Intelligent recommendation method and device for investment advisor service, equipment and medium

Abstract

The invention provides an intelligent recommendation method, device and equipment for an investment adviser service and a medium, and relates to the technical field of data processing.The method comprises the steps that a historical sample data set with a timestamp is obtained, the historical sample data set comprises a plurality of samples, and each sample is composed of a plurality of features, the historical sample data set comprises a time decay weight calculated for each sample; constructing a causal structure diagram of dynamic constraint; constructing a sparse deep neural network model of cross-layer jump connection according to the causal structure diagram of the dynamic constraint; calculating and outputting target response features according to target input acquired in real time by using the trained sparse deep neural network model; and based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool, screening out candidate products corresponding to a plurality of candidate features with the highest similarity as recommendation results. The problem of hysteresis caused by static modeling in a traditional recommendation system can be solved.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and specifically to an intelligent recommendation method, device, equipment and medium for investment advisory services. Background Art

[0002] Traditional investment advisory recommendation systems typically rely on historical transaction records, user profiles, and static risk preference data, using regression models, clustering algorithms, or classifiers to establish fixed-structure model mapping relationships. During the training phase, these models extract investment behavior patterns from known samples and apply the learned strategies to subsequent investment recommendations for users. Their core logic relies on fitting past behavior to infer possible future investment choices.

[0003] However, such systems are inherently static, lacking the ability to perceive market fluctuations, changes in user assets, or temporary adjustments in preferences. Lacking a real-time data-driven feedback mechanism, the model structure and parameters cannot be updated according to the dynamic evolution of the investment environment. Consequently, their recommendations often lag behind market realities, making it difficult to meet the increasing demands of users for personalized, dynamic, and scenario-specific capabilities in today's rapidly changing environment. Consequently, their practicality and investment-return matching are limited. Therefore, a method is needed to address the lag caused by static modeling in traditional recommendation systems. Summary of the Invention

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

[0005] In a first aspect of the present application, a method for intelligent recommendation of investment advisory services is provided, the method comprising: Acquire a historical sample dataset with a timestamp, where the historical sample dataset includes a plurality of samples, each sample is composed of a plurality of features, and the historical sample dataset includes a time decay weight calculated for each of the samples; Constructing a causal structure graph of dynamic constraints, wherein undirected edges are added between nodes of different types included in the causal structure graph, and no edges are constructed between nodes of the same type included in the causal structure graph, and the features correspond one-to-one to the nodes; Constructing a sparse deep neural network model with cross-layer skip connections based on the causal structure graph of the dynamic constraints; Utilize the trained sparse deep neural network model to calculate and output target response features based on the target input obtained in real time; Based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool, candidate products corresponding to several candidate features with the highest similarity are screened out as recommendation results.

[0006] Based on the above technical solution, preferably, after selecting candidate products corresponding to several candidate features with the highest similarity as recommendation results 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: For each recommended product included in the recommendation result, obtain a corresponding candidate feature vector from the sparse deep neural network model, where the candidate feature vector is a multidimensional vector, and each dimension corresponds to a candidate feature; For each dimension of the candidate features, based on the column vector corresponding to the dimension of the target response feature in the causal effect matrix, identify the dominant input feature with the maximum causal effect on the target response feature, where the dominant input feature is the input feature corresponding to the maximum value of the column where the environmental driving feature or the behavioral mediating feature is located in the causal effect matrix; Constructing a causal chain logic path based on 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 from the environmental driving feature to the target response feature. If the dominant input feature belongs to the behavioral mediating feature, a causal chain path is generated from the behavioral mediating feature to the target response feature, and the causal chain path includes the dominant input feature, the target response feature, and the causal effect value between the dominant input feature and the target response feature.

[0007] On the basis of the above technical solution, preferably, the sparse deep neural network model with cross-layer skip connections is constructed according to the causal structure graph of the dynamic constraints, specifically including: Constructing a structure mask matrix according to the directed edge relationship in the causal structure graph, wherein the structure mask matrix is ​​used to limit the connection legitimacy in the sparse deep neural network model, wherein the structure mask matrix includes a connection mask from the environment-driven feature to the behavior-mediated feature, a connection mask from the behavior-mediated feature to the target response feature, and a jump connection mask from the environment-driven feature to the target response feature, and the element values ​​corresponding to the unallowed connection positions in the structure mask matrix are fixed to zero; Mapping the nodes in the causal structure graph to the structure of the sparse deep neural network model, wherein the nodes corresponding to the environmental driving 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 items, including the first connection weight matrix and first bias item from the environment-driven feature to the behavior-mediated feature, the second connection weight matrix and second bias item from the behavior-mediated feature to the behavior-mediated feature, the third connection weight matrix and third bias item from the behavior-mediated feature to the target response feature, and the skip connection weight matrix and fourth bias item 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 is completed, 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.

[0008] During the forward propagation process of the sparse deep neural network model, an intermediate representation is calculated using an activation function 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 to obtain a first intermediate representation, where the first intermediate representation is an implicit activation vector of the second input layer; The first intermediate representation and the environment-driven feature jump connection jointly act on the output layer, and rely on the third connection weight matrix and the third bias item as well as the jump connection weight matrix and the fourth bias item to generate an output result.

[0009] On the basis of the above technical solution, preferably, the sparse deep neural network model that has been trained calculates and outputs the target feature vector according to the target input obtained in real time, specifically including: Receive target input acquired in real time, the target input including a set of real-time environment-driven features and a set of real-time behavior-mediating features, the real-time environment-driven features being input into the first input layer, and the real-time behavior-mediating features being input into the second input layer, the arrangement order and numerical dimensions of the real-time environment-driven features and the real-time behavior-mediating features being consistent with those in the training phase; Performing a matrix operation on the real-time environment driving feature, the first connection weight matrix, and the first bias term to obtain a first operation result, and performing a matrix operation on the real-time behavior mediating feature, the second connection weight matrix, and the second bias term to obtain a second operation result, adding the first operation result and the second operation result, and then calculating an intermediate representation through an activation function to obtain a second intermediate representation, where the second intermediate representation is an implicit activation vector of the second input layer; Performing matrix multiplication on the second intermediate representation and the third connection weight matrix to obtain a first product vector; Performing matrix multiplication calculation on the real-time environment driving feature and the skip connection weight matrix to obtain a second product vector; The first product vector and the second product vector are added together, and then added to the target response feature bias vector. The result is mapped through an activation function to obtain an output vector of the sparse deep neural network model, and the output vector is the target feature vector.

[0010] Based on the above technical solution, preferably, the construction of the causal structure diagram of the dynamic constraint specifically includes: Divide all features in the historical sample dataset into three types of feature types based on business attributes, namely, environment-driven features, behavior-mediating features, and target-response features. Map each feature to a node based on the division results, with the features corresponding to the nodes one-to-one. Initialize the causal structure graph. In the causal structure graph, add an undirected edge between two nodes of different types, and do not construct any edge for two nodes of the same feature type. The added undirected edges include the edge between the node corresponding to the environment-driven feature and the node corresponding to the behavior-mediating feature, the edge between the node corresponding to the behavior-mediating feature and the node corresponding to the target response feature, and the edge between the node corresponding to the environment-driven feature and the node corresponding to the target response feature.

[0011] Based on the above technical solution, preferably, the construction of the causal structure diagram of dynamic constraints further includes: A conditional independence test is performed on the target feature, 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, calculating the significance probability value corresponding to the weighted residual correlation between the target feature and the comparison feature under the conditional feature set, 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 among the multiple features; when there is any conditional feature set that makes the significance probability value exceed a preset significance threshold, deleting the undirected edge between the target feature and the comparison feature in the causal structure graph; On the causal structure graph obtained after the conditional independence test, all triples that simultaneously satisfy the following structural characteristics are identified: for each intermediate feature node, if there is no connecting edge between two adjacent feature nodes, the directions of both adjacent edges are directed toward the intermediate feature node, forming a unidirectional causal structure; Performing directional reasoning on the remaining undirected edges, and iteratively completing the direction determination of all the undirected edges; A unidirectional path constraint between feature types is imposed on the constructed directed graph. The unidirectional path constraint includes: retaining unidirectional edges from the node corresponding to the environment-driven feature to the node corresponding to the behavior-mediating feature, from the node corresponding to the behavior-mediating feature to the node corresponding to the target response feature, and from the node corresponding to the environment-driven feature to the node corresponding to the target response feature, and removing all edges that do not satisfy the unidirectional path constraint to generate a causal structure graph with dynamic constraints.

[0012] On the basis of the above technical solution, preferably, the method further comprises: The multiple features are divided into environment-driven features, behavior-mediating features and target-response features according to business attributes, wherein the environment-driven features include market features, the behavior-mediating features include customer features, and the target-response features include product features.

[0013] In a second aspect of the present application, there is provided an intelligent recommendation device for investment advisory services, the device being configured to execute any one of the above-described intelligent recommendation methods for investment advisory services. The device comprises an acquisition module, a processing module, and an output module, wherein: The acquisition module is configured to acquire a historical sample data set with a timestamp, wherein 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 includes a time decay weight calculated for each of the samples; The processing module is configured to construct a causal structure graph of dynamic constraints, wherein undirected edges are added between nodes of different types included in the causal structure graph, and no edges are constructed between nodes of the same type included in the causal structure graph, and the features correspond one-to-one to the nodes; The processing module 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 is used to calculate and output target response features based on the target input obtained in real time using the trained sparse deep neural network model; The output module is configured to screen out candidate products corresponding to several candidate features having the highest similarity as recommendation results based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool.

[0014] In the third aspect of the present application, an electronic device is provided, including a processor, a memory, a user interface and a network interface, the memory is used to store instructions, the user interface and the network interface are both used to communicate with other devices, and the processor is used to execute the instructions stored in the memory so that the electronic device performs any of the methods described above.

[0015] In a fourth aspect of the present application, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores instructions. When the instructions are executed, any one of the methods described above is executed.

[0016] In summary, one or more technical solutions provided in the embodiments of the present application have at least the following technical effects or advantages: 1. This application constructs a dynamically constrained causal structure graph and a sparse deep neural network model with cross-layer jump connections based on it, so that the recommendation model has the structural modeling ability of environmental driving characteristics and behavioral intermediary characteristics, and combines timestamps and time-decay weights to dynamically reflect the causal influence path of samples evolving over time, generates target response characteristics under real-time input and makes product recommendations based on them, so that the model structure and parameters can be adaptively adjusted with real-time changes in market conditions and user characteristics, effectively overcoming the limitations of traditional recommendation systems that rely on static modeling, response lags, lack timeliness and causal explanation capabilities, and significantly improve the dynamic adaptability and real-time accuracy of recommendations.

[0017] 2. By introducing a causal effect matrix to model the recommendation results explanatorily, it is possible to generate a causal chain path based on the dominant input features, thereby providing a clear causal reasoning basis for each candidate product, enhancing the transparency and explainability of the recommendation results, and helping to improve users' trust in and adoption rate of the recommended content.

[0018] 3. A structural mask matrix is ​​constructed through the causal structure graph and mapped to a network structure, effectively limiting the legitimacy of model connections. At the same time, combined with the initialization strategy, it ensures that unallowed connections are always invalid. The sparse structure is used to enhance causal prior injection, improve the convergence efficiency and generalization ability of model parameter learning, and avoid overfitting problems caused by redundant connections.

[0019] 4. Through the layered injection and structured calculation of real-time target input, the input consistency between the training and inference stages is maintained. The expressive power of the information pathway is enhanced by combining jump connections. Efficient target response feature calculation is achieved under the coordination of multi-source input, improving the real-time response capability and output accuracy of the recommendation model in dynamic environments.

[0020] 5. By classifying historical sample features and mapping them into causal structure graph nodes, the initial structural skeleton is constructed in combination with the connection constraints between feature types. The possible causal relationships between variables are clearly defined, avoiding the expansion of the invalid hypothesis space in the structural learning process, and improving the accuracy and efficiency of causal discovery.

[0021] 6. Through conditional independence tests, spurious correlation edges are weakened. By combining triple structure recognition and directional reasoning rules, the direction of causal edges is gradually clarified. Finally, business constraints are introduced to filter out illogical paths, thereby obtaining a dynamic causal graph with a reasonable structure and compliant constraints. This not only ensures the accuracy of causal identification, but also meets the specific semantic requirements of investment advisory scenarios, providing strong structural prior support for subsequent sparse neural network modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of an intelligent recommendation method for investment advisory services disclosed in an embodiment of the present application; Figure 2 This is a module diagram of an intelligent recommendation device for investment consulting services disclosed in an embodiment of the present application; Figure 3 This is a structural diagram of an electronic device disclosed in an embodiment of the present application.

[0023] Explanation of the reference numerals: 201, acquisition module; 202, processing module; 203, output module; 301, processor; 302, communication bus; 303, user interface; 304, network interface; 305, memory. DETAILED DESCRIPTION

[0024] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments.

[0025] In the description of the embodiments of this application, words such as "for example" or "for instance" are used to indicate examples, illustrations, or explanations. Any embodiment or design described as "for example" or "for instance" in the embodiments of this application should not be construed as being preferred or advantageous over other embodiments or designs. Rather, the use of words such as "for example" or "for instance" is intended to present the relevant concepts in a concrete manner.

[0026] In the description of the embodiments of the present application, the term "multiple" means two or more. For example, multiple systems refer to two or more systems, and multiple screen terminals refer to two or more screen terminals. In addition, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized.

[0027] Traditional investment advisory service recommendation systems rely on fixed data such as static risk preferences, historical transactions, and user profiles, and construct static mapping models through regression, clustering, or classification methods. Although they can reuse historical behavior patterns for investment recommendations, they lack the ability to perceive market fluctuations, user asset changes, and temporary preference adjustments in real time. In addition, the model structure cannot be adaptively updated as the environment evolves dynamically, resulting in delayed recommendation results and slow responses. This makes it difficult to meet the current actual needs for high-frequency changes, personalized decision-making, and scenario adaptation capabilities. There are obvious deficiencies in practicality and investment matching efficiency. Therefore, there is an urgent need to introduce a dynamic mechanism to solve the lag problem caused by static modeling.

[0028] This embodiment discloses an intelligent recommendation method for investment consulting services. Figure 1 , including the following steps S110-S150: S110, obtaining a historical sample data set with a timestamp.

[0029] The intelligent recommendation method for investment advisory services disclosed in the embodiments of this application is applied to a server. The server includes, but is not limited to, electronic devices such as mobile phones, tablet computers, wearable devices, and personal computers (PCs), and may also be a backend server that runs the intelligent recommendation method for investment advisory services. The server may be implemented as a standalone server or a server cluster consisting of multiple servers.

[0030] First, a historical sample dataset covering a period of time is collected from the investment advisory system. This dataset consists of multiple samples, each of which contains three types of feature variables: environmental driving features, behavioral mediating features, and target response features. Environmental driving features are preferably market features, behavioral mediating features are preferably customer features, and target response features are preferably product features. Each sample is accompanied by a unique timestamp to identify the time when it was recorded.

[0031] For each sample, its time decay weight is calculated, and an exponential decay function is used to express that the importance of the sample to causal structure learning decreases with time distance. The specific formula is as follows:

[0032] in, represents the time decay weight of the t-th sample, Represents the initial weight of the sample, which is a constant and usually set to 1. λ represents the time attenuation coefficient. , takes the value as a positive real number, controls the decay speed, and the typical value is λ=0.05. Δt represents the time difference between the current system time and the timestamp of the t-th sample.

[0033] The calculated time decay weight will be used as an additional attribute of each sample and will participate in the weighted calculation of the conditional independence test in the subsequent construction of the causal structure diagram, including the weighted correlation coefficient, weighted covariance, weighted mean and sample weighting processing in the Fisher Z test.

[0034] Time-decay weights are used to assign decreasing importance to samples within different time windows when processing historical sample data with timestamps, thereby controlling the model's sensitivity to recent data changes and strengthening temporal relationships in causal reasoning. Specifically, time-decay weights serve as weighting factors in conditional independence tests and residual correlation calculations, giving more weight to recent samples when evaluating causal dependencies between features. This helps capture dynamic time-varying patterns, improves the dynamic adaptability of causal structure diagrams, and enhances the accuracy of recommendation models' responses to current inputs, effectively alleviating the problem of traditional static modeling that ignores time evolution.

[0035] S120, construct a causal structure diagram of dynamic constraints.

[0036] In a possible implementation, constructing a causal structure diagram of dynamic constraints specifically includes: All features in the historical sample data set are divided into three types of feature types according to business attributes, namely, environment-driven features, behavior-mediating features, and target-response features. Each feature is mapped to a node based on the division results, and features correspond to nodes one-to-one. The causal structure graph is initialized. In the causal structure graph, undirected edges are added between two nodes of different types. No edges are constructed for two nodes of the same feature type. The added undirected edges include the edges between the nodes corresponding to the environment-driven features and the nodes corresponding to the behavior-mediating features, the edges between the nodes corresponding to the behavior-mediating features and the nodes corresponding to the target-response features, and the edges between the nodes corresponding to the environment-driven features and the nodes corresponding to the target-response features.

[0037] Specifically, implementing this technical solution for constructing a dynamically constrained causal structure graph requires first classifying the historical sample dataset into different types of features. This dataset contains multiple samples, each composed of several feature variables. Based on business semantics and causal logic, all features are divided into three types: environment-driven features, behavior-mediated features, and target-response features. Once this classification is complete, a unique graph node is established for each feature, forming a one-to-one correspondence between features and nodes, laying the structural foundation for the subsequent construction of the causal structure graph.

[0038] A causal structure graph is a directed graph structure constructed under the guidance of business semantics. It is used to depict the causal dependencies between three types of variables: environmental driving features, behavioral mediating features, and target response features. First, all features in historical sample data are divided into three feature types based on their business attributes. Each feature is mapped to a node in the graph, forming a one-to-one mapping between features and nodes. During graph initialization, undirected edges are added only between nodes of different feature types to construct potential causal paths. These include edges between environmental driving features and behavioral mediating features, between behavioral mediating features and target response features, and between environmental driving features and target response features. No edges are constructed between nodes of the same type. Subsequently, directed edges are derived through conditional independence tests and directional inference rules, further improving the causal structure graph. Unidirectional path constraints are imposed between feature types to ensure the semantically sound direction of the causal paths. Finally, this causal structure graph is embedded as a causal prior constraint in a sparse deep neural network model, providing a rigorous basis for the model's connectivity structure, information flow, and causal interpretability.

[0039] After completing the feature type division and generating the node set, the causal structure graph is initialized and constructed. During the initialization phase, graph connection restrictions must be imposed based on the causal direction constraints between feature types, explicitly allowing only connections between nodes of different types, and not between nodes of the same type. Specifically, undirected edges are only established between the following three feature type combinations: an undirected edge is established between the environment-driven feature node and the behavior-mediating feature node, representing the potential impact of market conditions on customer behavior; an undirected edge is established between the behavior-mediating feature node and the target-response feature node, representing the causal mediation effect of customer behavior on product selection results; and an undirected edge is established between the environment-driven feature node and the target-response feature node, representing the direct action path of external market factors on the final recommendation results. In this initialization graph, each edge is added as an undirected edge, representing a potential causal connection whose direction has not yet been clearly determined. This is used in the subsequent conditional independence test and directional reasoning phases to further identify the true causal direction.

[0040] This initialization construction process ensures that the causal structure diagram has the structural characteristics of heterogeneous node layering, controllable paths, and unidirectional constraints between types. It conforms to the three-layer causal modeling logic of market-customer-product in the financial investment advisory system, and at the same time provides a complete and controlled starting structure for the next step of skeleton diagram construction and collision structure identification based on weighted independence test.

[0041] In a possible implementation, constructing a causal structure diagram of dynamic constraints specifically includes: A conditional independence test is performed on the target feature, 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, calculating the significance probability value corresponding to the weighted residual correlation between the target feature and the comparison feature under the conditional feature set, 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 in the multiple features. When there is any conditional feature set that makes the significance probability value exceed the preset significance threshold, the undirected edge between the target feature and the comparison feature is deleted in the causal structure diagram; on the causal structure diagram obtained after the conditional independence test, all simultaneous A triplet that satisfies the following structural characteristics: for each intermediary feature node, when there is no connecting edge between two adjacent feature nodes, the directions of the two adjacent edges are directed to the intermediary feature node to form a one-way causal structure; directional reasoning is performed on the remaining undirected edges, and the directions of all undirected edges are determined iteratively; one-way path constraints are imposed on the constructed directed graph between feature types. 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-intermediary feature, from the corresponding node of the behavior-intermediary 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 eliminating all edges that do not satisfy the one-way path constraints to generate a dynamically constrained causal structure graph.

[0042] Specifically, when constructing a dynamically constrained causal structure graph, a conditional independence test must first be performed for each target feature. A target feature is any feature node that is fixed as the variable to be tested in the current test round. Its corresponding comparison feature must belong to a different type of feature node than the target feature. For each initial undirected edge between a target feature and a comparison feature, before independence is determined, a candidate conditional feature set is constructed for the target feature. This candidate conditional feature set consists of all nodes connected to the target feature but excluding the comparison feature. Based on this set, all non-empty subsets of the candidate conditional feature set are enumerated. Each subset is used as a conditional feature set, and the statistical independence between the target feature and the comparison feature under this conditional feature set is tested. This test incorporates a time-decay weight for each sample, and a weighted Pearson correlation analysis and Fisher Z transformation are performed based on residual correlation to obtain a significance probability value. If any conditional feature set among all non-empty conditional feature sets results in a significance probability value above a preset threshold (e.g., 0.05), the target feature and the comparison feature are considered independent under this condition, and the undirected edge between the two nodes is deleted from the causal structure graph.

[0043] After completing the conditional independence tests between all target features and the comparison features, a streamlined skeleton graph structure is obtained, upon which collision structure identification is performed. All intermediary feature nodes in the graph are traversed to determine whether the node is simultaneously connected to two unconnected feature nodes. If this condition is met, a typical collision structure is considered to exist. Within this structure, the two edges connecting the intermediary feature nodes are oriented to point simultaneously to the intermediary feature node, forming a standard causal structure of "external nodes pointing to intermediary nodes," i.e., a unidirectional causal unit.

[0044] After completing the collision structure processing, the remaining undirected edges are then inferred for direction. This inference process relies on three types of inference rules: path reachability rules, which state that if there is a directed path from node A to node B, then if an undirected edge A-B exists, it can be directed as A→B; shared source node structure rules, which state that when node A points to two nodes B and C, and an undirected edge exists between B and C, the direction can be inferred to be B→C or C→B; and non-connected path inference rules, which state that if there is no direct connection between a pair of nodes but other constrained paths exist, the edge direction is determined based on logical transitivity. These three types of rules are iteratively applied to the entire graph until all inferable edge directions are determined.

[0045] Finally, a structural pruning operation is performed on the complete directed graph based on the business path restrictions between feature types. The one-way path constraints between feature types are as follows: only one-way edges from environment-driven feature nodes to behavior-mediating feature nodes, one-way edges from behavior-mediating feature nodes to target-response feature nodes, and one-way edges from environment-driven feature nodes directly to target-response feature nodes are retained; all connecting edges that do not meet the above path constraints are deleted. This constraint operation ensures that the final constructed causal structure graph strictly conforms to the hierarchical modeling logic of market-driven behavior, intermediary conversion decision-making, and result response output. It serves as the prior input for the connection structure of the sparse deep neural network model, supporting the causal-driven product recommendation system to achieve high-confidence structural interpretability and causal consistency.

[0046] S130, constructing a sparse deep neural network model with cross-layer skip connections based on the dynamically constrained causal structure graph.

[0047] In one possible embodiment, a sparse deep neural network model with cross-layer jump connections is constructed based on a dynamically constrained causal structure graph, specifically including: constructing a structure mask matrix based on the directed edge relationship in the causal structure graph, the structure mask matrix is ​​used to limit the legitimacy of the connections in the sparse deep neural network model, wherein the structure mask matrix includes a connection mask from the environment-driven feature to the behavior-mediating feature, a connection mask from the behavior-mediating feature to the target response feature, and a jump connection mask from the environment-driven feature to the target response feature, and the element values ​​corresponding to the non-allowed connection positions in the structure mask matrix are fixed to zero; mapping the nodes in the causal structure graph to the structure of the sparse deep neural network model, wherein the nodes corresponding to the environment-driven feature are mapped to the first input layer of the sparse deep neural network model, the nodes corresponding to the behavior-mediating feature are mapped to the second input layer of the sparse deep neural network model, and the nodes corresponding to the target response feature are mapped to the output layer of the neural network model; initializing the connection weight matrix and bias parameter items, including the first connection weight matrix and the first bias item from the environment-driven feature to the behavior-mediating feature, and the first bias item from the behavior-mediating feature to the target response feature. The second connection weight matrix and the second bias term from the mediating feature to the behavioral mediating feature, the third connection weight matrix and the third bias term from the behavioral mediating feature to the target response feature, and the jump connection weight matrix and the fourth bias term from the environmental driving feature to the target response feature, wherein, for the 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, after the training is completed, the corresponding connection weight value is fixed to zero according to the zero value position of the structure mask matrix, and remains constant in the subsequent use stage; 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 term and the second input layer, the second connection weight matrix and the second bias term, the intermediate representation is calculated by the activation function to obtain the first intermediate representation, and the first intermediate representation is the implicit activation vector of the second input layer; according to the first intermediate representation and the environment driving feature jump connection, the output layer is jointly acted on, relying on the third connection weight matrix and the third bias term and the jump connection weight matrix and the fourth bias term to generate the output result.

[0048] Specifically, a structural mask matrix is ​​constructed based on the directed edge relationships in the causal structure graph. The structural mask matrix is ​​used to strictly limit the legitimacy of the inter-layer connections in the sparse deep neural network model. During the construction process, all edges in the causal structure graph that "direct from environmental driving features to behavioral mediating features" are mapped into the first mask matrix. , all edges of “behavior mediating features pointing to target response features” are mapped to the second mask matrix , all jump connections of “environment-driven features pointing to target response features” are mapped to the third mask matrix In the structure mask matrix, all connection positions that are not allowed by the causal structure diagram have their corresponding element values ​​forced to be set to zero, which will be used to mask illegal paths in the model connection later.

[0049] Next, the three types of nodes in the causal structure diagram are mapped to the hierarchical structure of a sparse deep neural network model: nodes corresponding to all environmental driving features are mapped to the model's first input layer, which receives the market environment feature vector; nodes corresponding to all behavioral mediating features are mapped to the model's second input layer, participating in the calculation of the intermediate representation; and nodes corresponding to all target response features are mapped to the output layer, which ultimately outputs the model's target feature vector. This structural mapping strictly adheres to the three-layer feature type division of the causal structure diagram, maintaining the semantic consistency of the causal driving path.

[0050] Then perform the initialization operation of the connection weight matrix and bias term. Set the first connection weight matrix and the first bias term , used to connect the environment-driven features to the behavior-mediating features; set the second connection weight matrix and the second bias term , used to map behavioral mediation features to behavioral mediation features; set the third connection weight matrix With the third bias term , used to connect behavioral intermediary features to target response features; set the jump connection weight matrix With the fourth bias term , achieving a direct impact of environmental driving features on 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 position of the structure mask matrix after training is completed, and remain constant in subsequent use stages.

[0051] In this embodiment, the construction of a sparse deep neural network model involves four sets of connection weight matrices and bias parameters, each used to express the causal relationship between environmental driving features, behavioral mediating features, and target response features. Each set of connection weight matrices and bias terms is strictly defined based on the directed edges retained in the causal structure graph, and is combined with a structural mask matrix to control the legitimacy of the connections.

[0052] The first group is the connection weight matrix and bias term from the environmental driving feature to the behavioral intermediary feature, which are respectively recorded as and This weight matrix is ​​used to describe the indirect impact path of market status on customer behavior, and is the backbone connection for the model to propagate from the first input layer to the second input layer. The second group is the connection weight matrix and bias term from behavioral intermediary features to behavioral intermediary features, which are respectively denoted as and The weight matrix is ​​used to extract features from customer behavior data, which is the feature processing of the second input layer by the model. The two sets of results are added together and the first intermediate representation is calculated through the activation function. The corresponding forward propagation calculation formula is:

[0053]

[0054] in, represents the environment driving feature input vector, Intermediate representations that represent intermediary features of the behavior, Represents a non-linear activation function such as ReLU or tanh.

[0055] The third group is the connection weight matrix and bias term from the behavior mediation feature to the target response feature, which are respectively recorded as and This weight matrix is ​​used to capture the transmission mechanism of customer behavior to the final recommendation results and is one of the core causal paths in the model structure. In the forward propagation stage of the model, its calculation formula is:

[0056] in, Represents the impact component of the behavior mediator feature on the target response feature as part of the final output.

[0057] The fourth group is the jump connection weight matrix and bias term from the environmental driving feature to the target response feature, which are recorded as and This jump connection represents the direct driving path of the environment-driven features to the recommended target, without passing through the behavioral intermediary features, and constitutes a cross-layer connection from the first input layer to the output layer. Its calculation expression in the forward propagation is:

[0058] Finally, the results of the concatenation of the third and fourth groups are added together, and the target response feature vector is output through the activation function:

[0059] Through the division of labor and coordination of the above four sets of connection weight matrices and bias terms, the sparse deep neural network model can flexibly express the direct and indirect driving relationships between the three types of characteristics: market, behavior, and results, while maintaining the consistency of the causal path structure, thereby improving recommendation accuracy and causal interpretability.

[0060] In the forward propagation process of the sparse deep neural network model, the first step is to input the vector based on the environment driving feature. , combined with the first connection weight matrix and the first bias term for linear transformation, and then according to the intermediate feature input vector , combine the second connection weight matrix and the second bias term to perform linear transformation for feature extraction, then add the two results and input them into the activation function (such as ReLU or tanh), to obtain the intermediate representation of the behavior-mediating features , the calculation expression is:

[0061] Intermediate representation based on behavioral intermediary features and the raw input of the environment-driven features , respectively, perform linear operations with the corresponding connection weight matrix, and generate the prediction results of the target response characteristics through weighted summation and bias superposition to form the model output vector , its calculation expression is:

[0062] The output vector This is the model's prediction of the target response characteristics, used as input for subsequent product recommendation and causal effect assessment. Through the constraints of the structural mask matrix, the design of cross-layer skip connections, and the strict mapping of the hierarchical structure, the model maintains a precise adherence to the causal path structure while achieving efficient representation of complex nonlinear causal relationships.

[0063] In one possible implementation, a sparse deep neural network model is constructed based on a dynamically constrained causal structure graph, employing cross-layer skip connections to enhance causal information transfer between features. The core goal of the training phase is to learn the predictive structure of target response features from environmental driving features and behavioral mediating features. While maintaining consistency in the causal structure, this approach maximizes prediction accuracy and improves the model's sparsity and generalization capabilities.

[0064] In the model structure, the intermediate representation of the behavior-mediating features is calculated by fusing the environment-driven features and the original behavior-mediating features, and its expression is:

[0065] in, The intermediate activation vector representing the behavior-mediating features, Input vector for the environment driving feature, is the behavioral mediating feature input vector, 、 are the first connection weight matrix and the second connection weight matrix respectively, is the second bias term, is a non-linear activation function.

[0066] The output of the product layer not only relies on the intermediate representation of the behavior-mediated features, but also integrates the environment-driven features of the jump connection to output the target response feature vector in the following way:

[0067] in, It is the output of the model, which represents the predicted value of product features in each dimension.

[0068] In order to supervise the training of the structure, the main loss function based on mean square error is introduced , where the goal is to minimize the distance between the predicted output and the sample label, which is defined as follows:

[0069] At the same time, in order to strengthen the sparse structure, suppress overfitting and further ensure the consistency of the causal structure, two regularization terms are introduced to form a composite regularization term On the one hand, the allowed connections in the structure are imposed The penalty encourages sparse weights; on the other hand, it imposes a penalty on prohibited connections in the structure (defined by the mask matrix). Regularization is used to further suppress the activation of non-causal connection weights. Its formula is as follows:

[0070] in represents the structure mask matrix, represents element-wise product, 、 is the regularization hyperparameter.

[0071] The training process includes the following steps: Step 1, preprocessing and initialization: constructing the structure mask matrix , the corresponding positions in the causal structure where connections are allowed are set to 1, and the rest are set to 0. Initialize all weight parameters.

[0072] The second step is to perform forward propagation: for each batch of samples , execute the above two-layer calculation process to get the model output .

[0073] The third step is to calculate the loss function: the batch loss function consisting of the sum of the mean square error and two regularization terms:

[0074] Step 4: Back propagation and parameter update: based on Perform gradient calculations, use optimizers (such as Adam) to update weights and biases, and perform mask corrections on weights to keep illegal connection weights zero.

[0075] The fifth step is dynamic sparse training: every several rounds of iterations, all weights are less than the preset threshold. The legal connection weights are pruned to zero; at the same time, illegal connections that are currently zero but have large gradients are reactivated and given small initial values ​​to dynamically adjust the connection structure.

[0076] Step 6: Iterate the training until the loss function converges or the predetermined number of iterations is reached.

[0077] Finally, in the prediction phase, the model receives new real-time environment-driven features Behavioral mediating characteristics , respectively input to the corresponding input layer to generate the intermediate representation:

[0078] And further output the product prediction vector:

[0079] This output can be used as a continuous product feature prediction, and can also be used to match the most similar product prototypes in the candidate product set to complete personalized recommendations, thereby achieving accurate investment advice generation driven by causality.

[0080] In one possible implementation, the causal effect matrix is ​​a structured tensor representation of the causal influence of market and customer characteristics on the output of each product feature, quantitatively estimating the strength of the causal impact of each feature through local gradient analysis, based on a sparse deep neural network model that utilizes cross-layer skip connections. Because this neural network model has a multi-layer nonlinear structure, traditional linear regression weights cannot directly explain causal effects. Therefore, the chain rule is used to combine 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 of each input feature at the local point of the current sample. This reflects its instantaneous response sensitivity to the output target, thereby achieving an approximate estimate of the causal impact.

[0081] First, the market characteristics are input into the vector and customer feature input vector As the input of the model, the intermediate representation is calculated through linear transformation and nonlinear activation. The activation representation of the customer feature intermediate layer is:

[0082]

[0083] in, is the connection weight matrix from market characteristics to customer characteristics, is the self-connection weight matrix between customer features, is the client layer bias, is the activation function. Then the customer feature intermediate representation and market characteristics input Input them into the product feature layer together to form a linear combination variable of the product layer:

[0084] in, is the connection weight from customer to product, is the jump connection weight from market to product, Bias for the product layer.

[0085] Based on the above structure, the chain rule is used to calculate the local derivative of the output layer results on each characteristic component of the input layer to obtain the causal sensitivity estimation of the input characteristics to the output product characteristics. , its product characteristics The local causal effect of is:

[0086] The customer characteristics The local causal effect on the product characteristics is:

[0087] Calculate the above partial derivative results for each sample separately, organize them according to the characteristic dimension, and define the causal effect matrix of market characteristics as , the causal effect matrix of customer characteristics is ,in 、 、 are the number of dimensions of market characteristics, customer characteristics, and product characteristics respectively.

[0088] Finally, the two causal effect matrices are spliced ​​along the input dimension to form a complete causal effect matrix , whose expression is:

[0089] matrix Each element Indicates the The input feature pair The local causal influence strength of a product feature, where , then it means the The market characteristics of The role of a product feature; if , then it means the Customer characteristics for the The role of product features.

[0090] This causal effect matrix not only supports causal explanations of recommendation outputs, but also serves as the basis for constructing causal chain paths for recommendation results. It is used to trace the causal driving mechanisms corresponding to the recommended items, thereby improving the interpretability and transparency of the system.

[0091] S140, using the trained sparse deep neural network model to calculate and output target response features based on the target input obtained in real time.

[0092] In a possible implementation, a trained sparse deep neural network model is used to calculate and output a target feature vector based on a target input obtained in real time, specifically including: receiving a target input obtained in real time, the target input including a set of real-time environment-driven features and a set of real-time behavior-mediating features, the real-time environment-driven features are input to a first input layer, the real-time behavior-mediating features are input to a second input layer, and the arrangement order and numerical dimensions of the real-time environment-driven features and the real-time behavior-mediating features are consistent with those in the training stage; performing a matrix operation on the real-time environment-driven features, the first connection weight matrix, and the first bias term to obtain a first operation result, and at the same time performing a matrix operation on the real-time behavior-mediating features and the second connection weight matrix. The first operation result is added to the second operation result, and the intermediate representation is calculated through the activation function to obtain the second intermediate representation, which is the implicit activation vector of the second input layer; the second intermediate representation is matrix multiplied by the third connection weight matrix to obtain a first product vector; at the same time, the real-time environment driving feature and the jump connection weight matrix are matrix multiplied 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 through the activation function to obtain the output vector of the sparse deep neural network model, and the output vector is the target feature vector.

[0093] Specifically, the model receives target input data acquired in real time. This input includes a set of real-time environmental driving features and a set of real-time behavioral mediating features. The environmental driving features represent variables related to the current market state, such as volatility and sector rotation intensity, while the behavioral mediating features represent current user profile information, such as risk level and asset allocation ratio. The environmental driving features are input as input tensors into the first input layer of the model, and the behavioral mediating features are input into the second input layer. The order and dimension settings of these two sets of input features must be exactly the same as those in the training phase to ensure consistency in model structure mapping and computational graph integrity.

[0094] In the forward propagation phase of the model, the real-time environment driving feature vector With the first connection weight matrix and the first bias term Perform linear combination operation to transform the behavior mediating feature vector With the second connection weight matrix and the second bias term Perform a linear combination operation, then add the two results and pass the activation function The implicit representation of the intermediate features of the computation behavior is the second intermediate representation. The computational expression for this process is:

[0095] in, The intermediate activation vector representing the mediating features of the behavior, Input vector for market characteristics, Input vector for customer features.

[0096] Subsequently, the intermediate representation of behavioral mediating features With the third connection weight matrix Perform matrix multiplication to represent the influence of behavior-mediated features on target response features, while the real-time environment-driven features and the skip connection weight matrix Perform matrix multiplication to represent the direct impact of the environmental driving features on the target response features. The two parts of the results and the third bias term and the skip connection bias Added and activated by the function The final output vector of the generated model after mapping , that is, the target feature vector, its calculation expression is:

[0097] The resulting output vector A set of multidimensional vectors represents the predicted results for each target response feature under the current market environment and combined with the current customer profile. This can be used to calculate similarity with the feature vectors of products in the candidate product pool to achieve personalized investment recommendations. This structure explicitly preserves the interaction paths between different feature types in the causal structure diagram and reflects the direct impact of market variables on response features through a skip connection mechanism, thereby improving the response sensitivity and causal interpretability of the recommendation model.

[0098] Correspondingly, in the stage of constructing the causal structure graph of dynamic constraints, all features in the historical sample data set are first divided into three types of feature types according to their business attributes, namely, environment-driven features, behavior-mediating features, and target-response features, to ensure that each feature is clearly attributed to its functional role; then, based on the division results, each feature is mapped to a node in the graph structure, maintaining a one-to-one correspondence between features and nodes, which facilitates subsequent structural constraints and graph modeling.

[0099] When initializing the graph, only undirected edges are allowed between nodes of different feature types. Connection paths between nodes of the same type are strictly prohibited, thereby satisfying the structural priors for causal modeling. Specifically, the undirected edges allowed include: edges between nodes corresponding to environmental driving features and nodes corresponding to behavioral mediating features, used to express the impact of the market environment on customer behavior; edges between nodes corresponding to behavioral mediating features and nodes corresponding to target response features, used to express the regulation of user behavior on product performance; and edges between nodes corresponding to environmental driving features and nodes corresponding to target response features, used to express the impact of market variables on the direct response of the product. All edges are undirected at initialization, indicating a state where causal directions are to be inferred, serving as the input structural basis for subsequent conditional independence tests and directional reasoning. Through this rigorous structural construction, a stable, compliant, and interpretable causal connection prior can be provided for the model, making the recommendation model highly credible in terms of structural consistency and semantic clarity.

[0100] S150 , based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool, select candidate products corresponding to several candidate features with the highest similarity as recommendation results.

[0101] Prepare a pool of candidate product features with standardized structures. Each candidate feature in the pool is a labeled product response feature vector. The candidate response feature vector and the target response feature vector have the same dimensions, feature arrangement order, and normalization scale, thus ensuring numerical comparability and semantic consistency in subsequent similarity calculations.

[0102] First, a trained sparse deep neural network model with cross-layer skip connections performs forward computations on real-time environmental driving features and behavioral mediating features, outputting a target response feature vector corresponding to the current user. This target response feature vector represents the product feature profile preferred by the current user under current market conditions and serves as the fundamental input for the subsequent screening phase.

[0103] In the similarity calculation phase, all candidate response feature vectors in the candidate product feature pool are traversed and their similarity with the target response feature vector is evaluated one by one. Similarity metrics can be cosine similarity, Euclidean distance, Mahalanobis distance, etc. Taking cosine similarity as an example, its calculation formula is:

[0104] in, is the target response feature vector, For the The candidate response feature vector of candidate products, symbol represents vector dot product, Represents the vector norm.

[0105] Based on the similarity calculated between all candidate response feature vectors and the target response feature vector, candidate products are ranked and the products with the highest similarity values ​​are selected as the final recommendation results. Each candidate product in the recommendation results is supported by a high match between its candidate response feature vector and the target response feature vector, accurately reflecting the current user's product preferences under the current environment, thereby achieving real-time, personalized, and causally consistent investment recommendation services. The entire process maintains the full-dimensional modeling and structural rigor of the response features, ensuring a high degree of consistency between the recommendation mechanism and the causal modeling process.

[0106] In one possible embodiment, after selecting candidate products corresponding to several candidate features with the highest similarity as recommendation results based on the similarity between the target response feature and the candidate features of candidate products in the candidate pool, the method further includes: obtaining a corresponding candidate feature vector from a sparse deep neural network model for each recommended product included in the recommendation result, where the candidate feature vector is a multidimensional vector, with each dimension corresponding to a candidate feature; for each dimensional candidate feature, identifying a dominant input feature with the maximum causal effect on the target response feature based on a column vector corresponding to the dimension of the target response feature in a causal effect matrix, where the dominant input feature is the input feature corresponding to the maximum value in the column containing the environmental driving feature or the behavioral mediating feature in the causal effect matrix; and constructing a causal chain logical path based on the type and causal effect value of the dominant input feature. If the dominant input feature is an environmental driving feature, a causal chain path is generated from the environmental driving feature to the target response feature; if the dominant input feature is a behavioral mediating feature, a causal chain path is generated from the behavioral mediating feature to the target response feature. The causal chain path includes the dominant input feature, the target response feature, and the causal effect value between the dominant input feature and the target response feature.

[0107] Specifically, for each selected candidate product in the recommendation results, the candidate response feature vector output by the sparse deep neural network model is extracted. This candidate response feature vector is a multidimensional vector with the same number of dimensions as the target response features at the product level. Each dimension represents the recommendation expression strength of the candidate product on a specific product feature indicator and serves as the target basis for subsequent causal path reasoning.

[0108] After obtaining the candidate response feature vector, each dimension of the candidate response feature vector is processed dimension by dimension. For the candidate response feature in the jth dimension, the jth column vector is extracted from the pre-generated causal effect matrix, representing the strength of the local causal effect of all input features on that candidate response feature. This causal effect matrix is ​​a structural derivative of the sparse deep neural network model. It contains the partial derivatives of all environmental driving features and behavioral mediating features with respect to each target response feature, and can numerically reflect the structural causal transmission between features.

[0109] In each column of the causal effect matrix, locate the input feature index corresponding to the maximum value, indicating that the input feature is the dominant input feature in the current product feature dimension. The dominant input feature may come from the environmental driving feature set or the behavioral mediating feature set. Its position in the overall input feature index determines its type.

[0110] Subsequently, a causal chain logic path is constructed based on the type of dominant input feature. If the dominant input feature belongs to the set of environmentally driven features, the constructed path structure is "environmentally driven feature → target response feature," indicating that changes in market conditions lead to fluctuations in the product's features through this feature. If the dominant input feature belongs to the set of behaviorally mediated features, the constructed path structure is "behavioral mediated feature → target response feature," indicating that customer behavioral preferences directly influence the inferred output of this 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 numerical value of the causal effect between the two. This structure can be rendered as an explanatory template statement or graphical result, enabling an interpretable presentation of the recommendation results.

[0111] Through the above process, each dimension of target response characteristics can be traced back to its most critical causal source characteristics, thereby establishing a transparent causal path from input to output, providing mechanism visibility and decision traceability for user recommendation results in investment advisory services, thereby enhancing the trust and practicality of financial intelligent systems.

[0112] This embodiment also discloses an intelligent recommendation device for investment consulting services. Figure 2 , comprising an acquisition module 201, a processing module 202 and an output module 203, the device is used to execute any of the above-mentioned intelligent recommendation methods for investment consulting services, wherein: The acquisition module 201 is used to acquire a historical sample data set with a timestamp. The historical sample data set includes multiple samples, each sample is composed of multiple features, and the historical sample data set includes a time decay weight calculated for each sample.

[0113] The processing module 202 is used to construct a causal structure graph of dynamic constraints, wherein undirected edges are added between nodes of different types included in the causal structure graph, and no edges are constructed between nodes of the same type included in the causal structure graph, and features correspond to nodes one by one.

[0114] The processing module 202 is used to construct a sparse deep neural network model with cross-layer skip connections based on the dynamically constrained causal structure graph.

[0115] The processing module 202 is used to calculate and output target response features based on the target input obtained in real time using the trained sparse deep neural network model.

[0116] The output module 203 is configured to screen out candidate products corresponding to several candidate features having the highest similarity as recommendation results based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool.

[0117] It should be noted that the above embodiments provide devices that implement their functions using only the division of the above functional modules as examples. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiments are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0118] This embodiment also discloses an electronic device, referring to Figure 3 The electronic device may 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 .

[0119] The communication bus 302 is used to implement the connection and communication between these components.

[0120] The user interface 303 may include a display screen (Display) and a camera (Camera). Optionally, the user interface 303 may also include a standard wired interface and a wireless interface.

[0121] The network interface 304 may optionally include a standard wired interface or a wireless interface (such as a WI-FI interface).

[0122] The processor 301 may include one or more processing cores. The processor 301 utilizes various interfaces and circuits to connect various components within the server. It executes instructions, programs, code sets, or instruction sets stored in the memory 305, as well as accesses data stored in the memory 305, to perform various server functions and process data. Optionally, the processor 301 may be implemented using at least one of the following hardware forms: a digital signal processing (DSP), a field-programmable gate array (FPGA), or a programmable logic array (PLA). The processor 301 may integrate one or a combination of a central processing unit (CPU), a graphics processing unit (GPU), and a modem. The CPU primarily handles the operating system, user interface, and application programs. The GPU is responsible for rendering and drawing content displayed on the display screen. The modem handles wireless communications. It is understood that the modem may not be integrated into the processor 301 but implemented as a separate chip.

[0123] Memory 305 may include random access memory (RAM) or read-only memory (ROM). Optionally, the memory may include non-transitory computer-readable storage medium. Memory 305 may be used to store instructions, programs, code, code sets, or instruction sets. Memory 305 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch control, sound playback, image playback, etc.), and instructions for implementing the aforementioned method embodiments. The data storage area may store data related to the aforementioned method embodiments. Memory 305 may also optionally be at least one storage device located remotely from the aforementioned processor 301. Memory 305, as a computer storage medium, may include an operating system, a network communication module, a user interface 303 module, and an application program for an intelligent recommendation method for investment advisory services.

[0124] exist Figure 3In the electronic device shown, user interface 303 is primarily used to provide an input interface for the user and to obtain user input data. Processor 301 can be used to invoke an application stored in memory 305 that describes an intelligent recommendation method for investment advisory services. When executed by one or more processors 301, the electronic device executes one or more of the methods described in the aforementioned embodiments.

[0125] It should be noted that for the aforementioned method embodiments, for simplicity of description, they are all expressed as a series of action combinations, but those skilled in the art should be aware that this application is not limited by the order of the actions described, because according to this application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required for this application.

[0126] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0127] In the several embodiments provided in this application, it should be understood that the disclosed devices can be implemented in other ways. For example, the device embodiments described above are merely schematic, such as the division of units, which is only a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some service interface, and the indirect coupling or communication connection of devices or units can be electrical or other forms.

[0128] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0129] In addition, the functional units in the various embodiments of the present application may be integrated into a single processing unit, or each unit may exist physically separately, or two or more units may be integrated into a single unit. The aforementioned integrated units may be implemented in the form of hardware or software functional units.

[0130] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable memory. Based on this understanding, the technical solution of this application, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory 305 and includes several instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the method of this application. The aforementioned memory 305 includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a magnetic disk, or an optical disk.

[0131] The present application also discloses a computer-readable storage medium storing instructions, which, when executed by one or more processors 301 , enable an electronic device to execute one or more methods in the above embodiments.

[0132] The above are merely exemplary embodiments of the present disclosure and are not intended to limit the scope of the present disclosure. That is, any equivalent changes and modifications made in accordance with the teachings of the present disclosure are still within the scope of the present disclosure. After considering the disclosure of the specification and the truth of practice, those skilled in the art will easily think of other embodiments of the present disclosure. This application is intended to cover any variations, uses or adaptations of the present disclosure, which follow the general principles of the present disclosure and include common knowledge or customary technical means in the art that are not recorded in the present disclosure. The description and examples are to be regarded as exemplary only, and the scope and spirit of the present disclosure are defined by the claims.

Claims

1. An intelligent recommendation method for investment advisory services, characterized in that: The method comprises: Acquire a historical sample dataset with a timestamp, where the historical sample dataset includes a plurality of samples, each sample is composed of a plurality of features, and the historical sample dataset includes a time decay weight calculated for each of the samples; Constructing a causal structure graph of dynamic constraints, wherein undirected edges are added between nodes of different types included in the causal structure graph, and no edges are constructed between nodes of the same type included in the causal structure graph, and the features correspond one-to-one to the nodes; Constructing a sparse deep neural network model with cross-layer skip connections based on the causal structure graph of the dynamic constraints; Utilize the trained sparse deep neural network model to calculate and output target response features based on the target input obtained in real time; Based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool, candidate products corresponding to several candidate features with the highest similarity are screened out as recommendation results.

2. The intelligent recommendation method for investment consulting services according to claim 1, characterized in that: After selecting candidate products corresponding to several candidate features having the highest similarity as recommendation results 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: For each recommended product included in the recommendation result, obtain a corresponding candidate feature vector from the sparse deep neural network model, where the candidate feature vector is a multidimensional vector, and each dimension corresponds to a candidate feature; For each dimension of the candidate features, based on the column vector corresponding to the dimension of the target response feature in the causal effect matrix, identify the dominant input feature with the maximum causal effect on the target response feature, where the dominant input feature is the input feature corresponding to the maximum value of the column where the environmental driving feature or the behavioral mediating feature is located in the causal effect matrix; According to the type and causal effect value of the dominant input feature, a causal chain logical path is constructed. If the dominant input feature belongs to the environmental driving feature, a causal chain path is generated from the environmental driving feature to the target response feature; if the dominant input feature belongs to the behavioral mediating feature, a causal chain path is generated from the behavioral mediating feature to the target response feature. The causal chain path includes 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 intelligent recommendation method for investment consulting services according to claim 1, characterized in that: The constructing of a sparse deep neural network model with cross-layer skip connections based on the causal structure graph of the dynamic constraints specifically includes: Constructing a structure mask matrix according to the directed edge relationship in the causal structure graph, wherein the structure mask matrix is ​​used to limit the connection legitimacy in the sparse deep neural network model, wherein the structure mask matrix includes a connection mask from the environment-driven feature to the behavior-mediated feature, a connection mask from the behavior-mediated feature to the target response feature, and a jump connection mask from the environment-driven feature to the target response feature, and the element values ​​corresponding to the unallowed connection positions in the structure mask matrix are fixed to zero; Mapping the nodes in the causal structure graph to the structure of the sparse deep neural network model, wherein the nodes corresponding to the environmental driving 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; Initializing the connection weight matrix and bias parameter items, including the first connection weight matrix and the first bias item from the environment-driven feature to the behavior-mediated feature, the second connection weight matrix and the second bias item from the behavior-mediated feature to the behavior-mediated feature, the third connection weight matrix and the third bias item from the behavior-mediated feature to the target response feature, and the jump connection weight matrix and the fourth bias item 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 jump connection weight matrix that are not allowed by the structure mask matrix, after training is completed, 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 process of the sparse deep neural network model, an intermediate representation is calculated using an activation function 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 to obtain a first intermediate representation, where the first intermediate representation is an implicit activation vector of the second input layer; The first intermediate representation and the environment-driven feature jump connection jointly act on the output layer, and rely on the third connection weight matrix and the third bias item as well as the jump connection weight matrix and the fourth bias item to generate an output result.

4. The intelligent recommendation method for investment consulting services according to claim 3, characterized in that: The trained sparse deep neural network model is used to calculate and output a target feature vector based on the target input obtained in real time, specifically including: Receive target input acquired in real time, the target input including a set of real-time environment-driven features and a set of real-time behavior-mediating features, the real-time environment-driven features being input into the first input layer, and the real-time behavior-mediating features being input into the second input layer, the arrangement order and numerical dimensions of the real-time environment-driven features and the real-time behavior-mediating features being consistent with those in the training phase; Performing a matrix operation on the real-time environment driving feature, the first connection weight matrix, and the first bias term to obtain a first operation result, and performing a matrix operation on the real-time behavior mediating feature, the second connection weight matrix, and the second bias term to obtain a second operation result, adding the first operation result and the second operation result, and then calculating an intermediate representation through an activation function to obtain a second intermediate representation, where the second intermediate representation is an implicit activation vector of the second input layer; Performing matrix multiplication on the second intermediate representation and the third connection weight matrix to obtain a first product vector; Performing matrix multiplication calculation on the real-time environment driving feature and the skip connection weight matrix to obtain a second product vector; The first product vector and the second product vector are added together, and then added to the target response feature bias vector. The result is mapped through an activation function to obtain an output vector of the sparse deep neural network model, and the output vector is the target feature vector.

5. The intelligent recommendation method for investment consulting services according to claim 1, characterized in that: The construction of the causal structure diagram of dynamic constraints specifically includes: Divide all features in the historical sample dataset into three types of feature types based on business attributes, namely, environment-driven features, behavior-mediating features, and target-response features. Map each feature to a node based on the division results, with the features corresponding to the nodes one-to-one. Initialize the causal structure graph. In the causal structure graph, add an undirected edge between two nodes of different types, and do not construct any edge for two nodes of the same feature type. The added undirected edges include the edge between the node corresponding to the environment-driven feature and the node corresponding to the behavior-mediating feature, the edge between the node corresponding to the behavior-mediating feature and the node corresponding to the target response feature, and the edge between the node corresponding to the environment-driven feature and the node corresponding to the target response feature.

6. The intelligent recommendation method for investment consulting services according to claim 5, characterized in that: The construction of the causal structure diagram of dynamic constraints specifically includes: A conditional independence test is performed on the target feature, 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, calculating the significance probability value corresponding to the weighted residual correlation between the target feature and the comparison feature under the conditional feature set, 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 among the multiple features; when there is any conditional feature set that makes the significance probability value exceed a preset significance threshold, deleting the undirected edge between the target feature and the comparison feature in the causal structure graph; On the causal structure graph obtained after the conditional independence test, all triples that simultaneously satisfy the following structural characteristics are identified: for each intermediate feature node, if there is no connecting edge between two adjacent feature nodes, the directions of both adjacent edges are directed toward the intermediate feature node, forming a unidirectional causal structure; Performing directional reasoning on the remaining undirected edges, and iteratively completing the direction determination of all the undirected edges; A unidirectional path constraint between feature types is imposed on the constructed directed graph. The unidirectional path constraint includes: retaining unidirectional edges from the node corresponding to the environment-driven feature to the node corresponding to the behavior-mediating feature, from the node corresponding to the behavior-mediating feature to the node corresponding to the target response feature, and from the node corresponding to the environment-driven feature to the node corresponding to the target response feature, and removing all edges that do not satisfy the unidirectional path constraint to generate a causal structure graph with dynamic constraints.

7. The intelligent recommendation method for investment consulting services according to claim 1, characterized in that: The method further comprises: The plurality of features are divided into environment-driven features, behavior-mediating features, and target-response features according to business attributes, wherein the environment-driven features include market features, the behavior-mediating features include customer features, and the target-response features include product features.

8. An intelligent recommendation device for investment consulting services, characterized in that: The device is used to execute an intelligent recommendation method for investment consulting services according to any one of claims 1 to 7, and the device comprises 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 data set with a timestamp, wherein 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 includes a time decay weight calculated for each of the samples; The processing module (202) is used to construct a causal structure graph of dynamic constraints, wherein undirected edges are added between nodes of different types included in the causal structure graph, and no edges are constructed between nodes of the same type included in the causal structure graph, and the features correspond one-to-one to 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 target input obtained in real time using the trained sparse deep neural network model; The output module (203) is used to screen out candidate products corresponding to several candidate features with the highest similarity as recommendation results based on the similarity between the target response feature and the candidate features of the candidate products in the candidate pool.

9. An electronic device, characterized in that: The electronic device comprises a processor (301), a communication bus (302), a user interface (303), a network interface (304) and a memory (305), wherein 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 connection and communication between components in the electronic device, and the processor (301) is used to execute the instructions stored in the memory (305) so that the electronic device executes the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores instructions, and when the instructions are executed, the method according to any one of claims 1 to 7 is executed.

Images