Recommendation method, system and device based on multi-objective and multi-task

Abstract

The present invention discloses a recommendation method, system and device based on multiple objectives and multiple tasks, the method comprising: obtaining relevant data to form a data set, preprocessing the data set to obtain a processed data set; forming a feature set based on user-related data and financial management-related data, the feature set comprising a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set; constructing a product recommendation pre-training model; training the product recommendation pre-training model based on the feature set to obtain a product recommendation model; the product recommendation model performs inference prediction on the target-related data to obtain a total probability and then obtain a financial management product recommendation result. The present invention constructs a multi-task and multi-objective product recommendation model by deeply analyzing the complete behavior chain of the user, accurately capturing and expressing the before-and-after dependencies between the behaviors in the chain, and providing users with more accurate behavior predictions.

Description

Technical Field

[0001] The present invention relates to the field of algorithm technology, and in particular to a recommendation method, system and device based on multi-objective and multi-task. Background Art

[0002] In the existing financial product recommendation field, recommendations are made through conversion rate estimation technology, which can accurately measure the possibility of users purchasing or investing in a certain financial product. Conversion rate estimation technology deeply mines user behavior data, analyzes users' financial preferences and intentions, and then provides personalized recommendation services. It not only improves the accuracy of recommendations, but also greatly enhances the user experience and can meet users' diverse investment needs.

[0003] However, the following problems still exist:

[0004] Although users are directed from the exposure page to the page for purchasing financial products, only a few users actually complete the purchase, resulting in a very unsatisfactory conversion rate in this link. From the perspective of modeling optimization, since the number of positive samples (i.e., actual purchase user data) collected in the conversion link is extremely limited, it is quite difficult to optimize the model to improve the purchase target.

[0005] There are significant differences in users' willingness to purchase financial products. Specifically, users who choose to complete transactions on the financial product page generally show a strong willingness to purchase financial products. Furthermore, even for the same user, their willingness to purchase different products will be different. For example, male users tend to prefer electronic devices, while female users prefer cosmetics. Therefore, if the user's willingness to purchase financial products can be accurately identified, this will undoubtedly greatly improve the user conversion rate. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a recommendation method, system and device based on multi-objective and multi-task.

[0007] In order to solve the above technical problems, the present invention is solved by the following technical solutions:

[0008] A recommendation method based on multi-objective and multi-task includes the following steps:

[0009] Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data;

[0010] Forming a feature set based on user-related data and financial management-related data, the feature set including a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set;

[0011] Constructing a product recommendation pre-training model, the product recommendation pre-training model includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer, the input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set, the shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, the subtask layer processes the shared feature set and the fused intention feature set respectively to obtain at least four probabilities and an intention probability, and the target fusion layer fuses all probabilities to obtain a total probability;

[0012] The product recommendation pre-trained model is trained based on the feature set to obtain a product recommendation model;

[0013] The product recommendation model performs inference and prediction on target-related data to obtain the total probability and then obtain the financial product recommendation result.

[0014] As an implementable method, a user feature set is obtained based on user-related data, a financial management-related feature set is obtained based on financial management-related data, and a cross-feature set is obtained based on the user feature set and the financial management-related data, wherein the financial management-related data includes click behavior labels, subscription button behavior labels, risk confirmation letter signing labels, and subscription behavior labels.

[0015] As an implementable method, obtaining a cross feature set based on a user feature set and financial management related data includes the following steps:

[0016] The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, and user financial management preference characteristics;

[0017] Based on user characteristics, the user's preference for financial management related data is obtained, and then the preference score is obtained;

[0018] Determine the cross-features of user characteristics and financial management related data based on the preference score;

[0019] The preference score is obtained in the following way:

[0020]

[0021] in, Indicates user Financial management related data The preference score of Represents an adjustment item. Indicates Weight coefficient, Function mapping representing user features, Function mapping representing financial management related data, Indicates the number of user features.

[0022] As an implementable method, the sequence feature is obtained by:

[0023] The financial management related data also includes sequence data, and the sequence data is converted through vector representation to obtain corresponding vector features, and the vector features are pooled and aggregated to form sequence features.

[0024] As an implementable method, the intention feature set includes at least personal asset characteristics, risk tolerance, investment preferences and investment experience, including the following process:

[0025] Obtain user behavior data for historical periods, including at least browsing behavior data, click behavior data, add behavior data, favorite behavior data, and purchase behavior data;

[0026] Sample user behavior data to obtain intention behavior data;

[0027] Analyze the intention behavior data to determine whether the user triggers the purchase action and analyze and count the number of times the user triggers the purchase action and the time of the purchase action;

[0028] Based on the number of times the user triggers the purchase action and the time of the purchase action, determine whether the user has the intention to manage the financial management. If the user views or understands the detailed information multiple times, it means that the user has the intention;

[0029] Build a user profile and obtain a set of user-related intention features.

[0030] As an implementable method, sampling the user behavior data to obtain the intended behavior data includes the following steps:

[0031] Determine whether there is abnormal data in user behavior data;

[0032] If it exists, the abnormal data is removed; if it does not exist, the sampling estimation value of the user behavior data is calculated based on the sampling estimation model, and the user behavior data with the sampling estimation value greater than the preset threshold is selected as the qualified intention behavior data;

[0033] The sampling estimation model is expressed as follows:

[0034]

[0035] in, represents the sample estimate, Indicates The sampling probability of user behavior data is Indicates the total amount of user behavior data. , , Indicates the time interval between user behavior data generation. Indicates the interval time, Indicates the average number of times user behavior data is generated per unit time; is the base of natural logarithms; Indicates the interval time User behavior data; is the weight of the i-th user behavior data.

[0036] As an implementable method, the shared feature structure layer processes the user feature set, the financial management related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, including the following steps:

[0037] The user feature set, the related feature set, the cross feature set and the sequence feature set are concatenated to obtain a shared feature set;

[0038] Fusing the shared feature set with the intention feature set to obtain a fused intention feature set;

[0039] The shared feature set is expressed as follows:

[0040]

[0041] The fusion intention feature set is expressed as follows:

[0042]

[0043] in, represents the user feature set, represents the set of financial management related features, represents the cross feature set, represents a sequence feature set, Indicates the intention feature set, represents the shared feature set, represents the fusion intention feature set, The vector representation of user features, A vector representation of financial management related features, represents the vector representation of the cross feature, Represents sequence characteristics, The vector representation of the sequence features, A vector representation of the intent features.

[0044] As an implementable method, the following steps are also included:

[0045] Construct a loss function to optimize the training process;

[0046] The loss function is expressed as follows:

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] in, , , , , and Represent the true labels of different target behaviors, , , , , and Represent the predicted probabilities of different target behaviors, Represents the number of samples of different target behaviors.

[0054] As an implementation method, the target fusion layer fuses all probabilities to obtain a total probability, including the following steps:

[0055] Fusing the first probability and the second probability to obtain a first fused probability;

[0056] The first fusion probability and the third probability are fused to obtain a second fusion probability;

[0057] The second fusion probability, the fourth probability and the intention probability are fused to obtain a total probability;

[0058] The first fusion probability is expressed as follows:

[0059]

[0060] The second fusion probability is expressed as follows:

[0061]

[0062] The total probability is expressed as follows:

[0063]

[0064] The total probability is expressed as follows:

[0065]

[0066] in, represents the first fusion probability, represents the second fusion probability, represents the total probability, represents the first probability, represents the second probability, represents the third probability, represents the fourth probability, Indicates the probability of intention.

[0067] As an implementable method, the following steps are also included:

[0068] Comparing the first likelihood ratio of the first fusion probability with the first preset threshold, and if the first likelihood ratio is less than the first preset threshold, adjusting the first cost function until the first likelihood ratio is greater than the first preset threshold;

[0069] Comparing the second likelihood ratio of the second fusion probability with the second preset threshold, and if the second likelihood ratio is less than the second preset threshold, adjusting the second cost function until the second likelihood ratio is greater than the second preset threshold;

[0070] Compare the likelihood ratio of the intention probability with a preset threshold, and if it is less than the preset threshold, adjust the cost function until the likelihood ratio is greater than the preset threshold;

[0071] The likelihood ratio is expressed as follows:

[0072]

[0073] The cost function is expressed as follows:

[0074]

[0075] in, represents the minimization of the overall expected value, Indicates that assuming there is actually a target Make a decision when true The price, In the assumption Make decisions when true The conditional probability of Indicates that it is assumed that there is actually a target The default probability, represents the i-th likelihood ratio, Indicates decision.

[0076] A recommendation system based on multi-objective and multi-task, comprising:

[0077] A data acquisition module, used to acquire relevant data to form a data set, and pre-process the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data;

[0078] A feature extraction module, which forms a feature set based on user-related data and financial management-related data, wherein the feature set includes a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set;

[0079] A model building module is used to build a product recommendation pre-training model, which includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer. The input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set. The shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set to obtain a shared feature set and a fused intention feature set. The subtask layer processes the shared feature set and the fused intention feature set to obtain at least four probabilities and an intention probability. The target fusion layer fuses all probabilities to obtain a total probability.

[0080] Model training module, which trains the product recommendation pre-trained model based on the feature set to obtain the product recommendation model;

[0081] The result recommendation module is used for the product recommendation model to perform inference and prediction on target-related data to obtain the total probability and then obtain the financial product recommendation result.

[0082] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following method is implemented:

[0083] Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data;

[0084] Forming a feature set based on user-related data and financial management-related data, the feature set including a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set;

[0085] Constructing a product recommendation pre-training model, the product recommendation pre-training model includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer, the input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set, the shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, the subtask layer processes the shared feature set and the fused intention feature set respectively to obtain at least four probabilities and an intention probability, and the target fusion layer fuses all probabilities to obtain a total probability;

[0086] The product recommendation pre-trained model is trained based on the feature set to obtain a product recommendation model;

[0087] The product recommendation model performs inference and prediction on target-related data to obtain the total probability and then obtain the financial product recommendation result.

[0088] A recommendation device based on multiple objectives and multiple tasks comprises a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the following method is implemented:

[0089] Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data;

[0090] Forming a feature set based on user-related data and financial management-related data, the feature set including a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set;

[0091] Constructing a product recommendation pre-training model, the product recommendation pre-training model includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer, the input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set, the shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, the subtask layer processes the shared feature set and the fused intention feature set respectively to obtain at least four probabilities and an intention probability, and the target fusion layer fuses all probabilities to obtain a total probability;

[0092] The product recommendation pre-trained model is trained based on the feature set to obtain a product recommendation model;

[0093] The product recommendation model performs inference and prediction on target-related data to obtain the total probability and then obtain the financial product recommendation result.

[0094] The present invention has significant technical effects due to the adoption of the above technical solution:

[0095] The present invention deeply analyzes the complete behavior chain of users, builds a multi-task and multi-objective product recommendation model, accurately captures and expresses the causal dependencies between various behaviors in the chain, and provides users with more accurate behavior predictions; the present invention can solve the problem of local data sparsity, and by modeling the sparse behaviors of the conversion links in the user behavior chain, it is possible to more accurately predict user behaviors and improve the prediction performance of the product recommendation model; the present invention also adds an intention feature set, so that the user conversion probability can be commissioned, thereby making the product recommendation model more accurate and providing users with more accurate and personalized financial product recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art 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 paying creative labor.

[0097] Figure 1 It is a schematic flow diagram of the method of the present invention;

[0098] Figure 2 It is a structural schematic diagram of the system of the present invention;

[0099] Figure 3 and Figure 4 is a schematic diagram corresponding to a specific embodiment;

[0100] Figure 5 It is a structural schematic diagram of the product recommendation model of the present invention. DETAILED DESCRIPTION

[0101] The present invention is further described in detail below in conjunction with embodiments. The following embodiments are for explanation of the present invention but the present invention is not limited to the following embodiments.

[0102] Embodiment 1:

[0103] A recommendation method based on multi-objective and multi-task, such as Figure 1 As shown, the following steps are included:

[0104] S100, acquiring relevant data to form a data set, and preprocessing the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data;

[0105] S200, forming a feature set based on user-related data and financial management-related data, wherein the feature set includes a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set;

[0106] S300, constructing a product recommendation pre-training model, the product recommendation pre-training model includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer, the input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set, the shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, the subtask layer processes the shared feature set and the fused intention feature set respectively to obtain at least four probabilities and an intention probability, and the target fusion layer fuses all probabilities to obtain a total probability;

[0107] S400, training a product recommendation pre-trained model based on the feature set to obtain a product recommendation model;

[0108] S500: The product recommendation model performs inference and prediction on target-related data to obtain a total probability and then obtain a financial product recommendation result.

[0109] The present invention deeply analyzes the complete behavior chain of users, builds a multi-task and multi-objective product recommendation model, accurately captures and expresses the causal dependencies between various behaviors in the chain, and provides users with more accurate behavior predictions; the present invention can solve the problem of local data sparsity, and by modeling the sparse behaviors of the conversion links in the user behavior chain, it is possible to more accurately predict user behaviors and improve the prediction performance of the product recommendation model; the present invention also adds an intention feature set, so that the user conversion probability can be commissioned, thereby making the product recommendation model more accurate and providing users with more accurate and personalized financial product recommendation results.

[0110] In one embodiment, a user feature set is obtained based on user-related data, a financial management-related feature set is obtained based on financial management-related data, and a cross feature set is obtained based on the user feature set and financial management-related data, wherein the financial management-related data includes a click behavior label, a purchase button behavior label, a risk confirmation letter signing label, and a purchase behavior label. That is to say, the user's reaction behavior data to the recommended financial management product is obtained from the log of the user behavior data, including whether there is a click, whether the purchase button is clicked, whether the risk confirmation letter is signed, whether the purchase is completed, etc., thereby generating data in the form of (user id, financial management product id, click behavior data, purchase button click behavior data, risk confirmation letter signing data, purchase behavior data) as a sample data set.

[0111] Here, the present invention also has financial intention data, and adds it to the existing data group to generate the final data group (user ID, financial product ID, click behavior data, click subscription button behavior data, signed risk confirmation data, purchase intention data, subscription behavior data), and the financial intention data is processed to form an intention feature set. Specifically, the intention feature set includes at least personal asset characteristics, risk tolerance, investment preferences and investment experience, including the following process;

[0112] Obtain user behavior data for historical periods, including at least browsing behavior data, click behavior data, add behavior data, favorite behavior data, and purchase behavior data;

[0113] Sample user behavior data to obtain intention behavior data;

[0114] Analyze the intention behavior data to determine whether the user triggers the purchase action and analyze and count the number of times the user triggers the purchase action and the time of the purchase action;

[0115] Based on the number of times the user triggers the purchase action and the time of the purchase action, determine whether the user has the intention to manage the financial management. If the user views or understands the detailed information multiple times, it means that the user has the intention;

[0116] Build a user profile and obtain a set of user-related intention features.

[0117] The sampling of user behavior data to obtain intention behavior data includes the following steps:

[0118] Determine whether there is abnormal data in user behavior data;

[0119] If it exists, the abnormal data is removed; if it does not exist, the sampling estimation value of the user behavior data is calculated based on the sampling estimation model, and the user behavior data with the sampling estimation value greater than the preset threshold is selected as the qualified intention behavior data;

[0120] The sampling estimation model is expressed as follows:

[0121]

[0122] in, represents the sample estimate, Indicates The sampling probability of user behavior data is Indicates the total amount of user behavior data. , , Indicates the time interval between user behavior data generation. Indicates the interval time, Indicates the average number of times user behavior data is generated per unit time; is the base of natural logarithms; Indicates the interval time User behavior data; is the weight of the ith user behavior data, which is generally determined based on experience.

[0123] This embodiment judges and samples the user behavior data, which makes the user behavior data more precise, and thus the accuracy of the intention feature set is higher. The judgment of abnormal data can rely on the existing technology to judge abnormal data.

[0124] In one embodiment, the user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, and user financial management preference characteristics. User characteristics mainly include user basic attributes and user tag preferences. User attributes include basic information such as user age, region, gender, and user financial management data, etc. User device characteristics include device operating system and device operator, etc. User activity characteristics include platform activity, scene activity, and user account level, etc. User financial management preference characteristics are calculated based on user clicks, browsing financial products, and other behaviors.

[0125] The method of obtaining a cross-feature set based on the user feature set and financial management related data comprises the following steps:

[0126] The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, and user financial management preference characteristics;

[0127] Based on user characteristics, the user's preference for financial management related data is obtained, and then the preference score is obtained;

[0128] Determine the cross-features of user characteristics and financial management related data based on the preference score;

[0129] The preference score is obtained in the following way:

[0130]

[0131] in, Indicates user Financial management related data The preference score of Represents an adjustment item. Indicates Weight coefficient, Function mapping representing user features, Function mapping representing financial management related data, Indicates the number of user features.

[0132] That is to say, the cross-feature is determined by the user's preference for financial management-related data. In other words, the user has a certain preference for financial management-related data, and has different preference scores for different financial management-related data. The user's preference score for financial management-related data in the financial management-related data set is considered to be the cross-feature of the user and the financial management-related data. Specifically, the cross-feature can be obtained in the following ways: Data collection: Collect historical user data, such as clicks, browsing, favorites, sharing, etc. Feature extraction: Extract features from user data, including basic attributes of users (such as age, gender, occupation) and behavioral characteristics (such as active time, preferred topics, etc.). Association analysis between features and labels: Analyze the relationship between user features and financial management-related data, and determine which user features are related to the preference for financial management-related data. Preference model construction: Use machine learning algorithms (such as logistic regression, random forest, gradient boosting tree, etc.) to build a model to learn the relationship between user features and label preferences. Preference score calculation: Use the constructed model to calculate the user's preference score for each label. Cross-feature determination: According to the preference score, determine the cross-features of user features and financial management-related data, that is, those features that have the greatest impact on the preference score. The above modeling process and training process can be implemented with the help of existing technologies and will not be described in detail here.

[0133] In addition, the sequence characteristics of the present invention are obtained by the following method:

[0134] The financial management related data also includes sequence data, and the sequence data is converted through vector representation to obtain corresponding vector features, and the vector features are pooled and aggregated to form sequence features.

[0135] In one embodiment, the shared feature structure layer processes the user feature set, the financial management related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, including the following steps:

[0136] The user feature set, the related feature set, the cross feature set and the sequence feature set are concatenated to obtain a shared feature set;

[0137] Fusing the shared feature set with the intention feature set to obtain a fused intention feature set;

[0138] The shared feature set is expressed as follows:

[0139]

[0140] The fusion intention feature set is expressed as follows:

[0141]

[0142] in, represents the user feature set, represents the set of financial management related features, represents the cross feature set, represents a sequence feature set, Indicates the intention feature set, represents the shared feature set, represents the fusion intention feature set, The vector representation of user features, A vector representation of financial management related features, represents the vector representation of the cross feature, Represents sequence characteristics, The vector representation of the sequence features, A vector representation of the intent features.

[0143] In other words, the shared feature processing layer mainly transforms the input features through feature embedding vectors, and then concatenates these features to form a sample feature, which is then input into the subtask layer as a shared feature.

[0144] In addition, in order to make the training results non-redundant, a loss function is constructed to optimize the training process;

[0145] The loss function is expressed as follows:

[0146]

[0147]

[0148]

[0149]

[0150]

[0151]

[0152] in, , , , , and Represent the true labels of different target behaviors, , , , , and Represent the predicted probabilities of different target behaviors, Represents the number of samples of different target behaviors.

[0153] In one embodiment, the target fusion layer fuses all probabilities to obtain a total probability, including the following steps:

[0154] Fusing the first probability and the second probability to obtain a first fused probability;

[0155] The first fusion probability and the third probability are fused to obtain a second fusion probability;

[0156] The second fusion probability, the fourth probability and the intention probability are fused to obtain a total probability;

[0157] The first fusion probability is expressed as follows:

[0158]

[0159] The second fusion probability is expressed as follows:

[0160]

[0161] The total probability is expressed as follows:

[0162]

[0163] The total probability is expressed as follows:

[0164]

[0165] in, represents the first fusion probability, represents the second fusion probability, represents the total probability, represents the first probability, represents the second probability, represents the third probability, represents the fourth probability, Indicates the probability of intention.

[0166] The calculated probability does not necessarily conform to the actual data or the actual situation of the user, so some additional steps are added, including the following steps:

[0167] Comparing the first likelihood ratio of the first fusion probability with the first preset threshold, and if the first likelihood ratio is less than the first preset threshold, adjusting the first cost function until the first likelihood ratio is greater than the first preset threshold;

[0168] Comparing the second likelihood ratio of the second fusion probability with the second preset threshold, and if the second likelihood ratio is less than the second preset threshold, adjusting the second cost function until the second likelihood ratio is greater than the second preset threshold;

[0169] Compare the likelihood ratio of the intention probability with a preset threshold, and if it is less than the preset threshold, adjust the cost function until the likelihood ratio is greater than the preset threshold;

[0170] The likelihood ratio is expressed as follows:

[0171]

[0172] The cost function is expressed as follows:

[0173]

[0174] in, represents the minimization of the overall expected value, Indicates that assuming there is actually a target Make a decision when true The price, In the assumption Make decisions when true The conditional probability of Indicates that it is assumed that there is actually a target The default probability, represents the i-th likelihood ratio, Indicates decision.

[0175] The cost function is used for adjustment each time comparison is made until the likelihood ratio meets the conditions. In this way, the probability obtained is close to the user's actual situation, and the final intention probability is more consistent with the user's intention.

[0176] Combined with Figure 3 , Figure 4 and Figure 5 To explain the practical application of the present invention in detail, the specific process is as follows:

[0177] This application is combined with a financial product recommendation system. For the financial product recommendation system, accurately identifying and recommending financial products that are most likely to trigger user purchases is a core issue that needs to be solved. Therefore, estimating the user's purchase probability becomes a crucial matter, and the purchase probability described here is the total probability of the present invention.

[0178] In order to make the total probability more accurate and closer to user portrait customization, there are currently two problems. The first is sample selection bias (SSB): the distribution sampling of the constructed training sample set deviates from the actual data distribution. Taking products as an example, the model is trained for clicked products, while most of the online data sets are not clicked, so the problem of sample selection bias will arise.

[0179] Second: Data sparsity (DS): The number of financial products displayed to users by the recommendation system is much larger than the number of financial products subscribed by users. At the same time, users who subscribe to the products only account for a small part of all users. Therefore, the sample space of users who subscribe to the products is very small compared to the entire sample space. Generally speaking, it is 1 to 3 orders of magnitude smaller.

[0180] Through in-depth analysis and construction of user behavior links, the present invention can cleverly express the conversion rate, that is, the total probability, as the product of the probability of each key behavior. This transformation not only breaks down the conversion rate problem that was originally complex and difficult to predict directly into several relatively simple and easy-to-handle behavior probability estimation problems, but also significantly improves the accuracy and reliability of the prediction. In this way, the financial product recommendation system can more accurately grasp the interests and needs of users, so as to recommend financial products that truly meet their subscription intentions to users. This not only improves the user experience, but also brings higher conversion rates and profits to related institutions.

[0181] Combined with the data of a certain APP: When the user enters the financial management page of a certain app, he will see a list of financial products recommended by the system. The user selects the financial product he is interested in to enter the financial product details page. The user will further explore the specific financial product by understanding the key information such as the product's rate of return, risk level, and investment period. If the user is interested in the financial product, he can click the "fixed investment / subscription" button at the bottom of the page. At this time, the user needs to complete the necessary identity verification, risk assessment and other processes to ensure the compliance and security of his investment behavior. When the user "confirms the signature", he will enter the final transaction page to complete the final payment action. This process can be referred to as follows Figure 3 shown.

[0182] based on Figure 3 The final "subscription" goal can be broken down into the following four goals: "click on the financial product", "click to subscribe / fixed investment", "sign the risk confirmation letter", and "subscribe", so as to design the corresponding multi-goals of subscription and financial management. For these four goals, design the corresponding model to estimate the conversion rate: the probability of a user clicking on a financial product in the exposed financial product list is called the click-through rate The probability of clicking the "Subscription / Fixed Investment" button on the financial product details page is called the click-through rate. (Click-to-Purchase Button Rate), the probability of signing the risk confirmation form on the risk confirmation page is called the signing rate (signature rate); the probability of completing the subscription on the trading page is called The probability of browsing the list of financial products and finally completing the purchase is called the financial conversion rate, where .

[0183] Throughout the entire process, if you want to accurately identify the user's financial management behavior, you need to deeply analyze the user's diverse portrait features and their behavior data. The user's financial management intention is closely related to his or her historical investment behavior and other personalized characteristics. For example, users with a rich investment history often show stronger financial management intentions than investment novices. Financial intention (FI) is a key variable that directly affects the user's behavioral choices after reaching the final transaction page. Specifically, when the user's financial management intention is high, they are more likely to complete the payment and successfully trade; conversely, if the financial management intention is insufficient, they are more likely to choose to abandon the transaction. In order to more accurately capture this trend and optimize the user experience, a financial management intention feature set is added. This measure not only makes the evaluation of the financial management conversion rate more detailed, but also provides strong data support for improving the transaction success rate and optimizing marketing strategies. By deeply understanding the user's financial management intention, it is possible to meet user needs more accurately, thereby promoting business growth. Therefore, the present invention introduces the user's financial management intention target in the last step of the transaction page to depict the financial management conversion rate more carefully. After adding the financial management intention target, the financial management conversion rate estimate can be expressed as , incorporating financial intention goals into multiple optimization goals.

[0184] In order to characterize the user's financial management intention, the present invention analyzes the differences in users' behavior patterns. In the e-commerce industry, users' willingness to buy can usually be seen from their browsing behavior. Users who show a strong willingness to buy tend to browse the favorite products and prices repeatedly, and even enter the transaction page multiple times to confirm the final transaction price. Each step they take reveals a strong interest in the product and a firm determination to buy. In contrast, users with weaker purchasing intentions may show completely different behavior patterns. They may only browse the products superficially and may not even go deep into the transaction page. Even if they accidentally enter the transaction page, they may quickly close and exit for various reasons, which is undoubtedly an obvious signal of lack of purchasing intention. Therefore, by accurately analyzing users' browsing and transaction behaviors, e-commerce platforms can more accurately identify which users have higher purchasing potential, thereby providing them with more personalized and accurate marketing and service strategies, thereby improving the overall transaction conversion rate and user satisfaction. The whole process is shown in the attached figure. Figure 4As shown. The intention feature set includes at least personal asset characteristics, risk tolerance, investment preferences and investment experience, including the following processes: obtaining user behavior data for a historical period, including at least browsing behavior data, clicking behavior data, adding behavior data, collecting behavior data and purchasing behavior data; sampling user behavior data to obtain intention behavior data; analyzing the intention behavior data to determine whether the user triggers a purchase action and analyzing and counting the number of times the user triggers the purchase action and the time of the purchase action; based on the number of times the user triggers the purchase action and the time of the purchase action, determining whether the user has financial management intentions, if the user views or understands the detailed information multiple times, it means that the user has intentions; constructing a user portrait to obtain a user-related intention feature set.

[0185] The features used in training the model also include the following:

[0186] User characteristics: including demographic characteristics (age, gender, city, etc.), user device characteristics (device operating system, device operator, etc.), user activity characteristics (platform activity, scenario activity, user account level, etc.), user financial management related characteristics (financial management style preferences, etc.), etc.

[0187] Financial management intention characteristics: user's personal financial status (such as investable assets, liabilities, etc.), risk tolerance (such as risk preference, risk tolerance, etc.), investment preferences (such as investment period, capital liquidity, etc.), user investment experience, etc.

[0188] Financial product features: including basic profile features such as financial product profitability, risk, and investment period, as well as pre-trained features of financial products (vector features of financial products trained based on co-occurrence data of financial products), etc.

[0189] Cross-characteristics: numerical characteristics of risk and investment period preferences between users and financial products, etc.

[0190] Sequence features: The financial products subscribed by a user within a period of time constitute a user sequence. In the training stage, the sequence passes through an end-to-end embedding layer to obtain the embedding features corresponding to the financial products, and then passes through a pooling layer (sum pooling is used in this invention) to aggregate into sequence features.

[0191] See attached Figure 5As shown, the product recommendation pre-training model constructed in the present application includes an input layer, a shared feature structure layer, a sub-task layer and a target fusion layer. The input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set. The shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set. The sub-task layer processes the shared feature set and the fused intention feature set respectively to obtain at least four probabilities and an intention probability. The target fusion layer fuses all probabilities to obtain a total probability.

[0192] The input layer receives represents the user feature set, represents the set of financial management related features, represents the cross feature set, represents a sequence feature set, Indicates the intention feature set, represents the shared feature set, Indicates fusion intention feature set, etc.

[0193] The shared feature layer processes the received feature set, and the shared feature set is represented as follows:

[0194]

[0195] The fusion intention feature set is expressed as follows:

[0196]

[0197] in, represents the user feature set, represents the set of financial management related features, represents the cross feature set, represents a sequence feature set, Indicates the intention feature set, represents the shared feature set, represents the fusion intention feature set, The vector representation of user features, A vector representation of financial management related features, represents the vector representation of the cross feature, Represents sequence characteristics, The vector representation of the sequence features, A vector representation of the intent features.

[0198] In other words, the shared feature processing layer mainly transforms the input features through feature embedding vectors, and then concatenates these features to form a sample feature, which is then input into the subtask layer as a shared feature.

[0199] At the subtask layer, four unique sub-models and one financial intention unit sub-model are built on the shared feature layer. These sub-models focus on different goals in the estimated behavior link, ensuring that each link to be estimated can be accurately predicted. As a sub-model , , , Input, As a sub-model Input. The core structure of each sub-model consists of a series of carefully connected fully connected layers, which are designed to refine and emphasize different feature weights through specific labels. In order to optimize the predictive ability of the model, a three-layer fully connected design was specially selected. In each layer, the data is first weighted and summed through the input layer. This process ensures that each input feature can be appropriately weighted according to its importance. Subsequently, the weighted data will pass through a relu activation function, which can introduce nonlinear factors to further enhance the expressiveness and flexibility of the model. After such processing, each layer can produce its unique output, which will then serve as the input of the next layer, step by step until the model produces the final prediction result. The output at the end of the full connection will be converted into a probability value output between 0 and 1 through the sigmoid function. The calculation formula of the sigmoid function is

[0200]

[0201] This architecture not only makes full use of the information in the shared feature layer, but also captures the subtle differences between different targets through four unique sub-models and one financial intention unit sub-model, thereby achieving more accurate and comprehensive predictions. This innovative design has shown significant advantages in multiple application scenarios and provides new ideas and solutions for the prediction of behavioral links.

[0202] In the target fusion layer, the output prediction values ​​of the five sub-models are , , , , , and the probability of users clicking the subscription button after entering the financial management page ,Depend on and Multiplying together gives:

[0203]

[0204] The probability that the user clicks the subscription button and signs the risk confirmation letter is ,Depend on and Multiply to get

[0205] The predicted probability value of a user having financial management intention and finally completing the subscription conversion is , from the above Multiply to get

[0206]

[0207] The final predicted subscription conversion probability value is:

[0208]

[0209] When training the model, the cross entropy loss function is selected as the optimization target. The cross entropy loss function is a widely used loss function that can effectively measure the difference between the probability distribution predicted by the model and the actual label distribution, and guide the learning process of the model by minimizing this difference. In combination with the model structure of the present invention, in order to make the training results non-redundant, a loss function is also constructed to optimize the training process;

[0210] The loss function is expressed as follows:

[0211]

[0212]

[0213]

[0214]

[0215]

[0216]

[0217] in, , , , , and Represent the true labels of different target behaviors, , , , , and Represent the predicted probabilities of different target behaviors, Represents the number of samples of different target behaviors.

[0218] In the model training stage of the present invention, AdamOptimizer is selected as the optimizer. The advantage of this optimizer is that it can adaptively adjust the learning rate and can effectively adjust the pace of parameter updates during the training process. Based on the LOSS value calculated in the forward propagation stage, the AdamOptimizer optimizer is used for back propagation, aiming to continuously iterate and optimize the model performance by calculating the gradient and updating the model parameters accordingly. At the same time, in order to ensure the efficiency and stability of the training process, a batch processing training strategy is adopted. That is, in each iteration, the entire data set will not be processed once, but will be divided into multiple small batches, each batch containing several samples. By optimizing each batch independently, the training process can be accelerated while ensuring the effective use of computing resources, and the risk of the model falling into a local optimal solution can be reduced. This iterative optimization process is continued until the model reaches the expected convergence state or meets the preset convergence conditions, ensuring that the model proposed by the present invention can remain efficient and accurate in a complex data environment.

[0219] The model inference process is as follows: When making online predictions, the model-related feature data will first be obtained. This data will be assembled strictly in the standard format required by the model to ensure data consistency and accuracy. Subsequently, this formatted data will be passed to the trained model, and the model will perform prediction operations based on this input data. During the prediction process, the model will output multiple key probability values Finally, in order to obtain a comprehensive prediction result, these five probability values ​​are multiplied. The calculation formula is as follows:

[0220]

[0221] Finally, the results are sorted in descending order according to the FICVR values, and the topn (highest ranking) results are intercepted and returned to the system to complete the model prediction.

[0222] Embodiment 2:

[0223] A recommendation system based on multi-objective and multi-task, such as Figure 2 As shown, including:

[0224] The data acquisition module 100 is used to acquire relevant data to form a data set, and pre-process the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data;

[0225] A feature extraction module 200 forms a feature set based on user-related data and financial management-related data, wherein the feature set includes a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set;

[0226] The model construction module 300 is used to construct a product recommendation pre-training model, which includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer. The input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set. The shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set to obtain a shared feature set and a fused intention feature set. The subtask layer processes the shared feature set and the fused intention feature set to obtain at least four probabilities and an intention probability. The target fusion layer fuses all probabilities to obtain a total probability.

[0227] A model training module 400 trains a product recommendation pre-trained model based on a feature set to obtain a product recommendation model;

[0228] The result recommendation module 500 is used for the product recommendation model to perform inference and prediction on target related data to obtain the total probability and then obtain the financial product recommendation result.

[0229] Various changes and modifications can be made without departing from the spirit and scope of the present invention, and all equivalent technical solutions also belong to the scope of the present invention.

[0230] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0231] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, devices 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 code.

[0232] The present invention is described with reference to the flowcharts and / or block diagrams of the method, terminal device (system), and computer program product according to the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0233] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0234] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0235] It should be noted that:

[0236] The "one embodiment" or "embodiment" mentioned in the specification means that the specific features, structures or characteristics described in conjunction with the embodiment are included in at least one embodiment of the present invention. Therefore, the phrases "one embodiment" or "embodiment" appearing in various places throughout the specification do not necessarily refer to the same embodiment.

[0237] In addition, it should be noted that the shapes and names of the parts and components of the specific embodiments described in this specification may be different. Any equivalent or simple changes made based on the structure, features and principles described in the patent concept of the present invention are included in the protection scope of the patent of the present invention. The technicians in the technical field of the present invention can make various modifications or supplements to the specific embodiments described or replace them in a similar manner, as long as they do not deviate from the structure of the present invention or exceed the scope defined by the claims, they should all fall within the protection scope of the present invention.

Claims

1. A recommendation method based on multi-objective and multi-task, characterized in that: The following steps are involved: Acquire relevant data to form a data set, and preprocess the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data; Forming a feature set based on user-related data and financial management-related data, the feature set including a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set; Constructing a product recommendation pre-training model, the product recommendation pre-training model includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer, the input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set, the shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, the subtask layer processes the shared feature set and the fused intention feature set respectively to obtain at least four probabilities and an intention probability, and the target fusion layer fuses all probabilities to obtain a total probability; The product recommendation pre-training model is trained based on the feature set to obtain a product recommendation model; The product recommendation model performs inference and prediction on target-related data to obtain the total probability and then obtain the financial product recommendation result; A user feature set is obtained based on user-related data, a financial management-related feature set is obtained based on financial management-related data, and a cross-feature set is obtained based on the user feature set and financial management-related data, wherein the financial management-related data includes click behavior labels, subscription button behavior labels, risk confirmation letter signing labels, and subscription behavior labels; The step of obtaining a cross-feature set based on the user feature set and financial management related data comprises the following steps: The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, and user financial management preference characteristics; Based on user characteristics, the user's preference for financial management related data is obtained, and then the preference score is obtained; Determine the cross-features of user characteristics and financial management related data based on the preference score; The preference score is obtained in the following way: in, Indicates user Financial management related data The preference score of Indicates the adjustment item. Indicates Weight coefficient, A function mapping representing user features, Function mapping representing financial management related data, Indicates the number of user features; Wherein, the sequence characteristics are obtained by the following method: The financial management related data also includes sequence data, the sequence data is converted through vector representation to obtain corresponding vector features, and the vector features are pooled and aggregated to form sequence features, wherein the sequence data is user sequence data composed of the user's subscription of financial management products within a period of time; Among them, the four probabilities include a first probability, a second probability, a third probability and a fourth probability, and the target fusion layer fuses all the probabilities to obtain a total probability, including the following steps: Fusing the first probability and the second probability to obtain a first fused probability; The first fusion probability and the third probability are fused to obtain a second fusion probability; The second fusion probability, the fourth probability and the intention probability are fused to obtain a total probability; The first fusion probability is expressed as follows: The second fusion probability is expressed as follows: The total probability is expressed as follows: The total probability is expressed as follows: in, represents the first fusion probability, represents the second fusion probability, represents the total probability, represents the first probability, represents the second probability, represents the third probability, represents the fourth probability, Indicates the probability of intention.

2. The multi-objective and multi-task recommendation method according to claim 1, characterized in that: The intention feature set includes at least personal asset characteristics, risk tolerance, investment preferences and investment experience, including the following process: Obtain user behavior data for historical periods, including at least browsing behavior data, click behavior data, add behavior data, favorite behavior data, and purchase behavior data; Sample user behavior data to obtain intention behavior data; Analyze the intention behavior data to determine whether the user triggers the purchase action and analyze and count the number of times the user triggers the purchase action and the time of the purchase action; Based on the number of times the user triggers the purchase action and the time of the purchase action, determine whether the user has the intention to manage the financial management. If the user views or understands the detailed information multiple times, it means that the user has the intention; Build a user profile and obtain a set of user-related intention features.

3. The multi-objective and multi-task recommendation method according to claim 2, characterized in that: The step of sampling the user behavior data to obtain the intended behavior data includes the following steps: Determine whether there is abnormal data in user behavior data; If it exists, the abnormal data is removed; if it does not exist, the sampling estimation value of the user behavior data is calculated based on the sampling estimation model, and the user behavior data with the sampling estimation value greater than the preset threshold is selected as the qualified intention behavior data; The sampling estimation model is expressed as follows: in, represents the sample estimate, Indicates The sampling probability of user behavior data is Indicates the total amount of user behavior data. , , Indicates the time interval between user behavior data generation. Indicates the interval time, Indicates the average number of times user behavior data is generated per unit time; is the base of natural logarithms; Indicates the interval time User behavior data; is the weight of the i-th user behavior data.

4. The multi-objective and multi-task recommendation method according to claim 1, characterized in that: The shared feature structure layer processes the user feature set, the financial management related feature set, the cross feature set, the sequence feature set and the intention feature set respectively to obtain a shared feature set and a fused intention feature set, including the following steps: The user feature set, the related feature set, the cross feature set and the sequence feature set are concatenated to obtain a shared feature set; Fusing the shared feature set with the intention feature set to obtain a fused intention feature set; The shared feature set is expressed as follows: The fusion intention feature set is expressed as follows: in, represents the user feature set, represents the set of financial management related features, represents the cross feature set, represents a sequence feature set, Indicates the intention feature set, represents the shared feature set, represents the fusion intention feature set, The vector representation of user features, A vector representation of financial management related features, represents the vector representation of the cross feature, Represents sequence characteristics, The vector representation of the sequence features, A vector representation of the intent features.

5. The multi-objective and multi-task recommendation method according to claim 1, characterized in that: The following steps are also included: Construct a loss function to optimize the training process; The loss function is expressed as follows: in, , , , , and Represent the true labels of different target behaviors, , , , , and Represent the predicted probabilities of different target behaviors, Represents the number of samples of different target behaviors.

6. The multi-objective and multi-task recommendation method according to claim 5, characterized in that: The following steps are also included: Comparing the first likelihood ratio of the first fusion probability with the first preset threshold, and if the first likelihood ratio is less than the first preset threshold, adjusting the first cost function until the first likelihood ratio is greater than the first preset threshold; Comparing the second likelihood ratio of the second fusion probability with the second preset threshold, and if the second likelihood ratio is less than the second preset threshold, adjusting the second cost function until the second likelihood ratio is greater than the second preset threshold; Compare the likelihood ratio of the intention probability with a preset threshold, and if it is less than the preset threshold, adjust the cost function until the likelihood ratio is greater than the preset threshold; The likelihood ratio is expressed as follows: The cost function is expressed as follows: in, represents minimizing the overall expected value, Indicates that assuming there is actually a target Make a decision when true The price, In the assumption Make decisions when true The conditional probability of Indicates that it is assumed that there is actually a target The default probability, represents the i-th likelihood ratio, Indicates decision.

7. A recommendation system based on multi-objective and multi-task, characterized in that: include: A data acquisition module, used to acquire relevant data to form a data set, and pre-process the data set to obtain a processed data set, wherein the relevant data includes user-related data and financial management-related data; A feature extraction module, which forms a feature set based on user-related data and financial management-related data, wherein the feature set includes a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set, and an intention feature set; A model building module is used to build a product recommendation pre-training model, which includes an input layer, a shared feature structure layer, a subtask layer and a target fusion layer. The input layer receives a user feature set, a financial management-related feature set, a cross feature set, a sequence feature set and an intention feature set. The shared feature structure layer processes the user feature set, the financial management-related feature set, the cross feature set, the sequence feature set and the intention feature set to obtain a shared feature set and a fused intention feature set. The subtask layer processes the shared feature set and the fused intention feature set to obtain at least four probabilities and an intention probability. The target fusion layer fuses all probabilities to obtain a total probability. Model training module, which trains the product recommendation pre-trained model based on the feature set to obtain the product recommendation model; A result recommendation module, which is used for the product recommendation model to perform inference and prediction on target-related data, obtain the total probability and then obtain the financial product recommendation result; A user feature set is obtained based on user-related data, a financial management-related feature set is obtained based on financial management-related data, and a cross-feature set is obtained based on the user feature set and financial management-related data, wherein the financial management-related data includes click behavior labels, subscription button behavior labels, risk confirmation letter signing labels, and subscription behavior labels; The step of obtaining a cross-feature set based on the user feature set and financial management related data comprises the following steps: The user characteristics include at least one or more of user attribute characteristics, user device characteristics, user activity characteristics, and user financial management preference characteristics; Based on user characteristics, the user's preference for financial management related data is obtained, and then the preference score is obtained; Determine the cross-features of user characteristics and financial management related data based on the preference score; The preference score is obtained in the following way: in, Indicates user Financial management related data The preference score of Indicates the adjustment item. Indicates Weight coefficient, A function mapping representing user features, Function mapping representing financial management related data, Indicates the number of user features; Wherein, the sequence characteristics are obtained by the following method: The financial management related data also includes sequence data, the sequence data is converted through vector representation to obtain corresponding vector features, and the vector features are pooled and aggregated to form sequence features, wherein the sequence data is user sequence data composed of the user's subscription of financial management products within a period of time; Among them, the four probabilities include a first probability, a second probability, a third probability and a fourth probability, and the target fusion layer fuses all the probabilities to obtain a total probability, including the following steps: Fusing the first probability and the second probability to obtain a first fused probability; The first fusion probability and the third probability are fused to obtain a second fusion probability; The second fusion probability, the fourth probability and the intention probability are fused to obtain a total probability; The first fusion probability is expressed as follows: The second fusion probability is expressed as follows: The total probability is expressed as follows: The total probability is expressed as follows: in, represents the first fusion probability, represents the second fusion probability, represents the total probability, represents the first probability, represents the second probability, represents the third probability, represents the fourth probability, Indicates the probability of intention.

8. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. A recommendation device based on multiple objectives and multiple tasks, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the method according to any one of claims 1 to 6 is implemented.

Images