A knowledge graph data classification and management method and system

Abstract

The present invention discloses a knowledge graph data classification and management method and system, which relates to the field of data management technology, and is used to solve the problem that when users search for new keywords, they do not take into account the possibility of user errors, thereby polluting user portraits. A long-term interest model and a short-term behavior model are established through the long-term and short-term keyword search frequencies and similarities, and user deep and shallow layer portraits and keyword overlap rates are generated. Fuzzy logic is established through the keyword overlap rate, the average deviation of the number of user click entity relationship layers, and the short-term bounce rate to classify and re-rank the strength of the relationship between entities. When the shallow layer portrait is saturated, an evaluation model is established by evaluating the viscosity, browsing time and access depth of the new keyword, and the new keyword evaluation coefficient is obtained and compared with the screening threshold to determine the screening result of the new keyword; thus, it is avoided that the user inputs keywords by mistake and causes the user portrait to be polluted, and the user experience is improved.

Description

Technical Field

[0001] The present invention relates to the field of data management technology, and more specifically, to a knowledge graph data classification and management method and system. Background Art

[0002] With the development of big data technology, knowledge graphs have become an important way of representing and organizing knowledge. Knowledge graphs are usually applicable in areas such as intelligent recommendation systems, search engines, and enterprise intelligent management. However, the classification of common knowledge graphs relies on the system analyzing the relationships between entities and then making recommendations based on the similarity of each relationship. Obviously, when users search for new keywords, only the strength of the connection between the entity expressed in the keyword and the other entities will be considered to sort and recommend, without considering the possibility of user errors, which will pollute the user portrait and affect the user's subsequent search experience. Therefore, a knowledge graph data classification and management method and system are proposed.

[0003] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a knowledge graph data classification and management method and system, which solves the problems raised in the above-mentioned background technology by using different product inspection methods.

[0005] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a knowledge graph data classification and management method, comprising S1: collecting behavior feature data and time feature data, and obtaining long-term and short-term keyword search frequencies, long-term and short-term keyword similarities, user click entity relationship layer deviation average value and short-term user access bounce rate through data processing;

[0006] S2: Obtain the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities to establish a user long-term interest model and a user short-term behavior model, and obtain a user deep and shallow profile based on the comparison, and calculate the keyword overlap rate for each keyword set;

[0007] S3: Obtain the keyword overlap rate, the average deviation of the number of levels of entity relationships clicked by users, and the short-term user visit bounce rate, use fuzzy logic to determine the entity ranking adjustment result, perform fuzzy reasoning based on fuzzy rules, and determine the entity change ranking;

[0008] S4: When the user's shallow profile is saturated, the new keyword is obtained to establish a data analysis model by obtaining the entity viscosity between the new keyword and the user's deep profile, the browsing time on the new keyword recommendation page, and the new keyword access depth to obtain the keyword evaluation coefficient;

[0009] S5: Obtain the keyword evaluation coefficient, and compare and analyze it with the screening threshold to obtain a comparison result, and determine a new keyword division result based on the comparison result.

[0010] In a preferred embodiment, by expressing a long-term and short-term time period set, long-term search records and short-term search records are obtained, and the frequency of the long-term and short-term keywords in the total data set is calculated to obtain the long-term and short-term keyword search frequency;

[0011] Through the long-term and short-term keyword sets, the keywords are vectorized, the word embedding model is used to obtain the word vector expression of the long-term keywords, and the cosine similarity method is selected to calculate the similarity of the long-term and short-term keywords;

[0012] By setting the number of relationship layers between entities in the knowledge graph, the number of relationship layers represents the path length between two entities. For each short-term click entity, the minimum number of relationship layers between it and the long-term interest entity is calculated, and the average deviation of the number of relationship layers of the user's click entity is obtained.

[0013] After defining the bounce behavior, the short-term user visit bounce rate Br is obtained by statistically calculating the ratio of the number of bounces in all user access records in a short period of time to the total number of visits.

[0014] In a preferred embodiment, the long-term and short-term keyword frequency matrix is ​​input into the LDA model to extract the topic distribution. The input of the LDA model is data in the bag-of-words model format. Each document d can be represented as a word frequency vector:

[0015] D=[n(w1,d),n(w2,d),…,n(w n , d)]

[0016] Among them, n(w1, d) represents the keyword w i The number of times it appears in document d;

[0017] The core idea of ​​the LDA model is to represent documents as a distribution of topics, and each topic is represented as a distribution of words. The training of the LDA model is as follows:

[0018] Let K be the number of topics, β k is the word distribution of the kth topic, θ l is the topic distribution of document d, then the goal of the LDA model is to maximize the likelihood function of the document:

[0019]

[0020] Among them, α and β are the hyperparameters of topic distribution and word distribution respectively;

[0021] Extract the topic distribution of each keyword from the LDA model, and the topic distribution of the long-term interest model is: long ={P(w1|T1), P(w2|T2),…,P(w n |T k )}, then the topic distribution of the short-term behavior model is: T short ={P(y1|T1), P(y2|T2),…,P(y m |T k )}.

[0022] In a preferred embodiment, a threshold Q is set, and for each keyword, its probability P in the long-term topic distribution is compared. long (w i |T) and the corresponding probability P under the short-term topic distribution short (y m |T) are all greater than Q;

[0023] If both are greater than Q, it is considered that the keyword has a high probability in both the long-term and short-term models, indicating that the keyword has a high match in the long-term and short-term models, and the keyword is classified into the user's deep profile. If any one of them is greater than Q, it can be considered that the keyword has a match in the long-term or short-term model, but it does not completely overlap, and the keyword is classified into the user's shallow profile.

[0024] According to the Jaccard similarity, the overlap rate SI of the keyword sets in the long-term model and the short-term model is calculated based on the formula:

[0025]

[0026] Among them, |A∩B| represents the number of keywords that appear together in the long-term and short-term models, and |A∪B| represents the number of all different keywords that appear in the long-term and short-term models.

[0027] In a preferred embodiment, the keyword overlap rate, the average deviation of the number of levels of entity relationships clicked by users, and the short-term user access bounce rate are defined as input variables, and they are divided into different fuzzy sets respectively;

[0028] The result of entity ranking adjustment is defined as output variable and divided into fuzzy sets;

[0029] Formulate fuzzy rules to describe the impact of keyword overlap rate, average deviation of the number of entity relationship layers clicked by users, and short-term user visit bounce rate on entity ranking adjustment results;

[0030] Fuzzy reasoning is performed based on fuzzy rules to determine the entity change order.

[0031] In a preferred embodiment, the memory saturation of the user's shallow profile is monitored, and if saturation occurs, new keywords in the user's shallow profile are evaluated.

[0032] In a preferred embodiment, the viscosity V between the new keyword and the entity in the user's deep profile is obtained by the relationship number and the relationship strength ratio. k ; Wherein, k = 1, 2, 3, 4, ... P, P is the total number of new keywords;

[0033] The browsing time Ti of the new keyword recommendation page is calculated by the difference between the page leaving time and the page loading time. k ;

[0034] The new keyword access depth De is obtained by adding the initial page depth and the number of clicks. k .

[0035] In a preferred embodiment, the keyword evaluation coefficient Si is generated based on the entity viscosity between the new keyword and the user's deep profile, the browsing time in the new keyword recommendation page, and the new keyword access depth. k ; The formula is as follows:

[0036]

[0037] In the formula, Sk is the keyword evaluation coefficient, as well as are the calculation weight coefficients of the entity viscosity between keywords and user deep portraits, the browsing time in the new keyword recommendation page, and the new keyword access depth, respectively. as well as Both are greater than 0.

[0038] In a preferred embodiment, after obtaining the keyword evaluation coefficient, the keyword evaluation coefficient is compared and analyzed with the continuously iterated screening threshold;

[0039] If the keyword evaluation coefficient is greater than or equal to the screening threshold, the new keyword is marked as a deep profile entity and a deep signal is generated;

[0040] If the keyword evaluation coefficient is less than the screening threshold, the new keyword is marked as a screened entity and a screening signal is generated.

[0041] A knowledge graph data classification and management system includes a collection module, a user portrait module, a fuzzy classification module, and an entity update module;

[0042] The collection module is used to collect behavioral feature data and time feature data, obtain long-term and short-term keyword search frequencies and long-term and short-term keyword similarities through data processing, and send them to the user portrait module. At the same time, the average deviation of the number of entity relationship layers clicked by users and the short-term user access bounce rate are sent to the fuzzy classification module;

[0043] The user portrait module is used to obtain the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities to establish the user's long-term interest model and the user's short-term behavior model, and obtain the user's deep and shallow portraits based on their comparison, and calculate the keyword overlap rate of each keyword set, and send it to the fuzzy classification module;

[0044] The fuzzy classification module is used to obtain the keyword overlap rate, the average deviation of the number of entity relationship layers clicked by users, and the short-term user access bounce rate, and use fuzzy logic to determine the adjustment results and send them to the entity update module;

[0045] The entity update module is used to obtain the adjustment results when the user's shallow portrait is saturated, and to establish a data analysis model based on the entity viscosity between the new keyword and the user's deep portrait, the browsing time on the new keyword recommendation page, and the new keyword access depth, to obtain the keyword evaluation coefficient, and compare it with the screening threshold to determine the new keyword division result.

[0046] Technical effects and advantages of the present invention:

[0047] 1. The present invention constructs a user long-term interest model and a short-term behavior model through the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities, obtains the user's deep portrait, the user's shallow portrait and the keyword overlap rate, and formulates a set of fuzzy rules for fuzzy reasoning with the average deviation of the number of levels of user-clicked entity relationships and the short-term user access bounce rate to determine the adjustment results, and classifies and re-sorts the strength of the relationship between entities based on the adjustment results, thereby avoiding the pollution of user portraits caused by users mistakenly entering keywords and improving the user experience.

[0048] 2. When the user's shallow profile is saturated, the present invention obtains a keyword evaluation coefficient through comprehensive analysis of the viscosity of new keywords and entities in the user's deep profile, the browsing time in the new keyword recommendation page, and the new keyword access depth, and compares it with the screening threshold to determine whether the new keyword is screened out or included in the user's deep profile, thereby avoiding poor operation caused by the oversaturation of the user's shallow profile. At the same time, the user's deep profile is updated, making the recommendation system based on the knowledge graph more personalized and smooth. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a method flow chart of a knowledge graph data classification and management method of the present invention.

[0050] Figure 2 This is a module schematic diagram of a knowledge graph data classification and management system of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] Example 1

[0053] Usually, the degree of connection between two entities is generally determined by the strength of the relationship and the similarity of the attributes, that is, the number of relationship layers between entities and the overlap rate of text information in the attributes. This is not user-friendly. However, the system usually uses long-term interest models to recommend to users to browse homepages. For content that users input by mistake, the long-term interest model cannot be used to recommend accurate content, and the user portrait is polluted. When the user browses next time, the content that the user inputs by mistake may be recommended first, thus affecting the user experience. Therefore, the following solutions are proposed:

[0054] Build a user's long-term interest model and short-term behavior model, compare the overlapping parts and mark them as the user's deep profile, and mark the non-overlapping parts as the user's shallow profile, and divide the knowledge graph according to the deep and shallow profiles;

[0055] See also Figure 1 , a knowledge graph data classification and management method, the specific operation process is as follows:

[0056] S1: Collect behavioral feature data and time feature data, and obtain long-term and short-term keyword search frequency, long-term and short-term keyword similarity, user click entity relationship layer deviation average value and short-term user visit bounce rate through data processing;

[0057] The definitions of long-term and short-term are set by the experimenters according to the specific implementation methods. Generally, the long-term is defined as the long-term set by the user's historical usage time, and the short-term is defined as the user's recent time, such as "24 hours", "one week" or "one month". The definition of long-term and short-term is set by the experimenters and will not be defined here.

[0058] The logic for obtaining the long-term and short-term keyword search frequencies is to give priority to expressing the long-term and short-term time period sets, that is, D long ={w1, w2, ..., w n} represents a long-term search record, where w iis the keyword searched for the i-th time; then D skort ={y1, y2, ..., y m} represents a short-term search record, where y j is the keyword searched for the jth time;

[0059] Define the function count(w, D) to represent the number of times the long and short keywords w and y appear in the data set D; then we get: C long (w) = count(w, D long ) and C short (y) = count(y, D short );

[0060] Calculate the frequency and define the function freq(w, D) to represent the frequency of long-term and short-term keywords w and y in the data set D, then we get:

[0061] and

[0062] in, Represents the total number of times all keywords appear in data set D;

[0063] The logic of obtaining the similarity of long-term and short-term keywords is to vectorize the keywords through the long-term and short-term keyword sets, and use the word embedding model to obtain the long-term keyword word vector expression V long ={vec(w1), vec(w2)..., vec(w n )}, each vec(w i ) is the keyword w i The vector representation of V short ={vec(y1), vec(y2)..., vec(y m )} where each vec(y1) is the keyword y m The cosine similarity method is used to calculate the similarity between each long-term and short-term keyword, based on the formula:

[0064]

[0065] Among them, V1 and V2 are vector representations of two keywords respectively;

[0066] The average deviation of the number of relationship layers of user-clicked entities is a measure of the difference in the relationship layers between the entities that users clicked in the short term during the usage period and the entities in the long-term interest model; this deviation reflects the consistency or deviation between the user's short-term behavior and long-term interests;

[0067] Set the number of relationship layers between entities in the knowledge graph. The number of relationship layers represents the path length between two entities. For example, the number of relationship layers between two directly related entities is 1, the number of relationship layers through an intermediary entity is 2, and so on;

[0068] Assume E long is the entity set in the user’s long-term interest model, E short A collection of entities that users click on in the short term, d9e i , e j ) represents entity e i and e j The number of relationship levels between them;

[0069] For each short-term click entity e short ∈E short , calculate its relationship with the long-term interest entity e long ∈E long The minimum number of relationship levels between:

[0070]

[0071] Then the average value of the relationship level deviation of all clicked entities is:

[0072]

[0073] The logic for obtaining the short-term user visit bounce rate is to define the bounce behavior, which is usually defined as the number of times a user directly leaves or closes a page without further interaction after visiting a page or entity. The short-term user visit bounce rate Br is obtained by statistically calculating the ratio of the number of bounces in all user visit records in a short period of time to the total number of visits.

[0074] S2: Obtain the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities to establish a user long-term interest model and a user short-term behavior model, and obtain a user deep and shallow profile based on the comparison, and calculate the keyword overlap rate for each keyword set;

[0075] Input the long-term and short-term keyword frequency matrix into the LDA model to extract the topic distribution. The input of the LDA model is data in the bag-of-words model format. Each document d can be represented as a word frequency vector:

[0076] D=[n(w1,d),n(w2,d),…,n(w n , d)]

[0077] Among them, n(w1, d) represents the keyword w i The number of times it appears in document d;

[0078] The core idea of ​​the LDA model is to represent documents as a distribution of topics, and each topic is represented as a distribution of words. The training of the LDA model is as follows:

[0079] Let K be the number of topics, β k is the word distribution of the kth topic, θ k is the topic distribution of document d, then the goal of the LDA model is to maximize the likelihood function of the document:

[0080]

[0081] Among them, α and β are the hyperparameters of topic distribution and word distribution respectively;

[0082] Extract the topic distribution of each keyword from the LDA model, and the topic distribution of the long-term interest model is: long ={P(w1|T1), P(w2|T2),…,P(w n |T k )}, then the topic distribution of the short-term behavior model is: T short ={P(y1|T1), P(y2|T2),…,P(y m |T k )};

[0083] Set a threshold Q, and for each keyword, compare its probability P in the long-term topic distribution long (w i |T) and the corresponding probability P under the short-term topic distribution short (y m |T) are all greater than Q;

[0084] If both are greater than Q, it is considered that the keyword has a high probability in both the long-term and short-term models, indicating that the keyword has a high match in the long-term and short-term models, and the keyword is classified into the user's deep profile. If any one of them is greater than Q, it can be considered that the keyword has a match in the long-term or short-term model, but it does not completely overlap, and the keyword is classified into the user's shallow profile.

[0085] It can be understood that n+m is the total number of all keywords that appear in the long term and the short term, but there is a possibility of duplication of keywords, so it is necessary to remove the duplicate keywords to obtain the total number of keywords;

[0086] According to the Jaccard similarity, the overlap rate of the keyword sets in the long-term model and the short-term model is calculated based on the formula:

[0087]

[0088] Among them, |A∩B| represents the number of keywords that appear together in the long-term and short-term models, and |A∪B| represents the number of all different keywords that appear in the long-term and short-term models;

[0089] S3: Obtain the keyword overlap rate, the average deviation of the number of levels of entity relationships clicked by users, and the short-term user visit bounce rate, use fuzzy logic to determine the entity ranking adjustment result, perform fuzzy reasoning based on fuzzy rules, and determine the entity change ranking;

[0090] For example, “high”, “low”, and “medium” for keyword overlap rate, average deviation of the number of entity relationship layers clicked by users, and bounce rate of short-term user visits;

[0091] Formulate a set of fuzzy rules to describe the impact of different input variables on output variables. The definition of rules can be based on professional knowledge or obtained through data analysis and experiments. For example:

[0092] The keyword overlap rate is marked as X, the average deviation of the number of levels of entity relationships clicked by users is marked as Y, the short-term user visit bounce rate is marked as Z, and the entity ranking adjustment result is marked as C_result;

[0093] Then we can define:

[0094] Rule 1: If (X is high) and (Y is high) and (Z is low) then (C_result changes)

[0095] Rule 2: If (X is low) and (Y is low) and (Z is high) then (C_result remains unchanged) ...

[0096] It should be noted that the division of fuzzy sets can be adjusted according to actual conditions. For example, although this embodiment takes three fuzzy sets as examples, in fact, the keyword overlap rate, the average deviation of the number of entity relationship levels clicked by users, and the short-term user access bounce rate can be divided into more than three sets to facilitate better precise adjustment according to different situations.

[0097] Furthermore, for the judgment of the high, medium and low values ​​of the keyword overlap rate, the average deviation of the number of user-click entity relationship levels and the short-term user access bounce rate, thresholds can be set according to actual conditions for judgment. For example, when the average deviation of the number of user-click entity relationship levels exceeds 75%, it is marked as "high"; when the keyword overlap rate is lower than 65%, it is marked as "low". I will not elaborate on this here.

[0098] List the initial rankings between entities from largest to smallest. The initial ranking is the ranking column obtained through the similarity of entity attributes and the close comprehensive analysis of relationships.

[0099] The number of entity relationship level deviations clicked by the user when C_result is "Change" is taken as the entity with the greatest entity association, and it is prioritized to be classified into the same category, and the entities associated with this keyword are re-sorted to the order at the top of the initial sorting;

[0100] If the result of C_ is "unchanged", it means that the current keyword cannot be counted in the shallow profile, and it may be clicked by the user by mistake. The recommendation order between entities does not need to be adjusted;

[0101] The present invention constructs a user's long-term interest model and a short-term behavior model through the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities, obtains the user's deep portrait, the user's shallow portrait and the keyword overlap rate, and formulates a set of fuzzy rules for fuzzy reasoning with the average value of the number of entity relationship layers clicked by the user and the short-term user access bounce rate to determine the adjustment result, and classifies and re-sorts the strength of the relationship between the entities based on the adjustment result, avoiding the pollution of the user portrait caused by the user's accidental input of keywords, thereby improving the user experience.

[0102] Example 2

[0103] In the first embodiment of the present invention, an example is given to construct a user long-term interest model and a short-term behavior model through the long-term and short-term keyword search frequency and the long-term and short-term keyword similarity, obtain a user deep profile, a user shallow profile and a keyword overlap rate, bring the newly input keywords of the user into the fuzzy rules for fuzzy reasoning, and determine the operation strategy for readjusting the relationship between shallow entities; however, in the first embodiment, it is only considered whether the new keywords can be put into the user shallow profile, and whether the new keywords can be put into the user deep profile or cleared in time is not considered. Obviously, although putting new keywords into the user shallow profile can avoid polluting the user deep profile, when too many new keywords are accumulated, it may cause the system to run too slowly and the user cannot browse in time; in view of the above problems, the second embodiment of the present invention is further refined;

[0104] Monitor the memory saturation of the user's shallow profile. If saturation occurs, evaluate the new keywords in the user's shallow profile.

[0105] S4: When the user's shallow profile is saturated, the new keyword is obtained to establish a data analysis model by obtaining the entity viscosity between the new keyword and the user's deep profile, the browsing time on the new keyword recommendation page, and the new keyword access depth to obtain the keyword evaluation coefficient;

[0106] Among them, the viscosity between new keywords and entities in the user's deep profile refers to the strength of association between new keywords and all entities in the user's deep profile. By setting S deep is the entity set in the user's deep profile, k is the new keyword, E(k) is the entity set related to the new keyword k, then r(ek , e d ) is entity e k and the entity e in the user's deep profile d The number of relationship layers between k , e d ) is entity e k and the entity e in the user's deep profile d The frequency between

[0107] For each entity e k In E(k), find its deep All entities in d The minimum relationship level between min (e k ) and the corresponding relationship strength w max (e k ):

[0108]

[0109] The ratio of the number of relationship layers and the relationship strength is used to obtain the entity viscosity V between the new keyword and the user's deep portrait. k ; Where k = 1, 2, 3, 4, ... P, P is the total number of new keywords;

[0110] The logic for obtaining the browsing time on the new keyword recommendation page is to record the timestamp of the user visiting the recommendation page and the timestamp of the user leaving the recommendation page, and calculate the difference between the page leaving time and the page loading time to get the browsing time Ti on the new keyword recommendation page. k ;

[0111] The logic of obtaining the new keyword access depth is to record the number of pages that users enter by clicking on the initial page of the recommended page. That is, the access depth increases by 1 every time the user clicks, and the new keyword access depth De is obtained by adding the initial page depth and the number of clicks. k ;

[0112] Generate the keyword evaluation coefficient Si based on the entity viscosity between the new keyword and the user's deep portrait, the browsing time on the new keyword recommendation page, and the new keyword access depth. k ; The formula is as follows:

[0113]

[0114] In the formula, Sk is the keyword evaluation coefficient, as well as are the calculation weight coefficients of the entity viscosity between keywords and user deep portraits, the browsing time in the new keyword recommendation page, and the new keyword access depth, respectively. as well as Both are greater than 0.

[0115] Among them, the entity viscosity between new keywords and the user's deep profile, the browsing time on the new keyword recommendation page, and the new keyword access depth are all data-based manifestations that directly express whether the new keyword is included in the user's deep profile;

[0116] From the formula, we can see that if the viscosity between new keywords and entities in the user's deep portrait, the browsing time in the new keyword recommendation page, and the new keyword access depth are greater, the greater the keyword evaluation coefficient is, which means that the new keyword is more closely related to the entity in the user's deep portrait, the greater the user's interest, and the more it needs to be put into the user's deep portrait in time. On the contrary, the smaller the keyword evaluation coefficient is, it means that the new keyword needs to be screened out of the user's shallow portrait in time;

[0117] S5: Obtain keyword evaluation coefficients, and compare and analyze them with the screening threshold to obtain comparison results, and determine new keyword division results based on the comparison results;

[0118] Specifically, set the screening threshold evaluation, which is obtained by continuously iterating the historical new keyword screening results;

[0119] After obtaining the keyword evaluation coefficient, compare and analyze the keyword evaluation coefficient with the continuously iterated screening threshold;

[0120] If the keyword evaluation coefficient is greater than or equal to the screening threshold, the new keyword is marked as a deep profile entity and a deep signal is generated;

[0121] If the keyword evaluation coefficient is less than the screening threshold, the new keyword is marked as a screened entity and a screening signal is generated;

[0122] When the user's shallow profile is saturated, the present invention obtains a keyword evaluation coefficient through comprehensive analysis of the viscosity of new keywords and entities in the user's deep profile, the browsing time in the new keyword recommendation page, and the new keyword access depth, and compares it with the screening threshold to determine whether the new keyword is screened out or included in the user's deep profile, thereby avoiding poor operation caused by the oversaturation of the user's shallow profile. At the same time, the user's deep profile is updated, making the recommendation system based on the knowledge graph more personalized and smooth.

[0123] Example 3

[0124] See also Figure 2 , a knowledge graph data classification and management system, including a collection module, a user portrait module, a fuzzy classification module, and an entity update module;

[0125] The collection module is used to collect behavioral feature data and time feature data, obtain long-term and short-term keyword search frequencies and long-term and short-term keyword similarities through data processing, and send them to the user portrait module. At the same time, the average deviation of the number of entity relationship layers clicked by users and the short-term user access bounce rate are sent to the fuzzy classification module;

[0126] The user portrait module is used to obtain the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities to establish the user's long-term interest model and the user's short-term behavior model, and obtain the user's deep and shallow portraits based on their comparison, and calculate the keyword overlap rate of each keyword set, and send it to the fuzzy classification module;

[0127] The fuzzy classification module is used to obtain the keyword overlap rate, the average deviation of the number of entity relationship layers clicked by users, and the short-term user access bounce rate, and use fuzzy logic to determine the adjustment results and send them to the entity update module;

[0128] The entity update module is used to obtain the adjustment results when the user's shallow portrait is saturated, and to establish a data analysis model based on the entity viscosity between the new keyword and the user's deep portrait, the browsing time on the new keyword recommendation page, and the new keyword access depth, to obtain the keyword evaluation coefficient, and compare it with the screening threshold to determine the new keyword division result.

[0129] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0130] The above embodiments can be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented by software, the above embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium, or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website site, computer, server or data center to another website site, computer, server or data center by wired or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state hard disk.

[0131] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0132] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0133] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0134] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0135] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0136] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0137] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memories (ROM), random access memories (RAM), magnetic disks or optical disks.

[0138] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A knowledge graph data classification and management method, characterized by: include S1: Collect behavioral feature data and time feature data, and obtain long-term and short-term keyword search frequency, long-term and short-term keyword similarity, user click entity relationship layer deviation average value and short-term user visit bounce rate through data processing; S2: Obtain the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities to establish a user long-term interest model and a user short-term behavior model, and obtain a user deep and shallow profile based on the comparison, and calculate the keyword overlap rate for each keyword set; S3: Obtain the keyword overlap rate, the average deviation of the number of levels of entity relationships clicked by users, and the short-term user visit bounce rate, use fuzzy logic to determine the entity ranking adjustment result, perform fuzzy reasoning based on fuzzy rules, and determine the entity change ranking; S4: When the user's shallow profile is saturated, the new keyword is obtained to establish a data analysis model by obtaining the entity viscosity between the new keyword and the user's deep profile, the browsing time on the new keyword recommendation page, and the new keyword access depth to obtain the keyword evaluation coefficient; S5: Obtain keyword evaluation coefficients, and compare and analyze them with the screening threshold to obtain comparison results, and determine new keyword division results based on the comparison results; By setting the number of relationship layers between entities in the knowledge graph, the number of relationship layers represents the path length between two entities. For each short-term click entity, the minimum number of relationship layers between it and the long-term interest entity is calculated, and the average deviation of the number of relationship layers of user-clicked entities is obtained.

2. A knowledge graph data classification and management method according to claim 1, characterized in that: By expressing the long-term and short-term time period set, long-term search records and short-term search records are obtained, and the frequency of the long-term and short-term keywords in the total data set is calculated to obtain the long-term and short-term keyword search frequency; Through the long-term and short-term keyword sets, the keywords are vectorized, the word embedding model is used to obtain the word vector expression of the long-term keywords, and the cosine similarity method is selected to calculate the similarity of the long-term and short-term keywords; After defining the bounce behavior, the short-term user visit bounce rate is obtained by statistically calculating the ratio of the number of bounces in all user access records in a short period of time to the total number of visits. .

3. A knowledge graph data classification and management method according to claim 2, characterized in that: Input the long-term and short-term keyword frequency matrix into the LDA model to extract the topic distribution. The input of the LDA model is data in the bag-of-words model format. Each document d can be represented as a word frequency vector: ； in, Indicates keywords The number of times it appears in document d; The logic for obtaining the long-term and short-term keyword search frequencies is to give priority to expressing the long-term and short-term time period sets, that is, Represents a long-term search record, where is the keyword searched for the i-th time; then Represents short-term search records, where is the keyword searched for the jth time; The core idea of ​​the LDA model is to represent documents as a distribution of topics, and each topic is represented as a distribution of words. The training of the LDA model is as follows: Let K be the number of topics, is the word distribution of the kth topic, is the topic distribution of document d, then the goal of the LDA model is to maximize the likelihood function of the document: ； in, and are the hyperparameters of topic distribution and word distribution respectively; Extract the topic distribution of each keyword from the LDA model, and the topic distribution of the long-term interest model is: , then the topic distribution of the short-term behavior model is: .

4. A knowledge graph data classification and management method according to claim 3, characterized in that: Set a threshold Q, and for each keyword, compare its probability in the long-term topic distribution And the corresponding probability under the short-term topic distribution Are they all greater than Q? If both are greater than Q, it is considered that the keyword has a high probability in both the long-term and short-term models, indicating that the keyword has a high match in the long-term and short-term models, and the keyword is classified into the user's deep profile. If any one of them is greater than Q, it can be considered that the keyword has a match in the long-term or short-term model, but it does not completely overlap, and the keyword is classified into the user's shallow profile. According to Jaccard similarity, calculate the overlap rate of keyword sets in long-term model and short-term model , based on the formula: ； in, To represent the number of keywords that co-occur in the long-term and short-term models, Represents the number of all distinct keywords that appear in the long-term and short-term models.

5. A knowledge graph data classification and management method according to claim 1, characterized in that: The keyword overlap rate, the average deviation of the number of entity relationship layers clicked by users, and the short-term user visit bounce rate are defined as input variables, and they are divided into different fuzzy sets respectively; The result of entity ranking adjustment is defined as output variable and divided into fuzzy sets; Formulate fuzzy rules to describe the impact of keyword overlap rate, average deviation of the number of entity relationship layers clicked by users, and short-term user visit bounce rate on entity ranking adjustment results; Fuzzy reasoning is performed based on fuzzy rules to determine the entity change order.

6. A knowledge graph data classification and management method according to claim 5, characterized in that: Monitor the memory saturation of the user's shallow profile. If saturation occurs, evaluate the new keywords in the user's shallow profile.

7. A knowledge graph data classification and management method according to claim 1, characterized in that: The viscosity of new keywords and entities in the user's deep portrait is obtained through the relationship number and relationship strength ratio ; Where k = 1, 2, 3, 4, ... P, P is the total number of new keywords; The browsing time on the new keyword recommendation page is calculated by the difference between the page leaving time and the page loading time. ; The new keyword access depth is obtained by adding the initial page depth and the number of clicks. .

8. A knowledge graph data classification and management method according to claim 7, characterized in that: Generate a keyword evaluation coefficient based on the entity viscosity between the new keyword and the user's deep portrait, the browsing time on the new keyword recommendation page, and the new keyword access depth. ; The formula is as follows: ； In the formula, is the keyword evaluation coefficient, , as well as are the calculation weight coefficients of the entity viscosity between keywords and user deep portraits, the browsing time in the new keyword recommendation page, and the new keyword access depth, respectively. , as well as Both are greater than 0.

9. A knowledge graph data classification and management method according to claim 8, characterized in that: After obtaining the keyword evaluation coefficient, compare and analyze the keyword evaluation coefficient with the continuously iterated screening threshold; If the keyword evaluation coefficient is greater than or equal to the screening threshold, the new keyword is marked as a deep profile entity and a deep signal is generated; If the keyword evaluation coefficient is less than the screening threshold, the new keyword is marked as a screened entity and a screening signal is generated.

10. A knowledge graph data classification and management system, used to implement a knowledge graph data classification and management method according to any one of claims 1 to 9, characterized in that: Including collection module, user portrait module, fuzzy classification module, entity update module; The collection module is used to collect behavioral feature data and time feature data, obtain long-term and short-term keyword search frequencies and long-term and short-term keyword similarities through data processing, and send them to the user portrait module. At the same time, the average deviation of the number of entity relationship layers clicked by users and the short-term user access bounce rate are sent to the fuzzy classification module; The user portrait module is used to obtain the long-term and short-term keyword search frequencies and the long-term and short-term keyword similarities to establish the user's long-term interest model and the user's short-term behavior model, and obtain the user's deep and shallow portraits based on their comparison, and calculate the keyword overlap rate of each keyword set, and send it to the fuzzy classification module; The fuzzy classification module is used to obtain the keyword overlap rate, the average deviation of the number of entity relationship layers clicked by users, and the short-term user access bounce rate, and use fuzzy logic to determine the adjustment results and send them to the entity update module; The entity update module is used to obtain the adjustment results when the user's shallow portrait is saturated, and to establish a data analysis model based on the entity viscosity between the new keyword and the user's deep portrait, the browsing time on the new keyword recommendation page, and the new keyword access depth, to obtain the keyword evaluation coefficient, and compare it with the screening threshold to determine the new keyword division result.

Images