Product recommendation method and device, electronic equipment and medium

Abstract

The embodiment of the invention provides a product recommendation method and device, electronic equipment and a medium, belongs to the technical field of artificial intelligence, and is suitable for financial scenes. The method comprises the following steps: performing feature alignment on dialogue features and behavior features extracted from user interaction data to obtain target alignment features; performing product demand prediction according to the target alignment feature to obtain a product demand probability; performing recommendation prediction according to the product demand probability and the target alignment feature to obtain target recommendation data; and generating a target product recommendation scheme according to the target recommendation product, and pushing the target product recommendation scheme at the product recommendation time. According to the embodiment of the invention, the product demand probability is predicted based on the dialogue characteristics and the behavior characteristics to generate the product recommendation scheme, and the scheme is pushed at the product recommendation time, so that the timeliness and effectiveness of the recommendation scheme can be ensured, the satisfaction degree and the transaction rate of the user are improved, and the product recommendation effect is improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, applied in financial scenarios, and particularly to a product recommendation method and apparatus, electronic device and medium. Background Technology

[0002] Traditional product recommendation methods typically involve customer service representatives making recommendations based on static customer profiles and customer questions. For example, in a financial scenario, if a customer inquires about "data security" or "hacking attacks," and the customer profile includes the tag "purchasing corporate property insurance," then "cybersecurity insurance" might be recommended. However, this method focuses solely on static customer tags and conversation data, failing to capture dynamic changes in customer characteristics and neglecting the customer's true product needs. This leads to a mismatch between recommended products and the customer's actual pain points, resulting in poor product recommendation effectiveness. Therefore, improving product recommendation effectiveness has become a pressing issue. Summary of the Invention

[0003] The main objective of this application is to provide a product recommendation method, apparatus, electronic device, and medium, which aims to improve the product recommendation effect.

[0004] To achieve the above objectives, a first aspect of this application proposes a product recommendation method, the method comprising: In response to a product recommendation request, user interaction data of the target user on a pre-built product platform is obtained; wherein, the user interaction data is multimodal data; Feature extraction is performed on the user interaction data to obtain user interaction features; wherein, the user interaction features include the dialogue features and behavioral features of the target user; The dialogue features and the behavior features are aligned to obtain the target alignment features; Based on the target alignment features, product demand is predicted to obtain the product demand probability. Recommendation prediction is performed based on the product demand probability and the target alignment features to obtain target recommendation data, wherein the target recommendation data includes target recommended products and the recommendation time of the target recommended products; A target product recommendation plan is generated based on the target recommended product, and the target product recommendation plan is pushed to the target user at the product recommendation time.

[0005] In some embodiments, the step of predicting product demand based on the target alignment features to obtain the product demand probability includes: The target alignment features are convolved to obtain user preference product features; Based on the target alignment features, a hypergraph transformation is performed to obtain the user preference product graph features; The product demand features are obtained by performing spatiotemporal convolution processing based on the user preference product features and the user preference product graph features. Based on the product demand characteristics, product demand is inferred to obtain the product demand probability.

[0006] In some embodiments, the step of inferring product demand based on the product demand characteristics to obtain the product demand probability includes: Based on a preset product recommendation strategy, the product demand characteristics are evaluated to obtain an initial product demand prediction probability. The product recommendation strategy is reinforced by the initial product demand prediction probability to obtain an updated product recommendation strategy. The initial product demand features are updated according to the updated product recommendation strategy to obtain updated product demand features. Based on the updated product demand characteristics, product demand is assessed to obtain the product demand probability.

[0007] In some embodiments, the step of performing recommendation prediction based on the product demand probability and the target alignment features to obtain target recommendation data includes: Based on the product demand probability and the target alignment features, a recommended product analysis is performed to obtain the target recommended product; The timing of product recommendations is determined by evaluating the probability of product demand and the target alignment features.

[0008] In some embodiments, the step of performing recommended product analysis based on the product demand probability and the target alignment features to obtain the target recommended product includes: The user preference product graph features are fused with a pre-constructed product knowledge graph to obtain a user-product association graph. Based on the product demand probability, product demand propagation is performed on the user product association graph to obtain candidate product nodes; The product relevance is calculated based on the candidate product nodes and the target users to obtain the candidate product relevance. The target recommended product is determined based on the relevance of the candidate products.

[0009] In some embodiments, the step of evaluating the recommendation timing based on the product demand probability and the target alignment features to obtain the product recommendation timing includes: Temporal encoding is performed based on the target alignment features to obtain temporal features; The time-series features are subjected to time-decay weighting to obtain weighted time-series features; The recommendation time sensitivity is obtained by calculating the recommendation time based on the product demand probability and the weighted time series features. Based on the recommended time sensitivity, the timing of the recommendation is analyzed to obtain the timing of the product recommendation.

[0010] In some embodiments, aligning the dialogue features and the behavior features to obtain target aligned features includes: The dialogue features and the behavior features are projected into a feature space to obtain initial alignment features; wherein, the initial alignment features include projected dialogue features and projected behavior features, and the projected dialogue features and the projected behavior features are located in the same vector dimension space; Target anchor point data is obtained based on the user interaction data, and feature extraction is performed on the target anchor point data to obtain target anchor point features; Alignment deviation detection is performed based on the target anchor point features, the projection dialogue features, and the projection behavior features to obtain the degree of alignment feature deviation. The projection dialogue features and the projection behavior features are subjected to feature distribution recognition to obtain alignment feature distribution information; The initial alignment features are updated based on the preset modal difference penalty coefficient, the degree of deviation of the alignment features, and the distribution information of the alignment features to obtain the target alignment features.

[0011] To achieve the above objectives, a second aspect of this application provides a product recommendation device, the device comprising: The user interaction data acquisition module is used to acquire user interaction data of the target user on a pre-built product platform in response to a product recommendation request; wherein, the user interaction data is multimodal data. The feature extraction module is used to extract features from the user interaction data to obtain user interaction features; wherein, the user interaction features include the dialogue features and behavioral features of the target user; The feature alignment module is used to align the dialogue features and the behavior features to obtain the target aligned features; The product demand prediction module is used to predict product demand based on the target alignment features and obtain the product demand probability. The recommendation prediction module is used to make recommendation predictions based on the product demand probability and the target alignment features to obtain target recommendation data, wherein the target recommendation data includes target recommended products and the recommendation time of the target recommended products; The product recommendation module is used to generate a target product recommendation plan based on the target recommended product, and push the target product recommendation plan to the target user at the product recommendation time.

[0012] To achieve the above objectives, a third aspect of the present application provides an electronic device, the electronic device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method described in the first aspect.

[0013] To achieve the above objectives, a fourth aspect of the present application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method of the first aspect described above.

[0014] The product recommendation method, apparatus, electronic device, and medium proposed in this application first extract user interaction features, including dialogue and behavioral features of the target user, from user interaction data. This dynamically captures user intent and dynamically changing real-time user behavior within the interaction data. The dialogue and behavioral features are aligned to ensure that different modal features can be compared and fused in a unified feature space, facilitating accurate reflection of the user's true product needs and preferences. Second, product demand is predicted based on the target alignment features, and recommendation prediction is performed based on the predicted product demand probability and the target alignment features. This yields target recommendation data, including the target recommended product and its recommendation time. This allows understanding the intensity of user demand for different products, deriving products that meet user needs, and ensuring that product recommendations are made when users are most likely to be interested or need them most. This avoids pushing products during busy times, which could lead to customer resentment or even complaints, thus improving the conversion rate of product recommendations. Finally, a target product recommendation plan is generated based on the target recommended product and pushed to the target user at the recommended time. This ensures the timeliness and effectiveness of the recommendation plan, improving user satisfaction and conversion rate, thereby enhancing the overall product recommendation effect. Attached Figure Description

[0015] Figure 1 This is a flowchart of the product recommendation method provided in the embodiments of this application; Figure 2 yes Figure 1 The flowchart of step S103 in the process; Figure 3 yes Figure 1 The flowchart of step S104 in the process; Figure 4 yes Figure 3 The flowchart of step S305 in the document; Figure 5 yes Figure 1 The flowchart of step S105 in the process; Figure 6 yes Figure 5The flowchart of step S501 in the text; Figure 7 yes Figure 5 The flowchart of step S502 in the document; Figure 8 This is a schematic diagram of the product recommendation device provided in the embodiments of this application; Figure 9 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0018] 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 belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0019] First, let's analyze some of the terms used in this application: Artificial intelligence (AI) is a new branch of computer science that studies, develops, and applies theories, methods, technologies, and systems to simulate, extend, and expand human intelligence. It aims to understand the essence of intelligence and produce intelligent machines that can react in a way similar to human intelligence. Research in this field includes robotics, speech recognition, image recognition, natural language processing, and expert systems. AI can simulate the information processes of human consciousness and thought. Furthermore, AI utilizes digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceiving the environment, acquiring knowledge, and using that knowledge to achieve optimal results.

[0020] This application provides a product recommendation method, apparatus, electronic device, and medium, aiming to improve the product recommendation effect.

[0021] The product recommendation method, apparatus, electronic device, and medium provided in this application are specifically described through the following embodiments. First, the product recommendation method in this application is described.

[0022] The embodiments of this application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use 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.

[0023] 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.

[0024] The product recommendation method provided in this application relates to the field of artificial intelligence technology. The product recommendation method provided in this application can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc.; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms; the software can be an application implementing the product recommendation method, but is not limited to the above forms.

[0025] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0026] Figure 1 This is an optional flowchart of the product recommendation method provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S101 to S106.

[0027] Step S101: In response to the product recommendation request, obtain user interaction data of the target user on the pre-built product platform; wherein, the user interaction data is multimodal data.

[0028] Step S102: Extract features from user interaction data to obtain user interaction features; wherein, user interaction features include the dialogue features and behavioral features of the target user.

[0029] Step S103: Align the dialogue features and behavioral features to obtain the target aligned features.

[0030] Step S104: Predict product demand based on target alignment features to obtain product demand probability.

[0031] Step S105: Based on the product demand probability and target alignment features, make recommendation predictions to obtain target recommendation data, which includes target recommended products and the recommendation time of target recommended products.

[0032] Step S106: Generate a target product recommendation plan based on the target recommended product, and push the target product recommendation plan to the target user at the product recommendation time.

[0033] Steps S101 to S106 as illustrated in this embodiment first extract user interaction features, including dialogue features and behavioral features of the target user, from user interaction data. This dynamically captures user intent and dynamically changing real-time user behavior within the user interaction data. The dialogue and behavioral features are then aligned to ensure that different modal features can be compared and fused within a unified feature space, facilitating accurate reflection of the user's true product needs and preferences. Second, product demand is predicted based on the target alignment features, and recommendation prediction is performed based on the predicted product demand probability and the target alignment features. This yields target recommendation data, including the target recommended product and its recommendation time. This allows understanding the intensity of user demand for different products, enabling the derivation of products that meet user needs. It also ensures that product recommendations are made when users are most likely to be interested or need them, avoiding pushing products during busy times to prevent customer aversion or complaints, thus improving the conversion rate of product recommendations. Finally, a target product recommendation plan is generated based on the target recommended product and pushed to the target user at the recommended product time. This ensures the timeliness and effectiveness of the recommendation plan, improving user satisfaction and conversion rate, thereby enhancing the product recommendation effect.

[0034] In step S101 of some embodiments, specifically, the pre-built product platform refers to a digital platform that integrates various product information, user account management, and interactive functions.

[0035] For example, in non-insurance scenarios, the product platform can serve as an application terminal for introducing various insurance products, including home insurance, accident insurance, health insurance, and pet insurance, providing terms and conditions, and premium calculation tools.

[0036] Specifically, user interaction data refers to multimodal data generated by target users on the insurance platform. This user interaction may include, but is not limited to, the target user's user information, voice dialogue data between the target user and insurance customer service, the target user's behavioral data on the platform, and product recommendation trigger data triggered by the target user on the platform.

[0037] For example, in the insurance scenario, user interaction data may include user information (such as occupation, age, income and medical history), user dialogue data with customer service (such as the specific content of a user's voice inquiry about "cybersecurity insurance"), user behavior data when browsing insurance product pages (such as viewing the terms of "corporate property insurance"), and product recommendation trigger data provided by users on the platform (such as user feedback on multiple hacker attacks or frequent fires in their city or time period during the month).

[0038] Specifically, when a product recommendation request is triggered, user interaction data of the target user can be extracted from the pre-built product platform database.

[0039] In this embodiment, by responding to product recommendation requests, multimodal user interaction data of target users on a pre-built product platform is obtained, which provides a rich data foundation for subsequent user interaction feature extraction and analysis, and helps to understand the specific needs of users from multiple dimensions.

[0040] In step S102 of some embodiments, specifically, the user interaction features refer to the vector representations of dialogue semantics, behavioral preferences, and product recommendation triggering events extracted from user interaction data. These user interaction features include the dialogue features and behavioral features of the target user. Specifically, the dialogue features refer to the dialogue semantic vector representation extracted from the dialogue data in the user interaction data; the behavioral features refer to the temporal behavioral vector representation extracted from the behavioral data in the user interaction data.

[0041] Specifically, for dialogue features, the BERT model can be used to encode the dialogue data to extract dialogue vector features; for behavioral features, the behavioral data is embedded to obtain a sequence of behavioral vectors, and then an RNN (Recurrent Neural Network) is used to perform behavioral analysis on the behavioral vector sequence step by step to obtain time-series-based behavioral features.

[0042] Please see Figure 2 In some embodiments, step S103 includes, but is not limited to, steps S201 to S205: Step S201: Project the dialogue features and behavior features into the feature space to obtain the initial alignment features; wherein, the initial alignment features include the projected dialogue features and the projected behavior features, and the projected dialogue features and the projected behavior features are located in the same vector dimension space.

[0043] Step S202: Obtain target anchor point data based on user interaction data, and extract features from the target anchor point data to obtain target anchor point features.

[0044] Step S203: Alignment deviation detection is performed based on target anchor point features, projection dialogue features, and projection behavior features to obtain the degree of alignment feature deviation.

[0045] Step S204: Perform feature distribution recognition on the projection dialogue features and projection behavior features to obtain alignment feature distribution information; Step S205: Update the initial alignment features according to the preset modal difference penalty coefficient, alignment feature deviation degree and alignment feature distribution information to obtain the target alignment features.

[0046] In step S201 of some embodiments, specifically, the initial alignment feature refers to the feature that is located in the same spatial dimension after being projected into the feature space. The initial alignment feature includes the projected dialogue feature and the projected behavior feature, and the projected dialogue feature and the projected behavior feature are located in the same vector dimension space.

[0047] Specifically, dialogue features can be linearly transformed using preset dialogue projection matrices and dialogue bias matrices to obtain projected dialogue features, and behavioral features can be linearly transformed using preset behavioral projection matrices and behavioral bias matrices to obtain projected behavioral features.

[0048] For example, for dialogue features with a feature dimension of 256, a linear transformation can be performed using a 256×1024-dimensional dialogue projection matrix and a 1×1024-dimensional dialogue bias matrix to output a 1×1024-dimensional projected dialogue feature; for behavioral features with a feature dimension of 1×768, a linear transformation can be performed using a 768×1024-dimensional behavioral projection matrix and a 1024×1-dimensional behavioral bias matrix to output a 1×1024-dimensional projected behavioral feature.

[0049] In this embodiment, by projecting dialogue features and behavioral features into the feature space, the dimensional differences between different modal features are eliminated, which helps to improve the accuracy of subsequent feature analysis and comparison.

[0050] In step S202 of some embodiments, specifically, the target anchor data refers to key semantic information extracted from user interaction data as a feature alignment benchmark.

[0051] Specifically, content with clear insurance business implications can be identified and extracted from dialogue data or product recommendation trigger data in user interaction data and used as target anchor data.

[0052] For example, in an insurance scenario, if a user inquires about "data center power outage liability," the "Service Interruption Liability Clause" insurance policy corresponding to "data center power outage liability" can be used as the target anchor data.

[0053] Specifically, target anchor features refer to the semantic vector representation of extracted target anchor data.

[0054] Specifically, the target anchor data can be encoded using the BERT model to generate corresponding text description anchor vectors, i.e., target anchor features.

[0055] In this embodiment, target anchor data is obtained based on user interaction data, and features are extracted from the target anchor data to obtain target anchor features. This provides a common and stable semantic reference benchmark for the feature alignment of subsequent dialogue features and behavioral features, thereby ensuring the accuracy of the alignment.

[0056] In step S203 of some embodiments, specifically, before performing alignment deviation detection, the product recommendation method further includes: performing feature enhancement processing on the dialogue features based on the projection behavior features to obtain enhanced dialogue features, and performing feature enhancement processing on the behavior features based on the projection dialogue features to obtain enhanced behavior features.

[0057] Specifically, the enhancement process for enhanced dialogue features can be performed by treating the projected dialogue features as queries and the projected behavioral features as keys and values ​​for attention processing. The attention weights of the projected behavioral features on the projected dialogue features are calculated, and the weighted fused behavioral features are determined based on these attention weights. Finally, the weighted fused behavioral features are summed with the projected dialogue features to obtain the enhanced dialogue features.

[0058] Specifically, the enhancement process for enhanced behavioral features can be performed by treating the projected behavioral features as queries and the projected dialogue features as keys and values ​​for attention processing. The attention weights of the projected dialogue features on the projected behavioral features are calculated, and the weighted fused dialogue features are determined based on the attention weights of the projected dialogue features on the projected behavioral features. Finally, the weighted fused dialogue features and the projected behavioral features are summed to obtain the enhanced behavioral features.

[0059] Specifically, the degree of alignment feature deviation refers to the degree of deviation between the enhanced dialogue features and enhanced behavior features and the target anchor features.

[0060] Specifically, the degree of deviation between the target anchor features and the enhanced dialogue features can be assessed by calculating the dialogue Euclidean distance between them. Alternatively, the degree of deviation between the target anchor features and the enhanced behavioral features can be assessed by calculating the behavioral Euclidean distance between them. The degree of deviation of the alignment features can be determined based on the degree of deviation of the dialogue features and the degree of deviation of the behavioral features. The larger the dialogue Euclidean distance, the higher the degree of feature deviation.

[0061] In this embodiment, alignment deviation detection is performed based on target anchor point features, projection dialogue features, and projection behavior features to obtain the degree of alignment feature deviation. This allows for an intuitive assessment of the matching degree between different modal features and the target anchor point (i.e., the user's potential core needs), which helps improve the accuracy of subsequent product demand prediction.

[0062] In step S204 of some embodiments, specifically, the alignment feature distribution information refers to the feature distribution differences between enhanced dialogue features and enhanced behavior features in the same vector space, such as the distribution differences of mean, variance, or higher-order statistics.

[0063] In the insurance scenario, if enhanced dialogue features indicate that "property insurance" is mentioned more frequently in user conversations, but enhanced behavioral features (such as clicks and browsing) are more frequently concentrated on "accident insurance" products, it indicates that there is a difference in the feature distribution of the two modalities, and this difference needs to be identified and quantified.

[0064] Specifically, the maximum mean distribution of enhanced dialogue features and enhanced behavior features can be calculated separately, and a distribution metric algorithm (such as MMD, Maximum Mean Discrepancy) can be used to compare the MMD distance values ​​between enhanced dialogue features and enhanced behavior features, so as to determine the MMD distance value as the alignment feature distribution information.

[0065] In this embodiment, by performing feature distribution recognition on the projection dialogue features and projection behavior features, alignment feature distribution information is obtained, which can help understand the inconsistency of the overall feature distribution of different modal features in the same space, providing data support for comprehensively optimizing the consistency of different modal features.

[0066] In step S205 of some embodiments, specifically, the preset modal difference penalty coefficient is a hyperparameter used to control the penalty intensity for the difference in feature distribution between modalities when updating the initial alignment features.

[0067] Specifically, the process of updating the initial alignment features can be achieved by combining the preset modality difference penalty coefficient, the degree of alignment feature deviation, and the alignment feature distribution information to construct a comprehensive alignment loss function. The initial alignment features are then optimized by minimizing the alignment loss function, so as to finally determine the target alignment features.

[0068] Specifically, the alignment loss function can be expressed by the following formula:

[0069] Specifically, This represents the alignment loss function, where M represents the number of modes. This represents the i-th modal feature (such as enhanced dialogue features and enhanced behavior features). Indicates the characteristics of the target anchor point. This represents the modal difference penalty coefficient. It represents the maximum mean difference measure between the m-th modal feature (such as enhanced dialogue features) and the n-th modal feature (such as enhanced behavior features), i.e., the alignment feature distribution information.

[0070] For example, if the Euclidean distance between the enhanced dialogue features and the target anchor features is large, or if the distribution of enhanced dialogue features differs significantly from the distribution of enhanced behavior features, a larger result will be obtained. Value, if The value does not meet the preset alignment loss condition (such as alignment loss). If the value is less than a preset loss threshold (e.g., 0.5), the feature weights of the enhanced dialogue features or enhanced behavior features are adjusted through backpropagation. This makes the updated enhanced dialogue features and enhanced behavior features closer to the target anchor features in the vector space, and their feature distributions become more similar, until the alignment loss is achieved. The value is less than the preset loss threshold (e.g., 0.5), thus obtaining the target alignment feature.

[0071] In this embodiment, the initial alignment features are updated based on the preset modal difference penalty coefficient, the degree of alignment feature deviation, and the alignment feature distribution information. By introducing modal difference penalty and distribution consistency measurement, the initial alignment features can be updated to effectively overcome the inherent heterogeneity and distribution offset problems of different modal data, and obtain target alignment features with more convergent distribution of different modalities. This helps to accurately reflect the user's real product needs and preferences in the future.

[0072] Steps S201 to S205 resolve the inconsistency in the expression and distribution of different modal features, enabling the fused target alignment features to more comprehensively and accurately capture users' real needs and preferences. This provides high-quality input for subsequent product demand prediction and recommendation prediction, thereby helping to improve the accuracy and reliability of product recommendations.

[0073] In an optional embodiment of this application, the user interaction feature may further include product recommendation trigger features and user information features. The product recommendation trigger feature refers to the vector representation extracted from the user recommendation trigger data in the user interaction data, and the user information feature refers to the vector representation extracted from the user interaction data. That is, in feature alignment, the product recommendation trigger feature and user information feature are further projected into the feature space to determine the projected product recommendation trigger feature and projected user information feature. Alignment deviation detection is performed based on the target anchor point feature, projected dialogue feature, projected behavior feature, product recommendation trigger feature, and projected user information feature to obtain the degree of alignment feature deviation. Feature distribution recognition is performed on the projected dialogue feature, projected behavior feature, product recommendation trigger feature, and projected user information feature to obtain alignment feature distribution information. The initial alignment feature is then updated based on the preset modal difference penalty coefficient, the degree of alignment feature deviation, and the alignment feature distribution information to obtain the target alignment feature.

[0074] Please see Figure 3In some embodiments, step S104 includes, but is not limited to, steps S301 to S304: Step S301: Perform convolution processing on the target alignment features to obtain user preference product features.

[0075] Step S302: Perform hypergraph transformation processing based on target alignment features to obtain user preference product graph features.

[0076] Step S303: Perform spatiotemporal convolution processing based on user preference product features and user preference product graph features to obtain product demand features.

[0077] Step S304: Infer product demand based on product demand characteristics to obtain product demand probability.

[0078] In step S301 of some embodiments, specifically, user preference product features refer to the relevant features that indicate a user's preference for a product.

[0079] Specifically, multiple meta-learning convolutional kernels are obtained, and each meta-learning convolutional kernel corresponds to a product category. The target alignment features are convolved through each meta-learning convolutional kernel to extract user preference product features related to each product type.

[0080] For example, in the insurance scenario, traditional convolution kernel parameters are usually fixed and cannot adapt to the huge differences in feature spaces between different types of insurance such as home insurance, liability insurance, accident insurance, and pet insurance. Meta-learning convolution kernels can dynamically adjust their own convolution parameters according to the input features to extract relevant features for different types of insurance products.

[0081] For example, if the target alignment features involve both pet liability insurance inquiries and frequent clicks on home insurance products, then the meta-learning convolutional kernel can identify keywords such as "pet" and "bite" appearing in the dialogue within the target alignment features. At the same time, combined with the user's browsing behavior of pet liability insurance terms within the target alignment features, the "pet insurance preference product features" related to "pet insurance," "bite," and "browsing pet liability insurance terms" can be extracted.

[0082] In this embodiment, by performing convolution processing on the target alignment features, user preference product features are obtained. By using meta-learning convolution kernels, the feature space of different product categories can be adapted to accurately capture user preferences.

[0083] In step S302 of some embodiments, specifically, the user preference product graph feature refers to the product preference feature of the user product preference graph.

[0084] Specifically, in the insurance scenario, the target user, the user's dialogue features, behavioral features, and insurance products can be identified as nodes. Based on the hyperedge weights between nodes, the nodes are connected by hyperedges to obtain a user product preference hypergraph. Then, user preference graph features are extracted from the user product preference hypergraph using graph neural networks (GNNs).

[0085] For example, in an insurance scenario, if the target user is a small or medium-sized business owner who inquires about "data recovery solutions after a hacker attack" and frequently browses cybersecurity insurance terms, the target user, the "data recovery" dialogue characteristics, the "browsing cybersecurity insurance" behavioral characteristics, the environmental characteristics of "three hacker attacks occurring in the user's region within a month," and the "cybersecurity insurance" product can be identified as independent nodes in a hypergraph. Weights are assigned based on the correlation strength between the node and the product. For instance, the hyperedge weight between the target user and the "data recovery" node is 0.7, the hyperedge weight between the target user and the "browsing cybersecurity insurance" behavioral node is 0.8, and the hyperedge weight between the "regional hacker attack" environmental characteristic node and the "cybersecurity insurance" product node, which is highly correlated, is 0.95. These nodes are connected to form a user product preference graph that includes the target user, dialogue characteristics, behavioral characteristics, environmental characteristics, and the product. Product preference graph features are then extracted using a graph neural network.

[0086] In this embodiment, hypergraph transformation is performed based on target alignment features to obtain user preference product graph features, which can comprehensively represent the relationship between users and products, providing a structured data foundation for subsequent inference of the intensity of user demand for products.

[0087] In step S303 of some embodiments, specifically, the product demand characteristics refer to the spatiotemporal characteristics of product demand that integrate user preference product characteristics and user preference product graph characteristics, which are used to reflect the actual needs of target users for the product.

[0088] Specifically, user preference product features and user preference product graph features can be aggregated to obtain aggregated preference product features. The influence of aggregated preference product features on product demand at different times can be adjusted according to a preset time decay factor to determine the final product demand features.

[0089] For example, in the insurance scenario, the user's occupational characteristics can be determined from the spatial dimension, and the user's characteristics of consulting about cybersecurity, browsing cybersecurity terms and conditions and their strong association with cybersecurity insurance products can be identified as aggregated product preferences. In addition, the user's browsing behavior of cybersecurity terms and conditions in the past three days and the number of cyberattacks in the city where the user is located in the past month can be given higher weights in the time dimension to determine the final cross-temporal and spatial product demand characteristics.

[0090] In this embodiment, spatiotemporal convolution processing is performed based on user preference product features and user preference product graph features. This process can simultaneously consider temporal dynamics and spatial structure, accurately capturing the dynamic changes in user product preferences and helping to improve the accuracy of subsequent product demand inference.

[0091] Please see Figure 4 In some embodiments, step S304 includes, but is not limited to, steps S401 to S404: Step S401: Based on the preset product recommendation strategy, evaluate the product demand characteristics to obtain the initial product demand prediction probability.

[0092] Step S402: Perform reinforcement learning on the product recommendation strategy based on the initial product demand prediction probability to obtain an updated product recommendation strategy.

[0093] Step S403: Update the initial product demand features according to the updated product recommendation strategy to obtain updated product demand features.

[0094] Step S404: Evaluate product demand based on updated product demand characteristics to obtain product demand probability.

[0095] In step S401 of some embodiments, specifically, the preset product recommendation strategy refers to the decision rule that transforms product demand characteristics into the probability of specific insurance product demand. This product recommendation strategy is determined based on the actual application scenario and is not limited here.

[0096] For example, in an insurance scenario, if a corporate user is a cybersecurity officer, and the cybersecurity officer's product demand characteristics include the dialogue keyword "cyberattack," and the cybersecurity officer's behavioral characteristics show that they have browsed the "cybersecurity insurance" product page, and the city where the cybersecurity officer is located has suffered three cyberattacks in the past month, then the probability of the cybersecurity officer having a demand for "cybersecurity insurance" increases by 20%.

[0097] Specifically, the initial product demand forecast probability refers to the intensity of target users' demand for the product.

[0098] Specifically, based on the product recommendation strategy, the feature weights of product demand characteristics can be determined, and by activating the product demand feature weights and product demand characteristics through an activation function, the initial product demand prediction probability can be obtained.

[0099] For example, if a user mentions the keyword "cyberattack" in their inquiry (feature value of 1, indicating the presence of this feature), the company size is 100 people (feature value of 100), and the user's location is an area that has experienced cyberattacks (feature value of 1), then based on the product recommendation strategy, the initial feature weights associated with "cybersecurity insurance" can be determined as follows: the feature weight for the user's inquiry involving cyberattacks is 0.6, the feature weight for the company size of 100 people is 0.1, and the feature weight for the user's location being an area that has experienced cyberattacks is 0.2. After weighting and summing the features with a Sigmoid activation function (0.6 for the user's inquiry involving cyberattacks, 0.1 for the company size of 100 people, and 0.2 for the user's location being an area that has experienced cyberattacks), and then activating them, the initial predicted probability of the user's demand for "cybersecurity insurance" can be obtained as approximately 0.85.

[0100] In step S402 of some embodiments, specifically, updating the product recommendation strategy refers to the updated recommendation strategy based on feedback of product demand probability.

[0101] Specifically, the action feedback signals of target users are obtained based on the initial product demand prediction probability. Based on the action feedback signals, the policy gradient output is determined by the PPO (Proximal Policy Optimization) algorithm. The influence weight of the policy gradient output on feature adjustment is controlled by the time decay factor to determine the updated product recommendation strategy.

[0102] For example, in an insurance scenario, if a product recommendation strategy predicts a user's demand for "cybersecurity insurance" to be 0.85, and recommends "cybersecurity insurance" to the user, and the user ultimately purchases the "cybersecurity insurance" product, then the user receives a positive reward from the product recommendation strategy. If the user explicitly rejects the "cybersecurity insurance" product or purchases other insurance products, then the user receives a negative reward from the product recommendation strategy. The strategy gradient output is determined using the PPO algorithm (e.g., if users who repeatedly "mention data breaches" tend to purchase "business interruption insurance" rather than "cybersecurity insurance," then the feature weight of the keyword "data breach" on "cybersecurity insurance" is reduced, or the feature association of "data breach" on "business interruption insurance" is increased). The strategy gradient output is then adjusted in conjunction with a time decay factor to update the product recommendation strategy.

[0103] In this implementation, reinforcement learning is applied to the product recommendation strategy based on the initial product demand prediction probability. This enables the product recommendation strategy to be adaptive, overcomes the static and empirical limitations that may exist in the product recommendation strategy, and helps to improve the accuracy of product recommendations and user satisfaction.

[0104] In step S403 of some embodiments, specifically, updating product demand characteristics refers to the updated demand characteristics based on the updated product recommendation strategy.

[0105] For example, in an insurance scenario, if users who repeatedly mention data breaches ultimately tend to purchase business interruption insurance rather than cybersecurity insurance, and it is found that the "cyberattack" feature is more predictive of users' demand for cybersecurity insurance than the "data breach" feature, then the dialogue feature weight of "cyberattack" will be increased in the updated product demand features, and the feature association between "data breach" and "business interruption insurance" will be increased to obtain updated product demand features.

[0106] Specifically, the updated product demand characteristics can be determined using the following formula:

[0107] in, This indicates an update to product demand characteristics. Let X represent the meta-learning convolutional kernel, which can adapt to the feature spaces of k different products, and let X represent the target alignment feature. Indicates aggregation processing. Let A denote the hypergraph transformation, and let A denote the adjacency matrix. Indicates the weight of the hyperedge between nodes. Indicates the time decay factor. This represents the policy gradient output.

[0108] Furthermore, through hypergraph transformation, the binary adjacency matrix can be upgraded into a multidimensional user preference product graph feature, i.e. This helps to better capture product features that reflect user preferences.

[0109] In step S404 of some embodiments, specifically, the updated product demand features can be reactivated through an activation function (such as the Sigmoid function) to output the final product demand probability.

[0110] For example, in an insurance scenario, if the weight of "number of employees" in the updated product demand features is increased, and a new feature, "data security concern index," is introduced, the updated product demand features are used to infer the probability that a cybersecurity officer's demand for "cybersecurity insurance" will be updated from the initial 0.85 to 0.9, while the probability values ​​of other unrelated insurance products will be reduced.

[0111] In this embodiment, product demand assessment is performed based on updated product demand characteristics. Through the dual optimization of strategy and characteristics, the final output product demand probability can more accurately reflect the user's potential needs, which helps to improve the conversion rate of insurance product recommendations in the future.

[0112] Through steps S401 to S404, the product recommendation strategy can be dynamically adjusted to adapt to changes in user needs. By updating product demand characteristics and conducting product demand assessment, the intensity of user demand for products can be accurately understood, providing users with more accurate product demand predictions, which helps to improve the conversion rate of product recommendations in the future.

[0113] Through steps S301 to S304, it is possible to accurately predict the intensity of users' demand for different products, dynamically capture users' product preferences, and provide more accurate and reliable data support for subsequent insurance product recommendations, thereby helping to improve user satisfaction and conversion rate.

[0114] Please see Figure 5 In some embodiments, step S105 includes, but is not limited to, steps S501 to S502: Step S501: Analyze recommended products based on product demand probability and target alignment features to obtain target recommended products.

[0115] Step S502: Evaluate the timing of product recommendations based on the product demand probability and target alignment characteristics to obtain the timing of product recommendations.

[0116] Please see Figure 6 In some embodiments, step S501 includes, but is not limited to, steps S601 to S604: Step S601: The user preference product graph features are fused with the pre-built product knowledge graph to obtain the user product association graph.

[0117] Step S602: Based on the product demand probability, propagate product demand into the user product association graph to obtain candidate product nodes.

[0118] Step S603: Calculate the product relevance based on the candidate product nodes and target users to obtain the candidate product relevance.

[0119] Step S604: Determine the target recommended product based on the relevance of the candidate products.

[0120] In step S601 of some embodiments, specifically, the product knowledge graph is a pre-built structured graph that stores product domain knowledge. The product knowledge graph includes knowledge graph nodes and knowledge graph node edges. The knowledge graph nodes represent different products, and the edges represent business relationships between different products (e.g., there is a "combined insurance" relationship between "cybersecurity insurance" and "data breach liability insurance").

[0121] Specifically, graph propagation algorithms can be used to align and connect the preferred product nodes in the user preference product graph with the corresponding knowledge graph nodes in the product knowledge graph, in order to form a user product association graph that includes both user preference features and global product relationships.

[0122] For example, in an insurance scenario, if a user's behavioral characteristic is "browsing corporate property insurance" and the associated product node is "corporate property insurance", then "browsing corporate property insurance" and "corporate property insurance" are associated with the knowledge graph node "corporate property insurance", as well as knowledge graph nodes such as "business interruption insurance" and "machine damage insurance" associated with "corporate property insurance". The resulting new graph will simultaneously contain the user's product preference (corporate property insurance) and potential associated insurance products (business interruption insurance and machine damage insurance, etc.).

[0123] In step S602 of some embodiments, specifically, a candidate product node refers to a product node with high recommendation potential that is selected from all possible products based on the user's product preferences.

[0124] Specifically, the product demand probability (e.g., the probability of a user demanding "cybersecurity insurance" is 0.9) can be determined as the source signal and loaded onto the "cybersecurity insurance" node corresponding to the user product association graph. Subsequently, through a graph propagation algorithm of random walk, the source signal is propagated along the edges of the user product association graph (e.g., from "cybersecurity insurance" to "data breach liability insurance"). Each propagation may weaken the source signal strength, but it may also strengthen it due to different paths. After propagation, user product association nodes in the graph that originally had weak or no demand signals will receive different levels of energy, so that the nodes that receive the original signal energy are determined as candidate product nodes.

[0125] For example, in an insurance scenario, the source signal is on the "cybersecurity insurance" node. After propagation, the energy value of the strongly correlated "data breach liability insurance" node is 0.65, and the energy value of the "cybersecurity emergency response service" node is 0.6 (greater than the preset energy threshold of 0.55). Then, the "cybersecurity insurance", "data breach liability insurance" and "cybersecurity emergency response service" nodes constitute the "candidate product node" set.

[0126] In this embodiment, product demand propagation is performed on the user product association graph based on product demand probability. This can identify product combinations that may meet user needs, rather than being limited to a single product. This provides a comprehensive selection of core product needs and related product needs for the final product recommendation, thereby helping to improve the effectiveness of subsequent product recommendations.

[0127] In step S603 of some embodiments, specifically, the candidate product relevance refers to the degree of matching between the candidate product and the target user's needs.

[0128] Specifically, multiple relevance evaluation dimensions can be determined for each candidate product node (such as "cybersecurity insurance" and "business interruption insurance"). These evaluation dimensions may include: product demand probability propagation value (i.e., product demand probability), the suitability of the target user industry to the product (e.g., the IT industry is more suitable for cybersecurity insurance), the matching degree between the user's company size and the product (e.g., large enterprises may need products with higher coverage), the user's historical similar product insurance records, etc. Feature values ​​of multiple evaluation dimensions are extracted from product demand characteristics to calculate the comprehensive relevance score of each candidate product.

[0129] For example, in an insurance scenario, regarding the relevance of a candidate product as "cybersecurity insurance," the probability of product demand (cybersecurity insurance demand characteristic value is 0.85, weight is 0.4) is 0.8 x 0.4 = 0.32; regarding industry suitability (IT industry characteristic value is 1.0, weight is 0.3), it is 1.0 x 0.3 = 0.3; regarding company size (more than 500 employees characteristic value is 0.8, weight is 0.2), it is 0.8 x 0.2 = 0.16; and regarding historical insurance (none, characteristic value is 0.5, weight is 0.1), it is 0.5 x 0.1 = 0.05. Therefore, the final candidate product relevance is 0.32 + 0.3 + 0.16 + 0.05 = 0.83. In step S604 of some embodiments, specifically, the target recommended product refers to a product that meets the user's current needs based on the user's needs characteristics and product needs preferences. The target recommended product can be a single product or a combination of products.

[0130] Specifically, all candidate products are sorted in descending order according to their relevance, and the final target recommended product is determined based on preset product recommendation rules (such as Top-N recommendation) or dynamic thresholds (such as selecting only products with a relevance greater than 0.7).

[0131] For example, in an insurance scenario, if the product recommendation rule is "recommend the top two products with the highest relevance among candidate products", then "cybersecurity insurance" with a candidate product relevance of 0.83 and "cybersecurity emergency response service" with a candidate product relevance of 0.75 will be determined as the final target recommended product combination.

[0132] Through steps S601 to S604, potential product needs of users can be explored based on their actual needs, providing users with a comprehensive selection of core and related product needs, thereby increasing the success rate of subsequent product recommendations and thus helping to improve the effectiveness of subsequent product recommendations.

[0133] Please see Figure 7In some embodiments, step S502 includes, but is not limited to, steps S701 to S704: Step S701: Encode temporal behavior based on target alignment features to obtain temporal features.

[0134] Step S702: Perform time decay weighting on the time series features to obtain weighted time series features.

[0135] Step S703: Calculate the recommendation time based on the product demand probability and weighted time series characteristics to obtain the recommendation time sensitivity.

[0136] Step S704: Analyze the timing of product recommendations based on the time sensitivity of recommendations to obtain the timing for product recommendations.

[0137] In step S701 of some embodiments, specifically, temporal features refer to the features related to time information extracted from the target alignment features. These features are used to reflect the time-series patterns of user characteristics.

[0138] For example, in an insurance scenario, if a user browsed the terms and conditions of "corporate property insurance" last week, checked the claims terms and conditions of "business operation interruption insurance" the day before yesterday, and clicked on the terms and conditions of "cybersecurity insurance" today, then the time points of occurrence of the behavioral features in the target alignment feature together constitute the temporal features of the behavior.

[0139] Specifically, Long Short-Term Memory (LSTM) networks can be used to extract time-stamped user behavior features from target alignment features (such as forming a time-based behavioral feature sequence from "browsing corporate property insurance, querying business interruption insurance, and clicking on cybersecurity insurance").

[0140] In step S702 of some embodiments, specifically, the weighted timing feature refers to the timing feature after being adjusted by time decay weight.

[0141] Specifically, a weight that is inversely proportional to the time distance can be assigned to each feature sequence in the time series features by a preset decay function (such as an exponential decay function), and then weighted processing is performed by combining the weights to obtain weighted time series features.

[0142] For example, in an insurance scenario, a user's action of clicking on the "cybersecurity insurance" claim terms three days ago can be assigned a higher weight (e.g., 0.9), while a user's action of "browsing property insurance" a week ago can be assigned a relatively lower weight (e.g., 0.5). The weights are then weighted and fused with the corresponding time-series features to generate weighted time-series features.

[0143] In step S703 of some embodiments, specifically, the recommendation time sensitivity is used to measure the probability of a user accepting a product when recommended product information is pushed to the user at different time periods.

[0144] Specifically, the probability of product demand can be used as a basic weight to reflect the overall intensity of users' demand for the product, and the weighted time series feature can be regarded as a time distribution function to reflect the user's behavioral activity and potential acceptance at different times of the day. The recommendation time is calculated by combining the probability of product demand with the various time window values ​​in the weighted time series feature.

[0145] For example, in an insurance scenario, if the probability of demand for cybersecurity insurance is 0.83, and the weighted time-series features show that user activity is 0.7 between 9-10 AM, 0.9 between 10-11 AM, 0.85 between 2-3 PM, and 0.75 between 3-4 PM, then the recommendation time sensitivity is 0.83 x 0.7 = 0.59 between 9-10 AM, 0.89 x 0.9 = 0.75 between 10-11 AM, 0.83 x 0.85 = 0.71 between 2-3 PM, and 0.83 x 0.75 = 0.62 between 3-4 PM. Therefore, the recommendation time sensitivity is highest at 0.75 between 10-11 AM.

[0146] In step S704 of some embodiments, specifically, the product recommendation timing refers to the specific time or time interval for recommending products to the user.

[0147] Specifically, the recommendation time sensitivity of each time window can be compared to select the time window with the highest recommendation time sensitivity as the optimal recommendation time. In addition, the actual situation of the user may also be considered during the recommendation process (such as avoiding push notifications during inappropriate time periods late at night, or making fine adjustments based on the user's current online status, whether they are currently interacting, etc.).

[0148] For example, if a user is currently (e.g., 10:05 AM) consulting about other insurance products, and the recommendation time sensitivity for this time period is 0.72 (the highest), the recommendation will be triggered immediately after the current conversation ends. If the user is not currently online, the recommendation time will be recorded as 10-11 AM the following Monday or 10-11 AM the next day.

[0149] By using steps S701 to S704, the best time for user decision-making can be accurately captured, ensuring that product recommendations are made when users are most likely to be interested or need them most. This avoids pushing products when customers are busy, which could cause customer resentment or even complaints, and helps to improve the conversion rate of product recommendations, thereby improving the effectiveness of product recommendations.

[0150] In step S106 of some embodiments, specifically, the target product recommendation scheme refers to the recommendation information generated based on the target recommended product. The target product recommendation scheme not only includes the core information of the target recommended product itself, but may also include the recommendation reasons, solutions for specific user needs, price information, preferential policies and insurance guidelines, etc.

[0151] Specifically, a target product recommendation scheme can be generated through the Transformer network.

[0152] For example, in an insurance scenario, if the target recommended product is "cybersecurity insurance" and the target user is a "software company," then by using Transformer and combining the user's previous inquiries about "data breach compensation," the target product recommendation plan would be generated as follows: "Dear software company, in view of your concern about the risk of customer data breach, we would like to recommend the 'cybersecurity insurance' product to you. This product can effectively cover legal litigation costs and data recovery costs caused by data breaches." The plan would also include a link to the specific terms and conditions of the "cybersecurity insurance" and a premium calculation entry point. This target product recommendation plan would be pushed to the user via internal message or SMS on the product platform (such as an insurance app) at 10:30 AM on Monday.

[0153] In this embodiment, a target product recommendation plan is generated based on the target recommended product, and the target product recommendation plan is pushed to the target user at the product recommendation time. This ensures the timeliness and effectiveness of the recommendation plan, thereby improving user satisfaction and conversion rate, and thus improving the product recommendation effect.

[0154] This application first extracts user interaction features, including dialogue and behavioral features of the target user, from user interaction data. This dynamically captures user intent and dynamically changing real-time user behavior within the interaction data. The dialogue and behavioral features are then aligned to ensure that different modal features can be compared and fused within a unified feature space, facilitating accurate reflection of the user's true product needs and preferences. Second, product demand is predicted based on the target alignment features, and recommendation prediction is performed based on the predicted product demand probability and the target alignment features. This yields target recommendation data, including the target recommended product and its recommendation time. This allows understanding the intensity of user demand for different products, enabling the derivation of products that meet user needs. It also ensures that product recommendations are made when users are most likely to be interested or need them, avoiding pushing products during busy times that could cause customer resentment or complaints, thus improving the conversion rate of product recommendations. Finally, a target product recommendation plan is generated based on the target recommended product and pushed to the target user at the recommended time. This ensures the timeliness and effectiveness of the recommendation plan, improving user satisfaction and conversion rate, thereby enhancing the overall product recommendation effect.

[0155] Please see Figure 8 This application also provides a product recommendation device that can implement the above-described product recommendation method. The device includes: The user interaction data acquisition module is used to respond to product recommendation requests and acquire user interaction data of target users on a pre-built product platform; wherein, the user interaction data is multimodal data. The feature extraction module is used to extract features from user interaction data to obtain user interaction features; among which, user interaction features include the dialogue features and behavioral features of the target user; The feature alignment module is used to align dialogue features and behavioral features to obtain target aligned features; The product demand forecasting module is used to forecast product demand based on target alignment features and obtain the product demand probability. The recommendation prediction module is used to make recommendation predictions based on the product demand probability and target alignment features to obtain target recommendation data, which includes the target recommended products and the recommendation time of the target recommended products. The product recommendation module is used to generate a target product recommendation plan based on the target recommended products, and push the target product recommendation plan to the target users at the product recommendation time.

[0156] The specific implementation of the product recommendation device is basically the same as the specific implementation of the product recommendation method described above, and will not be repeated here.

[0157] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the product recommendation method described above. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0158] Please see Figure 9 , Figure 9 The hardware structure of an electronic device according to another embodiment is illustrated. The electronic device includes: The processor 901 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this application. The memory 902 can be implemented as a read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory 902 can store the processing system and other application programs. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 902 and is called and executed by the processor 901 to execute the product recommendation method of the embodiments of this application. The input / output interface 903 is used to implement information input and output; The communication interface 904 is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.). Bus 905 transmits information between various components of the device (e.g., processor 901, memory 902, input / output interface 903, and communication interface 904); The processor 901, memory 902, input / output interface 903, and communication interface 904 are connected to each other within the device via bus 905.

[0159] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described product recommendation method.

[0160] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0161] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of this application, and do not constitute a limitation on the technical solutions provided in this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided in this application are also applicable to similar technical problems.

[0162] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0163] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0164] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0165] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0166] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0167] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. The coupling or direct coupling or communication connection between the shown or discussed units may be through some interfaces, or indirect coupling or communication connection between the apparatus or units, and may be electrical, mechanical, or other forms.

[0168] The units described above as separate components may or may not be physically separate. The 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 the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0169] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0170] 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 storage medium. Based on this understanding, the technical solution of this application, in essence, or the part 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 storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0171] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A product recommendation method characterized by, The method includes: In response to a product recommendation request, user interaction data of the target user on a pre-built product platform is obtained; wherein, the user interaction data is multimodal data; Feature extraction is performed on the user interaction data to obtain user interaction features; wherein, the user interaction features include the dialogue features and behavioral features of the target user; The dialogue features and the behavior features are aligned to obtain the target alignment features; Based on the target alignment features, product demand is predicted to obtain the product demand probability. Recommendation prediction is performed based on the product demand probability and the target alignment features to obtain target recommendation data, wherein the target recommendation data includes target recommended products and the recommendation time of the target recommended products; A target product recommendation plan is generated based on the target recommended product, and the target product recommendation plan is pushed to the target user at the product recommendation time.

2. The method of claim 1, wherein, The step of predicting product demand based on the target alignment features to obtain the product demand probability includes: The target alignment features are convolved to obtain user preference product features; Based on the target alignment features, a hypergraph transformation is performed to obtain the user preference product graph features; The product demand features are obtained by performing spatiotemporal convolution processing based on the user preference product features and the user preference product graph features. Based on the product demand characteristics, product demand is inferred to obtain the product demand probability.

3. The method of claim 2, wherein, The step of inferring product demand based on the product demand characteristics to obtain the product demand probability includes: Based on a preset product recommendation strategy, the product demand characteristics are evaluated to obtain an initial product demand prediction probability. The product recommendation strategy is reinforced by the initial product demand prediction probability to obtain an updated product recommendation strategy. The initial product demand features are updated according to the updated product recommendation strategy to obtain updated product demand features. Based on the updated product demand characteristics, product demand is assessed to obtain the product demand probability.

4. The method of claim 2, wherein, The step of making recommendation predictions based on the product demand probability and the target alignment features to obtain target recommendation data includes: Based on the product demand probability and the target alignment features, a recommended product analysis is performed to obtain the target recommended product; The timing of product recommendations is determined by evaluating the probability of product demand and the target alignment features.

5. The method of claim 4, wherein, The step of performing product recommendation analysis based on the product demand probability and the target alignment features to obtain target recommended products includes: The user preference product graph features are fused with a pre-constructed product knowledge graph to obtain a user-product association graph. Based on the product demand probability, product demand propagation is performed on the user product association graph to obtain candidate product nodes; The product relevance is calculated based on the candidate product nodes and the target users to obtain the candidate product relevance. The target recommended product is determined based on the relevance of the candidate products.

6. The method of claim 4, wherein, The step of evaluating the timing of product recommendations based on the product demand probability and the target alignment features to obtain the product recommendation timing includes: Temporal encoding is performed based on the target alignment features to obtain temporal features; The time-series features are subjected to time-decay weighting to obtain weighted time-series features; The recommendation time sensitivity is obtained by calculating the recommendation time based on the product demand probability and the weighted time series features. Based on the recommended time sensitivity, the timing of the recommendation is analyzed to obtain the timing of the product recommendation.

7. The method according to any one of claims 1 to 6, characterized in that, The step of aligning the dialogue features and the behavior features to obtain the target aligned features includes: The dialogue features and the behavior features are projected into a feature space to obtain initial alignment features; wherein, the initial alignment features include projected dialogue features and projected behavior features, and the projected dialogue features and the projected behavior features are located in the same vector dimension space; Target anchor point data is obtained based on the user interaction data, and feature extraction is performed on the target anchor point data to obtain target anchor point features; Alignment deviation detection is performed based on the target anchor point features, the projection dialogue features, and the projection behavior features to obtain the degree of alignment feature deviation. The projection dialogue features and the projection behavior features are subjected to feature distribution recognition to obtain alignment feature distribution information; The initial alignment features are updated based on the preset modal difference penalty coefficient, the degree of deviation of the alignment features, and the distribution information of the alignment features to obtain the target alignment features.

8. A product recommendation device, characterized in that, The device includes: The user interaction data acquisition module is used to acquire user interaction data of the target user on a pre-built product platform in response to a product recommendation request; wherein, the user interaction data is multimodal data. The feature extraction module is used to extract features from the user interaction data to obtain user interaction features; wherein, the user interaction features include the dialogue features and behavioral features of the target user; The feature alignment module is used to align the dialogue features and the behavior features to obtain the target aligned features; The product demand prediction module is used to predict product demand based on the target alignment features and obtain the product demand probability. The recommendation prediction module is used to make recommendation predictions based on the product demand probability and the target alignment features to obtain target recommendation data, wherein the target recommendation data includes target recommended products and the recommendation time of the target recommended products; The product recommendation module is used to generate a target product recommendation plan based on the target recommended product, and push the target product recommendation plan to the target user at the product recommendation time.

9. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the product recommendation method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the product recommendation method according to any one of claims 1 to 7.

Images