An intelligent recommendation method, device and equipment of a consultant service and a medium

Abstract

The application provides an intelligent recommendation method, device and equipment for a consultant service and a medium, and relates to the technical field of data processing.The method comprises the following steps: obtaining a historical sample data set with a time stamp, the historical sample data set comprising a plurality of samples, each sample being composed of a plurality of features, and the historical sample data set containing a time decay weight calculated for each sample; constructing a dynamically constrained causal structure diagram; constructing a sparse deep neural network model with cross-layer jump connection according to the dynamically constrained causal structure diagram; calculating and outputting a target response feature according to a real-time target input by using the trained sparse deep neural network model; and selecting a plurality of candidate products corresponding to candidate features with the highest similarity as a recommendation result based on the similarity between the target response feature and candidate features of candidate products in a candidate pool.The application can solve the problem of hysteresis caused by static modeling in a traditional recommendation system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, and in particular to an intelligent recommendation method and device for a consultant service, an equipment and a medium. BACKGROUND

[0002] Traditional consultant service recommendation systems usually rely on historical transaction records, user portrait information and static risk preference data to establish a fixed structure model mapping relationship through regression models, clustering algorithms or classifiers. Such models extract investment behavior patterns using known samples in the training phase, and use the learned strategies for subsequent investment recommendations for users. The core logic lies in fitting past behavior to infer future possible investment choices.

[0003] However, such systems are essentially static and lack the ability to perceive market fluctuations, user asset changes or temporary preference adjustments. Due to the lack of real-time data-driven feedback mechanisms, model structures and parameters cannot be updated according to the dynamic evolution of the investment environment, resulting in a lag between the recommended results and the actual market, making it difficult to meet the higher requirements of users for personalization, dynamization and scene adaptation in the current high-frequency changing environment, thus limiting the practicality and investment yield matching degree. Therefore, a method is needed to solve the problem of lag caused by static modeling in traditional recommendation systems. SUMMARY

[0004] The present application provides an intelligent recommendation method, device, equipment and medium for a consultant service, which can solve the problem of lag caused by static modeling in traditional recommendation systems.

[0005] In a first aspect of the present application, an intelligent recommendation method for a consultant service is provided, the method comprising:

[0006] Obtaining a historical sample data set with a time stamp, the historical sample data set comprising a plurality of samples, each sample consisting of a plurality of features, and the historical sample data set containing a time decay weight calculated for each sample;

[0007] Building a dynamically constrained causal structure graph, wherein undirected edges are added between different types of nodes contained in the causal structure graph, and no edges are built between nodes of the same type contained in the causal structure graph, and the features correspond one-to-one to the nodes;

[0008] Building a sparse deep neural network model with cross-layer skip connections according to the dynamically constrained causal structure graph;

[0009] Using the trained sparse deep neural network model to calculate and output target response features according to real-time target inputs;

[0010] screening candidate products corresponding to several candidate features with the highest similarity as the recommendation result based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool.

[0011] Based on the above technical solutions, preferably, after the candidate products corresponding to several candidate features with the highest similarity are screened as the recommendation result based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool, the method further comprises:

[0012] For each recommended product included in the recommendation result, a corresponding candidate feature vector is obtained from the sparse deep neural network model, the candidate feature vector being a multi-dimensional vector, each dimension corresponding to a candidate feature;

[0013] For each dimension of the candidate feature, a dominant input feature with the largest causal effect on the target response feature is identified according to a column vector in the causal effect matrix corresponding to the dimension of the target response feature, the dominant input feature being an input feature corresponding to the maximum value in the column of the environmental driving feature or the behavior intermediary feature in the causal effect matrix;

[0014] According to the type and causal effect value of the dominant input feature, a causal chain logical path is constructed, and if the dominant input feature belongs to the environmental driving feature, a causal chain path of the environmental driving feature pointing to the target response feature is generated;

[0015] If the dominant input feature belongs to the behavior intermediary feature, a causal chain path of the behavior intermediary feature pointing to the target response feature is generated, the causal chain path including the dominant input feature, the target response feature, and the causal effect value between the dominant input feature and the target response feature.

[0016] Based on the above technical solutions, preferably, the sparse deep neural network model with cross-layer jump connection constructed according to the dynamic constraint causal structure diagram specifically comprises:

[0017] According to the directed edge relationship in the causal structure diagram, a structure mask matrix is constructed, which is used to limit the connection legality in the sparse deep neural network model, wherein the structure mask matrix includes a connection mask of the environmental driving feature to the behavior intermediary feature, a connection mask of the behavior intermediary feature to the target response feature, and a jump connection mask of the environmental driving feature to the target response feature, and the element value corresponding to the connection position not allowed in the structure mask matrix is fixed as zero;

[0018] mapping the nodes in the causal structure diagram as the structure of the sparse deep neural network model, wherein the nodes corresponding to the environmental driving features are mapped as a first input layer of the sparse deep neural network model, the nodes corresponding to the behavior intermediary features are mapped as a second input layer of the sparse deep neural network model, and the nodes corresponding to the target response features are mapped as an output layer of the neural network model;

[0019] initializing a connection weight matrix and a bias parameter item, including a first connection weight matrix and a first bias item of the environmental driving features to the behavior intermediary features, a second connection weight matrix and a second bias item of the behavior intermediary features to the behavior intermediary features, a third connection weight matrix and a third bias item of the behavior intermediary features to the target response features, and a skip connection weight matrix and a fourth bias item of the environmental driving features to the target response features. For positions in the first connection weight matrix, the third connection weight matrix and the skip connection weight matrix that are not allowed by the structure mask matrix, the corresponding connection weight values are fixed to zero according to the zero value positions of the structure mask matrix after the training is completed, and remain unchanged in the subsequent use stage.

[0020] in the forward propagation process of the sparse deep neural network model, according to the first input layer, the first connection weight matrix and the first bias item, and the second input layer, the second connection weight matrix and the second bias item, an intermediate representation is calculated through an activation function to obtain a first intermediate representation, and the first intermediate representation is a hidden activation vector of the second input layer;

[0021] According to the first intermediate representation and the environmental driving feature skip connection acting on the output layer, relying on the third connection weight matrix and the third bias item and the skip connection weight matrix and the fourth bias item, an output result is generated.

[0022] On the basis of the above technical solutions, preferably, the sparse deep neural network model trained is used to calculate and output a target feature vector according to a real-time acquired target input, specifically including:

[0023] receiving a real-time acquired target input, the target input including a group of real-time environmental driving features and a group of real-time behavior intermediary features, the real-time environmental driving features being input to the first input layer, the real-time behavior intermediary features being input to the second input layer, and the arrangement order and numerical dimension of the real-time environmental driving features and the real-time behavior intermediary features being consistent with those in the training stage;

[0024] matrix operation of the real-time environment driving feature, the first connection weight matrix and the first bias term, to obtain a first operation result, and matrix operation of the real-time behavior intermediary feature, a second connection weight matrix and a second bias term to obtain a second operation result, addition of the first operation result and the second operation result, and calculation of an intermediate representation via an activation function to obtain a second intermediate representation, the second intermediate representation being an implicit activation vector of a second input layer;

[0025] matrix multiplication calculation of the second intermediate representation and the third connection weight matrix to obtain a first product vector;

[0026] matrix multiplication calculation of the real-time environment driving feature and the skip connection weight matrix to obtain a second product vector;

[0027] addition of the first product vector and the second product vector, and addition of a target response feature bias vector, and mapping of a result via an activation function to obtain an output vector of the sparse deep neural network model, the output vector being a target feature vector.

[0028] On the basis of the above technical solutions, preferably, the constructing a dynamically constrained causal structure diagram specifically comprises:

[0029] dividing all features in the historical sample data set into three feature types according to business attributes, the three feature types being an environment driving feature, a behavior intermediary feature and a target response feature, and mapping each feature according to a division result as a node, the feature and the node corresponding one-to-one;

[0030] initializing the causal structure diagram, adding an undirected edge between two nodes belonging to different types in the causal structure diagram, and not constructing any edge for two nodes belonging to the same feature type, wherein the added undirected edge includes an edge between an environment driving feature corresponding node and a behavior intermediary feature corresponding node, an edge between a behavior intermediary feature corresponding node and a target response feature corresponding node, and an edge between an environment driving feature corresponding node and a target response feature corresponding node.

[0031] On the basis of the above technical solutions, preferably, the constructing a dynamically constrained causal structure diagram specifically further comprises:

[0032] For the target feature, conditional independence test is performed, specifically including: constructing a candidate conditional feature set and enumerating all non-empty subsets, for each non-empty subset as a conditional feature set, based on the time decay weight of each feature in the sample, the significance probability value corresponding to the weighted residual correlation of the target feature and the contrast feature under the conditional feature set is calculated, wherein the target feature is any one of the plurality of features, the contrast feature is a feature of a different type from the target feature in the plurality of features, when there is any one conditional feature set that makes the significance probability value exceed a preset significance threshold, the undirected edge between the target feature and the contrast feature is deleted in the causal structure graph;

[0033] On the causal structure graph obtained after the conditional independence test processing, all triplets that simultaneously satisfy the following structural characteristics are identified: for each intermediate feature node, if there is no connection edge between the two adjacent feature nodes, the directions of the two adjacent edges are both directed to the intermediate feature node, forming a unidirectional causal structure;

[0034] Direction reasoning is performed on the remaining undirected edges, and the direction determination of all the undirected edges is iteratively completed;

[0035] The unidirectional path constraint between feature types is applied to the constructed directed graph, the unidirectional path constraint includes: retaining the unidirectional edges of the environmental driving feature corresponding node pointing to the behavior intermediate feature corresponding node, the behavior intermediate feature corresponding node pointing to the target response feature corresponding node, and the environmental driving feature corresponding node pointing to the target response feature corresponding node, and eliminating all edges that do not satisfy the unidirectional path constraint, to generate a dynamically constrained causal structure graph.

[0036] On the basis of the above technical solutions, preferably, the method further includes:

[0037] The plurality of features are divided into environmental driving features, behavior intermediate features, and target response features according to business attributes, wherein the environmental driving features include market features, the behavior intermediate features include customer features, and the target response features include product features.

[0038] In a second aspect of the present application, an intelligent recommendation device for a consultant service is provided, which is used to execute any one of the intelligent recommendation methods for a consultant service described above, and the device includes an acquisition module, a processing module, and an output module, wherein:

[0039] The acquisition module is used to acquire a historical sample data set with a time stamp, the historical sample data set includes a plurality of samples, each sample is composed of a plurality of features, and the historical sample data set contains a time decay weight calculated for each sample;

[0040] The processing module is configured to construct a dynamic constraint causal structure graph, wherein a non-directed edge is added between different types of nodes in the causal structure graph, and no edge is constructed between nodes of the same type in the causal structure graph, and the features correspond one-to-one to the nodes;

[0041] The processing module is configured to construct a sparse deep neural network model with cross-layer jump connections according to the dynamic constraint causal structure graph.

[0042] The processing module is configured to calculate and output target response features according to real-time target inputs by using the trained sparse deep neural network model.

[0043] The output module is configured to filter out candidate products corresponding to candidate features with the highest similarity as the recommendation results based on the similarity between the target response features and the candidate features of the candidate products in a candidate pool.

[0044] In a third aspect of the present application, an electronic device is provided, which includes a processor, a memory, a user interface, and a network interface, the memory is configured to store instructions, the user interface and the network interface are configured to communicate with other devices, and the processor is configured to execute the instructions stored in the memory to enable the electronic device to perform the method according to any one of the preceding aspects.

[0045] In a fourth aspect of the present application, a computer readable storage medium is provided, which stores instructions, and when the instructions are executed, the method according to any one of the preceding aspects is performed.

[0046] In summary, the one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages:

[0047] 1. In the embodiments of the present application, a dynamic constraint causal structure graph is constructed, and a sparse deep neural network model with cross-layer jump connections is constructed according to the causal structure graph, so that the recommendation model has a structural modeling capability for environmental driving features and behavior intermediary features, and the causal influence path of samples evolving over time is dynamically reflected by combining time stamps and time decay weights, target response features are generated under real-time input, and product recommendation is performed according to the target response features, so that the model structure and parameters can be adaptively adjusted according to the real-time changes of market states and user features, effectively overcoming the limitations of traditional recommendation systems, such as dependence on static modeling, response lag, lack of timeliness and causal explanation capability, and significantly improving the dynamic adaptability and real-time accuracy of recommendation.

[0048] 2. By introducing the causal effect matrix to explain the modeling of the recommendation result, the causal chain path based on the dominant input feature can be generated, so as to provide clear causal reasoning basis for each candidate product, enhance the transparency and explainability of the recommendation result, and help to improve the trust and adoption rate of the user to the recommended content.

[0049] 3. By constructing the structure mask matrix through the causal structure diagram and mapping it to the network structure, the legality of the model connection is effectively limited, and the initialization strategy is combined to ensure that the unallowed connection is always invalid. The sparse structure is used to enhance the causal prior injection, improve the convergence efficiency and generalization ability of the model parameter learning, and avoid the overfitting problem caused by the redundant connection.

[0050] 4. By hierarchical injection and structured calculation of real-time target input, the consistency of input in the training and reasoning stages is maintained, the expression ability of the information channel is enhanced by combining the skip connection, and efficient target response feature calculation is realized under the cooperation of multiple source inputs, thereby improving the real-time response ability and output accuracy of the recommendation model in the dynamic environment.

[0051] 5. By dividing the historical sample features into types and mapping them to the nodes of the causal structure diagram, the initial structure skeleton is constructed by combining the connection constraints between feature types, the possible causal relationship between variables is clearly limited, the invalid hypothesis space expansion in the structure learning process is avoided, and the accuracy and efficiency of causal discovery are improved.

[0052] 6. By conditional independence test to weaken the pseudo-correlation edge, combined with the ternary structure identification and direction reasoning rule to gradually determine the causal edge direction, and finally introducing the business constraint to filter the illogical path, a dynamic causal graph with reasonable structure and constraint compliance is obtained, which not only guarantees the accuracy of causal identification, but also meets the specific semantic needs of the investment scenario, and provides strong structure prior support for subsequent sparse neural network modeling. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 is a flowchart of an intelligent recommendation method for an investment service disclosed by an embodiment of the present application;

[0054] Figure 2 is a module schematic diagram of an intelligent recommendation device for an investment service disclosed by an embodiment of the present application;

[0055] Figure 3 is a structural schematic diagram of an electronic device disclosed by an embodiment of the present application.

[0056] Explanation of reference signs: 201, acquisition module; 202, processing module; 203, output module; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. DETAILED DESCRIPTION

[0057] In order for those skilled in the art to better understand the technical solutions in the specification, the technical solutions in the specification will be clearly and completely described below in combination with the drawings in the specification. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments.

[0058] In the description of the embodiments of the present application, the words such as "for example" or "for instance" are used to represent an example, illustration or description. Any embodiment or design scheme described as "for example" or "for instance" in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the words such as "for example" or "for instance" are intended to present the relevant concept in a specific manner.

[0059] In the description of the embodiments of the present application, the term "a plurality of" means two or more. For example, a plurality of systems means two or more systems, and a plurality of screen terminals means two or more screen terminals. In addition, the terms "first" and "second" are used only for the purpose of description, and should not be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Therefore, the features defined with "first" and "second" can explicitly or implicitly include one or more of the features. The terms "include", "contain", "have" and their variants mean "include but are not limited to", unless otherwise specifically emphasized.

[0060] The traditional investment advisor recommendation system relies on static risk preference, historical transactions and user portrait fixed data, and builds a static mapping model through regression, clustering or classification method. Although the historical behavior patterns can be reused for investment recommendation, due to the lack of real-time perception ability of market volatility, user asset changes and temporary preference adjustment, and the model structure cannot be updated adaptively with the dynamic evolution of the environment, the recommendation result is lagging and the response is slow, which is difficult to meet the current actual needs of high-frequency changes, personalized decision-making and scene adaptation ability, and there is obvious deficiency in practicability and investment matching efficiency. Therefore, it is urgent to introduce a dynamic mechanism to solve the lag problem caused by static modeling.

[0061] The embodiment discloses an intelligent recommendation method of investment advisor service, referring to Figure 1 , comprising the following steps S110-S150:

[0062] S110, acquiring a historical sample data set with a time stamp.

[0063] The intelligent recommendation method for a consultant service disclosed by the embodiments of the present application is applied to a server. The server includes but is not limited to electronic devices such as a mobile phone, a tablet computer, a wearable device, a PC (Personal Computer), and the like, and can also be a background server running an intelligent recommendation method for a consultant service. The server can be implemented by an independent server or a server cluster composed of multiple servers.

[0064] First, a historical sample data set in a time range is collected from a consultant system. The historical sample data set is composed of multiple samples. Each sample contains three types of feature variables: environment driving features, behavior intermediary features, and target response features. The environment driving features are preferably market features, the behavior intermediary features are preferably customer features, and the target response features are preferably product features. Each sample is attached with a unique time stamp for identifying a recording time point of the sample.

[0065] For each sample, a time decay weight is calculated. An exponential decay function is used to represent that the importance of a sample to causal structure learning decreases with the time distance. The specific formula is as follows:

[0066]

[0067] wherein, the time decay weight of the t th sample is denoted as w t, the sample initial weight is denoted as w 0, which is a constant and is usually set to 1, λ represents a time decay coefficient, Δt represents a time difference between the current system time and the time stamp of the t th sample.

[0068] The calculated time decay weight will be used as an additional attribute of each sample and participate in the weighted calculation of the conditional independence test in the subsequent causal structure graph construction process, including the weighted correlation coefficient, the weighted covariance, the weighted mean, and the sample weighting in the Fisher Z test.

[0069] The time decay weight is used to give samples in different time windows decreasing importance when processing historical sample data with time stamps, so as to realize the sensitivity control of the latest data change of the model and the strengthening of the time sequence relationship in causal reasoning. Specifically, the time decay weight is used as a weighting factor in the conditional independence test and residual correlation calculation, so that recent samples have a greater weight in evaluating the causal dependence relationship between features, which helps to capture dynamic time-varying rules, improve the dynamic adaptability of the causal structure graph, and improve the response accuracy of the recommendation model to the current input, thereby effectively alleviating the problem of ignoring time evolution in traditional static modeling.

[0070] In S120, a dynamic constraint causal structure graph is constructed.

[0071] In a possible implementation, the dynamic constraint causal structure graph is constructed, specifically comprising:

[0072] All features in the historical sample data set are divided into three types of feature types according to business attributes, which are environment driven features, behavior intermediary features and target response features. Each feature is mapped to a node according to the division result, and the feature and the node correspond one by one. The causal structure graph is initialized. In the causal structure graph, an undirected edge is added between two nodes belonging to different types. No edge is constructed between two nodes belonging to the same feature type. The added undirected edge includes the edge between the environment driven feature corresponding node and the behavior intermediary feature corresponding node, the edge between the behavior intermediary feature corresponding node and the target response feature corresponding node, and the edge between the environment driven feature corresponding node and the target response feature corresponding node.

[0073] Specifically, when implementing the technical solution of constructing the dynamic constraint causal structure graph, first, the historical sample data set needs to be divided into feature types. The historical sample data set contains multiple samples, and each sample is composed of several feature variables. According to business semantics and causal logic, all features are divided into three types of feature types: environment driven features, behavior intermediary features and target response features. After the division is completed, a unique corresponding graph node is established for each feature, forming a one-to-one correspondence between features and nodes, laying a structural foundation for subsequent causal structure graph construction.

[0074] The causal structure graph is a directed graph structure constructed under the guidance of business semantics, which is used to depict the causal dependence relationship between the three types of variables: environment driven features, behavior intermediary features and target response features. First, all features in the historical sample data are divided into three types of feature types according to their business attributes, and each feature is mapped to a node in the graph, forming a one-to-one correspondence between features and nodes. Then, in the graph initialization process, only undirected edges are added between nodes of different feature types to construct potential causal paths, including edges between environment driven features and behavior intermediary features, edges between behavior intermediary features and target response features, and edges between environment driven features and target response features. No edge is constructed between nodes of the same type. Subsequently, through conditional independence test and direction reasoning rules, directed edges are derived to further improve the causal structure graph and impose one-way path constraints between feature types to ensure that the semantic direction of the causal path is reasonable. Finally, the causal structure graph is embedded into the sparse deep neural network model as a causal prior constraint, providing a rigorous basis for the connection structure, information flow and causal explanation of the model.

[0075] After completing the feature type division and generating the node set, the causal structure graph is initialized. The initialization stage needs to restrict the graph connection according to the causal action direction constraint between feature types, and clearly only allows to build connection edges between nodes of different types, and does not allow to build edges between nodes of the same type. Specifically, only the following three feature type combinations are established between the undirected edges: the establishment of undirected edges between the environment-driven feature nodes and the behavior intermediary feature nodes represents the potential influence of market state on customer behavior; the establishment of undirected edges between the behavior intermediary feature nodes and the target response feature nodes represents the causal intermediary role of customer behavior on product selection results; the establishment of undirected edges between the environment-driven feature nodes and the target response feature nodes represents the direct action path of external market factors on the final recommendation results. In this initialization graph, each edge is added in the form of an undirected edge, representing a potential causal relationship connection that has not yet been determined direction, for further identification of the true causal direction in the subsequent conditional independence test and direction reasoning stage.

[0076] The initialization graphing process ensures that the causal structure graph has the characteristics of heterogeneous node layering, path controllability, and type single-directional constraint in structure, conforms to the market-customer-product three-layer causal modeling logic in the financial investment system, and at the same time provides a complete and controlled starting structure for the next step of skeleton graph construction and collision structure identification based on weighted independence test.

[0077] In one possible implementation, the dynamic constraint causal structure graph is constructed, and specifically further includes:

[0078] For the target feature, conditional independence test is performed, specifically including: constructing a candidate condition feature set and enumerating all non-empty subsets, for each non-empty subset as a condition feature set, based on the time decay weight of each feature in the sample, the significance probability value corresponding to the weighted residual correlation of the target feature and the contrast feature under the condition feature set is calculated, wherein the target feature is any one of the plurality of features, and the contrast feature is a feature of a different type from the target feature, when there is any one condition feature set that makes the significance probability value exceed the preset significance threshold, the undirected edge between the target feature and the contrast feature in the causal structure graph is deleted; on the causal structure graph obtained after the conditional independence test, all triplets that simultaneously satisfy the following structural characteristics are identified: for each intermediate feature node, if there is no connection edge between the two adjacent feature nodes, the directions of the two adjacent edges are both directed to the intermediate feature node, forming a unidirectional causal structure; the direction reasoning is performed on the remaining undirected edges, and the direction determination of all undirected edges is iteratively completed; the unidirectional path constraint between feature types is applied to the constructed directed graph, the unidirectional path constraint includes: retaining the unidirectional edges of the environment driving feature corresponding node pointing to the behavior intermediate feature corresponding node, the behavior intermediate feature corresponding node pointing to the target response feature corresponding node, and the environment driving feature corresponding node pointing to the target response feature corresponding node, and eliminating all edges that do not satisfy the unidirectional path constraint, to generate a dynamically constrained causal structure graph.

[0079] Specifically, in the process of constructing the dynamically constrained causal structure graph, first, conditional independence test needs to be performed for each target feature. The target feature refers to any feature node fixed as a variable to be detected in the current test round, and the corresponding contrast feature must belong to a feature node of a different type from the target feature. For each initial undirected edge between the target feature and the contrast feature, before independence judgment, a candidate condition feature set of the target feature is constructed, which consists of all other nodes connected to the target feature but excluding the contrast feature. On this basis, all non-empty subsets of the candidate condition feature set are enumerated, and each subset is taken as a condition feature set to test the statistical independence of the target feature and the contrast feature under the condition feature set. When testing, the time decay weight of each sample is introduced, the weighted Pearson correlation analysis and Fisher Z transformation are performed based on the residual correlation, and the significance probability value is obtained. If there is any one condition feature set in all non-empty condition feature sets that makes the significance probability value higher than the preset threshold (for example, 0.05), it is judged that the target feature and the contrast feature are independent under the condition, and the undirected edge between the two nodes in the causal structure graph is deleted.

[0080] After the conditional independence test between all target features and contrast features is completed, a simplified skeleton graph structure is obtained, and then a collision structure recognition operation is performed on the structure. All intermediate feature nodes in the graph are traversed to determine whether the node is connected to two feature nodes that have no connection between each other. If the condition is met, it is considered that a typical collision structure exists. In this structure, the edges connecting the intermediate feature nodes are oriented to point to the intermediate feature nodes at the same time, forming a standard causal structure of "external node pointing to intermediate node", that is, a unidirectional causal structure unit is formed.

[0081] After the collision structure processing is completed, the direction reasoning of the remaining undirected edges is continued. The reasoning process relies on three types of reasoning rules: path reachability rule, that is, if node A has a directed path to node B, the undirected edge A-B can be oriented as A→B when it exists; shared source node structure rule, that is, when node A points to two nodes B and C, and there is an undirected edge between B and C, the direction can be inferred as B→C or C→B; and non-connection path reasoning rule, that is, if there is no direct connection between a node pair, but there are other constraint paths, the edge direction is determined according to logical transitivity. The above three types of rules are iteratively applied in the whole graph until all inferable edge directions are determined.

[0082] Finally, a structure pruning operation is performed on the complete directed graph according to the business path restrictions between feature types. The unidirectional path constraints between feature types are as follows: only the unidirectional edges from the environment-driven feature node to the behavior intermediate feature node, the unidirectional edges from the behavior intermediate feature node to the target response feature node, and the unidirectional edges from the environment-driven feature node to the target response feature node are reserved; all connection edges that do not meet the above path constraints are deleted. The constraint operation ensures that the causal structure graph finally constructed strictly conforms to the hierarchical modeling logic of market-driven behavior, intermediate transformation decision, and result response output, and serves as a priori input of the connection structure of the sparse deep neural network model, supporting the implementation of the product recommendation system with high credibility of structural interpretability and causal consistency.

[0083] S130, constructing a sparse deep neural network model with cross-layer jump connection according to the dynamic constraint causal structure graph.

[0084] In a possible implementation, a sparse deep neural network model of cross-layer jump connection is constructed according to a dynamic constraint causal structure diagram, and specifically includes: constructing a structure mask matrix according to a directed edge relationship in the causal structure diagram, the structure mask matrix being used to limit connection legality in the sparse deep neural network model, wherein the structure mask matrix includes a connection mask of an environmental driving feature to a behavior intermediary feature, a connection mask of the behavior intermediary feature to a target response feature, and a jump connection mask of the environmental driving feature to the target response feature, and an element value corresponding to an unallowed connection position in the structure mask matrix is fixed as zero; mapping nodes in the causal structure diagram to a structure of the sparse deep neural network model, wherein a node corresponding to the environmental driving feature is mapped to a first input layer of the sparse deep neural network model, a node corresponding to the behavior intermediary feature is mapped to a second input layer of the sparse deep neural network model, and a node corresponding to the target response feature is mapped to an output layer of the neural network model; initializing a connection weight matrix and a bias parameter term, including a first connection weight matrix and a first bias term of the environmental driving feature to the behavior intermediary feature, a second connection weight matrix and a second bias term of the behavior intermediary feature to the behavior intermediary feature, a third connection weight matrix and a third bias term of the behavior intermediary feature to the target response feature, and a jump connection weight matrix and a fourth bias term of the environmental driving feature to the target response feature, wherein for positions in the first connection weight matrix, the third connection weight matrix, and the jump connection weight matrix that are not allowed by the structure mask matrix, the corresponding connection weight values are fixed as zero according to zero position of the structure mask matrix after training is completed, and remain unchanged in a subsequent use stage; in a forward propagation process of the sparse deep neural network model, an intermediate representation is calculated by an activation function according to the first input layer, the first connection weight matrix, and the first bias term, and the second input layer, the second connection weight matrix, and the second bias term, to obtain a first intermediate representation, which is an implicit activation vector of the second input layer; the first intermediate representation and the environmental driving feature jump connection jointly act on the output layer, and generate an output result relying on the third connection weight matrix and the third bias term, and the jump connection weight matrix and the fourth bias term.

[0085] Specifically, a structure mask matrix is constructed according to the oriented edge relationship in the causal structure diagram, and the structure mask matrix is used to strictly limit the legality of inter-layer connection in the sparse deep neural network model. In the construction process, all edges of “environmental driving feature pointing to behavior intermediary feature” in the causal structure diagram are mapped to a first mask matrix , all edges of “behavior intermediary feature pointing to target response feature” are mapped to a second mask matrix , and all jump connections of “environmental driving feature pointing to target response feature” are mapped to a third mask matrix In the structure mask matrix, all the connection positions not allowed by the causal structure diagram are forced to set the corresponding element value to zero, which will be used to shield illegal paths in the model connection in the subsequent.

[0086] Next, the three types of nodes in the causal structure diagram are mapped into the hierarchical structure of the sparse deep neural network model: all the nodes corresponding to the environmental driving features are mapped into the first input layer of the model, which is used to receive the market environment feature vector; all the nodes corresponding to the behavior intermediary features are mapped into the second input layer of the model, which participates in the calculation of the intermediate representation; all the nodes corresponding to the target response features are mapped into the output layer, which finally outputs the target feature vector of the model. This structure mapping strictly follows the three-layer feature type division of the causal structure diagram, and maintains the semantic consistency of the causal driving path.

[0087] Then, the initialization operation of the connection weight matrix and the bias term is performed. The first connection weight matrix and the first bias term are set for connecting the environmental driving features to the behavior intermediary features; the second connection weight matrix and the second bias term are set for mapping the behavior intermediary features to the behavior intermediary features; the third connection weight matrix and the third bias term are set for connecting the behavior intermediary features to the target response features; the jump connection weight matrix and the fourth bias term are set to realize the direct influence of the environmental driving features on the target response features. For the positions in the first connection weight matrix, the third connection weight matrix and the jump connection weight matrix not allowed by the structure mask matrix, the corresponding connection weight values are fixed to zero according to the zero position of the structure mask matrix after the training is completed, and remain unchanged in the subsequent use stage.

[0088] In the embodiments of the present application, four groups of connection weight matrices and bias parameter terms are involved in the construction process of the sparse deep neural network model, which are respectively used to express the causal connection relationship between the environmental driving features, the behavior intermediary features and the target response features. Each group of connection weight matrix and bias term is strictly defined according to the directed edges retained in the causal structure diagram, and the legality of the connection is realized in combination with the structure mask matrix.

[0089] The first group is the connection weight matrix and the bias term from the environmental driving features to the behavior intermediary features, which are denoted as and respectively. The weight matrix is used to describe the indirect influence path of the market state on the customer behavior, which is the main connection of the model from the first input layer to the second input layer. The second group is the connection weight matrix and the bias term from the behavior intermediary features to the behavior intermediary features, which are denoted as and The weight matrix is used for feature extraction of customer behavior data, which is the feature processing of the second input layer of the model. The results of the two groups are added, and the first intermediate representation is calculated through the activation function

[0090] The corresponding forward propagation calculation formula is:

[0091]

[0092]

[0093] wherein, represents the environment-driven feature input vector, represents the intermediate representation of the behavior intermediary feature, represents a nonlinear activation function such as ReLU or tanh.

[0094] The third group is the connection weight matrix and bias term of the behavior intermediary feature to the target response feature, denoted as and The weight matrix is used to capture the conduction mechanism of customer behavior on the final recommendation result, which is one of the core causal paths in the model structure. In the forward propagation stage of the model, its calculation formula is:

[0095]

[0096] wherein, represents the influence component of the behavior intermediary feature on the target response feature, which is part of the final output.

[0097] The fourth group is the skip connection weight matrix and bias term of the environment-driven feature to the target response feature, denoted as and The skip connection represents the direct driving path of the environment-driven feature to the recommendation target, without passing through the behavior intermediary feature, constituting a cross-layer connection from the first input layer to the output layer. Its calculation expression in the forward propagation is:

[0098]

[0099] Finally, the results of the third group and the fourth group are added, and the target response feature vector is output through the activation function:

[0100]

[0101] Through the division and cooperation of the above four groups of connection weight matrices and bias terms, the sparse deep neural network model can maintain the consistency of the causal path structure while flexibly expressing the direct and indirect driving relationships between the market, behavior, and result features, thereby improving the recommendation accuracy and causal explainability.

[0102] In the forward propagation process of the sparse deep neural network model, first, the environment-driven feature input vector is linearly transformed in combination with the first connection weight matrix and the first bias term, and then the intermediate feature input vector is linearly transformed in combination with the second connection weight matrix and the second bias term for feature extraction. Then, the results of the two are added and input into the activation function (such as ReLU or tanh), to obtain the intermediate representation of the behavior-intermediate feature , and the calculation expression is:

[0103]

[0104] Based on the intermediate representation of the behavior-intermediate feature and the original input of the environment-driven feature , linear operations are respectively performed on the corresponding connection weight matrices, and the prediction result of the target response feature is generated through weighted summation and bias superposition to form the model output vector , and the calculation expression is:

[0105]

[0106] The output vector is the prediction result of the target response feature by the model, which is used as the input for subsequent product selection and causal effect evaluation. Through the constraint of the structure mask matrix, the design of the cross-layer jump connection, and the strict mapping of the hierarchical structure, the model accurately follows the causal path structure while efficiently expressing complex nonlinear causal relationships.

[0107] In one possible implementation, the sparse deep neural network model is constructed based on a dynamically constrained causal structure diagram, and a cross-layer jump connection structure is used to enhance the causal information transmission between features. The core goal of the training stage is to learn the prediction structure of the environment-driven feature and the behavior-intermediate feature on the target response feature, maximize the prediction accuracy, and improve the sparsity and generalization ability of the model while maintaining the consistency of the causal structure.

[0108] In the model structure, the intermediate representation of the behavior-intermediate feature is calculated by fusing the environment-driven feature and the original behavior-intermediate feature, and the expression is:

[0109]

[0110] wherein represents the intermediate activation vector of the behavior-intermediate feature, is the environment-driven feature input vector, is the behavior-intermediate feature input vector, , a first connection weight matrix and a second connection weight matrix, respectively, is a second bias term, is a nonlinear activation function.

[0111] The product layer output not only depends on the intermediate representation of the behavior mediator feature, but also fuses the environment-driven feature of the skip connection, and outputs the target response feature vector in the following way:

[0112]

[0113] wherein, is the output of the model, representing the predicted value of each dimension of the product feature.

[0114] In order to supervise the training of the structure, a main loss function based on mean square error is introduced , wherein the target is to minimize the distance between the predicted output and the sample label, which is defined as follows:

[0115]

[0116] At the same time, in order to strengthen the sparse structure, suppress overfitting and further ensure the consistency of the causal structure, two regular terms are introduced to form a composite regularization term . On the one hand, a penalty is imposed on the allowed connections in the structure to encourage weight sparsity; on the other hand, a regularization is imposed on the forbidden connections (defined by the mask matrix) in the structure to further suppress the activation of non-causal connection weights, and its formula is as follows:

[0117]

[0118] wherein represents the structure mask matrix, represents the element-wise product, , is a regularization hyperparameter.

[0119] The training process includes the following steps:

[0120] Step 1, preprocessing and initialization: construct the structure mask matrix , which is 1 for the allowed connections in the causal structure and 0 for the rest. Initialize all weight parameters.

[0121] Step 2, perform forward propagation: for each batch of samples , perform the above two-layer calculation process to obtain the model output .

[0122] Third step, loss function calculation: the batch loss function is composed of the sum of mean square error and two regularization terms:

[0123]

[0124] Fourth step, back propagation and parameter update: based on Perform gradient calculation, use optimizer (such as Adam) to update weights and biases, and mask correction to keep illegal connection weights to zero.

[0125] Fifth step, dynamic sparse training: every few iterations, prune and set to zero the legal connection weights in all weights that are less than the preset threshold ; At the same time, re-activate the illegal connection that is currently zero but has a large gradient to give it a small initial value to dynamically adjust the connection structure.

[0126] Sixth step, iterative training until the loss function converges or reaches the predetermined number of iterations.

[0127] Finally, in the prediction stage, the model receives new real-time environment driving features and behavior intermediary features , which are input into the corresponding input layer to generate intermediate representations:

[0128]

[0129] And further output the product prediction vector:

[0130]

[0131] This output result can be used as continuous product feature prediction, or used to match the most similar product prototype in the candidate product set to complete personalized recommendation, so as to realize accurate investment suggestion generation under causal driving.

[0132] In one possible implementation, the causal effect matrix is a structured tensor representation of the strength of the causal impact of market features and customer features on each product feature output, estimated quantitatively based on a sparse deep neural network model with cross-layer skip connections through local gradient analysis. Because the neural network model has a multi-layer nonlinear structure, traditional linear regression weights cannot directly explain the causal effect, so the chain rule is used in combination with the derivatives of the linear combination variables and activation functions of each layer in the forward propagation of the model to calculate the partial derivative value of each input feature at the local point of the current sample to reflect its instantaneous response sensitivity to the output target, thereby achieving approximate estimation of causal impact.

[0133] First, input the market feature vector and customer feature input vector As input to the model, the intermediate representation is computed by linear transformation and non-linear activation respectively. The customer feature intermediate layer activation is denoted as:

[0134]

[0135]

[0136] where, is the connection weight matrix from market features to customer features, is the self-connection weight matrix among customer features, is the customer layer bias, is the activation function. Then the customer feature intermediate representation is input to the product feature layer together with the market feature input to form the linear combination variable of the product layer:

[0137]

[0138] where, is the connection weight from customer to product, is the skip-connection weight from market to product, is the product layer bias.

[0139] Based on the above structure, the chain rule is used to compute the partial derivative of the output layer result with respect to each feature component in the input layer, to obtain the causal sensitivity estimation of the input feature to the output product feature. For each market feature , the partial causal effect on the product feature is:

[0140]

[0141] And the partial causal effect of the customer feature on the product feature is:

[0142]

[0143] The above partial derivative results are computed for each sample respectively, and are organized by feature dimension. The causal effect matrix of the market feature is defined as , and the causal effect matrix of the customer feature is , where , , are the dimension numbers of the market feature, customer feature and product feature respectively.

[0144] Finally, the two causal effect matrices are spliced along the input dimension direction to form the complete causal effect matrix , whose expression is:

[0145]

[0146] matrix Each element Indicates the first The input feature for the th _th The strength of the local causal influence of a product feature, where if Then it means the first The first market characteristic affects the second The role of each product feature; if Then it means the first The first customer characteristic is related to the second The role of each product feature.

[0147] This causal effect matrix not only supports the causal explanation of the recommendation output, but also serves as the basis for constructing the causal chain path of the recommendation results. It can be used to trace back the causal driving mechanism corresponding to the recommendation item, thereby improving the interpretability and transparency of the system.

[0148] S140 uses a trained sparse deep neural network model to calculate and output target response features based on the target input acquired in real time.

[0149] In one possible implementation, a trained sparse deep neural network model is used to calculate and output a target feature vector based on real-time acquired target input. Specifically, this includes: receiving real-time acquired target input, which includes a set of real-time environment-driven features and a set of real-time behavior-mediated features; inputting the real-time environment-driven features to a first input layer and the real-time behavior-mediated features to a second input layer; maintaining the same order and numerical dimensions as during the training phase for the real-time environment-driven features and the real-time behavior-mediated features; performing matrix operations on the real-time environment-driven features with a first connection weight matrix and a first bias term to obtain a first operation result; and simultaneously performing matrix operations on the real-time behavior-mediated features with a second connection weight matrix. The matrix operation is performed on the first and second bias terms to obtain the second operation result. The first and second operation results are added together and then the intermediate representation is calculated through the activation function to obtain the second intermediate representation, which is the hidden activation vector of the second input layer. The second intermediate representation is multiplied by the third connection weight matrix to obtain the first product vector. At the same time, the real-time environment driving features are multiplied by the skip connection weight matrix to obtain the second product vector. The first and second product vectors are added together and then added to the target response feature bias vector. The result is mapped by the activation function to obtain the output vector of the sparse deep neural network model, which is the target feature vector.

[0150] Specifically, real-time target input data is received, which includes a set of real-time environment-driven features and a set of real-time behavioral intermediary features. The environment-driven features represent variables related to the current market state, such as volatility, sector rotation strength, etc. The behavioral intermediary features represent current user profile information such as risk level, asset allocation ratio, etc. The environment-driven features are input as an input tensor to the first input layer of the model, and the behavioral intermediary features are input to the second input layer. The arrangement order and dimension setting of the two sets of input features must be completely consistent with the training stage to ensure the consistency of model structure mapping and the integrity of the computational graph.

[0151] In the forward propagation stage of the model, the real-time environment-driven feature vector is first linearly combined with the first connection weight matrix and the first bias term The behavioral intermediary feature vector is linearly combined with the second connection weight matrix and the second bias term Then the two results are added and the hidden representation of the behavioral intermediary feature, i.e. the second intermediate representation, is calculated through the activation function The calculation expression of this process is:

[0152]

[0153] where, is the intermediate activation vector of the behavioral intermediary feature, is the market feature input vector, is the customer feature input vector.

[0154] Subsequently, the intermediate representation of the behavioral intermediary feature is matrix multiplied with the third connection weight matrix to represent the influence of the behavioral intermediary feature on the target response feature, while the real-time environment-driven feature is matrix multiplied with the skip connection weight matrix to represent the direct influence of the environment-driven feature on the target response feature. The results of the two parts are added with the third bias term and the skip connection bias term and mapped through the activation function to generate the final output vector of the model, i.e. the target feature vector, whose calculation expression is:

[0155]

[0156] The output vector is a set of multi-dimensional vectors, representing the prediction results of each target response feature under the current market environment combined with the current customer portrait, which can be used for similarity calculation with the product feature vectors in the candidate product pool to achieve personalized investment recommendation. This structure explicitly preserves the action paths between different feature types in the causal structure diagram, and reflects the direct impact of market variables on response features through the jump connection mechanism, thereby improving the response sensitivity and causal interpretability of the recommendation model.

[0157] Correspondingly, in the stage of constructing the dynamic constraint causal structure diagram, first, all features in the historical sample data set are divided into three types of feature types according to their business attributes, namely environment-driven features, behavior-intermediate features and target response features, to ensure that each feature is clearly attributed to its functional role. Then, according to the division result, each feature is mapped to a node in the graph structure, maintaining a one-to-one correspondence between the feature and the node, facilitating subsequent structure constraints and graph modeling.

[0158] When initializing the graph, only undirected edges between nodes belonging to different feature types are allowed, and connections between nodes of the same type are strictly prohibited, thereby meeting the structural prior of causal modeling. The undirected edges allowed to be added include: edges between environment-driven feature corresponding nodes and behavior-intermediate feature corresponding nodes, used to express the impact of market environment on customer behavior; edges between behavior-intermediate feature corresponding nodes and target response feature corresponding nodes, used to express the regulation of user behavior on product performance; edges between environment-driven feature corresponding nodes and target response feature corresponding nodes, used to express the impact of market variables on product direct response. All edges are undirected edges at initialization, indicating that the causal direction is to be inferred, serving as the input structure basis for subsequent conditional independence test and direction reasoning. Through the above strict structure construction, a stable, compliant and interpretable causal connection prior can be provided for the model, making the recommendation model highly reliable in terms of structural consistency and semantic clarity.

[0159] S150, based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool, filtering out the candidate products corresponding to the candidate features with the highest similarity as the recommendation results.

[0160] Prepare a set of structure-standardized candidate product feature pool. Each candidate feature in the candidate product feature pool is a labeled product response feature vector. The candidate response feature vector and the target response feature vector have the same dimension, the same feature arrangement order and consistent normalization scale, thereby ensuring the numerical comparability and semantic consistency of subsequent similarity calculation.

[0161] First, based on the cross-layer jump connection of the sparse deep neural network model completed training, the real-time environment driving features and behavior intermediary features are calculated forwardly, and the target response feature vector corresponding to the current user is output. The target response feature vector represents the product feature portrait preferred by the current user under the current market conditions, and is the basic input for the subsequent screening stage.

[0162] In the similarity calculation stage, all candidate response feature vectors in the candidate product feature pool are traversed, and similarity evaluation is performed one by one with the target response feature vector. The similarity measure can be selected in the form of cosine similarity, Euclidean distance, Mahalanobis distance, etc. Taking cosine similarity as an example, the calculation formula is:

[0163]

[0164] Wherein, is the target response feature vector, is the candidate response feature vector of the th candidate product, and the symbol represents vector dot product, represents vector norm.

[0165] According to the similarity calculation results between all candidate response feature vectors and the target response feature vector, the candidate products are sorted, and several candidate products with the highest similarity value are selected and output as the final recommendation result. Each candidate product in the recommendation result is supported by the high matching between its candidate response feature vector and the target response feature vector, and can accurately reflect the product preference of the current user under the current environmental conditions, thereby realizing the real-time, personalized and causal consistency of the investment recommendation service. The whole process maintains the full-dimensional modeling and structural rigor of the response feature, ensuring the high unity of the recommendation mechanism and the causal modeling process.

[0166] In a possible implementation, after screening out candidate products corresponding to a plurality of candidate feature pairs with the highest similarity as the recommendation result based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool, the method further includes: obtaining, for each recommended product included in the recommendation result, a corresponding candidate feature vector from the sparse deep neural network model, the candidate feature vector being a multi-dimensional vector, each dimension corresponding to a candidate feature; identifying, for each dimensional candidate feature, a dominant input feature having the largest causal effect on the target response feature according to a column vector corresponding to the dimension of the target response feature in the causal effect matrix, the dominant input feature being an input feature corresponding to the maximum value in the column where the environmental driving feature or the behavioral intermediary feature is located in the causal effect matrix; and constructing a causal chain logical path according to the type and the causal effect value of the dominant input feature, generating an environmental driving feature pointing to the target response feature causal chain path if the dominant input feature belongs to the environmental driving feature, or generating a behavioral intermediary feature pointing to the target response feature causal chain path if the dominant input feature belongs to the behavioral intermediary feature, the causal chain path including the dominant input feature, the target response feature, and the causal effect value between the dominant input feature and the target response feature.

[0167] Specifically, for each selected candidate product in the recommendation result, a candidate response feature vector output by the sparse deep neural network model is extracted. The candidate response feature vector is a multi-dimensional vector, the number of dimensions being equal to the number of product layer target response features, and each dimension representing the recommended expression strength of the candidate product on a product feature indicator, which is the basis for subsequent causal path reasoning.

[0168] After obtaining the candidate response feature vector, each dimensional response feature of the candidate response feature vector is processed. For the j-th dimensional candidate response feature, the j-th column vector is extracted from the pre-generated causal effect matrix, representing the local causal effect strength of all input features on the candidate response feature. The causal effect matrix is a structural derivative product of the sparse deep neural network model, including the partial derivative calculation results of all environmental driving features and behavioral intermediary features on each target response feature, and can reflect the structural causal transmission between features in numerical form.

[0169] In each column of the causal effect matrix, the input feature index corresponding to the maximum value is located, which indicates that the input feature is the dominant input feature in the current product feature dimension. The dominant input feature can come from the environmental driving feature set or the behavioral intermediary feature set, and the type of the dominant input feature is determined by the position of the dominant input feature in the overall input feature index.

[0170] Subsequently, a causal chain logical path is constructed according to the type of the dominant input feature. If the dominant input feature belongs to the set of environment-driven features, the path structure is constructed as “environment-driven feature → target response feature”, indicating that the change in market conditions dominates the fluctuation of the product feature through the feature; if the dominant input feature belongs to the set of behavior-mediated features, the path structure is constructed as “behavior-mediated feature → target response feature”, indicating that the customer behavior preference directly affects the inferred output of the product feature. Each causal chain path contains three elements: the name of the dominant input feature, the name of the target response feature, and the causal effect value between them. The structure can be rendered as an explanatory template sentence or a diagram result, realizing the interpretable presentation of the recommendation result.

[0171] Through the above process, each dimension of the target response feature can trace back to its most critical causal source feature, and a transparent causal path from input to output is established, providing mechanism visibility and decision traceability for user recommendation results in the advisory service, thereby improving the trustworthiness and practicality of the financial intelligent system.

[0172] The embodiment also discloses an intelligent recommendation device for an advisory service, referring to Figure 2 , comprising an acquisition module 201, a processing module 202, and an output module 203, and the device is used for executing any one of the above intelligent recommendation methods for an advisory service, wherein:

[0173] The acquisition module 201 is used for acquiring a historical sample data set with a time stamp, the historical sample data set includes a plurality of samples, each sample is composed of a plurality of features, and the historical sample data set contains a time decay weight calculated for each sample.

[0174] The processing module 202 is used for constructing a dynamically constrained causal structure diagram, wherein a non-directed edge is added between different types of nodes contained in the causal structure diagram, no edge is constructed between the same types of nodes contained in the causal structure diagram, and the features and the nodes are in one-to-one correspondence.

[0175] The processing module 202 is used for constructing a sparse deep neural network model with cross-layer jump connection according to the dynamically constrained causal structure diagram.

[0176] The processing module 202 is used for calculating and outputting the target response feature according to the real-time acquired target input by using the trained sparse deep neural network model.

[0177] The output module 203 is used for filtering a plurality of candidate products corresponding to the candidate features with the highest similarity as the recommendation result based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool.

[0178] It should be noted that the device provided in the above embodiment is only exemplified by the above division of functional modules when realizing its functions, and in actual application, the above functions can be completed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to complete all or part of the above described functions. In addition, the device and method embodiments provided in the above embodiments belong to the same concept, and the specific implementation process is detailed in the method embodiments, which will not be repeated here.

[0179] The embodiment also discloses an electronic device, which refers to Figure 3 The electronic device can include at least one processor 301, at least one communication bus 302, a user interface 303, a network interface 304, and at least one memory 305.

[0180] The communication bus 302 is used to realize the connection and communication between the components.

[0181] The user interface 303 can include a display screen (Display) and a camera (Camera), and the optional user interface 303 can also include a standard wired interface and a wireless interface.

[0182] The network interface 304 can optionally include a standard wired interface and a wireless interface (such as a WI-FI interface).

[0183] The processor 301 can include one or more processing cores. The processor 301 connects various parts of the server through various interfaces and lines, executes various functions of the server and processes data by running or executing instructions, programs, code sets or instruction sets stored in the memory 305, and calling data stored in the memory 305. Optionally, the processor 301 can be realized in at least one of the hardware forms of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The processor 301 can integrate a combination of one or several of central processing units (Central Processing Unit, CPU), graphics processing units (Graphics Processing Unit, GPU), and modems. The CPU is mainly used to process operating systems, user interfaces, and application programs. The GPU is used to render and draw the content to be displayed on the display screen. The modem is used to process wireless communication. It can be understood that the above-mentioned modem can also not be integrated into the processor 301, but can be realized by a separate chip.

[0184] The memory 305 can include a Random Access Memory (RAM) and can also include a Read-Only Memory (ROM). Optionally, the memory includes a non-transitory computer-readable storage medium. The memory 305 can be used to store instructions, programs, codes, code sets, or instruction sets. The memory 305 can include a program storage area and a data storage area, where the program storage area can store instructions for implementing an operating system, instructions for at least one function (such as a touch function, a sound playing function, an image playing function, etc.), instructions for implementing the various method embodiments described above, etc. The data storage area can store data involved in the various method embodiments described above, etc. The memory 305 can also optionally be at least one storage device located away from the aforementioned processor 301. The memory 305, as a computer storage medium, can include an operating system, a network communication module, a user interface 303 module, and an application program of an intelligent recommendation method of a financial advisor service.

[0185] In Figure 3 In the electronic device shown, the user interface 303 is mainly used to provide an interface for user input and obtain data input by the user. The processor 301 can be used to call an application program of an intelligent recommendation method of a financial advisor service stored in the memory 305, and when executed by one or more processors 301, the electronic device performs the method of one or more of the above embodiments.

[0186] It should be noted that, for the above-mentioned method embodiments, in order to simply describe, they are all expressed as a combination of a series of actions, but those skilled in the art should know that the present application is not limited by the order of the described actions, because according to the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0187] In the above embodiments, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0188] In several embodiments provided in the present application, it should be understood that the disclosed apparatus can be implemented in other manners. For example, the division of the apparatus embodiments is merely illustrative, and the division of units can be changed according to actual needs. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between units can be indirect coupling or communication connection through some intervening interface, device or unit, and can be electrical or other forms.

[0189] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one place or distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0190] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0191] If the integrated unit is realized in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium 305 and includes a number of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the methods of the embodiments. The aforementioned storage medium 305 includes: a U disk, a mobile hard disk, a magnetic or optical disk, and various program code storage media.

[0192] The present application also discloses a computer readable storage medium, which stores instructions. When executed by one or more processors 301, the electronic device executes one or more methods as described in the above embodiments.

[0193] The above merely show example embodiments of the present disclosure, and cannot limit the scope of the present disclosure. That is, equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon considering the specification and practice of the present disclosure. The present application is intended to cover any variations, uses, or adaptations of the present disclosure following the general principles thereof and including such departures from the present disclosure that come within known or customary practice in the art to which the present disclosure pertains. The specification and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are to be limited only by the claims.

Claims

1. An intelligent recommendation method of a portfolio service, characterized by, The method comprises: acquiring a historical sample data set with a time stamp, the historical sample data set comprising a plurality of samples, each sample consisting of a plurality of features, the historical sample data set containing a time decay weight calculated for each sample, each sample comprising three types of feature variables: environmental driving features, behavioral intermediary features, and target response features, wherein the environmental driving features represent variables related to the current market state, including volatility and sector rotation intensity, the behavioral intermediary features represent current user portrait information, including risk level and asset allocation ratio, and the target response features are product features, representing the product feature portrait preferred by the current user under the current market conditions; building a dynamically constrained causal structure diagram, wherein undirected edges are added between different types of nodes contained in the causal structure diagram, and no edges are built between nodes of the same type contained in the causal structure diagram, and the features correspond one-to-one to the nodes; building a sparse deep neural network model with cross-layer jump connections according to the dynamically constrained causal structure diagram; calculating and outputting the target response features according to real-time target inputs using the trained sparse deep neural network model; based on the similarity between the target response features and the candidate features of the candidate products in the candidate pool, filtering out the candidate products corresponding to the candidate features with the highest similarity as the recommendation results. 2.The method of claim 1, wherein, After filtering out the candidate products corresponding to the candidate features with the highest similarity as the recommendation results based on the similarity between the target response features and the candidate features of the candidate products in the candidate pool, the method further comprises: for each recommended product contained in the recommendation results, obtaining a corresponding candidate feature vector from the sparse deep neural network model, the candidate feature vector being a multi-dimensional vector, each dimension corresponding to a candidate feature; for each dimension of the candidate feature, identifying a dominant input feature with the largest causal effect on the target response feature according to the column vector corresponding to the dimension of the target response feature in the causal effect matrix, the dominant input feature being the input feature corresponding to the maximum value in the column of the environmental driving feature or the behavioral intermediary feature in the causal effect matrix; building a causal chain logic path according to the type and causal effect value of the dominant input feature, if the dominant input feature belongs to the environmental driving feature, generating a causal chain path of the environmental driving feature pointing to the target response feature, if the dominant input feature belongs to the behavioral intermediary feature, generating a causal chain path of the behavioral intermediary feature pointing to the target response feature, the causal chain path comprising the dominant input feature, the target response feature, and the causal effect value between the dominant input feature and the target response feature. 3.The method of claim 1, wherein, The method of building a sparse deep neural network model with cross-layer jump connections according to the dynamically constrained causal structure diagram specifically comprises: Based on the directed edge relationships in the causal structure graph, a structure mask matrix is ​​constructed. The structure mask matrix is ​​used to limit the legality of connections in the sparse deep neural network model. The structure mask matrix includes a connection mask from environment-driven features to behavioral mediator features, a connection mask from behavioral mediator features to target response features, and a skip connection mask from environment-driven features to target response features. The element values ​​corresponding to the unallowed connection positions in the structure mask matrix are fixed to zero. The nodes in the causal structure graph are mapped to the structure of the sparse deep neural network model, wherein the nodes corresponding to the environment-driven features are mapped to the first input layer of the sparse deep neural network model, the nodes corresponding to the behavioral mediation features are mapped to the second input layer of the sparse deep neural network model, and the nodes corresponding to the target response features are mapped to the output layer of the neural network model. Initialize the connection weight matrix and bias parameter terms, including the first connection weight matrix and first bias term from the environment-driven feature to the behavior mediation feature, the second connection weight matrix and second bias term from the behavior mediation feature to the behavior mediation feature, the third connection weight matrix and third bias term from the behavior mediation feature to the target response feature, and the skip connection weight matrix and fourth bias term from the environment-driven feature to the target response feature. For positions in the first connection weight matrix, the third connection weight matrix, and the skip connection weight matrix that are not allowed by the structure mask matrix, after training, the corresponding connection weight values ​​are fixed to zero according to the zero value position of the structure mask matrix, and remain constant in subsequent use stages. During the forward propagation of the sparse deep neural network model, based on the first input layer, the first connection weight matrix and the first bias term, and the second input layer, the second connection weight matrix and the second bias term, an intermediate representation is calculated through an activation function to obtain a first intermediate representation, which is the hidden activation vector of the second input layer. The first intermediate representation and the environment-driven feature jump connection work together in the output layer, and the output result is generated based on the third connection weight matrix and the third bias term, as well as the jump connection weight matrix and the fourth bias term.

4. The method of claim 3, wherein the method further comprises: The process of using a trained sparse deep neural network model to calculate and output a target feature vector based on the real-time acquired target input specifically includes: The system receives a target input acquired in real time. The target input includes a set of real-time environment-driven features and a set of real-time behavior-mediated features. The real-time environment-driven features are input to the first input layer, and the real-time behavior-mediated features are input to the second input layer. The arrangement order and numerical dimensions of the real-time environment-driven features and the real-time behavior-mediated features are consistent with those of the training phase. The real-time environment-driven features are subjected to matrix operations with the first connection weight matrix and the first bias term to obtain a first operation result. At the same time, the real-time behavior mediation features are subjected to matrix operations with the second connection weight matrix and the second bias term to obtain a second operation result. The first operation result and the second operation result are added together and then the intermediate representation is calculated through the activation function to obtain a second intermediate representation. The second intermediate representation is the hidden activation vector of the second input layer. The second intermediate representation is multiplied by the third connection weight matrix to obtain the first product vector; The real-time environment-driven features are multiplied with the jump connection weight matrix to obtain a second product vector; The first product vector is added to the second product vector, and then added to the target response feature bias vector. The result is mapped by an activation function to obtain the output vector of the sparse deep neural network model, which is the target feature vector.

5. The method of claim 1, wherein the method further comprises: The construction of the causal structure graph of dynamic constraints specifically includes: All features in the historical sample dataset are divided into three feature types according to business attributes: environment-driven features, behavioral mediation features, and target response features. Based on the division results, each feature is mapped to a node, and the feature and the node correspond one-to-one. Initialize the causal structure graph. In the causal structure graph, add undirected edges between two nodes of different types. Do not construct any edges between two nodes of the same feature type. The added undirected edges include the edges between the nodes corresponding to the environment-driven feature and the nodes corresponding to the behavior-mediating feature, the edges between the nodes corresponding to the behavior-mediating feature and the nodes corresponding to the target response feature, and the edges between the nodes corresponding to the environment-driven feature and the nodes corresponding to the target response feature.

6. The method of claim 5, wherein the method further comprises: The construction of the causal structure graph with dynamic constraints further includes: For the target feature, a conditional independence test is performed, specifically including: constructing a candidate conditional feature set and enumerating all non-empty subsets; for each non-empty subset as a conditional feature set, calculating the significance probability value corresponding to the weighted residual correlation between the target feature and the comparison feature under the conditional feature set based on the time decay weight of each feature in the sample; wherein the target feature is any one of the multiple features, and the comparison feature is a feature of a different type from the target feature; when there exists any conditional feature set that makes the significance probability value exceed a preset significance threshold, the undirected edge between the target feature and the comparison feature is deleted from the causal structure graph. On the causal structure graph obtained after the conditional independence test, all triples that simultaneously satisfy the following structural features are identified: for each intermediate feature node, if there is no connecting edge between two adjacent feature nodes, the directions of the two adjacent edges are both directed towards the intermediate feature node, forming a unidirectional causal structure. Perform direction reasoning on the remaining undirected edges, and iterate to determine the direction of all the undirected edges. One-way path constraints between feature types are applied to the constructed directed graph. The one-way path constraints include: retaining one-way edges from the corresponding node of the environment-driven feature to the corresponding node of the behavior-mediating feature, from the corresponding node of the behavior-mediating feature to the corresponding node of the target response feature, and from the corresponding node of the environment-driven feature to the corresponding node of the target response feature; and removing all edges that do not satisfy the one-way path constraints to generate a causal structure graph with dynamic constraints.

7. The method of claim 1, wherein the method further comprises: The method further includes: The features are categorized into environment-driven features, behavioral mediator features, and target response features based on their business attributes. The environment-driven features include market features, the behavioral mediator features include customer features, and the target response features include product features.

8. An intelligent recommendation device for investment advisory services, characterized in that, The device is used to execute a smart recommendation method for investment advisory services as described in any one of claims 1-7, the device comprising an acquisition module (201), a processing module (202), and an output module (203), wherein: The acquisition module (201) is used to acquire a historical sample dataset with timestamps. The historical sample dataset includes multiple samples, each sample is composed of multiple features, and the historical sample dataset contains time decay weights calculated for each sample. Each sample includes three types of feature variables: environment-driven features, behavioral mediator features, and target response features. The environment-driven features represent variables related to the current market state, including volatility and sector rotation intensity. The behavioral mediator features represent current user profile information, including risk level and asset allocation ratio. The target response features are product features, representing the product feature profile preferred by the current user under the current market conditions. The processing module (202) is used to construct a causal structure graph with dynamic constraints, wherein undirected edges are added between nodes of different types contained in the causal structure graph, and no edges are constructed between nodes of the same type contained in the causal structure graph, and the features correspond one-to-one with the nodes; The processing module (202) is used to construct a sparse deep neural network model with cross-layer skip connections based on the causal structure graph of the dynamic constraints. The processing module (202) is used to calculate and output target response features based on the target input acquired in real time using the trained sparse deep neural network model; The output module (203) is used to select the candidate products corresponding to the most similar candidate features as the recommendation results based on the similarity between the target response features and the candidate features of the candidate products in the candidate pool.

9. An electronic device, comprising: The device includes a processor (301), a communication bus (302), a user interface (303), a network interface (304), and a memory (305). The memory (305) is used to store instructions. The user interface (303) and the network interface (304) are both used to communicate with other devices. The communication bus (302) is used to realize the connection and communication between the components within the electronic device. The processor (301) is used to execute the instructions stored in the memory (305) so that the electronic device performs the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed, perform the method as described in any one of claims 1-7.

Images