A Fund Recommendation Method

By combining fund classification based on thematic sector and deep reinforcement learning model, the problem of unreasonable existing fund recommendation system is solved, intelligent recommendation and investment advice for funds are achieved, and the accuracy of investors' return prediction is improved.

CN112948700BActive Publication Date: 2024-11-12QIRE (BEIJING) DIGITAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110400043.9
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-04-14
Publication Date
2024-11-12
Estimated Expiration
2041-04-14

AI Technical Summary

Technical Problem

The existing fund recommendation system is unreasonable and utilitarian, and it is impossible to effectively screen out high-quality funds, making it difficult to meet the vital interests of investors.

Method used

By establishing a fund classification table based on thematic sectors, high-quality funds are screened using the initial screening function and semantic judgment model, and using the deep reinforcement learning model to predict the operational strategies of the funds, providing investment advice.

Benefits of technology

It realizes intelligent, reasonable and fair recommendations of funds, screens out high-value funds, and improves investors' investment experience and return prediction accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112948700B_ABST
    Figure CN112948700B_ABST
Patent Text Reader

Abstract

The present invention discloses a fund recommendation method, which includes establishing a fund classification table based on theme plates, selecting preliminary screening funds from the fund classification table according to a preliminary screening function, generating a preliminary screening fund table, calculating the quality of the funds in the preliminary screening fund table based on a second parameter, and filling the fund identification corresponding to the high-quality funds into the fund recommendation table; establishing and training a deep reinforcement learning model based on fund operation strategies, adding the predicted operation strategies as investment suggestions to the fund recommendation table, and recommending the fund recommendation table to users. The method of the embodiment of the present invention proposes a fund recommendation method, which can screen high-value funds and recommend them to customers, improve customer investment experience, and provide operation suggestions for funds. The prediction accuracy is high, and investment guidance can be provided for users, especially novice users. The recommendation method is intelligent, reasonable and fair, and can solve the unreasonable and utilitarian fund recommendations of the existing fund recommendation system. Therefore, the vital interests of investors can be better guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to the technical field of financial risk management and control, and in particular to a fund recommendation method. [Background technology]

[0002] In the past decade, financial products have developed rapidly in my country. For example, the public offering fund products with the highest conversion rate have more than 100 fund companies, more than 6,000 fund products, and an asset scale of more than 8 trillion yuan. Various types of fund products have been formed, such as stock funds, bond funds, money market funds, FOF funds, etc. From the historical performance of the fund market, most funds have good annual returns, but most consumers have not been able to make profits in the fund market. The reason is that there is a lack of professional and systematic analysis of fund products, and it is difficult to select high-quality funds from a large number of fund products.

[0003] The existing fund recommendation system is neither reasonable nor objective in its fund recommendations and is often influenced by stakeholders. Therefore, it cannot guarantee the vital interests of investors. [Summary of the invention]

[0004] In view of this, an embodiment of the present invention provides a fund recommendation method.

[0005] An embodiment of the present invention provides a fund recommendation method, the method comprising:

[0006] S1. Establish a fund classification table based on the theme sector, obtain the first parameter of the fund within the preset time period, select the initial screening fund from the fund classification table according to the initial screening function, and fill the fund identifier corresponding to the initial screening fund into the initial screening fund table;

[0007] S2. Determine the data source platform according to the preset data collection list, obtain the characteristic data of the funds in the preliminary screening fund table within the preset time period from the data source platform, pre-process the characteristic data and generate the second parameter based on the semantic judgment model classification;

[0008] S3. Calculate the quality Q of the funds in the initial screening fund table based on the second parameter, and fill the fund identifiers corresponding to the funds with quality Q>quality threshold Q0 into the fund recommendation table;

[0009] S4. Establish and train a deep reinforcement learning model based on fund operation strategies, and use the trained deep reinforcement learning model to predict the operation strategies of the target funds in the fund recommendation table;

[0010] S5. Add the predicted operation strategy as investment advice to the fund recommendation table, and recommend the fund recommendation table to the user.

[0011] According to the aspects and any possible implementations described above, there is further provided an implementation, wherein the first parameter includes: number of visits, number of holders and net worth.

[0012] According to the above aspects and any possible implementation, an implementation is further provided, wherein the initial screening function is defined as:

[0013]

[0014] Among them, F(x) represents the initial screening function, α i represents the average number of visits to the ith fund in the xth theme section within the preset time period, α0 represents the visit threshold set according to the actual situation; β i represents the average number of holders of the ith fund in the xth theme sector within the preset time period, β0 represents the threshold number of holders set according to actual conditions; γ i represents the average net asset value of the ith fund in the xth theme sector within the preset time period, γ0 represents the net asset value threshold set according to the actual situation; n x represents the number of funds in the xth theme sector; w1, w2, w3 represent weights, satisfying w1, w2, w3∈[0,1] and w1+w2+w3=1;

[0015] From each theme sector, select M funds with the largest values ​​of the initial screening function F(x) as the initial screening funds.

[0016] According to the aspects described above and any possible implementation method, an implementation method is further provided, wherein the characteristic data of the funds in the initial screening fund table include: fund mention data, fund comment data, fund comment like data, fund manager mention data, fund manager comment data, and fund manager comment like data.

[0017] According to the above aspects and any possible implementation manner, an implementation manner is further provided, wherein the preprocessing of the feature data specifically includes:

[0018] Set preprocessing priorities for the acquired fund feature data, where text comments are the first priority, voice comments are the second priority, and picture comments are the third priority;

[0019] First priority processing: determine whether the text review matches the positive word library. If the match is successful, determine whether the comment IP overlap degree corresponding to the text review exceeds the overlap threshold. If it exceeds, abandon all potential positive marks of the comment IP; if it does not exceed, perform potential positive marks and enter the second priority processing;

[0020] Second priority processing: convert the voice comments into text comments, determine whether the text comments match the positive word library, and if the match is successful, determine whether the comment IP overlap degree corresponding to the text comments exceeds the overlap degree threshold. If it exceeds, abandon all potential positive marks of the comment IP; if it does not exceed, mark the potential positive marks and enter the third priority processing;

[0021] The third priority processing: determine whether the number of identical pictures in the picture comments exceeds the quantity threshold. If it does not exceed the quantity threshold, identify the text comments in the picture comments, and further determine whether the text comments match the positive review vocabulary. If the match is successful, mark it as a potential positive review. If it exceeds the quantity threshold, further determine whether the IP overlap of the picture comments exceeds the overlap threshold. If it exceeds the overlap threshold, abandon all potential positive review marks for the comment IPs that exceed the overlap threshold. If it does not exceed the overlap threshold, identify the text comments in the picture comments, and further determine whether the text comments match the positive review vocabulary. If the match is successful, mark it as a potential positive review.

[0022] According to the above aspects and any possible implementation manner, an implementation manner is further provided, wherein the generating the second parameter based on the semantic judgment model classification specifically includes:

[0023] After pre-processing the characteristic data of the fund, a positive evaluation tendency mark is generated;

[0024] Construct a semantic judgment model, and its construction method is as follows:

[0025] Building the optimal reward model: Among them, E represents the expected value, λ represents the discount factor, λ∈[0,1]; s0 represents the initial state, R represents the reward function, π(s t ) represents a strategy for mapping states to operations;

[0026] Define the Q function: Where πi represents the current strategy for determining the Q value according to the equation, R represents the function, λ represents the discount factor, and p(s,a,s * ) indicates that action a transitions from state s to s * The transition probability, T πi represents the reward obtained in iterative step i;

[0027] The iterative update of the new strategy is as follows:

[0028] π(i+1)(s)=arg max Q(s,a), define the ε-greedy behavior strategy, and use the ε-greedy behavior strategy to determine the behavior of the current state, where each action is based on some predefined fixed probability Randomly selected.

[0029] The Q value is obtained by learning iteratively to approximate the optimal strategy;

[0030] The semantic judgment model is used to perform reinforcement learning on feature data that carry positive review tendency indicators to generate positive review classification results.

[0031] According to the above aspects and any possible implementation, an implementation is further provided, wherein the calculation formula of the quality Q is as follows:

[0032]

[0033] Among them, Q represents the quality, p1 represents the number of positive reviews of the fund, m1 represents the number of mentions of the fund, c1 represents the number of likes for the positive reviews of the fund, p2 represents the number of positive reviews of the fund corresponding to the fund manager, m2 represents the number of mentions of the fund corresponding to the fund manager, c2 represents the number of likes for the positive reviews of the fund corresponding to the fund manager, t represents the time in days; k1, k2 represent adjustment coefficients, satisfying k1, k2∈[0,1] and k1+k2=1.

[0034] According to the above aspects and any possible implementation, an implementation is further provided, wherein the establishment and training of a deep reinforcement learning model based on a fund operation strategy specifically includes:

[0035] Obtain the historical operation strategy data of multiple funds, sum and average them for input, predict the fund operation strategy, and establish its corresponding Markov decision process model. The action is represented by a, which includes buy, sell and keep. The state is represented by s, which is the fund price information generated by the behavioral strategy. The reward is represented by R, which is the change in the portfolio value when the state changes.

[0036] Training data, constantly updating the value function V π (s,a), until the value function V π (s,a) converges and obtains the optimal value function V * (s,a);

[0037] Optimal value function V * (s,a) is expressed as follows:

[0038]

[0039] Among them, V * (s,a) represents the optimal value function, s′∈S represents the state instance, a∈A represents the action instance, γ represents the discount factor, R represents the reward function, which specifies the reward, and P represents the transition function, which specifies the state transition probability;

[0040] Based on the above optimal value function V * (s,a), the optimal strategy π *(s) can be obtained:

[0041]

[0042] Among them, π * (s) represents the optimal strategy, P sa (s′,a) represents the transition probability from state s taking action a to the next state s′, a∈A represents the action instance, and γ represents the discount factor;

[0043] A recurrent neural network is used as the Q value network with a parameter of θ;

[0044] H t =f(u×x t +w×H t-1 +b1),

[0045] Q t =f(v×H t-1 +b2),

[0046] L=Q t -y t ,

[0047] Among them, H t represents the hidden state at time t, H t-1 represents the hidden state at time t-1, Q t represents the output of the current layer at time t, L represents the error, x t represents the training data input at time t, y t represents the original output of the training data, the original output of the training data, f represents the activation function of the hidden layer, u, w and v represent the weights shared by the recurrent neural network, b1 and b2 represent the thresholds shared by the recurrent neural network;

[0048] Define the loss function L(θ) in Q value;

[0049] The batch gradient descent method is used to train the parameters of the recurrent neural network. As the number of training times increases, the action with the largest Q value is selected through the Q value output by the network, and finally converges to the optimal strategy;

[0050] During the update cycle, the historical operation strategy data that has been pre-divided into test sets will be used to test the trained model.

[0051] According to the above aspects and any possible implementation, an implementation is further provided, wherein the loss function L(θ) is expressed by the following formula:

[0052]

[0053] Among them, L(θ) represents the loss function, r represents the reward value, θ and θ′ represent the neural network weights, represents the target Q function value, Q(s,a,θ) represents the predicted Q function value, and γ represents the discount factor.

[0054] One of the above technical solutions has the following beneficial effects:

[0055] The method of the embodiment of the present invention proposes a fund recommendation method. By selecting preliminary screening funds from a fund classification table through a preliminary screening function, funds with higher activity and greater growth potential can be initially filtered out for investors, and a secondary evaluation standard based on the dual indicators of funds and fund managers can be obtained, so as to screen out funds with a quality Q that meets the requirements and recommend them to users; a deep reinforcement learning model is used to predict the operating strategies of target funds in the fund recommendation table, and operation suggestions for the funds are provided. The prediction has high accuracy and can provide investment guidance for users, especially novice users; the recommendation method is intelligent, reasonable and fair, and can screen high-value funds and recommend them to customers, thereby improving customer investment experience and solving the unreasonable and utilitarian nature of fund recommendation in existing fund recommendation systems. Therefore, the vital interests of investors can be better guaranteed.

Brief Description of the Drawings

[0056] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 It is a flow chart of S1-S5 of the fund recommendation method provided in an embodiment of the present invention;

[0058] Figure 2 It is a schematic diagram of a flow chart of preprocessing feature data provided by an embodiment of the present invention. [Specific implementation method]

[0059] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. 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.

[0060] Please refer to Figure 1 , Figure 1 This is a flow chart of fund recommendation methods S1-S5. Figure 1As shown, the method comprises the following steps:

[0061] S1. Establish a fund classification table based on the theme section, obtain the first parameter of the fund within the preset time period, select the initial screening fund from the fund classification table according to the initial screening function, and fill the fund identifier corresponding to the initial screening fund into the initial screening fund table. Among them, the first parameter includes: visit volume, number of holders and net value.

[0062] S2. Determine the data source platform according to the preset data collection list, obtain the characteristic data of the funds in the preliminary screening fund table within the preset time period from the data source platform, pre-process the characteristic data and generate the second parameter based on the semantic judgment model classification. Among them, the characteristic data of the funds in the preliminary screening fund table include: fund mention data, fund comment data, fund comment like data, fund manager mention data, fund manager comment data, fund manager comment like data.

[0063] S3. Calculate the quality Q of the funds in the initial screening fund table based on the second parameter, and fill the fund identifiers corresponding to the funds with quality Q>quality threshold Q0 into the fund recommendation table.

[0064] S4. Establish and train a deep reinforcement learning model based on fund operation strategies, and use the trained deep reinforcement learning model to predict the operation strategies of the target funds in the fund recommendation table.

[0065] S5. Add the predicted operation strategy as investment advice to the fund recommendation table, and recommend the fund recommendation table to the user.

[0066] The present invention establishes fund recommendations based on theme sectors, which is convenient for investors to quickly find funds and is also beneficial for initial contacts of funds to quickly understand the funds; by selecting initial screening funds from a fund classification table through an initial screening function, funds with higher activity and greater growth potential can be initially filtered out for investors; then, based on the initial screening, the quality Q of the funds in the initial screening fund table is calculated based on a second parameter, and the fund identifiers corresponding to the funds with quality Q>quality threshold Q0 are filled into the fund recommendation table, the characteristic data of the funds are analyzed, and a secondary evaluation standard based on the dual indicators of funds and fund managers is obtained, and funds with quality Q that meet the requirements are screened out for recommendation to users; in addition, a deep reinforcement learning model based on fund operation strategies is established and trained, and the operation strategies of target funds in the fund recommendation table are predicted using the trained deep reinforcement learning model, so that while recommending funds to users, the operation strategies of target funds in the fund recommendation table are predicted using a reinforcement learning model, and operation suggestions for the funds are provided, so as to provide investment guidance for users, especially novice users. The fund recommendation method of the present invention is intelligent, reasonable and fair, can screen high-value funds and recommend them to customers, improve customer investment experience, and can solve the unreasonable and utilitarian fund recommendations of existing fund recommendation systems. Therefore, it can better guarantee the vital interests of investors.

[0067] The primary screening function of the embodiment of the present invention is defined as:

[0068]

[0069] Among them, F(x) represents the initial screening function, α i represents the average number of visits to the ith fund in the xth theme section within the preset time period, α0 represents the visit threshold set according to the actual situation; β i represents the average number of holders of the ith fund in the xth theme sector within the preset time period, β0 represents the threshold number of holders set according to actual conditions; γ i represents the average net asset value of the ith fund in the xth theme sector within the preset time period, γ0 represents the net asset value threshold set according to the actual situation; n x represents the number of funds in the xth theme sector; w1, w2, w3 represent weights, satisfying w1, w2, w3∈[0,1] and w1+w2+w3=1;

[0070] From each theme sector, select M funds with the largest values ​​of the initial screening function F(x) as the initial screening funds.

[0071] Please refer to Figure 2 , which is Figure 2 FIG. 1 is a flow chart of preprocessing feature data in an embodiment of the present invention. Figure 2 As shown, the method comprises the following steps:

[0072] The preprocessing of the feature data specifically includes:

[0073] Set preprocessing priorities for the acquired fund feature data, where text comments are the first priority, voice comments are the second priority, and picture comments are the third priority;

[0074] First priority processing: determine whether the text review matches the positive word library. If not, directly enter the second priority processing; if the match is successful, determine whether the comment IP overlap corresponding to the text review exceeds the overlap threshold. If exceeded, abandon all potential positive marks of the comment IP; if not exceeded, perform potential positive marks and enter the second priority processing;

[0075] Second priority processing: convert the voice comments into text comments, determine whether the text comments match the positive word library, if not, directly enter the third priority processing; if the match is successful, determine whether the comment IP overlap degree corresponding to the text comments exceeds the overlap degree threshold, if exceeded, abandon all potential positive comments marks of the comment IP; if not exceeded, perform potential positive comments marks, and enter the third priority processing;

[0076] The third priority processing: determine whether the number of identical pictures in the picture comments exceeds the quantity threshold. If it does not exceed the quantity threshold, identify the text comments in the picture comments, and further determine whether the text comments match the positive review vocabulary. If the match is not successful, terminate directly; if the match is successful, mark the potential positive review; if it exceeds the quantity threshold, further determine whether the IP overlap of the picture comments exceeds the overlap threshold. If it exceeds the overlap threshold, abandon all potential positive review marks of the comment IPs that exceed the overlap threshold. If it does not exceed the overlap threshold, identify the text comments in the picture comments, and further determine whether the text comments match the positive review vocabulary. If the match is successful, mark the potential positive review.

[0077] The embodiment of the present invention pre-processes the feature data text comments, voice comments, and picture comments as the third priority, and can fully obtain the comment data, so that the system can make more accurate judgments on funds and fund managers. In addition, by judging the overlap of comment IPs and the number of identical pictures in picture comments, water army comments can be efficiently eliminated, making the data more real and accurate. Marking potential positive reviews in advance can reduce the task processing volume of the semantic judgment model and further improve the accuracy of the semantic judgment model classification.

[0078] The second parameter is generated based on the semantic judgment model classification, specifically including:

[0079] After pre-processing the characteristic data of the fund, a positive evaluation tendency mark is generated;

[0080] Construct a semantic judgment model, and its construction method is as follows:

[0081] Building the optimal reward model: Among them, E represents the expected value, λ represents the discount factor, λ∈[0,1]; s0 represents the initial state, R represents the reward function, π(s t ) represents a strategy for mapping states to operations;

[0082] Define the Q function: Where πi represents the current strategy for determining the Q value according to the equation, R represents the function, λ represents the discount factor, and p(s,a,s * ) indicates that action a transitions from state s to s * The transition probability, T πi represents the reward obtained in iteration step i;

[0083] The iterative update of the new strategy is as follows:

[0084] π(i+1)(s)=arg max Q(s,a), define the ε-greedy behavior strategy, and use the ε-greedy behavior strategy to determine the behavior of the current state, where each action is based on some predefined fixed probability Randomly selected.

[0085] The Q value is obtained by learning iteratively to approximate the optimal strategy;

[0086] The semantic judgment model is used to perform reinforcement learning on feature data that carry positive review tendency indicators to generate positive review classification results.

[0087] The semantic judgment model of the embodiment of the present invention can realize accurate classification of favorable comments. After training the RL-based algorithm, it can classify favorable comments and other comments with a higher confidence. The model has good robustness and can efficiently and accurately extract the second parameter from the characteristic data of the fund.

[0088] The calculation formula of the quality Q of the embodiment of the present invention is as follows:

[0089]

[0090] Among them, Q represents the quality, p1 represents the number of positive reviews of the fund, m1 represents the number of mentions of the fund, c1 represents the number of likes for the positive reviews of the fund, p2 represents the number of positive reviews of the fund corresponding to the fund manager, m2 represents the number of mentions of the fund corresponding to the fund manager, c2 represents the number of likes for the positive reviews of the fund corresponding to the fund manager, t represents the time in days; k1, k2 represent adjustment coefficients, satisfying k1, k2∈[0,1] and k1+k2=1.

[0091] The calculation formula for quality Q can establish a secondary evaluation standard based on the dual indicators of funds and fund managers, and establish a time decay model based on logarithmic functions to efficiently and accurately select reliable funds and fund managers.

[0092] The establishment and training of a deep reinforcement learning model based on a fund operation strategy in an embodiment of the present invention specifically includes:

[0093] Obtain the historical operation strategy data of multiple funds, sum and average them for input, predict the fund operation strategy, and establish its corresponding Markov decision process model. The action is represented by a, which includes buy, sell and keep. The state is represented by s, which is the fund price information generated by the behavioral strategy. The reward is represented by R, which is the change in the portfolio value when the state changes.

[0094] Training data, constantly updating the value function V π (s,a), until the value function V π (s,a) converges and obtains the optimal value function V * (s,a);

[0095] Optimal value function V * (s,a) is expressed as follows:

[0096]

[0097] Among them, V * (s,a) represents the optimal value function, s′∈S represents the state instance, a∈A represents the action instance, γ represents the discount factor, R represents the reward function, which specifies the reward, and P represents the transition function, which specifies the state transition probability;

[0098] Based on the above optimal value function V * (s,a), the optimal strategy π * (s) can be obtained:

[0099]

[0100] Among them, π * (s) represents the optimal strategy, P sa (s′,a) represents the transition probability from state s taking action a to the next state s′, a∈A represents the action instance, and γ represents the discount factor;

[0101] A recurrent neural network is used as the Q value network with a parameter of θ;

[0102] H t =f(u×x t +w×H t-1 +b1),

[0103] Q t=f(v×H t-1 +b2),

[0104] L=Q t -y t ,

[0105] Among them, H t represents the hidden state at time t, H t-1 represents the hidden state at time t-1, Q t represents the output of the current layer at time t, L represents the error, x t represents the training data input at time t, y t represents the original output of the training data, the original output of the training data, f represents the activation function of the hidden layer, u, w, and v represent the weights shared by the recurrent neural network, and b1 and b2 represent the thresholds shared by the recurrent neural network.

[0106] The recursive neural network of the embodiment of the present invention has excellent perception and feature extraction capabilities, and its key aspects are the representation of actual features, layer-by-layer self-learning, and sparse constraints that limit the parameter space to prevent overfitting.

[0107] Define the loss function L(θ) in Q value;

[0108] The batch gradient descent method is used to train the parameters of the recurrent neural network. As the number of training times increases, the action with the largest Q value is selected through the Q value output by the network, and finally converges to the optimal strategy;

[0109] During the update cycle, the historical operation strategy data that has been pre-divided into test sets will be used to test the trained model.

[0110] The loss function L(θ) of the embodiment of the present invention is expressed by the following formula:

[0111]

[0112] Among them, L(θ) represents the loss function, r represents the reward value, θ and θ′ represent the neural network weights, represents the target Q function value, Q(s,a,θ) represents the predicted Q function value, and γ represents the discount factor.

[0113] The deep reinforcement learning model constructed in the embodiment of the present invention is used for recursive Q network, and self-exploration and experience replay are realized for prediction based on combined feature processing of RNN and self-experimental analysis based on DQN. The deep reinforcement learning model predicts funds accurately and has good model robustness.

[0114] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0115] The present invention may be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform specific tasks or implement specific abstract data types. The present invention may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communication network. In a distributed computing environment, program modules may be located in local and remote computer storage media, including storage devices.

[0116] Each embodiment of the present invention is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0117] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A fund recommendation method, It is characterized in that The method comprises: S1. Establish a fund classification table based on the theme sector, obtain the first parameter of the fund within the preset time period, select the initial screening fund from the fund classification table according to the initial screening function, and fill the fund identifier corresponding to the initial screening fund into the initial screening fund table; S2. Determine the data source platform according to the preset data collection list, obtain the characteristic data of the funds in the preliminary screening fund table within the preset time period from the data source platform, pre-process the characteristic data and generate the second parameter based on the semantic judgment model classification; S3. Calculate the quality Q of the funds in the initial screening fund table based on the second parameter 1 , the quality Q 1 >Quality threshold Q 0 The fund ID corresponding to the fund is filled into the fund recommendation table; S4. Establish and train a deep reinforcement learning model based on fund operation strategies, and use the trained deep reinforcement learning model to predict the operation strategies of the target funds in the fund recommendation table; S5. Add the predicted operation strategy as investment advice to the fund recommendation table, and recommend the fund recommendation table to the user; The generating the second parameter based on the semantic judgment model classification specifically includes: After pre-processing the characteristic data of the fund, a positive evaluation tendency mark is generated; Construct a semantic judgment model, and its construction method is as follows: Building the optimal reward model: Where E represents the expected value, λ represents the discount factor, λ∈[0,1]; s 0 represents the initial state, R represents the reward function, π(s t ) represents a strategy for mapping states to operations; Define the Q function: Where πi represents the current strategy for determining the Q function value according to the equation, R represents the reward function, λ represents the discount factor, and p(s,a,s * ) indicates that action a moves from state s to state s * The transition probability, T πi represents the reward obtained in iteration step i; The iterative update of the new strategy is as follows: π(i+1)(s)=argmaxQ(s,a), define the ε-greedy behavior strategy, and use the ε-greedy behavior strategy to determine the behavior of the current state, where each action is based on some predefined fixed probability randomly selected; The Q function value is obtained by learning iteratively to approximate the optimal strategy; Through the semantic judgment model, the feature data carrying the positive review tendency mark is reinforced to generate the positive review classification result; The quality Q 1 The calculation formula is as follows: Among them, Q 1 Indicates quality, p 1 Indicates the number of favorable comments of the fund, m 1 represents the number of fund mentions, c 1 Indicates the number of likes for the fund's favorable comments, p 2 Indicates the number of positive reviews of the fund manager corresponding to the fund, m 2 represents the number of mentions of the fund manager corresponding to the fund, c 2 represents the number of likes for the fund manager's favorable comments, t represents the time in days; k 1 , k 2 Represents the adjustment coefficient, satisfying k 1 , k 2 ∈[0,1] and k 1 +k 2 =1.

2. The method according to claim 1, It is characterized in that The first parameters include: number of visits, number of holders and net worth.

3. The method according to claim 2, It is characterized in that The initial screening function is defined as: Among them, F(x) represents the initial screening function, α i represents the average number of visits to the ith fund in the xth theme section within the preset time period, α 0 Indicates the visit volume threshold set according to the actual situation; β i represents the average number of holders of the ith fund in the xth theme sector within the preset time period, β 0 Indicates the number of holders threshold set according to actual conditions; γ i represents the average net asset value of the ith fund in the xth theme sector within the preset time period, γ 0 Indicates the net value threshold set according to actual conditions; n x represents the number of funds in the xth theme sector; w 1 ,w 2 , w 3 Represents the weight, satisfying w 1 ,w 2 , w 3 ∈[0,1] and w 1 +w 2 +w 3 =1; From each theme sector, select M funds with the largest values ​​of the initial screening function F(x) as the initial screening funds.

4. The method according to claim 1, It is characterized in that The characteristic data of the funds in the preliminary screening fund table include: fund mention data, fund comment data, fund comment like data, fund manager mention data, fund manager comment data, and fund manager comment like data.

5. The method according to claim 4, It is characterized in that The preprocessing of the feature data specifically includes: Set preprocessing priorities for the acquired fund feature data, where text comments are the first priority, voice comments are the second priority, and picture comments are the third priority; First priority processing: determine whether the text review matches the positive word library. If the match is successful, determine whether the comment IP overlap degree corresponding to the text review exceeds the overlap threshold. If it exceeds, abandon all potential positive marks of the comment IP; if it does not exceed, perform potential positive marks and enter the second priority processing; Second priority processing: convert the voice comments into text comments, determine whether the text comments match the positive word library, and if the match is successful, determine whether the comment IP overlap degree corresponding to the text comments exceeds the overlap degree threshold. If it exceeds, abandon all potential positive marks of the comment IP; if it does not exceed, mark the potential positive marks and enter the third priority processing; The third priority processing: determine whether the number of identical pictures in the picture comments exceeds the quantity threshold. If it does not exceed the quantity threshold, identify the text comments in the picture comments, and further determine whether the text comments match the positive review vocabulary. If the match is successful, mark it as a potential positive review. If it exceeds the quantity threshold, further determine whether the IP overlap of the picture comments exceeds the overlap threshold. If it exceeds the overlap threshold, abandon all potential positive review marks for the comment IPs that exceed the overlap threshold. If it does not exceed the overlap threshold, identify the text comments in the picture comments, and further determine whether the text comments match the positive review vocabulary. If the match is successful, mark it as a potential positive review.

6. The method according to claim 1, It is characterized in that The establishment and training of a deep reinforcement learning model based on fund operation strategies specifically includes: Obtain the historical operation strategy data of multiple funds, sum and average them for input, predict the fund operation strategy, and establish its corresponding Markov decision process model. The action is represented by a, which includes buy, sell and keep. The state is represented by s, which is the fund price information generated by the behavioral strategy. The reward is represented by R, which is the change in the portfolio value when the state changes. Training data, constantly updating the value function V π (s,a), until the value function V π (s,a) converges and obtains the optimal value function V * (s,a); Optimal value function V * (s,a) is expressed as follows: Among them, V * (s,a) represents the optimal value function, s′∈S represents the state instance, a∈A represents the action instance, γ represents the discount factor, R represents the reward function, P represents the transition function, and specifies the state transition probability; Based on the above optimal value function V * (s,a), the optimal strategy π * (s) can be obtained: Among them, π * (s) represents the optimal strategy, P sa (s′,a) represents the transition probability from state s taking action a to the next state s′, a∈A represents the action instance, and γ represents the discount factor; A recurrent neural network is used as the network of the Q function value, with the parameter θ; H t =f(u×x t +w×H t-1 +b 1 ), Q t =f(v×H t-1 +b 2 ), L=Q t -y t , Among them, H t represents the hidden state at time t, H t-1 represents the hidden state at time t-1, Q t represents the output of the current layer at time t, L represents the error, x t represents the training data input at time t, y t represents the original output of the training data, the original output of the training data, f represents the activation function of the hidden layer, u, w and v represent the weights shared by the recurrent neural network, b 1 and b 2 represents the threshold shared by the recurrent neural network; Define the loss function L(θ) in the Q function value; The batch gradient descent method is used to train the parameters of the recurrent neural network. As the number of training times increases, the action with the largest Q function value is selected through the Q function value output by the network, and finally converges to the optimal strategy; During the update cycle, the historical operation strategy data that has been pre-divided into test sets will be used to test the trained model.

7. The method according to claim 6, It is characterized in that The loss function L(θ) is expressed as follows: Among them, L(θ) represents the loss function, r represents the reward value, θ and θ′ represent the neural network weights, represents the target Q function value, Q(s,a,θ) represents the predicted Q function value, and γ represents the discount factor.