Product recommendation method and device based on artificial intelligence, computer device and medium

Abstract

The application belongs to the technical field of artificial intelligence, and relates to a product recommendation method based on artificial intelligence, which comprises the following steps: acquiring multi-source data of a user from a business system; calling a recommendation model; performing feature extraction on the multi-source data based on a feature encoder to obtain corresponding target features; performing feature fusion on the target features based on an adaptive fusion module to obtain corresponding fusion features; performing decision processing on the fusion features based on a path decision module to obtain corresponding comprehensive features; performing scoring and prediction processing on the comprehensive features and candidate products based on a scoring module to obtain scores of the candidate products; screening all the candidate products based on the scores to obtain a recommended product list, and pushing the recommended product list to the user. The application also provides a product recommendation device based on artificial intelligence, a computer device and a storage medium. The application can be applied to product recommendation business scenarios in the fields of financial technology and medical health, and effectively improves the accuracy of product recommendation.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology and can be applied to fields such as fintech and healthcare, particularly to artificial intelligence-based product recommendation methods, devices, computer equipment, and storage media. Background Technology

[0002] In traditional insurance service models, insurance product recommendations primarily rely on rule engines and static customer tagging systems, resulting in low accuracy. Specifically, traditional methods typically match products based on manually pre-set hard rules (such as "recommend driver and passenger accident insurance to customers who buy car insurance") and static customer profiles (such as age, gender, and existing policy types), lacking the ability to analyze dynamic customer behavior and deeper intentions. This crude recommendation approach struggles to capture real-time changes in customer needs and fails to identify personalized insurance demands in specific scenarios, leading to insufficient matching between recommended products and actual customer needs, thus reducing customer purchase intention and service experience.

[0003] For example, in the home insurance recommendation scenario within the financial insurance sector, if a customer has recently inquired about home insurance via customer service, a traditional rule engine might not recommend it because the customer already has "home insurance" listed in their tag. This overlooks the customer's potential need to increase coverage or insure a newly purchased property, resulting in a lost recommendation opportunity. Similarly, in the health management service recommendation scenario within the healthcare sector, if a customer has recently been frequently browsing content related to hypertension management and repeatedly inquiring about chronic disease medications, traditional methods might only recommend routine physical examination packages based on their age and existing basic medical insurance tags, failing to identify their urgent need for specialized chronic disease management services. This leads to a mismatch between the recommended health management plan and the customer's current health condition.

[0004] Therefore, there is an urgent need to provide an intelligent product recommendation method to improve the accuracy of insurance product recommendations, meet customers' personalized needs, and enhance the overall efficiency of insurance services. Summary of the Invention

[0005] The purpose of this application is to propose a product recommendation method, apparatus, computer equipment, and storage medium based on artificial intelligence, so as to solve the technical problem of low accuracy in product recommendations in traditional insurance service models.

[0006] Firstly, an artificial intelligence-based product recommendation method is provided, including: The system obtains multi-source data of users from a pre-defined business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data. A preset recommendation model is invoked; wherein the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; Based on the feature encoder, feature extraction is performed on the multi-source data to obtain the corresponding target features; Based on the adaptive fusion module, feature fusion is performed on the target features to obtain the corresponding fused features; The fused features are processed by a preset path decision module to obtain the corresponding comprehensive features; The scoring module performs score prediction processing on the comprehensive features and candidate products to obtain the scores of the candidate products; wherein, the number of candidate products is multiple. Based on the rating, all candidate products are filtered to obtain a recommended product list, and the recommended product list is pushed to the user.

[0007] Secondly, an artificial intelligence-based product recommendation device is provided, including: The acquisition module is used to acquire multi-source data of users from a preset business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data; The calling module is used to call a preset recommendation model; wherein, the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; The extraction module is used to extract features from the multi-source data based on the feature encoder to obtain the corresponding target features; The fusion module is used to perform feature fusion on the target features based on the adaptive fusion module to obtain the corresponding fused features; The decision module is used to perform decision processing on the fused features based on the preset path decision module to obtain the corresponding comprehensive features; The prediction module is used to perform score prediction processing on the comprehensive features and candidate products based on the scoring module to obtain the score of the candidate products; wherein, the number of candidate products is multiple; The recommendation module is used to filter all the candidate products based on the rating to obtain a list of recommended products, and push the list of recommended products to the user.

[0008] Thirdly, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described artificial intelligence-based product recommendation method.

[0009] Fourthly, a computer-readable storage medium is provided, which stores a computer program that, when executed by a processor, implements the steps of the above-described artificial intelligence-based product recommendation method.

[0010] In the above-mentioned solution implemented by the AI-based product recommendation method, apparatus, computer equipment, and storage medium, firstly, multi-source data of the user is obtained from a preset business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data; then, a preset recommendation model is invoked; wherein, the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; subsequently, feature extraction is performed on the multi-source data based on the feature encoder to obtain corresponding target features; further, feature fusion is performed on the target features based on the adaptive fusion module to obtain corresponding fused features; and decision processing is performed on the fused features based on the preset path decision module to obtain corresponding comprehensive features; subsequently, score prediction processing is performed on the comprehensive features and candidate products based on the scoring module to obtain the scores of the candidate products; wherein, the number of candidate products is multiple; finally, all candidate products are filtered based on the scores to obtain a recommended product list, and the recommended product list is pushed to the user. Based on the above automated processing flow, this application provides an intelligent product recommendation method. It obtains structured data, unstructured text data, and time-series behavioral sequence data from the business system, and calls a recommendation model containing a feature encoder, adaptive fusion module, dual-path decision module, and scoring module. The method sequentially performs feature extraction, feature fusion, decision processing, and score prediction on the multi-source data, ultimately generating a list of recommended products and pushing it to the user. This method overcomes the shortcomings of traditional methods that rely on rule engines and static labels, failing to analyze dynamic customer behavior and deep intent. It can fully utilize multi-source heterogeneous data to capture customers' real-time changing needs and personalized insurance demands, effectively improving the matching degree between recommended products and customers' actual needs. This significantly improves the accuracy of product recommendations, meets customers' personalized needs, and contributes to enhancing the overall efficiency of insurance services. Attached Figure Description

[0011] To more clearly illustrate the solutions in this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0012] Figure 1 This is an exemplary system architecture diagram to which this application can be applied; Figure 2This is a flowchart of an embodiment of the AI-based product recommendation method according to this application; Figure 3 This is a schematic diagram of a structure of an embodiment of the artificial intelligence-based product recommendation device according to this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the computer device according to this application. Detailed Implementation

[0013] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein in the specification of the application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings of this application, are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings of this application are used to distinguish different objects, not to describe a particular order.

[0014] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0015] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0016] like Figure 1 As shown, system architecture 100 may include terminal device 101, network 102, and server 103. Terminal device 101 may be a laptop 1011, tablet 1012, or mobile phone 1013. Network 102 is used as a medium to provide a communication link between terminal device 101 and server 103. Network 102 may include various connection types, such as wired, wireless communication links, or fiber optic cables, etc.

[0017] Users can use terminal device 101 to interact with server 103 via network 102 to receive or send messages, etc. Various communication client applications can be installed on terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0018] Terminal device 101 can be various electronic devices with a display screen and support web browsing. In addition to laptops 1011, tablets 1012, or mobile phones 1013, terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 player (Moving Picture Experts Group Audio Layer IV), a laptop computer, and a desktop computer, etc.

[0019] Server 103 can be a server that provides various services, such as a backend server that provides support for the pages displayed on terminal device 101.

[0020] It should be noted that the AI-based product recommendation method provided in this application is generally executed by a server / terminal device, and correspondingly, the AI-based product recommendation device is generally located in the server / terminal device.

[0021] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0022] Continue to refer to Figure 2 This document illustrates a flowchart of an embodiment of the AI-based product recommendation method according to this application. The order of steps in the flowchart can be changed, and some steps can be omitted, depending on different requirements. The AI-based product recommendation method provided in this application can be applied to any scenario requiring product recommendations, and thus can be applied to systems in these scenarios, such as recommendation systems in the financial insurance and healthcare fields. The AI-based product recommendation method includes the following steps: Step S201: Obtain multi-source data of users from a preset business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data.

[0023] In this embodiment, the AI-based product recommendation method operates on an electronic device (e.g., Figure 1The server / terminal device shown can acquire multi-source data from users via wired or wireless connections. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, WiFi connections, Bluetooth connections, WiMAX connections, Zigbee connections, UWB (ultrawideband) connections, and other currently known or future-developed wireless connection methods. The implementing entity of this application is specifically a product recommendation system, which can be simply referred to as the system.

[0024] This involves retrieving three types of raw data (i.e., multi-source data) from various business systems of the insurance company for user i. The first type is structured data, which can be extracted from the policy management system, including fields such as the user's age, number of policies, policy type (car insurance / health insurance / home insurance, etc.), total premiums, and historical claims, stored in structured tables. The second type is unstructured text data, which can be extracted from the customer service call center system and online customer service system, containing the user's most recent N conversation records. Each record includes information such as conversation time, conversation content, and communication channel (telephone / online / email). The third type is time-series behavioral sequence data, which can be extracted from the user behavior log system, containing user behavior records arranged chronologically. Each record includes a behavior timestamp, behavior type (browsing / consulting / searching / paying / favoriting, etc.), and behavior object (e.g., which product page was viewed, which question was consulted). Additionally, current context information can be obtained from the system configuration, including the current month (used to infer seasonal characteristics), the ID and type of the currently ongoing marketing campaign, etc.

[0025] Furthermore, the obtained multi-source data can be preprocessed. Specifically, 1) for structured data... Numerical features (such as age, total premium, and number of claims) are normalized and mapped to the [0,1] interval to eliminate dimensional differences; categorical features (such as policy type) are embedded or one-hot encoded and converted into a vector form that the model can process. 2) For unstructured text data Perform basic cleaning: remove meaningless special characters and emoticons, perform word segmentation, and then sort in reverse chronological order (most recent conversations first), ensuring that conversations that are more recent in time have higher weight. 3) Process time-series behavioral sequence data. Each behavior type (browsing / consulting / searching / paying / favoriting) is predefined as a fixed behavior embedding vector. The original behavior records are converted into a time-ordered sequence of embedded vectors. , ,..., It also checks if there are any records in the sequence with abnormal timestamps (such as time reversal), and removes them if so.

[0026] Data acquisition is the starting point of the entire recommendation process, requiring the collection of three types of heterogeneous data—structured, unstructured, and time-series behavioral data—from multiple business systems. Preprocessing aims to transform raw data from different sources and formats into a standardized input format acceptable to the model, while cleaning away obviously erroneous data to ensure the input quality for subsequent feature extraction. The acquisition and preprocessing of the three types of data are conducted independently, but they are ultimately fused during the feature extraction stage.

[0027] Step S202: Invoke the preset recommendation model; wherein the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module.

[0028] In this embodiment, the recommendation model described above is not a single neural network, but a complete system composed of multiple modules connected in series. The complete components of this recommendation model are as follows: First group of components: three feature encoders (each independent), the first encoder (i.e., the first encoder itself) is a fully connected layer that processes structured data, containing a weight matrix. and bias Its function is to transform tabular data such as age, number of policies, and premiums into a vector. The second encoder is a lightweight BERT + attention pooling layer, containing all the parameters of the BERT model and the pooling layer. Its function is to transform the user's customer service dialogue text into a vector. The third encoder (the third encoder) is the gate timing unit (GTU), which contains the gate update mechanism. and Reset the door and Candidate state and Its function is to transform the user's behavior sequence arranged by time into a vector. .

[0029] The second set of components: the adaptive fusion module, which contains a small neural network. Its function is to automatically calculate the weight of each of the three types of features based on the current user's data, and output three attention coefficients. , , Then, the three vectors are concatenated according to their weights into a unified fusion representation. .

[0030] The third component: a dual-path decision module, which contains two independent path networks and a gating mechanism. The first path is... Transforming the fusion representation into a macro-preference vector The second path is Transforming the same fusion representation into micro-intent vectors Then there is a gating layer, which contains and Combining macroscopic vectors, microscopic vectors, and contextual features Input all values ​​together and calculate a scalar between 0 and 1. Finally, the product enhancement formula is used to fuse the outputs of the two paths into the final user representation. .

[0031] The fourth component: the scoring module, which contains a learnable bilinear matrix M and a balancing hyperparameter β, is used to process user representations. embedding vectors of each candidate product Perform a match and calculate a score. The score is obtained by adding the cosine similarity and the bilinear term.

[0032] The fifth component: three loss functions. Main loss. This is a BPR loss with negative sampling, used to ensure that customers rate products they've purchased higher than products they haven't purchased. Orthogonal loss. It is the absolute value of the inner product of the output vectors of the two paths, used to force the two paths to learn different information. Variance loss Used to force the gate value Reasonable fluctuations exist between different samples to avoid using the same proportion for all customers when mixing dual paths.

[0033] Furthermore, the model training process of the aforementioned recommendation model will be described in more detail in subsequent specific embodiments of this application, and will not be elaborated upon here.

[0034] Step S203: Based on the feature encoder, perform feature extraction on the multi-source data to obtain the corresponding target features.

[0035] In this embodiment, the specific implementation process of extracting features from the multi-source data based on the feature encoder to obtain the corresponding target features will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0036] Step S204: Based on the adaptive fusion module, feature fusion is performed on the target features to obtain the corresponding fused features.

[0037] In this embodiment, the specific implementation process of performing feature fusion on the target features based on the adaptive fusion module to obtain the corresponding fused features will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0038] Step S205: Based on the preset path decision module, the fused features are processed to obtain the corresponding comprehensive features.

[0039] In this embodiment, the specific implementation process of the above-mentioned decision-making module based on the preset path decision module to perform decision processing on the fused features to obtain the corresponding comprehensive features will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0040] Step S206: Based on the scoring module, perform score prediction processing on the comprehensive features and candidate products to obtain the score of the candidate products; wherein, the number of candidate products is multiple.

[0041] In this embodiment, the specific implementation process of performing score prediction processing on the comprehensive features and candidate products based on the scoring module to obtain the score of the candidate product will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0042] Step S207: Based on the rating, all candidate products are filtered to obtain a recommended product list, and the recommended product list is pushed to the user.

[0043] In this embodiment, the specific implementation process of filtering all candidate products based on the rating to obtain a recommended product list and pushing the recommended product list to the user will be described in more detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0044] This application first obtains multi-source data of users from a pre-set business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data; then, it calls a pre-set recommendation model; wherein, the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; then, based on the feature encoder, it extracts features from the multi-source data to obtain corresponding target features; further, based on the adaptive fusion module, it performs feature fusion on the target features to obtain corresponding fused features; and based on the pre-set path decision module, it performs decision processing on the fused features to obtain corresponding comprehensive features; subsequently, based on the scoring module, it performs score prediction processing on the comprehensive features and candidate products to obtain the scores of the candidate products; wherein, the number of candidate products is multiple; finally, based on the scores, it filters all the candidate products to obtain a recommended product list, and pushes the recommended product list to the user. Based on the above automated processing flow, this application provides an intelligent product recommendation method. It obtains structured data, unstructured text data, and time-series behavioral sequence data from the business system, and calls a recommendation model containing a feature encoder, adaptive fusion module, dual-path decision module, and scoring module. The method sequentially performs feature extraction, feature fusion, decision processing, and score prediction on the multi-source data, ultimately generating a list of recommended products and pushing it to the user. This method overcomes the shortcomings of traditional methods that rely on rule engines and static labels, failing to analyze dynamic customer behavior and deep intent. It can fully utilize multi-source heterogeneous data to capture customers' real-time changing needs and personalized insurance demands, effectively improving the matching degree between recommended products and customers' actual needs. This significantly improves the accuracy of product recommendations, meets customers' personalized needs, and contributes to enhancing the overall efficiency of insurance services.

[0045] In some optional implementations, the feature encoder includes a first feature encoder, a second feature encoder, and a third feature encoder; step S203 includes the following steps: Step S2031: Based on the first feature encoder, perform feature extraction on the structured data to obtain the corresponding structured features.

[0046] In this embodiment, the first feature encoder can specifically employ a pre-trained fully connected layer (including a pre-trained weight matrix). and bias vector The obtained structured data can be processed... The input is fed into the fully connected layer using the trained weight matrix W_s and bias vector. After performing a linear transformation and adding a bias, nonlinearity is introduced by using the ReLU activation function: .in, This represents matrix multiplication between the weight matrix and the input vector. Set all negative values ​​to 0, and leave positive values ​​unchanged. Output: It is a dense vector (i.e., structured feature) of the user's structured profile, whose dimensions implicitly encode the non-linear interaction relationships between features such as age, number of policies, and premiums.

[0047] Step S2032: Based on the second feature encoder, feature extraction is performed on the unstructured text data to obtain the corresponding text features.

[0048] In this embodiment, the second feature encoder can specifically be a pre-trained lightweight BERT model. This is achieved by processing the obtained unstructured text data. The trained lightweight BERT model is input, and it outputs a sentence vector for each dialogue segment. These sentence vectors are then fed into a trained attention pooling layer, which calculates an attention weight for each dialogue segment. The weights reflect the importance of that dialogue segment to the current recommendation; more recent and semantically relevant dialogues receive higher weights. Finally, these weights are used to perform a weighted sum of the sentence vectors. Output It is a dense vector of user text semantics, i.e., text features, whose semantics are dominated by the content of high-weighted dialogue.

[0049] Step S2033: Based on the third feature encoder, feature extraction is performed on the time-series behavior sequence data to obtain the corresponding behavior sequence features.

[0050] In this embodiment, the third feature encoder can specifically be a gated temporal unit (GTU), which includes trained parameters. (include , , )and (include , , By obtaining the temporal behavior sequence data. Input the trained GTU. Initial hidden state. Let it be the zero vector. The embedding is given for the t-th row of the sequence. (t ranges from 1 to T), calculate sequentially: Update Gate:

[0051] Reset Door:

[0052] Candidate state:

[0053] Final state:

[0054] Where σ is the sigmoid function, For element-wise multiplication, This represents concatenation; after calculating up to t=T, the last hidden state is taken to obtain the final representation of the sequence. Unlike LSTM / GRU, GTU introduces an additional behavior-type dependent parameter into its gating signal, enabling the model to more finely distinguish the impact of different behavior types, such as "browsing," "consulting," and "purchasing," on state transitions. The output... It is a dense vector of user temporal behavior patterns, i.e., behavioral sequence features.

[0055] Step S2034: Integrate the structured features, the text features, and the behavior sequence features to obtain the corresponding integrated features.

[0056] In this embodiment, the extracted structured features, text features, and behavioral sequence features are integrated, and the resulting integrated features are used as the corresponding target features.

[0057] Step S2035: The integrated feature is used as the target feature.

[0058] In this embodiment, the parameters of the three encoders have already learned knowledge of their respective domains during training: and By learning the interaction patterns of structured features, BERT and the pooling layer learn the logic for extracting textual semantics and determining importance, while GTU's gating parameters learn the influence of different behavior types on state transitions. During inference, these parameters are completely fixed and only perform forward computation.

[0059] This application extracts features from the structured data using the first feature encoder to obtain corresponding structured features; simultaneously extracts features from the unstructured text data using the second feature encoder to obtain corresponding text features; and extracts features from the temporal behavior sequence data using the third feature encoder to obtain corresponding behavior sequence features. Then, the structured features, text features, and behavior sequence features are integrated to obtain integrated features; subsequently, these integrated features are used as the target features. Based on this processing flow, the three feature encoders of this application each use pre-trained fixed parameters, enabling automatic and accurate mapping of three types of heterogeneous raw data into dense vectors of a unified dimension, ensuring the accuracy of the generated target features.

[0060] In some optional implementations of this embodiment, step S204 includes the following steps: Step S2041: Based on the adaptive fusion module, the structured features, the text features, and the behavior sequence features are concatenated to obtain the corresponding concatenated data.

[0061] In this embodiment, the above-mentioned structured features can be used. Text features and behavioral sequence features Perform splicing processing to obtain spliced ​​data. .

[0062] Step S2042: Perform attention calculation on the spliced ​​data to obtain the corresponding attention weights.

[0063] In this embodiment, the process of performing attention calculation on the above-mentioned spliced ​​data includes: inputting a pre-trained small neural network. :

[0064] in, It is a pre-trained small, fully connected network of two or three layers, used to learn the relative importance weights α of the three types of features under the current user i's data, and ensured by Softmax normalization. + + =1. These three weights are determined by the current user's data itself: if the user's recent conversations show strong demand signals, It will automatically increase; if the behavioral sequence is very rich. It will automatically increase in height; if the structured image is particularly clear... It will rise automatically.

[0065] Step S2043: Perform weighted concatenation processing on the structured features, the text features, the behavior sequence features, and the attention weights to obtain the corresponding generated features.

[0066] In this embodiment, the three vectors are weighted separately using the calculated attention weights and then concatenated: .

[0067] in, This indicates vector concatenation. The output is... It is a deep fusion feature of user i, which is a unified dense vector that contains information in three dimensions: structured, textual, and behavioral. The contribution of each dimension is adaptively determined by the data.

[0068] Step S2044: Use the generated feature as the fused feature.

[0069] This application uses the adaptive fusion module to concatenate the structured features, text features, and behavioral sequence features to obtain concatenated data. Then, attention is calculated on the concatenated data to obtain corresponding attention weights. Next, the structured features, text features, behavioral sequence features, and attention weights are weighted and concatenated to obtain corresponding generated features. These generated features are then used as the fused features. Based on this processing flow, this application, through the use of the adaptive fusion module, can automatically and accurately integrate three types of heterogeneous features into a unified representation to obtain a highly condensed comprehensive customer profile, which is beneficial for providing accurate data input for subsequent dual-path decision-making.

[0070] In some optional implementations, the path decision module includes a preferred path and an intent path; step S205 includes the following steps: Step S2051: Process the fusion features based on the preference path to obtain the output preference features.

[0071] In this embodiment, the process of processing the fusion feature based on the preference path includes: processing the fusion feature... Input trained : .in, It is a trained multi-layer fully connected network, and its output is... (i.e., preference features) represent a user's long-term stable preference vector. This path learns the user's consistent interests across different time periods.

[0072] Step S2052: Process the fused features based on the intent path to obtain the output intent features.

[0073] In this embodiment, the process of processing the fused features based on the intent path to obtain the output intent features includes: processing the fused features... Input trained : .in, Is with A separate, multi-layered, fully connected network with parameters different from... Sharing, output (i.e., intent features) represent the user's short-term, immediate intent vector. This path learns the user's strongest demand signals over a recent period.

[0074] Step S2053: Process the preference features and intent features based on a preset enhanced fusion strategy to obtain the corresponding processing results.

[0075] In this embodiment, the specific implementation process of processing the preference features and intent features based on the preset enhanced fusion strategy to obtain the corresponding processing results will be further described in detail in subsequent specific embodiments of this application, and will not be elaborated on here.

[0076] Step S2054: The processing result is used as the comprehensive feature.

[0077] This application processes the fusion features based on the preference path to obtain output preference features; then processes the fusion features based on the intent path to obtain output intent features; subsequently, it processes the preference features and intent features based on a preset enhanced fusion strategy to obtain corresponding processing results; finally, the processing results are used as the comprehensive features. Based on the above processing flow, the dual paths of this application use the same input but different network parameters. and This ensures that complementary information is learned. Furthermore, by using an enhanced fusion strategy to process the preference and intent features generated from the dual paths, the recommendations can always stay within the customer's long-term interests, while short-term intents can be precisely adjusted in a specific direction, effectively guaranteeing the accuracy of the generated comprehensive features.

[0078] In some alternative implementations, step S2053 includes the following steps: Step S20531: Obtain preset context information and extract features from the context information to obtain corresponding context features.

[0079] In this embodiment, the current context information, including the current month (used to infer seasonal characteristics, such as summer corresponding to peak travel demand), the ID and type of the currently ongoing marketing campaign (such as "Summer Travel Guarantee Season"), is read from the system configuration and encoded into a fixed-dimensional vector. That is, contextual features.

[0080] Step S20532: The preference feature, the intent feature, and the context feature are concatenated to obtain the corresponding concatenated feature.

[0081] In this embodiment, by using preference features The intent features and the context features By splicing the components together, the corresponding splicing features can be obtained. .

[0082] Step S20533: Calculate and process the splicing features based on a preset gating layer to obtain the corresponding generated data.

[0083] In this embodiment, the aforementioned gating layer is specifically a trained Adaptive Enhancement Gate (AEG). The gating mechanism of this gating layer can dynamically determine whether to rely more on macro preferences or micro intentions in the current recommendation decision.

[0084] in, It is the sigmoid function. It is a scalar between 0 and 1. When A value close to 1 indicates that the system's judgment should rely more on micro-level intentions (strong immediate demand); a value close to 0 indicates that it should rely more on macro-level preferences (stable demand). Context The introduction of this feature allows gating to take external environmental factors into account, such as the context of the "summer travel season". It will be driven up.

[0085] By inputting the above splicing features into the gating layer, the output is processed and then... That is, the data generated above.

[0086] Step S20534: Calculate and process the preference features, intent features, and generated data based on a preset enhancement formula to obtain corresponding enhanced data.

[0087] In this embodiment, the above-mentioned enhancement formula is specifically as follows:

[0088] Among them, the microscopic intention here Influencing macroeconomic preferences through product enhancement Rather than a simple addition, this means that micro-intentions, based on macro-preferences, "amplify" or "suppress" specific dimensions, which aligns better with the cognitive logic that "short-term intentions play a moderating role within the scope of long-term interests." This is fundamentally different from existing attention mechanisms. Furthermore, the aforementioned preference features, intention features, and generated data can be substituted into the above enhancement formula for calculation, and the resulting enhanced data can be used as the corresponding processing result.

[0089] Step S20535: The enhanced data is used as the processing result.

[0090] This application obtains preset context information and extracts features from the context information to obtain corresponding context features. Then, it concatenates the preference features, intent features, and context features to obtain corresponding concatenated features. Next, it calculates and processes the concatenated features based on a preset gating layer to obtain corresponding generated data. Subsequently, it calculates and processes the preference features, intent features, and generated data based on a preset enhancement formula to obtain corresponding enhanced data. Finally, it uses the enhanced data as the processing result. Based on the above processing flow, this application, by using a gating layer, dynamically calculates the mixing ratio based on macro-vectors, micro-vectors, and context, and then enhances and fuses preference features, intent features, and generated data based on the enhancement formula. This helps ensure that recommendations always remain within the customer's long-term interest range, thereby improving the accuracy and adaptability of product recommendations.

[0091] In some optional implementations of this embodiment, step S206 includes the following steps: Step S2061: Invoke a preset candidate product library; wherein the candidate product library includes multiple candidate products.

[0092] In this embodiment, the system pre-builds a candidate product library, which includes multiple candidate products, and each candidate product j has embedded features learned together with the recommendation model during the training phase. It remains constant during reasoning.

[0093] Step S2062: Obtain the embedded features corresponding to the specified candidate product.

[0094] In this embodiment, the designated candidate product is any one of all the candidate products.

[0095] Step S2063: Based on the scoring module, a preset scoring algorithm is used to calculate and process the comprehensive feature and the embedded feature to obtain the corresponding calculation result.

[0096] In this embodiment, for each candidate product j, a predicted score is calculated using a trained bilinear matrix M and a balanced hyperparameter β:

[0097] in, As a comprehensive feature, The embedded features corresponding to the candidate products, =( · ) / (|| ||·|| ||) Measures the directional consistency between user representations and product representations, i.e., overall similarity. The second item... The terms are bilinear, M is a trained learnable matrix (used to capture more complex, asymmetric cross-features; for example, a certain user feature may be particularly inclined towards a certain type of product feature, but the reverse is not necessarily true), and β is a trained scalar (i.e., a balanced hyperparameter). Both capture asymmetric cross-features, such as a specific strong correlation between "users with health insurance" and "travel insurance including acute illness medical care." Furthermore, the above-mentioned comprehensive features and embedded features can be substituted into the scoring algorithm for calculation, and the result can be used as the specified score for a given candidate product.

[0098] Step S2064: Use the calculation result as the designated score for the designated candidate product.

[0099] Based on the above processing flow, this application calls a preset candidate product library, which includes multiple candidate products; then obtains the embedded features corresponding to a specified candidate product, where the specified candidate product is any one of all the candidate products; subsequently, based on the scoring module, a preset scoring algorithm is used to calculate and process the comprehensive features and the embedded features to obtain the corresponding calculation results; the calculation results are then used as the specified score for the specified candidate product, thereby automatically and accurately calculating the score of the candidate product based on the use of the scoring module, ensuring the accuracy of the obtained candidate product scores.

[0100] In some optional implementations of this embodiment, step S207 includes the following steps: Step S2071: Sort all the candidate products according to the score from high to low to obtain the corresponding sorting result.

[0101] In this embodiment, all candidate products are categorized according to... Sort from highest to lowest to obtain the sorted result.

[0102] Step S2072: Extract a specified number of specified products from all the sorting results.

[0103] In this embodiment, the selection of a specified number of Top-K is not specifically limited and can be set according to actual business needs, for example, it can be set to 3.

[0104] Step S2073: Construct a list of all the specified products to obtain the corresponding recommended product list.

[0105] In this embodiment, the selected products can be filled into a preset blank list in descending order of their ratings, and the product ID, product name, and rating of each product can be entered accordingly to obtain a corresponding list of recommended products.

[0106] Step S2074: Push the recommended product list to the user based on a preset push method.

[0107] In this embodiment, the generated list of recommended products can be pushed to the aforementioned users through push methods such as email, message, or interface display.

[0108] Based on the above processing flow, this application sorts all candidate products according to their scores from high to low to obtain corresponding sorting results. Then, it extracts a specified number of specified products from all the sorting results and constructs a list of all the specified products to obtain a corresponding recommended product list. This enables the automatic and intelligent construction of a recommended product list that meets user needs. Subsequently, the recommended product list is pushed to the user based on a preset push method, thereby improving the accuracy of product recommendations and enhancing the user experience.

[0109] In some optional implementations, the model training process of the above recommendation model includes: I. Training Data Preparation. Extract multi-source data for each customer i from the historical business system, including structured data. Unstructured text Temporal Behavior Sequence Context The data includes a list of value-added service products that the customer has actually purchased (positive samples). Simultaneously, products that the customer has not purchased are randomly selected from all products as negative samples, forming training positive and negative sample pairs.

[0110] II. Complete Training Process.

[0111] Step 1: Forward computation. This involves executing the entire inference process for customer i's data: using the current parameters (initially random values), sequentially calculating the three types of feature codes. , , (Calculated step by step by GTU), and then obtained through adaptive fusion. Then, through two paths, we obtain , Then through AEG gating Then, through product enhancement fusion, we obtain Finally, a score is calculated for each candidate product j. .

[0112] Step 2: Calculate the main loss By employing Bayesian Personalized Ranking (BPR) loss with negative sampling, the goal is to ensure that positive sample scores are higher than negative sample scores. For customer i, one product is drawn from each of its positive and negative sample sets, and the following calculation is performed: =-∑lnσ( (just)- (Negative) Here, σ is the sigmoid function, which sums over all positive and negative sample pairs. The loss is smaller when the positive sample score is higher than the negative sample score, and larger when the negative sample score is lower. The training objective is to minimize this loss so that purchased products are ranked higher than unpurchased products.

[0113] Step 3: Calculate the orthogonal loss By obtaining the macro-preference vector of customer i. With micro-intention vector Calculate the absolute value of their inner product: =| ^T· This loss penalizes situations where the output vectors of the two paths are in similar directions. A larger absolute value of the inner product indicates more redundant information learned by the two paths; a smaller absolute value (approaching orthogonality) indicates stronger complementarity and independent learning by the two paths. This loss forces the two paths to learn different information.

[0114] Step 4: Calculate the variance loss By taking the threshold values ​​of a batch of customers (e.g., 256). Calculate this batch variance: =Var( )=(1 / N)∑( -mean(g))^2. If the variance is too small, it indicates that all customers... If the variances converge (e.g., all close to 0.5), then AEG gating loses its adaptive significance. This loss penalizes cases where the variance is too small, forcing different samples to... Maintaining reasonable differences allows the gating mechanism to truly play an adaptive adjustment role.

[0115] Step 5: Calculate the total loss L. This is done by weighted summing of the three losses: .in, and These are hyperparameters that control the influence of orthogonal loss and variance loss, respectively. If the value is too large, the dual paths will be over-regularized, and it may be unable to learn effective information. If the value is too small, the dual paths may become redundant. If the value is too large, the gating variance will be excessively compressed; If the value is too small, the gating may collapse to a fixed value. The two hyperparameters are typically determined through validation set tuning.

[0116] Step Six: Backpropagation and Parameter Update. Based on the total loss L, calculate the gradient of the loss with respect to all parameters using gradient descent optimizers such as Adam, and synchronously update all parameters along the reverse gradient direction, including: , , , , , , , BERT parameters, pooling layer parameters parameter, parameter, parameter, , The matrix M and β are used. All parameters are updated simultaneously, with the update direction reducing the total loss L in the next forward computation.

[0117] Step 7: Iterate training until convergence. Divide the training data into batches (e.g., 256 customers per batch) and repeatedly execute steps 1 through 6, traversing the entire training set multiple times (e.g., 50 times). After each round, evaluate the recommendation accuracy using the validation set. If the validation set metric no longer improves, stop training.

[0118] At the end of training, all parameters converge to their optimal or near-optimal state; this set of fixed parameters constitutes the trained recommendation model. Subsequent inference phases directly use this set of parameters to perform forward computation on new customer data, without involving any further parameter updates or loss calculations.

[0119] In some optional implementations, the system also has the function of generating interpretable text, and the specific implementation process includes: 1) Backtracking gate value: calculated during the read operation. Value. If >0.7, judged as "micro-intention-driven recommendation"; if If the value is less than 0.3, it is classified as "macro-preference-driven recommendation"; if 0.3 ≤ If the value is ≤0.7, it is judged as "comprehensive recommendation".

[0120] 2) Backtracking fusion attention weights: Read the calculated values , , This determines which of the three feature classes contributes the most to the recommendation. At its highest level, this indicates that the text dialogue is the primary basis; if The highest level indicates that the behavioral sequence is the primary basis; if The highest score indicates that structured profiling is the primary basis for analysis.

[0121] 3) Backtrack the attention weights within the text: Read the attention weights calculated by the attention pooling layer for each dialogue, find the 1-2 dialogues with the highest weights, and extract the key semantics (such as "travel", "renovation", "children going to school" etc.).

[0122] 4) Text Combination: Combine the above backtracking information into natural language according to a preset template. The template structure is: "The system detected your recent [key text content] ([text feature contribution]%), combined with your long-term [macro-preference content] (macro-preference path), and in the current [context content] context, the system determines your current [intent strength description] (intent gating value g_i=[numerical value]), and recommends [product name] ([driving path description]) for you."

[0123] The purpose of interpretability generation is to help customers understand why they received a recommendation, increasing transparency and trust. By retrospectively analyzing gating values, it can be determined whether the recommendation is driven by long-term preferences or recent intentions; by retrospectively analyzing attention weights, it can be determined which data played a key role in the recommendation. This interpretability is not artificially added after the fact, but rather extracted directly from intermediate variables within the model, ensuring consistency between the explanation and the recommendation result.

[0124] In some alternative implementations, the user information obtained is subject to user consent and complies with relevant laws and policies.

[0125] Furthermore, any software tools or components not belonging to our company that appear in the embodiments of this application are merely illustrative examples and do not represent actual use.

[0126] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0127] It should be emphasized that, to further ensure the privacy and security of the above-mentioned recommended product list, the recommended product list can also be stored in a blockchain node.

[0128] The blockchain referred to in this application is a novel application model of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanisms, and encryption algorithms. Essentially, a blockchain is a decentralized database, a chain of data blocks linked together using cryptographic methods. Each data block contains information about a batch of network transactions, used to verify the validity of the information (anti-counterfeiting) and generate the next block. A blockchain can include an underlying blockchain platform, a platform product service layer, and an application service layer.

[0129] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0130] Foundational technologies for artificial intelligence generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies mainly encompass computer vision, robotics, biometrics, speech processing, natural language processing, and machine learning / deep learning.

[0131] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0132] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0133] Further reference Figure 3 As a response to the above Figure 2 The implementation of the method shown in this application provides an embodiment of an artificial intelligence-based product recommendation device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0134] like Figure 3As shown, the AI-based product recommendation device 300 described in this embodiment includes: an acquisition module 301, a retrieval module 302, an extraction module 303, a fusion module 304, a decision-making module 305, a prediction module 306, and a recommendation module 307. Wherein: The acquisition module 301 is used to acquire multi-source data of users from a preset business system; wherein, the multi-source data includes structured data, unstructured text data and time-series behavioral sequence data; Module 302 is used to invoke a preset recommendation model; wherein the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module. Extraction module 303 is used to extract features from the multi-source data based on the feature encoder to obtain the corresponding target features; The fusion module 304 is used to perform feature fusion on the target features based on the adaptive fusion module to obtain corresponding fused features; The decision module 305 is used to perform decision processing on the fused features based on a preset path decision module to obtain the corresponding comprehensive features; The prediction module 306 is used to perform score prediction processing on the comprehensive features and candidate products based on the scoring module to obtain the score of the candidate products; wherein, the number of candidate products is multiple. The recommendation module 307 is used to filter all the candidate products based on the rating to obtain a recommended product list, and push the recommended product list to the user.

[0135] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based product recommendation method in the aforementioned implementation method, and will not be repeated here.

[0136] In some optional implementations of this embodiment, the feature encoder includes a first feature encoder, a second feature encoder, and a third feature encoder; the extraction module 303 includes: The first extraction submodule is used to extract features from the structured data based on the first feature encoder to obtain the corresponding structured features; The second extraction submodule is used to extract features from the unstructured text data based on the second feature encoder to obtain the corresponding text features; The third extraction submodule is used to extract features from the time-series behavioral sequence data based on the third feature encoder to obtain the corresponding behavioral sequence features. The integration submodule is used to integrate the structured features, the text features, and the behavior sequence features to obtain the corresponding integrated features; The first determining submodule is used to use the integrated feature as the target feature.

[0137] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based product recommendation method in the aforementioned implementation method, and will not be repeated here.

[0138] In some optional implementations of this embodiment, the fusion module 304 includes: The first splicing submodule is used to splice the structured features, text features and behavior sequence features based on the adaptive fusion module to obtain the corresponding spliced ​​data; The first calculation submodule is used to perform attention calculation on the spliced ​​data to obtain the corresponding attention weights; The second splicing submodule is used to perform weighted splicing processing on the structured features, the text features, the behavior sequence features, and the attention weights to obtain the corresponding generated features; The second determining submodule is used to use the generated features as the fused features.

[0139] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based product recommendation method in the aforementioned implementation method, and will not be repeated here.

[0140] In some optional implementations of this embodiment, the decision module 305 includes: The first processing submodule is used to process the fused features based on the preference path to obtain the output preference features; The second processing submodule is used to process the fused features based on the intent path to obtain the output intent features; The third processing submodule is used to process the preference features and the intent features based on a preset enhanced fusion strategy to obtain the corresponding processing results; The third determining submodule is used to use the processing result as the comprehensive feature.

[0141] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based product recommendation method in the aforementioned implementation method, and will not be repeated here.

[0142] In some optional implementations of this embodiment, the third processing submodule includes: An extraction unit is used to acquire preset context information and perform feature extraction on the context information to obtain corresponding context features; The splicing unit is used to splice the preference feature, the intent feature, and the context feature to obtain the corresponding spliced ​​feature; The first computing unit is used to perform calculations on the splicing features based on a preset gating layer to obtain the corresponding generated data; The second calculation unit is used to perform calculations on the preference features, the intent features, and the generated data based on a preset enhancement formula to obtain corresponding enhanced data. A determining unit is used to take the enhanced data as the processing result.

[0143] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based product recommendation method in the aforementioned implementation method, and will not be repeated here.

[0144] In some optional implementations of this embodiment, the prediction module 306 includes: The calling submodule is used to call a preset candidate product library; wherein, the candidate product library includes multiple candidate products; The acquisition submodule is used to acquire the embedded features corresponding to the specified candidate product; wherein, the specified candidate product is any one of all the candidate products; The second calculation submodule is used to perform calculations on the comprehensive features and the embedded features based on the scoring module using a preset scoring prediction algorithm to obtain the corresponding calculation results. The fourth determination submodule is used to use the calculation result as the designated score of the designated candidate product.

[0145] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based product recommendation method in the aforementioned implementation method, and will not be repeated here.

[0146] In some optional implementations of this embodiment, the recommended module 307 includes: The sorting submodule is used to sort all the candidate products in descending order of their scores to obtain the corresponding sorting results. The fourth extraction submodule is used to extract a specified number of specified products from all the sorting results; A construction submodule is used to build a list of all the specified products to obtain the corresponding recommended product list; The push module is used to push the recommended product list to the user based on a preset push method.

[0147] In this embodiment, the operations performed by the above modules or units correspond one-to-one with the steps of the artificial intelligence-based product recommendation method in the aforementioned implementation method, and will not be repeated here.

[0148] To address the aforementioned technical problems, embodiments of this application also provide a computer device. Please refer to [link / reference] for details. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0149] The computer device 4 includes a memory 41, a processor 42, and a network interface 43 that are interconnected via a system bus. It should be noted that only the computer device 4 with components 41-43 is shown in the figure; however, it should be understood that it is not required to implement all the shown components, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), digital signal processors (DSPs), embedded devices, etc.

[0150] The computer device can be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device can interact with the user via a keyboard, mouse, remote control, touchpad, or voice control.

[0151] The memory 41 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 41 may be an internal storage unit of the computer device 4, such as the hard disk or memory of the computer device 4. In other embodiments, the memory 41 may also be an external storage device of the computer device 4, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the computer device 4. Of course, the memory 41 may also include both the internal storage unit and its external storage device of the computer device 4. In this embodiment, the memory 41 is typically used to store the operating system and various application software installed on the computer device 4, such as computer-readable instructions for product recommendation methods based on artificial intelligence. In addition, the memory 41 can also be used to temporarily store various types of data that have been output or will be output.

[0152] In some embodiments, the processor 42 may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chip. The processor 42 is typically used to control the overall operation of the computer device 4. In this embodiment, the processor 42 is used to execute computer-readable instructions stored in the memory 41 or to process data, for example, to execute computer-readable instructions for the AI-based product recommendation method.

[0153] The network interface 43 may include a wireless network interface or a wired network interface, which is typically used to establish communication connections between the computer device 4 and other electronic devices.

[0154] Compared with the prior art, the embodiments of this application have the following beneficial effects: In this embodiment, firstly, multi-source data of the user is obtained from a preset business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data; then, a preset recommendation model is invoked; wherein, the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; then, feature extraction is performed on the multi-source data based on the feature encoder to obtain corresponding target features; further, feature fusion is performed on the target features based on the adaptive fusion module to obtain corresponding fused features; and decision processing is performed on the fused features based on the preset path decision module to obtain corresponding comprehensive features; subsequently, score prediction processing is performed on the comprehensive features and candidate products based on the scoring module to obtain the scores of the candidate products; wherein, the number of candidate products is multiple; finally, all candidate products are filtered based on the scores to obtain a recommended product list, and the recommended product list is pushed to the user. Based on the above automated processing flow, this application provides an intelligent product recommendation method. It obtains structured data, unstructured text data, and time-series behavioral sequence data from the business system, and calls a recommendation model containing a feature encoder, adaptive fusion module, dual-path decision module, and scoring module. The method sequentially performs feature extraction, feature fusion, decision processing, and score prediction on the multi-source data, ultimately generating a list of recommended products and pushing it to the user. This method overcomes the shortcomings of traditional methods that rely on rule engines and static labels, failing to analyze dynamic customer behavior and deep intent. It can fully utilize multi-source heterogeneous data to capture customers' real-time changing needs and personalized insurance demands, effectively improving the matching degree between recommended products and customers' actual needs. This significantly improves the accuracy of product recommendations, meets customers' personalized needs, and contributes to enhancing the overall efficiency of insurance services.

[0155] This application also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the artificial intelligence-based product recommendation method described above.

[0156] Compared with the prior art, the embodiments of this application have the following main advantages: In this embodiment, firstly, multi-source data of the user is obtained from a preset business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data; then, a preset recommendation model is invoked; wherein, the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; then, feature extraction is performed on the multi-source data based on the feature encoder to obtain corresponding target features; further, feature fusion is performed on the target features based on the adaptive fusion module to obtain corresponding fused features; and decision processing is performed on the fused features based on the preset path decision module to obtain corresponding comprehensive features; subsequently, score prediction processing is performed on the comprehensive features and candidate products based on the scoring module to obtain the scores of the candidate products; wherein, the number of candidate products is multiple; finally, all candidate products are filtered based on the scores to obtain a recommended product list, and the recommended product list is pushed to the user. Based on the above automated processing flow, this application provides an intelligent product recommendation method. It obtains structured data, unstructured text data, and time-series behavioral sequence data from the business system, and calls a recommendation model containing a feature encoder, adaptive fusion module, dual-path decision module, and scoring module. The method sequentially performs feature extraction, feature fusion, decision processing, and score prediction on the multi-source data, ultimately generating a list of recommended products and pushing it to the user. This method overcomes the shortcomings of traditional methods that rely on rule engines and static labels, failing to analyze dynamic customer behavior and deep intent. It can fully utilize multi-source heterogeneous data to capture customers' real-time changing needs and personalized insurance demands, effectively improving the matching degree between recommended products and customers' actual needs. This significantly improves the accuracy of product recommendations, meets customers' personalized needs, and contributes to enhancing the overall efficiency of insurance services.

[0157] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0158] Obviously, the embodiments described above are only some embodiments of this application, not all embodiments. The accompanying drawings show preferred embodiments of this application, but do not limit the patent scope of this application. This application can be implemented in many different forms; rather, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of this application. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this application's specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the scope of patent protection of this application.

[0159] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

Claims

1. A product recommendation method based on artificial intelligence, characterized in that, Includes the following steps: The system obtains multi-source data of users from a pre-defined business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data. A preset recommendation model is invoked; wherein the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; Based on the feature encoder, feature extraction is performed on the multi-source data to obtain the corresponding target features; Based on the adaptive fusion module, feature fusion is performed on the target features to obtain the corresponding fused features; The fused features are processed by a preset path decision module to obtain the corresponding comprehensive features; The scoring module performs score prediction processing on the comprehensive features and candidate products to obtain the scores of the candidate products; wherein, the number of candidate products is multiple. Based on the rating, all candidate products are filtered to obtain a recommended product list, and the recommended product list is pushed to the user.

2. The product recommendation method based on artificial intelligence according to claim 1, characterized in that, The feature encoder includes a first feature encoder, a second feature encoder, and a third feature encoder; the step of extracting features from the multi-source data based on the feature encoder to obtain the corresponding target features specifically includes: Based on the first feature encoder, feature extraction is performed on the structured data to obtain the corresponding structured features; Based on the second feature encoder, feature extraction is performed on the unstructured text data to obtain the corresponding text features; Based on the third feature encoder, feature extraction is performed on the time-series behavioral sequence data to obtain the corresponding behavioral sequence features; The structured features, text features, and behavioral sequence features are integrated to obtain the corresponding integrated features; The integrated feature is used as the target feature.

3. The product recommendation method based on artificial intelligence according to claim 2, characterized in that, The step of fusing the target features based on the adaptive fusion module to obtain the corresponding fused features specifically includes: The structured features, text features, and behavior sequence features are concatenated using the adaptive fusion module to obtain the corresponding concatenated data. Attention calculations are performed on the spliced ​​data to obtain the corresponding attention weights; The structured features, text features, behavior sequence features, and attention weights are weighted and concatenated to obtain the corresponding generated features; The generated features are used as the fused features.

4. The product recommendation method based on artificial intelligence according to claim 1, characterized in that, The path decision module includes a preference path and an intent path; the step of processing the fused features based on the preset path decision module to obtain the corresponding comprehensive features specifically includes: The fusion features are processed based on the preference path to obtain the output preference features; The fused features are processed based on the intent path to obtain the output intent features; The preference features and intent features are processed based on a preset enhanced fusion strategy to obtain the corresponding processing results; The processing result is used as the comprehensive feature.

5. The product recommendation method based on artificial intelligence according to claim 4, characterized in that, The step of processing the preference features and intent features based on a preset enhanced fusion strategy to obtain the corresponding processing results specifically includes: Obtain preset context information, and extract features from the context information to obtain corresponding context features; The preference feature, the intent feature, and the context feature are concatenated to obtain the corresponding concatenated feature; The splicing features are calculated and processed based on a preset gating layer to obtain the corresponding generated data; The preference features, intent features, and generated data are calculated and processed based on a preset enhancement formula to obtain corresponding enhanced data. The enhanced data is used as the processing result.

6. The product recommendation method based on artificial intelligence according to claim 1, characterized in that, The step of performing score prediction processing on the comprehensive features and candidate products based on the scoring module to obtain the score of the candidate products specifically includes: Call a preset candidate product library; wherein, the candidate product library includes multiple candidate products; Obtain the embedded features corresponding to a specified candidate product; wherein, the specified candidate product is any one of all the candidate products; Based on the scoring module, a preset scoring prediction algorithm is used to calculate and process the comprehensive features and the embedded features to obtain the corresponding calculation results; The calculation result is used as the designated score for the specified candidate product.

7. The product recommendation method based on artificial intelligence according to claim 1, characterized in that, The step of filtering all candidate products based on the rating to obtain a recommended product list and pushing the recommended product list to the user specifically includes: All candidate products are sorted according to their scores from highest to lowest to obtain the corresponding sorting results; Extract a specified number of the specified products from all the sorting results; A list is constructed for all the specified products to obtain the corresponding recommended product list; The recommended product list is pushed to the user based on a preset push method.

8. A product recommendation device based on artificial intelligence, characterized in that, include: The acquisition module is used to acquire multi-source data of users from a preset business system; wherein, the multi-source data includes structured data, unstructured text data, and time-series behavioral sequence data; The calling module is used to call a preset recommendation model; wherein, the recommendation model includes a feature encoder, an adaptive fusion module, a dual-path decision module, and a scoring module; The extraction module is used to extract features from the multi-source data based on the feature encoder to obtain the corresponding target features; The fusion module is used to perform feature fusion on the target features based on the adaptive fusion module to obtain the corresponding fused features; The decision module is used to perform decision processing on the fused features based on the preset path decision module to obtain the corresponding comprehensive features; The prediction module is used to perform score prediction processing on the comprehensive features and candidate products based on the scoring module to obtain the score of the candidate products; wherein, the number of candidate products is multiple; The recommendation module is used to filter all the candidate products based on the rating to obtain a list of recommended products, and push the list of recommended products to the user.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the artificial intelligence-based product recommendation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the artificial intelligence-based product recommendation method as described in any one of claims 1 to 7.

Images