Marketing recommendation method and system for cultural products in electronic commerce

By constructing a multi-dimensional user cultural affiliation matrix and a fuzzy embedding deep learning model, and dynamically modulating the attention mechanism to generate a culturally sensitive vector matrix, the accuracy problem of cultural product recommendations in e-commerce platforms is solved, achieving efficient user screening and recommendation.

CN121883098APending Publication Date: 2026-04-17HANGZHOU NORMAL UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610015278.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-01-07
Publication Date
2026-04-17

AI Technical Summary

Technical Problem

Existing e-commerce platform recommendation systems lack the ability to quantify users' cultural identity when recommending cultural products. They struggle to distinguish users' ambiguous and multi-layered cultural preferences, leading to decreased recommendation accuracy, wasted resources, and user resentment.

Method used

A multi-dimensional user culture affiliation matrix is ​​constructed, and a fuzzy embedding deep learning model is combined with a dynamic modulation attention mechanism to generate a culture-sensitive vector matrix, thereby accurately screening out potential user groups with high cultural matching.

Benefits of technology

It improved the accuracy and relevance of cultural product marketing recommendations, increased user satisfaction, and reduced the waste of resources from irrelevant recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121883098A_ABST
    Figure CN121883098A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of marketing, and discloses a marketing recommendation method and system for cultural products in electronic commerce. The method comprises the following steps: constructing a user semantic vector and a user culture membership matrix based on original data of a target user; generating a product culture vector based on the original information of the to-be-recommended culture product; jointly inputting the user semantic vector and the user culture membership matrix into a configured fuzzy embedded deep learning model, and outputting a culture sensitive vector matrix capable of reflecting deep culture preferences of the user; screening and sorting the users based on the comprehensive similarity between the culture sensitive vector matrix and the product culture vector, and generating an initial recommendation candidate set; and fusing the initial recommendation candidate set with an operation strategy rule of the platform, and determining and pushing the to-be-recommended cultural product to the user in the final recommendation object set. According to the method, the potential user groups with high culture matching degree with the to-be-recommended culture products can be accurately screened out from massive users.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of marketing technology, specifically to a marketing recommendation method and system for cultural products in e-commerce. Background Technology

[0002] With the deepening development of the internet economy, e-commerce has become an important sales channel for cultural products, such as intangible cultural heritage creative products, regional handicrafts, and festival gifts. However, the consumption decision-making logic of cultural products differs significantly from that of traditional standardized products. Their value is not solely determined by physical attributes, but is deeply rooted in complex, subjective, and multi-layered cultural connotations and user emotional identification.

[0003] Existing e-commerce platform recommendation systems widely employ recommendation algorithms based on collaborative filtering or content embedding. These methods are typically built on the assumption that user preferences are explicit, static, and can be fully expressed by high-dimensional dense vectors. When applied to cultural product recommendation scenarios, these traditional technical solutions expose numerous limitations, leading to a significant decrease in recommendation accuracy, specifically in the following aspects: (1) Existing technologies lack the ability to quantify the core motivation of users' cultural identity. Cultural identity is inherently vague and continuous. For example, users may highly identify with the aesthetic design (such as patterns and colors) of a certain type of cultural product, but have a low degree of identification with the traditional etiquette or philosophical concepts it embodies. Traditional user profiling models, whether based on tag systems or embedded vectors, cannot effectively represent this partial or hierarchical identification relationship between different dimensions within the same cultural object, thus leading to distortion in user interest modeling.

[0004] (2) While mainstream deep learning recommendation models can capture complex nonlinear features, their core components, such as attention mechanisms, are not specifically designed to address the ambiguity and hierarchy of cultural preferences. When processing user behavior data, these models struggle to distinguish between superficial clicks and deep cultural resonance, failing to incorporate users' cultural affiliation as a modulation signal into the model's computation process. Consequently, the generated user vectors fail to fully reflect their deep and sensitive cultural preferences.

[0005] Due to the aforementioned shortcomings, when identifying potential users interested in target cultural products, platforms struggle to accurately identify those with a high degree of cultural identification from a massive user base. They are also prone to incorrectly including users who interacted only by chance, or those with similar cultural backgrounds but low actual cultural identification, in their recommendations. This not only wastes marketing computing and traffic resources but may also cause user resentment due to frequent irrelevant recommendations, damaging the user experience and ultimately hindering the conversion efficiency and market vitality of cultural products on e-commerce platforms.

[0006] Therefore, how to further improve the accuracy and effectiveness of cultural product marketing recommendations is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] The purpose of this invention is to provide a marketing recommendation method, system, electronic device, readable storage medium, and computer program product for cultural products in e-commerce, so as to solve the problems mentioned in the background art.

[0008] This invention provides a marketing recommendation method for cultural products in e-commerce, comprising the following steps: Step S10: Obtain the original data of the target user, and construct a user semantic vector and a user cultural membership matrix based on the original data; wherein, the user cultural membership matrix contains multiple levels, which are used to characterize the degree of continuous identification of the user with cultural elements at different cultural levels; Step S20: Obtain the original information of the cultural product to be recommended, and generate a product culture vector based on the original information; Step S30: Input the user semantic vector and the user cultural membership matrix into the configured fuzzy embedding deep learning model. Use the user cultural membership matrix to dynamically modulate the attention mechanism in the fuzzy embedding deep learning model to output a culturally sensitive vector matrix that reflects the user's deep cultural preferences. Step S40: Calculate the comprehensive similarity between the cultural sensitivity vector matrix and the product cultural vector; filter and sort users based on the comprehensive similarity to generate an initial recommendation candidate set. Step S50: The initial recommendation candidate set is merged with the platform's operation strategy rules. Based on the fusion result, the final recommendation object set is determined, and the recommended cultural products are pushed to users in the recommendation object set.

[0009] The present invention also provides a marketing recommendation system for cultural products in e-commerce, the system comprising a first acquisition unit, a second acquisition unit, a user analysis unit, a first filtering processing unit, and a second filtering processing unit; The first acquisition unit is configured to: acquire the original data of the target user, and construct a user semantic vector and a user cultural membership matrix based on the original data; wherein, the user cultural membership matrix contains multiple levels, which are used to characterize the degree of continuous identification of the user with cultural elements at different cultural levels; The second acquisition unit is configured to: acquire the original information of the cultural product to be recommended, and generate a product culture vector based on the original information; The user analysis unit is configured to: input the user semantic vector and the user cultural membership matrix into a configured fuzzy embedding deep learning model, and use the user cultural membership matrix to dynamically modulate the attention mechanism in the fuzzy embedding deep learning model to output a culturally sensitive vector matrix that reflects the user's deep cultural preferences. The first filtering processing unit is configured to: calculate the comprehensive similarity between the cultural sensitivity vector matrix and the product cultural vector, filter and sort users based on the comprehensive similarity, and generate an initial recommendation candidate set; The second filtering processing unit is configured to: merge the initial recommendation candidate set with the platform's operation strategy rules, determine the final recommendation object set based on the fusion result, and push the recommended cultural products to users in the recommendation object set.

[0010] The present invention also provides an electronic device including a processor and a memory, the memory storing a program or instructions executable on the processor, the program or instructions, when executed by the processor, implementing the steps of the method as described in any of the preceding claims.

[0011] The present invention also provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method as described in any of the preceding claims.

[0012] The present invention also provides a computer program product stored in a storage medium, the program product being executed by at least one processor to implement the steps of the method as described in any of the preceding claims.

[0013] This invention constructs a multi-dimensional user cultural affiliation matrix, achieving a precise mathematical representation of users' fuzzy, hierarchical, and dynamically changing cultural identity tendencies. This overcomes the inaccuracy problem of traditional user profiling models in cultural product recommendations. Furthermore, through a designed fuzzy embedding deep learning model and its dynamic modulation mechanism, the aforementioned cultural affiliation relationships are deeply integrated as modulation signals into the model's attention calculation process. This enhances the cultural context sensitivity of user semantic vectors, generating a culturally sensitive vector matrix that better reflects users' deep cultural preferences. Ultimately, this invention can accurately select potential user groups with high cultural matching to the recommended cultural products from a massive user base, improving the accuracy, targeting, and user satisfaction of cultural product marketing recommendations. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a marketing recommendation method for cultural products in e-commerce, as disclosed in an embodiment of the present invention. Figure 2This is a schematic diagram of the structure of the fuzzy embedding deep learning model disclosed in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a marketing recommendation system for cultural products in e-commerce, as disclosed in an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] Please see Figure 1 This invention provides a marketing recommendation method 100 for cultural products in e-commerce, comprising the following steps: Step S10: Obtain the original data of the target user, and construct a user semantic vector and a user cultural membership matrix based on the original data; wherein, the user cultural membership matrix contains multiple levels, which are used to characterize the degree of continuous identification of the user with cultural elements at different cultural levels; In this step, the terminal (such as the recommendation server of an e-commerce platform) receives and acquires the target user's raw data on the platform. This raw data is a collection of the user's digital footprint, including but not limited to: the user's explicit historical behavioral data, such as browsing, collecting, purchasing, and searching records of cultural products; implicit content data actively published by the user, such as comments on cultural products and shared content; and the user's static attribute data, such as interest tags and geographic information filled in during registration.

[0017] Based on the aforementioned multi-source heterogeneous raw data, user semantic vectors are constructed. This construction process utilizes Natural Language Processing (NLP) techniques, particularly pre-trained language models, to perform deep semantic analysis and encoding on user behavior sequences, comment texts, etc., mapping them into dense vectors in a high-dimensional space. Understandably, this dense vector aims to capture and condense the semantic information behind user behavior, forming a distributed representation of users' general interests and preferences.

[0018] A user cultural affiliation matrix is ​​also constructed. A predefined set of cultural elements is used, comprising multiple levels (e.g., Level 1 cultural categories: Eastern culture, Western culture; Level 2 cultural subcategories: traditional Chinese culture, Japanese Zen culture; Level 3 cultural elements: calligraphy, ink painting, embroidery, blue and white porcelain, etc.). Then, by analyzing the association between users' raw data and these cultural elements, a continuously varying affiliation degree within the range [0,1] is calculated for each user for each cultural element. The affiliations of all users collectively constitute a multidimensional user cultural affiliation matrix. Understandably, this matrix mathematically represents the user's fuzzy, non-binary level of identification with different cultural elements at multiple levels. For example, a user might have a 0.9 affiliation degree (high identification) for the "blue and white porcelain" pattern, but only a 0.3 affiliation degree (low identification) for the related cultural connotation of "traditional official kiln system."

[0019] Step S20: Obtain the original information of the cultural product to be recommended, and generate a product culture vector based on the original information; This step involves obtaining the original information of the cultural products to be recommended, including but not limited to the product title, text description, multi-angle images, video introduction, and cultural tags marked by the seller.

[0020] Based on the aforementioned raw information, a product culture vector is generated. This generation process employs multimodal information fusion technology, specifically: using computer vision (CV) models to analyze product images and extract features such as visual style, color scheme, and patterns; and using natural language processing (NLP) models to understand the cultural semantics in the product's text description. Finally, these cultural features extracted from different modalities are aligned and fused to form a unified product culture vector that comprehensively represents the cultural attributes inherent in the product.

[0021] Step S30: Input the user semantic vector and the user cultural membership matrix into the configured fuzzy embedding deep learning model. Use the user cultural membership matrix to dynamically modulate the attention mechanism in the fuzzy embedding deep learning model to output a culturally sensitive vector matrix that reflects the user's deep cultural preferences. In this step, a fuzzy embedding deep learning model is pre-constructed, preferably a structured Transformer model. The inputs to this fuzzy embedding deep learning model are: a user semantic vector, used to provide the user's basic interest profile; and a user cultural membership matrix, used to provide refined, multi-layered cultural preference information.

[0022] In the process of processing user semantic vectors and calculating attention weights in fuzzy embedding deep learning models (especially their self-attention layers), the user cultural membership matrix is ​​not treated as a simple feature concatenation, but rather as a modulation signal that finely guides and constrains the attention allocation process. Specifically, this dynamic modulation can be manifested as follows: Hierarchical gating: Based on the weights of different cultural levels in the membership matrix, the control model focuses on semantic information related to the user's high membership cultural level.

[0023] Uncertainty Modulation: For cultural elements with low user affiliation or vagueness, the model introduces uncertainty when allocating attention to avoid overemphasizing users' uncertain preferences.

[0024] Through this deep fusion modulation, the fuzzy embedding deep learning model ultimately outputs a cultural sensitivity vector matrix that more accurately reflects users' complex and deep-seated cultural preferences. Understandably, this cultural sensitivity vector matrix can be seen as a refinement and sublimation of user interests within a specific cultural context.

[0025] Step S40: Calculate the comprehensive similarity between the cultural sensitivity vector matrix and the product cultural vector; filter and sort users based on the comprehensive similarity to generate an initial recommendation candidate set. In this step, a suitable similarity calculation algorithm (such as cosine similarity, matching degree calculation based on neural networks, etc.) is used to calculate the comprehensive similarity between the cultural sensitivity vector matrix output in step S30 and the product culture vector generated in step S20 in the vector space.

[0026] This comprehensive similarity score takes into account the degree of alignment between the product's cultural attributes and users' deep-seated, multi-level cultural preferences. Then, based on this comprehensive similarity score, all target users to be screened are sorted in descending order, and a similarity threshold is set. Users with comprehensive similarity scores higher than this threshold are selected to form an initial recommendation candidate set. Thus, this initial recommendation candidate set contains the potential user group most likely to develop cultural identification and purchase interest in the current cultural products.

[0027] Step S50: The initial recommendation candidate set is merged with the platform's operation strategy rules. Based on the fusion result, the final recommendation object set is determined, and the recommended cultural products are pushed to users in the recommendation object set.

[0028] In this step, in order to ensure that the recommendation results are accurate while also meeting the platform's operational needs, the initial recommendation candidate set generated in step S40, which is entirely based on algorithm matching, is further integrated with various operational strategy rules preset by the platform.

[0029] The aforementioned operational strategies and rules include, but are not limited to: overall budget constraints, specific merchant constraints (such as traffic support for new merchants), platform exposure rules (such as preventing overexposure), and ad priorities. The integration process can employ multi-objective optimization decisions, such as prioritizing budget control and merchant fairness while ensuring high similarity.

[0030] Finally, based on the fusion decision results, the final set of recommended objects is determined from the initial recommendation candidate set, and the recommended cultural products are accurately pushed to users in this set through the platform's message push system, personalized display interface and other means, thus completing the entire marketing recommendation process.

[0031] This invention constructs a multi-dimensional user cultural affiliation matrix, achieving a precise mathematical representation of users' fuzzy, hierarchical, and dynamically changing cultural identity tendencies. This overcomes the inaccuracy problem of traditional user profiling models in cultural product recommendations. Furthermore, through a designed fuzzy embedding deep learning model and its dynamic modulation mechanism, the aforementioned cultural affiliation relationships are deeply integrated as modulation signals into the model's attention calculation process. This enhances the cultural context sensitivity of user semantic vectors, generating a culturally sensitive vector matrix that better reflects users' deep cultural preferences. Ultimately, this invention can accurately select potential user groups with high cultural matching to the recommended cultural products from a massive user base, improving the accuracy, targeting, and user satisfaction of cultural product marketing recommendations.

[0032] As an example, the user culture membership matrix is ​​constructed using a fuzzy comprehensive evaluation model, specifically: The fuzzy comprehensive evaluation model integrates multi-source behavioral data and contextual data from the user's original data to output a membership degree that continuously varies within the interval [0,1]. The multi-source behavioral data includes explicit behavioral data and implicit behavioral data that reflects the intensity of interest, and the contextual data includes at least time, location, and social context.

[0033] In this implementation, the fuzzy comprehensive evaluation model integrates multi-source behavioral data and contextual data from the user's original data, outputting a membership degree that continuously varies within the interval [0,1]. The multi-source behavioral data includes not only explicit user behavior data, such as actions explicitly expressing preferences like purchasing or collecting cultural products, but more importantly, it also includes implicit behavioral data that more deeply reflects the intensity of the user's true interests, such as the duration of time spent on specific cultural product detail pages, the sequence of repeated browsing of a certain type of cultural product, and the emotional polarity of comments on cultural products. It is understood that this implicit behavioral data can effectively reveal a user's latent but strong cultural interests, even if the user has not made a direct purchase.

[0034] Meanwhile, the input of this fuzzy comprehensive evaluation model also includes contextual data, including at least time context (such as traditional festivals and seasonal changes), location context (such as the user's location and whether they are near cultural attractions), and social context (such as whether they are recommending a friend or preparing to give a gift).

[0035] By comprehensively considering the aforementioned multi-source behavioral and contextual data, the fuzzy comprehensive evaluation model can output more accurate and dynamic membership values, which continuously represent the strength of a user's identification with a particular cultural element in a specific context. It is understandable that the fuzzy comprehensive evaluation model can employ fuzzy inference algorithms, and its processing generally involves: 1. Define input variables and fuzzy sets: Define the key dimensions in the multi-source data as input variables (such as "browsing duration", "comment sentiment score", "festival proximity"). Define 2-3 fuzzy sets for each input variable, such as "short, medium, long", "negative, neutral, positive", "non-festival period, festival period".

[0036] 2. Define the output variable and fuzzy set: Define the degree of user identification with a specific cultural element as the output variable, and its fuzzy set can be defined as "low, medium, high".

[0037] 3. Construct a fuzzy rule base: Establish a series of fuzzy rules in the form of "IF-THEN". For example: "IF browsing time is long AND comment sentiment is positive AND it is a festive period, THEN the degree of recognition of the 'Chinese red' element is high". The rule base can be initially constructed based on business experience or analysis of sample data and can be optimized.

[0038] 4. Perform fuzzy inference: For each user, their specific multi-source data (precise values) are converted into membership degrees to each input fuzzy set through a fuzzification interface. Then, through a fuzzy inference engine (such as Mamdani or Sugeno type), the relevant fuzzy rules are activated and executed, and finally, through a defuzzification interface, a precise membership value in the interval [0,1] is obtained.

[0039] 5. Matrix Generation: For each element in the cultural element set, repeat steps 2-4 to obtain the user's complete cultural membership vector. Perform this process for all users to generate the user cultural membership matrix.

[0040] As an example, the fuzzy embedding deep learning model includes a semantic encoding module, a culture-modulated attention module, and a vector generation module; The semantic encoding module is used to convert the user semantic vector into an initial hidden state sequence; The culture-modulated attention module includes a multi-head self-attention layer, which receives the initial hidden state sequence and dynamically modulates the attention distribution of the multi-head self-attention layer using the user culture membership matrix, and outputs a hidden state sequence enhanced with cultural context. The vector generation module is used to aggregate or select the cultural context-enhanced hidden state sequence to generate the culturally sensitive vector matrix.

[0041] Please see Figure 2 The fuzzy embedding deep learning model is defined as consisting of three core modules: a semantic encoding module, a culture-modulated attention module, and a vector generation module.

[0042] In this embodiment, the fuzzy embedding deep learning model includes the three modules mentioned above. The semantic encoding module, as the front end of the model, performs preliminary feature transformation and enhancement on the input user semantic vector. This semantic encoding module consists of one or more feedforward neural networks or initial Transformer layers, and is responsible for converting the original user semantic vector into an initial hidden state sequence rich in semantic information.

[0043] The culture-modulated attention module embeds a multi-head self-attention layer to receive the initial hidden state sequence. Its key role is that, instead of directly performing conventional attention calculations, it uses the user's cultural membership matrix as a guiding signal to dynamically modulate the attention distribution of the multi-head self-attention layer. This modulation process allows the fuzzy embedding deep learning model to selectively emphasize or ignore certain semantic information based on the user's cultural preferences when allocating attention, thus outputting a hidden state sequence enhanced with cultural context. Each hidden state in this sequence is imbued with a specific cultural semantic connotation.

[0044] The vector generation module, serving as the backend of the model, is tasked with integrating the culturally context-enhanced sequence of hidden states. This module utilizes aggregation or selection operations, such as global average pooling and selecting hidden states corresponding to specific flags (e.g., [CLS]), to aggregate the sequence of hidden states into a fixed-dimensional, culturally sensitive vector matrix that comprehensively represents the user's deep cultural preferences.

[0045] As an example, the dynamic modulation of the attention mechanism in the fuzzy embedding deep learning model using the user culture membership matrix includes: The user culture membership matrix is ​​projected into a culture modulation vector through a fully connected network; In this step, the user cultural membership matrix is ​​a multidimensional tensor that contains the degree of continuous identification of users with multiple cultural elements at different cultural levels. If it is directly operated with the attention weight matrix in the model, it will be difficult to achieve due to the mismatch between the two dimensions and structures.

[0046] Therefore, this invention projects the user's cultural membership matrix using a fully connected network. This fully connected network is a non-linear transformer containing a trainable weight matrix and bias terms, connected to an activation function. By inputting the user's cultural membership matrix (which can be flattened or pooled before being fed into the network) into this fully connected network, the network can automatically learn to extract the most critical features for the current task from high-dimensional cultural preference information and output a cultural modulation vector that matches the dimension of the model's hidden layers.

[0047] Understandably, this projection process essentially transforms structured cultural affiliations into a concise, model-understandable distributed vector representation. Simultaneously, the nonlinear activation function in the fully connected network enables the model to capture the complex nonlinear impact of user cultural preferences on attention allocation.

[0048] After calculating the attention weight matrix, the cultural modulation vector is fused with the attention weight matrix to generate a modulated attention weight matrix, thereby reshaping the attention distribution.

[0049] In this step, within the multi-head self-attention layer of the fuzzy embedding deep learning model, after calculating the original attention weight matrix based on the query matrix and key matrix, the culture modulation vector generated in the previous step is further fused with the original attention weight matrix. This fusion operation is key to achieving dynamic modulation. Its purpose is to reshape the original attention distribution, which is based solely on content similarity, according to the user preference information carried by the culture modulation vector, in order to generate a modulated attention weight matrix.

[0050] In practice, this fusion operation can be performed using any of the following methods: Additive fusion: After broadcasting the cultural modulation vector dimensionally, it is weighted with a learnable scaling factor and then added to the original attention log (its), followed by Softmax normalization. This approach is equivalent to adding a bias reward to the semantic relevance related to user cultural preferences, giving it a higher weight in the final attention distribution. It is suitable for scenarios that require gentle enhancement while preserving the original semantic relevance.

[0051] Multiplicative fusion (gated modulation): The cultural modulation vector is transformed into a gated vector with values ​​between [0,1] using the Sigmoid function, and then broadcast and multiplied element-wise with the original attention weight matrix. This approach is more refined and aggressive, acting like a cultural filter that directly suppresses attention streams to content with low cultural affiliation to the user, achieving precise cultural-level information filtering. For example, when a user has a high affinity for "minimalism," even if the semantics contain descriptions related to "complexity," the model can significantly reduce attention to such information through the gating mechanism.

[0052] Through the above fusion operation, the culture modulation attention module outputs a hidden state sequence enhanced with cultural context. Its information aggregation process has been deeply integrated with the user's cultural preference guidance, which is conducive to the subsequent generation of a high-quality culture-sensitive vector matrix.

[0053] As an example, projecting the user culture membership matrix into a culture modulation vector through a fully connected network includes: In the first path of the pre-constructed dual-path modulation vector generation network, the user culture membership matrix is ​​directly input into the fully connected network to generate the basic culture modulation vector. In the second path of the dual-path modulation vector generation network, a contextual cultural anchor vector related to the current cultural product to be recommended is introduced. The interaction matrix between the user's cultural membership matrix and the contextual cultural anchor vector is calculated, and the interaction matrix is ​​input into another fully connected network to generate a context-aware cultural modulation vector. In this step, the impact of a user's cultural preferences on their attention allocation is not constant across different recommendation contexts. To address this issue, this invention further projects the user's cultural membership matrix into a cultural modulation vector through a fully connected network, specifically executed within a pre-constructed dual-path modulation vector generation network. This network comprises two parallel processing paths, ultimately fusing the outputs of the two paths. The first path extracts the user's inherent, static cultural preference base; the second path captures the dynamic cultural preference focus elicited by the currently recommended product. This dual-path design effectively decouples the user's stable cultural identity from context-dependent preference focus, enabling the generated modulation signal to more accurately reflect the user's psychological state in specific human-product interaction scenarios.

[0054] First Path: Extraction of Basic Cultural Modulation Vectors In the first path of the dual-path modulation vector generation network, the user culture membership matrix is ​​directly input into a fully connected network to generate the basic culture modulation vector.

[0055] The purpose of this path design is to ensure that the user's long-term, stable core cultural inclinations (e.g., being an overall enthusiast of Eastern culture) are preserved as the foundation for modulation, preventing the model from completely deviating from the user's fundamental interests due to excessive focus on the current product context. Understandably, the basic cultural modulation vector generated by this first path is a universal, decontextualized generalization of the user's cultural profile.

[0056] Second approach: Generation of context-aware cultural modulation vectors The second path of the dual-path modulation vector generation network aims to achieve micro-interactions between preferences and context, specifically: First, a contextual cultural anchor vector related to the currently recommended cultural product is introduced. This contextual cultural anchor vector originates from the product cultural vector generated in step S20 and can be further refined through a lightweight network layer to extract the core cultural characteristics of the product. Understandably, the introduction of this anchor vector transforms the modulation signal generation process from a one-way application of user preferences to a two-way dialogue between the user and the product's cultural attributes.

[0057] Next, the interaction matrix between the user's cultural membership matrix and the contextual cultural anchor vector is calculated. In a preferred implementation, an attention mechanism can be used for this calculation. Specifically, the contextual cultural anchor vector is used as the query, and the user's cultural membership matrix (or its transformed form) is used as the key and value. Through an attention layer, the relevance weight between the product anchor and the user's cultural preferences is calculated, and the user preference information is weighted and aggregated accordingly. This process can be understood as the product proactively querying the user's preference matrix to explore the points of matching and conflict between the two in terms of cultural characteristics.

[0058] Then, the calculated interaction matrix is ​​input into another fully connected network to generate a context-aware culture modulation vector. This culture modulation vector specifically characterizes the state in which user cultural preferences are activated or suppressed under the stimulation of the cultural attributes of the currently recommended product, indicating the specific direction in which user cultural preferences exert their influence when faced with this particular product.

[0059] The basic cultural modulation vector and the context-aware cultural modulation vector are weighted and fused to generate the final cultural modulation vector.

[0060] In this step, the basic cultural modulation vector and the context-aware cultural modulation vector are weighted and fused to generate the final cultural modulation vector. This weighted fusion process integrates the user's long-term cultural identity and short-term contextual focus. The weighting coefficients can be preset fixed values ​​or dynamically generated by the network based on the current context.

[0061] The final cultural modulation vector generated by the above integration not only maintains the overall consistency of user cultural preferences, but also has dynamic adaptability to specific recommended objects, which is conducive to improving the accuracy of the attention modulation process.

[0062] Please see Figure 3 The present invention also provides a marketing recommendation system 200 for cultural products in e-commerce, the system including a first acquisition unit 201, a second acquisition unit 202, a user analysis unit 203, a first filtering processing unit 204, and a second filtering processing unit 205; The first acquisition unit 201 is configured to: acquire the original data of the target user, and construct a user semantic vector and a user cultural membership matrix based on the original data; wherein, the user cultural membership matrix contains multiple levels, which are used to characterize the degree of continuous identification of the user with cultural elements at different cultural levels; The second acquisition unit 202 is configured to: acquire the original information of the cultural product to be recommended, and generate a product culture vector based on the original information; The user analysis unit 203 is configured to: input the user semantic vector and the user cultural membership matrix into a configured fuzzy embedding deep learning model, and use the user cultural membership matrix to dynamically modulate the attention mechanism in the fuzzy embedding deep learning model to output a culturally sensitive vector matrix that reflects the user's deep cultural preferences. The first filtering processing unit 204 is configured to: calculate the comprehensive similarity between the cultural sensitivity vector matrix and the product cultural vector, filter and sort users based on the comprehensive similarity, and generate an initial recommendation candidate set; The second filtering processing unit 205 is configured to: merge the initial recommendation candidate set with the platform's operation strategy rules, determine the final recommendation object set based on the fusion result, and push the recommended cultural products to users in the recommendation object set.

[0063] The present invention also provides an electronic device including a processor and a memory, the memory storing a program or instructions executable on the processor, the program or instructions being executed by the processor to implement the steps of the method as described in any of the foregoing embodiments.

[0064] The present invention also provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method as described in any of the foregoing embodiments.

[0065] The present invention also provides a computer program product stored in a storage medium, the program product being executed by at least one processor to implement the steps of the method as described in any of the foregoing embodiments.

[0066] The above description is only a part of the embodiments of this application and does not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A marketing recommendation method for cultural products in e-commerce, characterized in that, The methods and steps include the following: Step S10: Obtain the original data of the target user, and construct a user semantic vector and a user cultural membership matrix based on the original data; wherein, the user cultural membership matrix contains multiple levels, which are used to characterize the degree of continuous identification of the user with cultural elements at different cultural levels; Step S20: Obtain the original information of the cultural product to be recommended, and generate a product culture vector based on the original information; Step S30: Input the user semantic vector and the user cultural membership matrix into the configured fuzzy embedding deep learning model. Use the user cultural membership matrix to dynamically modulate the attention mechanism in the fuzzy embedding deep learning model to output a culturally sensitive vector matrix that reflects the user's deep cultural preferences. Step S40: Calculate the comprehensive similarity between the cultural sensitivity vector matrix and the product cultural vector; filter and sort users based on the comprehensive similarity to generate an initial recommendation candidate set. Step S50: The initial recommendation candidate set is merged with the platform's operation strategy rules. Based on the fusion result, the final recommendation object set is determined, and the recommended cultural products are pushed to users in the recommendation object set.

2. The marketing recommendation method for cultural products in e-commerce according to claim 1, characterized in that: The user culture membership matrix is ​​constructed using a fuzzy comprehensive evaluation model, specifically: The fuzzy comprehensive evaluation model integrates multi-source behavioral data and contextual data from the user's original data and outputs a membership degree that varies continuously within the interval [0,1]. The multi-source behavioral data includes explicit behavioral data and implicit behavioral data that reflects the intensity of interest, and the contextual data includes at least time, location, and social context.

3. The marketing recommendation method for cultural products in e-commerce according to claim 1, characterized in that: The fuzzy embedding deep learning model includes a semantic encoding module, a culture-modulated attention module, and a vector generation module; The semantic encoding module is used to convert the user semantic vector into an initial hidden state sequence; The culture-modulated attention module includes a multi-head self-attention layer, which receives the initial hidden state sequence and dynamically modulates the attention distribution of the multi-head self-attention layer using the user culture membership matrix, and outputs a hidden state sequence enhanced with cultural context. The vector generation module is used to aggregate or select the cultural context-enhanced hidden state sequence to generate the culturally sensitive vector matrix.

4. The marketing recommendation method for cultural products in e-commerce according to claim 3, characterized in that: The attention mechanism in the fuzzy embedding deep learning model is dynamically modulated using the user culture membership matrix, including: The user culture membership matrix is ​​projected into a culture modulation vector through a fully connected network; After calculating the attention weight matrix, the cultural modulation vector is fused with the attention weight matrix to generate a modulated attention weight matrix, thereby reshaping the attention distribution.

5. The marketing recommendation method for cultural products in e-commerce according to claim 4, characterized in that: Projecting the user culture membership matrix into a culture modulation vector through a fully connected network includes: In the first path of the pre-constructed dual-path modulation vector generation network, the user culture membership matrix is ​​directly input into the fully connected network to generate the basic culture modulation vector. In the second path of the dual-path modulation vector generation network, a contextual cultural anchor vector related to the current cultural product to be recommended is introduced. The interaction matrix between the user's cultural membership matrix and the contextual cultural anchor vector is calculated, and the interaction matrix is ​​input into another fully connected network to generate a context-aware cultural modulation vector. The basic cultural modulation vector and the context-aware cultural modulation vector are weighted and fused to generate the final cultural modulation vector.

6. A marketing recommendation system for cultural products in e-commerce, characterized in that: The system includes a first acquisition unit, a second acquisition unit, a user analysis unit, a first filtering processing unit, and a second filtering processing unit; The first acquisition unit is configured to: acquire the original data of the target user, and construct a user semantic vector and a user cultural membership matrix based on the original data; wherein, the user cultural membership matrix contains multiple levels, which are used to characterize the degree of continuous identification of the user with cultural elements at different cultural levels; The second acquisition unit is configured to: acquire the original information of the cultural product to be recommended, and generate a product culture vector based on the original information; The user analysis unit is configured to: input the user semantic vector and the user cultural membership matrix into a configured fuzzy embedding deep learning model, and use the user cultural membership matrix to dynamically modulate the attention mechanism in the fuzzy embedding deep learning model to output a culturally sensitive vector matrix that reflects the user's deep cultural preferences. The first filtering processing unit is configured to: calculate the comprehensive similarity between the cultural sensitivity vector matrix and the product cultural vector, filter and sort users based on the comprehensive similarity, and generate an initial recommendation candidate set; The second filtering processing unit is configured to: merge the initial recommendation candidate set with the platform's operation strategy rules, determine the final recommendation object set based on the fusion result, and push the recommended cultural products to users in the recommendation object set.

7. A marketing recommendation system for cultural products in e-commerce according to claim 6, characterized in that: The user culture membership matrix is ​​constructed using a fuzzy comprehensive evaluation model, specifically: The fuzzy comprehensive evaluation model integrates multi-source behavioral data and contextual data from the user's original data and outputs a membership degree that varies continuously within the interval [0,1]. The multi-source behavioral data includes explicit behavioral data and implicit behavioral data that reflects the intensity of interest, and the contextual data includes at least time, location, and social context.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the steps of the method as described in any one of claims 1 to 5.

9. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 5.

10. A computer program product, characterized in that, The program product is stored in a storage medium and is executed by at least one processor to implement the steps of the method as claimed in any one of claims 1 to 5.