A standard think tank knowledge recommendation method and system based on big data

By filtering data through semantic analysis and time decay mechanisms, and combining it with historical user data to build a recommendation model, the problem of insufficient timeliness and accuracy of recommended content in existing systems is solved, and dynamic updates and efficient recommendations are achieved.

CN118193845BActive Publication Date: 2024-11-05CHINA NAT INST OF STANDARDIZATION
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202410431659.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-11
Publication Date
2024-11-05
Estimated Expiration
2044-04-11

AI Technical Summary

Technical Problem

Existing recommendation systems struggle to effectively combine user historical data with standard think tank data, resulting in insufficient timeliness and accuracy of recommended content.

Method used

By using semantic analysis and data type classification, combined with a time decay mechanism to filter data, a recommendation model is built. By utilizing users' historical click data and rating information, the recommended content is dynamically updated, optimizing the mapping process from state to action.

Benefits of technology

It improves the accuracy and coverage of recommendations, ensures the timeliness of recommended content, and has a strong dynamic update capability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118193845B_ABST
    Figure CN118193845B_ABST
Patent Text Reader

Abstract

The present invention discloses a standard think tank knowledge recommendation method and system based on big data, including obtaining standard think tank data and historical valid data of users, the historical valid data including historical click data of users and historical ratings of standard think tank data; classifying standard think tank data according to semantic analysis and data type to obtain first classification results and second classification results, filtering the first classification results based on time decay to obtain incremental data; matching the first classification results according to historical valid data to obtain first data, matching the second classification results according to historical valid data to obtain second data, and obtaining recommended categories based on the first data and the second data; establishing a recommendation model through historical valid data, incremental data and recommended categories, inputting the data to be recommended into the recommendation model and outputting the recommendation results. The present invention can more accurately understand user needs and provide more personalized knowledge recommendation services through big data technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of big data analysis and recommendation systems, and in particular to a standard think tank knowledge recommendation method and system based on big data. Background Art

[0002] With the rapid development of the Internet, the amount of data has surged, and big data analysis technology has emerged. It can process and analyze massive amounts of data and mine valuable information. As a form of information filtering, the recommendation system provides personalized content to users by analyzing user behavior and preferences. It is widely used in e-commerce, social networking and other platforms. Natural language processing technology plays a key role in this process. It helps the system better understand and organize standard think tank data through means such as semantic analysis and text classification. In addition, machine learning algorithms are the basis for building recommendation models. By learning from historical valid data, it predicts content that users may be interested in, thereby improving the relevance and accuracy of recommendations. The combination of these technologies provides strong support for building efficient and intelligent knowledge recommendation systems. Summary of the invention

[0003] The purpose of the present invention is to provide a standard think tank knowledge recommendation method and system based on big data.

[0004] To achieve the above object, the present invention is implemented according to the following technical solutions:

[0005] The present invention comprises the following steps:

[0006] A obtains standard think tank data and historical valid data of the user, wherein the historical valid data includes the historical click data of the user and the historical rating of the standard think tank data;

[0007] B. classifying the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result, and performing data screening on the first classification result based on time decay to obtain incremental data;

[0008] C. matching the first classification result according to the historical valid data to obtain first data, matching the second classification result according to the historical valid data to obtain second data, and obtaining a recommended category based on the first data and the second data;

[0009] D establishes a recommendation model through the historical valid data, the incremental data and the recommended categories, inputs the data to be recommended into the recommendation model and outputs the recommendation result.

[0010] Furthermore, the method of classifying the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result includes:

[0011] Use TF-IDF to convert the standard think tank data into a numerical vector, normalize the numerical vector as a data point, use the data point as the initial cluster, and use cosine similarity as the spacing of the data points:

[0012]

[0013] Among them C S (h,b) is the cosine similarity of data points h and b, h and b are two n-dimensional vectors, h i represents the component of vector h in the i-th dimension. The value range of the cosine similarity is [-1, 1], where -1 means completely dissimilar and 1 means completely similar. The two clusters with the smallest distance between the data points in the initial cluster are merged to construct a new cluster, and the inter-cluster distance of the cluster is calculated:

[0014]

[0015] Where D U (A m ,A n ) is A m and A n The inter-cluster spacing, A m and A n are different clusters, x and y are different data points, |A m | and |A n | for A m and A n The number of the data points included, merging the clusters corresponding to the shortest inter-cluster distance, repeatedly calculating the inter-cluster distance and merging until the number of the remaining clusters is the same as the preset classification number l, one cluster corresponds to one classification category, and the first classification result is obtained;

[0016] According to the characteristics of the data type, the feature vectors of the standard think tank data are extracted, the feature vectors are normalized, and the distance between the feature vectors is calculated using the Euclidean distance:

[0017]

[0018] Where d(p,q) is the distance between the feature vectors p and q, p k represents the vector value of the feature vector p of the kth dimension, s k is the standard deviation of the kth dimension, and the feature vector is also used as the initial cluster, and the clusters corresponding to the shortest inter-cluster distance are repeatedly merged until the number of the remaining clusters is the same as the preset classification number η, thereby obtaining the second classification result.

[0019] Furthermore, the method of performing data screening on the first classification result based on time decay to obtain incremental data includes:

[0020] The standard think tank data is sorted according to time to obtain time series data, and the time weight under the first classification result is calculated:

[0021]

[0022] where w i is the time weight of the ith data point, B(α,β) is the beta function, Γ is the gamma function, α and β are shape parameters that determine the rate and shape of the time weight changing over time, t i is the time point of the i-th data point, t0 is the reference time point, is the latest time point of the category to which the data point belongs, τ is the time scale parameter, and the incremental data is obtained by removing the time series data whose time weight is less than the preset threshold. The preset threshold is determined according to the update frequency of different categories and is the time weight of the update frequency.

[0023] Furthermore, the method of obtaining the first data by matching the first classification result according to the historical valid data includes:

[0024] Calculate the similarity of the user according to the historical rating and the first classification result:

[0025]

[0026] Where sim(u,v) represents the similarity between user u and user v, P uv represents the set of categories rated by user u and user v, R ua represents the rating of user u on category a, represents the average rating of user u for all categories, t ua is the time when user u scores category a, λ is the time decay factor, exp represents an exponential function, a similarity matrix is ​​constructed according to the similarity, the users whose similarity in the similarity matrix is ​​greater than a preset similarity threshold are taken as collaborative users, and the value range of the similarity threshold is [0.4, 0.9], and the click rates of the historical click data of the user and the collaborative user for the first classification result are weighted and fused to obtain the first data:

[0027]

[0028] Where P ui is the first data, indicating the predicted preference of user u for category i, i represents the category in the first classification result, w is the weight coefficient, p ui represents the click rate of user u on category i, Represents the click rate of the collaborative users of user u on category i.

[0029] Furthermore, the method of obtaining the second data by matching the second classification result according to the historical valid data includes:

[0030] The click rate and average score of the second classification result are calculated based on the historical valid data of the user and normalized preprocessed, and the weighted average of the click rate and average score of the second classification result is used as the second data:

[0031]

[0032] Where P uj is the second data, indicating the predicted preference of user u for category j, j indicates the category in the second classification result, w1 is the weight of the click rate of the second classification result, w2 is the weight of the average score, p uj represents the click rate of user u on category j, A uj is the average rating of user u for category j.

[0033] Furthermore, a method for obtaining a recommended category based on the first data and the second data includes:

[0034] A data set is established using the first data and the second data, each data vector in the data set is composed of the first data and the second data, and two distribution models are trained using the data set, representing the recommended category and other categories respectively, and the probability density function of the distribution model is:

[0035]

[0036] Among them, γ1 and γ2 are mixing coefficients, represents a normal distribution, μ1 and μ2 are the mean vectors of the recommended category and the other categories, Ω1 and Ω2 are the covariance matrices of the recommended category and the other categories, and the data vector is classified into the category with the largest posterior probability. The posterior probability of the distribution model is:

[0037]

[0038] Where p(c=k|x) is the posterior probability, which indicates the probability that the data vector x belongs to category k, c indicates the category, k indicates the index of the category (k=1 or 2), γ k To update the mixing coefficient, μ k is the updated mean, Ω k To update the covariance matrix, the calculation formulas for the updated mixing coefficient, the updated mean and the updated covariance matrix are:

[0039]

[0040] Wherein N represents the number of the data vectors in the data set, and T represents the transpose of the matrix.

[0041] Furthermore, the method of establishing a recommendation model using the historical valid data, the incremental data and the recommendation categories includes:

[0042] A state space is constructed according to the incremental data and the recommended categories, and the recommended data is used as an action space. The action space is mapped from the state space, and the actual feedback of the historical valid data corresponding to the recommended data is used as a reward signal to establish the recommendation model. When the recommendation model is output, the historical valid data in the recommended data is removed, and the recommendation model is updated according to the reward signal to optimize the mapping from state to action.

[0043] Update of value function:

[0044]

[0045] Where V α (S) is the value function of state S, μ(A|S) is the strategy for selecting action A in state S, R(S,A) is the immediate reward obtained when taking action A, α is the discount factor of the value function, which is used to adjust the current value of future rewards, and P SS′ (A) is the probability of executing action A from state S to state S′, and the update of Q function:

[0046]

[0047] Where Q β (S, A; φ) is the Q function of state S and action A, is the expected estimate of taking action A in state S, β is the discount factor of the Q function, φ is the parameter of the main network used to parameterize the Q function, E is the expected value, which represents the average value under the probability distribution, and the policy gradient formula for updating φ is:

[0048]

[0049] Where J(φ) is the objective function, is the expected return, B is the currently selected action, γ represents the discount factor of the policy gradient, and π φ (B|S) is the parameterized strategy, the target Q function:

[0050]

[0051] Where Q γ(S, B; θ) is the Q function approximated by the neural network, which is a stable estimate used to calculate the future reward part of the expected return, C is the best action in the next state, θ- represents the parameters of the target network, which is regularly updated from the main network parameters θ,

[0052] The optimization is continuously iterated until the strategy converges, and the recommendation model uses the preference category filled in by the user when the user logs in for the first time as the recommendation category.

[0053] A standard think tank knowledge recommendation system based on big data, characterized by including:

[0054] A collection module is used to obtain standard think tank data and historical valid data of users, wherein the historical valid data includes the historical click data of the users and the historical ratings of the standard think tank data;

[0055] A classification module, which classifies the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result, and obtains a recommended category based on the first data and the second data;

[0056] a calculation module, which performs data screening on the first classification result based on time decay to obtain incremental data, matches the first classification result according to the historical valid data to obtain first data, matches the second classification result according to the historical valid data to obtain second data, and establishes a recommendation model through the historical valid data, the incremental data and the recommended category;

[0057] The output module inputs the data to be recommended into the recommendation model and outputs the recommendation result.

[0058] The beneficial effects of the present invention are:

[0059] Through the time decay mechanism, the system of the present invention can dynamically update the recommended content to ensure the timeliness of the recommendation, and combine the multi-dimensional analysis of the user's historical valid data and standard think tank data to improve the accuracy and coverage of the recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 This is a flowchart of the steps of a standard think tank knowledge recommendation method and system based on big data in the present invention. DETAILED DESCRIPTION

[0061] The present invention is further described below by means of specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention but are not intended to limit the present invention.

[0062] The present invention provides a method and system for recommending standard think tank knowledge based on big data, comprising the following steps:

[0063] like Figure 1 As shown, in this embodiment, the following steps are included:

[0064] A obtains standard think tank data and historical valid data of the user, wherein the historical valid data includes the historical click data of the user and the historical rating of the standard think tank data;

[0065] B. classifying the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result, and performing data screening on the first classification result based on time decay to obtain incremental data;

[0066] C. matching the first classification result according to the historical valid data to obtain first data, matching the second classification result according to the historical valid data to obtain second data, and obtaining a recommended category based on the first data and the second data;

[0067] D establishes a recommendation model through the historical valid data, the incremental data and the recommended categories, inputs the data to be recommended into the recommendation model and outputs the recommendation result.

[0068] In this embodiment, the historical valid data are click data and rating data with a browsing time of more than 2 minutes, the value range of the historical rating is [1, 5], and the historical valid data obtained are: user U1: "The rise of smart cities" 4.5 points-2023-11-10, "Sustainable development of future cities" 5 points-2023-12-15; user U2: "Application of artificial intelligence in the medical field" 4 points-2023-10-05, "Sustainable development of future cities" 4.5 points-2023-12-15; the standard think tank data obtained are: "Application of artificial intelligence in the medical field"-academic research, text report-2023-10-05, "The rise of smart cities"-policy analysis, data set-2023-11-10, "Global City Quality of Life Index"-statistical data, data set-2023-11-20, "Sustainable development of future cities"-industry application, text report-2023-12-15.

[0069] In this embodiment, the method of classifying the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result includes:

[0070] Use TF-IDF to convert the standard think tank data into a numerical vector, normalize the numerical vector as a data point, use the data point as the initial cluster, and use cosine similarity as the spacing of the data points:

[0071]

[0072] Among them CS (h,b) is the cosine similarity of data points h and b, h and b are two n-dimensional vectors, h i represents the component of vector h in the i-th dimension. The value range of the cosine similarity is [-1, 1], where -1 means completely dissimilar and 1 means completely similar. The two clusters with the smallest distance between the data points in the initial cluster are merged to construct a new cluster, and the inter-cluster distance of the cluster is calculated:

[0073]

[0074] Where D U (A m ,A n ) is A m and A n The inter-cluster spacing, A m and A n are different clusters, x and y are different data points, |A m | and |A n | for A m and A n The number of the data points included, merging the clusters corresponding to the shortest inter-cluster distance, repeatedly calculating the inter-cluster distance and merging until the number of the remaining clusters is the same as the preset classification number l, l is 2, one cluster corresponds to one classification category, and the first classification result is obtained;

[0075] According to the characteristics of the data type, the feature vectors of the standard think tank data are extracted, the feature vectors are normalized, and the distance between the feature vectors is calculated using the Euclidean distance:

[0076]

[0077] Where d(p,q) is the distance between the feature vectors p and q, p k represents the vector value of the feature vector p of the kth dimension, s k is the standard deviation of the kth dimension, and the feature vector is also used as the initial cluster, and the clusters corresponding to the shortest inter-cluster distance are repeatedly merged until the number of the remaining clusters is the same as the preset classification number η, η is 2, and the second classification result is obtained.

[0078] In this embodiment, the first classification results are: Medical and Artificial Intelligence: "Application of Artificial Intelligence in the Medical Field", Urban Development: "The Rise of Smart Cities", "Sustainable Development of Future Cities", and "Global City Quality of Life Index"; the second classification results are: Text Report: "Application of Artificial Intelligence in the Medical Field" and "Sustainable Development of Future Cities", Data Sets: "The Rise of Smart Cities" and "Global City Quality of Life Index".

[0079] In this embodiment, the method of performing data screening on the first classification result based on time decay to obtain incremental data includes:

[0080] The standard think tank data is sorted according to time to obtain time series data, and the time weight under the first classification result is calculated:

[0081]

[0082] where w i is the time weight of the ith data point, B(α,β) is the beta function, Γ is the gamma function, α and β are shape parameters that determine the rate and shape of the time weight changing over time, t i is the time point of the i-th data point, t0 is the reference time point, is the latest time point of the category to which the data point belongs, τ is the time scale parameter, and the incremental data is obtained by removing the time series data whose time weight is less than the preset threshold. The preset threshold is determined according to the update frequency of different categories and is the time weight of the update frequency.

[0083] In this embodiment, the update frequency of the medical and artificial intelligence categories is monthly, and the update frequency of the urban development category is quarterly. The incremental data are: Medical and artificial intelligence: "Application of artificial intelligence in the medical field"-2023-10-05; Urban development: "The rise of smart cities"-2023-11-10, "Global City Quality of Life Index"-2023-11-20, "Sustainable development of future cities"-2023-12-15.

[0084] In this embodiment, the method of obtaining the first data by matching the first classification result according to the historical valid data includes:

[0085] Calculate the similarity of the user according to the historical rating and the first classification result:

[0086]

[0087] Where sim(u,v) represents the similarity between user u and user v, P uv represents the set of categories rated by user u and user v, R ua represents the rating of user u on category a, represents the average rating of user u for all categories, t uais the time when user u scores category a, λ is the time decay factor, exp represents an exponential function, a similarity matrix is ​​constructed according to the similarity, and the users whose similarity in the similarity matrix is ​​greater than a preset similarity threshold are taken as collaborative users. The value range of the similarity threshold is [0.4, 0.9], and 0.5 is taken. The click rates of the historical click data of the user and the collaborative user for the first classification result are weighted and fused to obtain the first data:

[0088]

[0089] Where P ui is the first data, indicating the predicted preference of user u for category i, i represents the category in the first classification result, w is the weight coefficient, taking w=0.6, p ui represents the click rate of user u on category i, Represents the click rate of the collaborative users of user u on category i.

[0090] In this embodiment, the method of obtaining the second data by matching the second classification result according to the historical valid data includes:

[0091] The click rate and average score of the second classification result are calculated based on the historical valid data of the user and normalized preprocessed, and the weighted average of the click rate and average score of the second classification result is used as the second data:

[0092]

[0093] Where P uj is the second data, indicating the predicted preference of user u for category j, j indicates the category in the second classification result, w1 is the weight of the click rate of the second classification result, w1=0.6, w2 is the weight of the average score, w2=0.4, p uj represents the click rate of user u for category j, A uj is the average rating of user u for category j.

[0094] In this embodiment, the method for obtaining a recommended category based on the first data and the second data includes:

[0095] A data set is established using the first data and the second data, each data vector in the data set is composed of the first data and the second data, and two distribution models are trained using the data set, representing the recommended category and other categories respectively, and the probability density function of the distribution model is:

[0096]

[0097] Among them, γ1 and γ2 are mixing coefficients, represents a normal distribution, μ1 and μ2 are the mean vectors of the recommended category and the other categories, Ω1 and Ω2 are the covariance matrices of the recommended category and the other categories, and the data vector is classified into the category with the largest posterior probability. The posterior probability of the distribution model is:

[0098]

[0099] Where p(c=k|x) is the posterior probability, which indicates the probability that the data vector x belongs to category k, c indicates the category, k indicates the index of the category (k=1 or 2), γ k To update the mixing coefficient, μ k is the updated mean, Ω k To update the covariance matrix, the calculation formulas for the updated mixing coefficient, the updated mean and the updated covariance matrix are:

[0100]

[0101] Wherein N represents the number of the data vectors in the data set, and T represents the transpose of the matrix.

[0102] In this embodiment, the first data: medical and artificial intelligence: user 1, 0.2, user 2, 0.3, urban development: user 1, 0.8, user 2, 0.7; the second data: text report: user 1, 0.66, user 2, 0.62, data set: user 1, 0.7, user 2, 0.66; the recommended categories are: user 1, urban development text report and urban development data set, user 2, urban development text report and urban development data set.

[0103] In this embodiment, the method of establishing a recommendation model using the historical valid data, the incremental data, and the recommended categories includes:

[0104] A state space is constructed according to the incremental data and the recommended categories, and the recommended data is used as an action space. The action space is mapped from the state space, and the actual feedback of the historical valid data corresponding to the recommended data is used as a reward signal to establish the recommendation model. When the recommendation model is output, the historical valid data in the recommended data is removed, and the recommendation model is updated according to the reward signal to optimize the mapping from state to action.

[0105] Update of value function:

[0106]

[0107] Where V α(S) is the value function of state S, μ(A|S) is the strategy for selecting action A in state S, R(S,A) is the immediate reward obtained when taking action A, α is the discount factor of the value function, which is used to adjust the current value of future rewards, and P SS′ (A) is the probability of executing action A from state S to state S′, and the update of Q function:

[0108]

[0109] Where Q β (S, A; φ) is the Q function of state S and action A, is the expected estimate of taking action A in state S, β is the discount factor of the Q function, φ is the parameter of the main network used to parameterize the Q function, E is the expected value, which represents the average value under the probability distribution, and the policy gradient formula for updating φ is:

[0110]

[0111] Where J(φ) is the objective function, is the expected return, B is the currently selected action, γ represents the discount factor of the policy gradient, and π φ (B|S) is the parameterized strategy, the target Q function:

[0112]

[0113] Where Q γ (S, B; θ) is the Q function approximated by the neural network, which is a stable estimate used to calculate the future reward part of the expected return, C is the best action in the next state, and θ - represents the parameters of the target network, which are regularly updated from the main network parameters θ.

[0114] The optimization is continuously iterated until the strategy converges, and the recommendation model uses the preference category filled in by the user when the user logs in for the first time as the recommendation category.

[0115] In this embodiment, the recommended data are: User 1: "Global City Life Quality Index", User 2: "Global City Life Quality Index" and "Sustainable Development of Future Cities".

[0116] A standard think tank knowledge recommendation system based on big data, characterized by including:

[0117] A collection module is used to obtain standard think tank data and historical valid data of users, wherein the historical valid data includes the historical click data of the users and the historical ratings of the standard think tank data;

[0118] A classification module, which classifies the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result, and obtains a recommended category based on the first data and the second data;

[0119] a calculation module, which performs data screening on the first classification result based on time decay to obtain incremental data, matches the first classification result according to the historical valid data to obtain first data, matches the second classification result according to the historical valid data to obtain second data, and establishes a recommendation model through the historical valid data, the incremental data and the recommended category;

[0120] The output module inputs the data to be recommended into the recommendation model and outputs the recommendation result.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A standard think tank knowledge recommendation method based on big data, It is characterized in that The following steps are involved: A obtains standard think tank data and historical valid data of the user, wherein the historical valid data includes the historical click data of the user and the historical rating of the standard think tank data; B classifies the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result, Use TF-IDF to convert the standard think tank data into a numerical vector, normalize the numerical vector as a data point, use the data point as the initial cluster, and use cosine similarity as the spacing of the data points: Among them C S (h, b) is the cosine similarity of data points h and b, where h and b are two n-dimensional vectors, h i represents the component of vector h in the i-th dimension. The value range of the cosine similarity is [-1, 1], where -1 means completely dissimilar and 1 means completely similar. The two clusters with the smallest distance between the data points in the initial cluster are merged to construct a new cluster, and the inter-cluster distance of the cluster is calculated: Where D U (A m ,A n ) is A m and A n The inter-cluster spacing, A m and A n are different clusters, x and y are different data points, |A m | and |A n | for A m and A n The number of the data points included, merging the clusters corresponding to the shortest inter-cluster distance, repeatedly calculating the inter-cluster distance and merging until the number of the remaining clusters is the same as the preset classification number l, one cluster corresponds to one classification category, and the first classification result is obtained; According to the characteristics of the data type, the feature vectors of the standard think tank data are extracted, the feature vectors are normalized, and the distance between the feature vectors is calculated using the Euclidean distance: Where d(p,q) is the distance between the feature vectors p and q, p k represents the vector value of the feature vector p of the kth dimension, s k is the standard deviation of the kth dimension, and the feature vector is also used as the initial cluster, and the clusters corresponding to the shortest inter-cluster distance are repeatedly merged until the number of the remaining clusters is the same as the preset classification number η, to obtain the second classification result, Performing data screening on the first classification result based on time decay to obtain incremental data, The standard think tank data is sorted according to time to obtain time series data, and the time weight under the first classification result is calculated: where w i is the time weight of the ith data point, B(α,β) is the beta function, Γ is the gamma function, α and β are shape parameters that determine the rate and shape of the time weight changing over time, t i is the time point of the i-th data point, t 0 is the reference time point, is the latest time point of the category to which the data point belongs, τ is the time scale parameter, and the incremental data is obtained by removing the time series data whose time weight is less than a preset threshold. The preset threshold is determined according to the update frequency of different categories, and is the time weight of the update frequency; C. matching the first classification result according to the historical valid data to obtain first data, matching the second classification result according to the historical valid data to obtain second data, and obtaining a recommended category based on the first data and the second data; D. establishes a recommendation model based on the historical valid data, the incremental data and the recommended categories, A state space is constructed according to the incremental data and the recommended categories, and the recommended data is used as an action space. The action space is mapped from the state space, and the actual feedback of the historical valid data corresponding to the recommended data is used as a reward signal to establish the recommendation model. When the recommendation model is output, the historical valid data in the recommended data is removed, and the recommendation model is updated according to the reward signal to optimize the mapping from state to action. Update of value function: Where V α (S) is the value function of state S, μ(A|S) is the strategy for selecting action A in state S, R(S,A) is the immediate reward obtained when taking action A, α is the discount factor of the value function, which is used to adjust the current value of future rewards, and P SS′ (A) is the probability of executing action A from state S to state S′, and the update of Q function: Where Q β (S, A; φ) is the Q function of state S and action A, is the expected estimate of taking action A in state S, β is the discount factor of the Q function, φ is the parameter of the main network used to parameterize the Q function, E is the expected value, which represents the average value under the probability distribution, and the policy gradient formula for updating φ is: Where J(φ) is the objective function, is the expected return, B is the currently selected action, γ represents the discount factor of the policy gradient, and π φ (B|S) is the parameterized strategy, the target Q function: Where Q γ (S, B; θ) is the Q function approximated by the neural network, which is a stable estimate used to calculate the future reward part of the expected return, C is the best action in the next state, and θ - represents the parameters of the target network, which are regularly updated from the main network parameters θ. The optimization is continuously iterated until the strategy converges. The recommendation model uses the preference category filled in by the user when the user logs in for the first time as the recommendation category. The data to be recommended is input into the recommendation model to output the recommendation result.

2. According to the standard think tank knowledge recommendation method based on big data as described in claim 1, It is characterized in that The method of obtaining first data by matching the first classification result according to the historical valid data includes: Calculate the similarity of the user according to the historical rating and the first classification result: Where sim(u,v) represents the similarity between user u and user v, P uv represents the set of categories rated by user u and user v, R ua represents the rating of user u on category a, represents the average rating of user u for all categories, t ua is the time when user u scores category a, λ is the time decay factor, exp represents an exponential function, a similarity matrix is ​​constructed according to the similarity, the users whose similarity in the similarity matrix is ​​greater than a preset similarity threshold are taken as collaborative users, and the value range of the similarity threshold is [0.4, 0.9], and the click rates of the historical click data of the user and the collaborative user for the first classification result are weighted and fused to obtain the first data: Where P ui is the first data, indicating the predicted preference of user u for category i, i represents the category in the first classification result, w is the weight coefficient, p ui represents the click rate of user u on category i, Represents the click rate of the collaborative users of user u on category i.

3. According to the standard think tank knowledge recommendation method based on big data as described in claim 1, It is characterized in that The method for obtaining second data by matching the second classification result according to the historical valid data includes: The click rate and average score of the second classification result are calculated based on the historical valid data of the user and normalized preprocessed, and the weighted average of the click rate and average score of the second classification result is used as the second data: Where P uj is the second data, indicating the predicted preference of user u for category j, where j represents the category in the second classification result, and w 1 is the weight of the click rate of the second classification result, w 2 is the weight of the average score, p uj represents the click rate of user u for category j, A uj is the average rating of user u for category j.

4. According to the standard think tank knowledge recommendation method based on big data as described in claim 1, It is characterized in that The method for obtaining a recommended category based on the first data and the second data includes: A data set is established using the first data and the second data, each data vector in the data set is composed of the first data and the second data, and two distribution models are trained using the data set, representing the recommended category and other categories respectively, and the probability density function of the distribution model is: Among them, γ 1 and γ 2 is the mixing coefficient, represents normal distribution, μ 1 and μ 2 is the mean vector of the recommended category and the other categories, Ω 1 and Ω 2 is the covariance matrix of the recommended category and the other categories, classifying the data vector into the category with the largest posterior probability, and the posterior probability of the distribution model: Where p(c=k|x) is the posterior probability, which indicates the probability that the data vector x belongs to category k, c represents the category, k represents the index of the category, and γ k To update the mixing coefficient, μ k is the updated mean, Ω k To update the covariance matrix, the calculation formulas for the updated mixing coefficient, the updated mean and the updated covariance matrix are: Wherein N represents the number of the data vectors in the data set, and T represents the transpose of the matrix.

5. A standard think tank knowledge recommendation system based on big data, used to execute the standard think tank knowledge recommendation method based on big data according to any one of claims 1 to 4, It is characterized in that include: A collection module is used to obtain standard think tank data and historical valid data of users, wherein the historical valid data includes the historical click data of the users and the historical ratings of the standard think tank data; A classification module, which classifies the standard think tank data according to semantic analysis and data type to obtain a first classification result and a second classification result, and obtains a recommended category based on the first data and the second data; a calculation module, which performs data screening on the first classification result based on time decay to obtain incremental data, matches the first classification result according to the historical valid data to obtain first data, matches the second classification result according to the historical valid data to obtain second data, and establishes a recommendation model through the historical valid data, the incremental data and the recommended category; The output module inputs the data to be recommended into the recommendation model and outputs the recommendation result.

Citation Information

Patent Citations

  • Patent recommendation method and device, computer equipment and storage medium

    CN112434151A

  • Think tank knowledge recommendation method, device and equipment and medium

    CN114707068A

  • Footwear fashion trend prediction recommendation method and system based on user transaction data

    CN115713390A

  • Content recommendation method, system and equipment and storage medium

    CN116521985A

  • System and method for smart categorization of content in a content management system

    WO2022067236A1