the attention of the target user to each target group.

[0180] In one embodiment, the path analysis unit 402 is specifically configured to:

[0181] Based on the target user's behavior browsing data, second public opinion data browsed by the target user is filtered from the first public opinion data; the first public opinion data includes data of hot public opinions within a first time period; the second public opinion data includes at least one event;

[0182] Performing semantic analysis on the second public opinion data to obtain information about the subject person of each event;

[0183] Based on the subject person information of all events, determine the relevance between each target group of the target users and each event;

[0184] Based on the relevance of each target group to each event and the popularity of each event, the correlation score between each target group and public opinion is determined; the correlation score formula is expressed as:

[0185]

[0186] Among them, S g represents the correlation score between the target group g and public opinion, B g,i represents the relevance of the target group g to the event i in the first time period, β i represents the heat coefficient of event i, β0 is the correction coefficient, G is the target group collection, This is the second collection of public opinion data.

[0187] In one embodiment, the apparatus further includes a configuration unit; the configuration unit is configured to:

[0188] Obtain product labels of all preset recommended products and generate a first triple set including a label-recommended product relationship; the product label includes the type of the recommended product;

[0189] Based on the applicable populations of all preset recommended products, a second triplet set including recommended products and intended groups is generated;

[0190] Based on historical behavior data of at least one user including the target user, generating a third triplet set including user-tag relationships, a fourth triplet set including user-recommended keyword relationships, and a fifth triplet set including user-recommended webpage relationships;

[0191] A product knowledge graph is generated based on the first triple set, the second triple set, the third triple set, the fourth triple set, and the fifth triple set.

[0192] It should be noted that the above-mentioned embodiment provides an information recommendation device based on a knowledge graph, and only uses the division of the above-mentioned program modules as an example to illustrate when performing information recommendation based on a knowledge graph. In actual applications, the above-mentioned processing can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program modules to complete all or part of the processing described above. In addition, the information recommendation device based on a knowledge graph provided in the above-mentioned embodiment and the embodiment of the information recommendation method based on a knowledge graph are of the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0193] It should be noted that: "first", "second", etc. are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0194] In addition, the technical solutions described in the embodiments of the present application can be arbitrarily combined without conflict.

[0195] The above description is merely a preferred embodiment of the present application and is not intended to limit the scope of protection of the present application.

Claims

1. An information recommendation method based on knowledge graph, characterized in that: The method comprises: Based on the target user's behavioral data, a preference analysis is performed on the target user to generate a user profile of the target user; the user profile includes the target user's positive preferences and negative preferences; Based on the user profile, path analysis is performed in a preset product knowledge graph to generate at least one candidate path from the user node to the recommended product node; the candidate path includes a first candidate path and a second candidate path, the first candidate path is determined based on the positive preference, and the second candidate path is determined based on the negative preference; Score each candidate path to obtain a path score for each candidate path; selecting a target path from the at least one candidate path based on the path score of each candidate path; The recommended product corresponding to the target path is used as the target product to be recommended; wherein, The method of performing path analysis in a preset product knowledge graph based on the user profile to generate at least one candidate path from the user node to the recommended product node includes: Based on the user portrait, path analysis is performed in a preset product knowledge graph to generate at least one first candidate path and at least one second candidate path from the user node to the recommended product node; based on the family relationship information of the target user, the target user's preference for each of at least one preset recommended product is determined; based on the preference, a target product is determined from the at least one recommended product; based on the target product, path analysis is performed in a preset product knowledge graph to generate at least one third candidate path from the user node to the recommended product node; based on the first candidate path, the second candidate path, and the third candidate path, at least one candidate path from the user node to the recommended product node is obtained; The determining, based on the target user's family relationship information, the target user's preference for each of at least one preset recommended product includes: Based on the family relationship information of the target user, an intention group matching the target user is determined from at least one intention group associated with the recommended product to obtain at least one target group of the target user; based on the consumption data of the target user, the target user's attention to each target group is determined; based on the at least one target group and the intention group associated with each recommended product in at least one preset recommended product, an association relationship among the target user, the target group and the product is established; the association relationship includes the association degree between each product and each target group; based on the target user's attention to each target group and the association degree between each product and each target group, the target user's preference for each recommended product in the at least one recommended product is determined.

2. The method according to claim 1, characterized in that The behavior data includes user sessions; performing preference analysis on the target user based on the behavior data of the target user to generate a user profile of the target user includes: Determining, based on the session data, recommended keywords presented to the target user in each interaction and reply content of the target user in response to each recommended keyword; Determining the target user's response type to each recommended keyword based on the reply content; the response type includes a clear response, an ambiguous response, a clear rejection, and ending the conversation; When the response type is a clear response or an ambiguous response, the corresponding recommended keyword is used as a positive preference; When the response type is explicit rejection and session termination, the corresponding recommended keyword is used as a negative preference.

3. Based on the method of claim 2, scoring each candidate path to obtain a path score for each candidate path includes: Score each candidate path to obtain an initial score for each candidate path; Based on the response type corresponding to each path, the initial score of each candidate path is updated to obtain the path score of each candidate path.

4. The method according to claim 1, wherein The behavior data includes browsing data; performing preference analysis on the target user based on the behavior data of the target user to generate a user profile of the target user includes: Filtering data of target users browsing preset recommended web pages from the browsing data to obtain target browsing data; different recommended web pages are associated with different recommended products; Determining, based on the target browsing data, the type of browsing behavior of the target user during each browsing of a preset recommended web page; the type of browsing behavior includes interested browsing and uninterested browsing; When the browsing behavior type is interested browsing, the theme of the corresponding recommended web page is used as a positive preference; When the type of the browsing behavior is browsing without interest, the theme of the corresponding recommended web page is regarded as a negative preference.

5. The method according to claim 1, wherein The determining, based on the target user's consumption data, the target user's attention to each target group includes: Determining the target user's basic attention to each target group based on the target user's consumption data; Based on the public opinion analysis within the first period, determine the correlation score between each target group and public opinion; The corresponding basic attention is weighted using the association score of each target group to obtain the attention of the target user to each target group.

6. The method according to claim 5, characterized in that The analysis of public opinion within the first time period is based on determining the correlation score between each target group and public opinion, including: Based on the target user's behavior browsing data, second public opinion data browsed by the target user is filtered from the first public opinion data; the first public opinion data includes data of hot public opinions within a first time period; the second public opinion data includes at least one event; Performing semantic analysis on the second public opinion data to obtain information about the subject person of each event; Based on the subject person information of all events, determine the relevance between each target group of the target users and each event; Based on the relevance of each target group to each event and the popularity of each event, the correlation score between each target group and public opinion is determined; the correlation score formula is expressed as: ； in, Indicates the target group The correlation score with public opinion, Indicates the target group In the first period and events The correlation of Representing an event The heat coefficient, is the correction factor, For the target group collection, This is the second collection of public opinion data.

7. The method according to claim 1, characterized in that The method further includes: configuring a product knowledge graph; configuring the product knowledge graph includes: Obtain product labels of all preset recommended products and generate a first triple set including a label-recommended product relationship; the product label includes the type of the recommended product; Based on the applicable populations of all preset recommended products, a second triplet set including recommended products and intended groups is generated; Based on historical behavior data of at least one user including the target user, generating a third triplet set including user-tag relationships, a fourth triplet set including user-recommended keyword relationships, and a fifth triplet set including user-recommended webpage relationships; A product knowledge graph is generated based on the first triple set, the second triple set, the third triple set, the fourth triple set, and the fifth triple set.

8. An information recommendation device based on knowledge graph, characterized in that: The device comprises: A data analysis unit, configured to perform a preference analysis on the target user based on the target user's behavior data, and generate a user profile of the target user; the user profile includes the target user's positive preferences and negative preferences; a path analysis unit configured to perform path analysis in a preset product knowledge graph based on the user profile to generate at least one candidate path from a user node to a recommended product node; the candidate path comprising a first candidate path and a second candidate path, the first candidate path being determined based on the positive preference and the second candidate path being determined based on the negative preference; scoring each candidate path to obtain a path score for each candidate path; and selecting a target path from the at least one candidate path based on the path score for each candidate path; The processing unit is configured to use the recommended product corresponding to the target path as the target product to be recommended; wherein, The path analysis unit is used to: Based on the user portrait, path analysis is performed in a preset product knowledge graph to generate at least one first candidate path and at least one second candidate path from the user node to the recommended product node; based on the family relationship information of the target user, the target user's preference for each of at least one preset recommended product is determined; based on the preference, a target product is determined from the at least one recommended product; based on the target product, path analysis is performed in a preset product knowledge graph to generate at least one third candidate path from the user node to the recommended product node; based on the first candidate path, the second candidate path, and the third candidate path, at least one candidate path from the user node to the recommended product node is obtained; The path analysis unit is used to: Based on the family relationship information of the target user, an intention group matching the target user is determined from at least one intention group associated with the recommended product to obtain at least one target group of the target user; based on the consumption data of the target user, the target user's attention to each target group is determined; based on the at least one target group and the intention group associated with each recommended product in at least one preset recommended product, an association relationship among the target user, the target group and the product is established; the association relationship includes the association degree between each product and each target group; based on the target user's attention to each target group and the association degree between each product and each target group, the target user's preference for each recommended product in the at least one recommended product is determined.

Images