Methods and related products for predicting user consumption behavior under cross-tenant privacy protection

Abstract

This application discloses a method and related products for predicting user consumption behavior under cross-tenant privacy protection. Upon receiving a tenant's consumption prediction request for a target user, the method first acquires the target user's interaction data and current geographic location data. Then, using a large language model, the interaction data is transformed into a first semantic note describing the user's consumption motivation, and the geographic location data is abstracted into regional commercial attribute tags without specific street information, resulting in a first tag. The first semantic note and the first tag are then vectorized and concatenated to form a target feature vector. Next, reference vectors with high similarity to the target feature vector are retrieved from a public vector database, and combined with the group characteristics associated with the reference vectors, inference analysis is performed using a large language model to finally obtain the predicted consumption intention of the target user. This application can effectively utilize cross-brand data for precise marketing while ensuring user privacy and security.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, specifically to a method and related products for predicting user consumption behavior under cross-tenant privacy protection. Background Technology

[0002] When new users first encounter a brand, the brand often struggles to accurately predict their purchase intentions due to a lack of historical consumption data. Although these users may have established consumption patterns with other brands, information that could serve as crucial predictive data cannot be effectively utilized due to data silos between brands.

[0003] To achieve data interoperability between different brands, it is common practice to rely on matching "hard identifiers" such as phone numbers and Identifier for Advertisers (IDFA) to obtain cross-brand user behavior data and build more complete user profiles based on this data. However, this interoperability method based on "hard identifiers" essentially falls under the category of personal information processing. It not only requires explicit authorization from users, thus increasing the complexity and cost of business implementation, but also poses significant privacy and security risks.

[0004] If this core identity data is leaked, it could directly link to specific individuals, leading to the exposure of their identities. Furthermore, user profiles often include detailed geographic location information, such as street names and house numbers. This type of information has strong deanonymization capabilities, allowing for the derivation of specific individuals from addresses, which also exposes user identities and violates fundamental principles of personal information protection.

[0005] Therefore, how to effectively predict users' purchase intentions across brands while ensuring user privacy and security is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] To address the aforementioned issues, this application provides a method and related products for predicting user consumption behavior under cross-tenant privacy protection, which can effectively utilize cross-brand data for precise marketing while ensuring user privacy and security.

[0007] The embodiments of this application disclose the following technical solutions:

[0008] A method for predicting user consumption behavior with cross-tenant privacy protection, the method comprising:

[0009] In response to a consumption prediction request for a target user issued by any tenant, the system acquires the target user's interaction data and the target user's current geographical location data; the interaction data includes product click records and product browsing records.

[0010] The interaction data is transformed into a natural language description of user consumption motivation using a large language model to obtain a first semantic note, and the current geographical location data is abstracted into a regional commercial attribute label that does not contain specific street information to obtain a first label.

[0011] The first semantic note and the first tag are vectorized and concatenated to obtain the target feature vector;

[0012] Retrieve at least one reference vector from a pre-built public vector database that has a similarity to the target feature vector greater than or equal to a similarity threshold;

[0013] Based on the at least one reference vector and the group features associated with the reference vector, as well as the target feature vector, the large language model is used for inference analysis to obtain the consumption intention prediction result of the target user.

[0014] In one possible implementation, the construction process of the public vector database includes:

[0015] The system retrieves second semantic notes and corresponding second tags from multiple historical users uploaded by any tenant. The second semantic note represents the user's consumption motivation, obtained by converting the consumption behavior data of each historical user into natural language descriptions using the large language model. The second tag corresponding to the second semantic note is a regional commercial attribute tag that abstracts the historical geographical location data of each historical user without containing specific street information. The consumption behavior data is log data describing user consumption behavior after the tenant has anonymized the data, including product click records, product browsing records, product order records, and customer service records.

[0016] The second semantic notes and second tags of each historical user are vectorized and concatenated to obtain the behavioral feature vectors of each historical user. The behavioral feature vectors of each historical user are then clustered to obtain multiple vector clusters.

[0017] Cluster center vectors or typical sample vectors are extracted from each vector cluster, and statistical analysis is performed based on multiple behavioral feature vectors in each vector cluster to obtain the group characteristics of each vector cluster; the group characteristics are used to describe the common and distinctive consumption trends and preferences of users in the vector cluster.

[0018] The common vector database is obtained by associating and storing each cluster center vector or typical sample vector with the corresponding group features.

[0019] In one possible implementation, the extraction process of cluster center vectors in each vector cluster includes:

[0020] Using a pre-defined vector clustering algorithm, the spatial mean vector of all behavioral feature vectors within each vector cluster is calculated, and this mean vector is determined as the cluster center vector of the corresponding vector cluster.

[0021] The cluster center vector is used to characterize the core semantic features of all behavioral feature vectors within the corresponding vector cluster.

[0022] In one possible implementation, the step of obtaining the group characteristics of each vector cluster by performing statistical analysis based on multiple behavioral feature vectors in each vector cluster includes:

[0023] For each vector cluster, text semantic mining is performed on the second semantic note corresponding to each behavioral feature vector in the vector cluster to extract common features of user consumption preferences and obtain common consumption preferences. Furthermore, attribute summarization is performed on the second label corresponding to each behavioral feature vector in the vector cluster to extract common features of the user's consumption environment and obtain common consumption environment.

[0024] Statistically analyze the conversion probability of consumption behavior for each vector cluster under the corresponding common consumption preferences and common consumption environment;

[0025] By integrating the common consumption preferences, common consumption environments, and consumption behavior conversion probabilities of each vector cluster, the group characteristics of each vector cluster are obtained.

[0026] In one possible implementation, the step of abstracting the current geographic location data into a regional commercial attribute label that does not contain specific street information to obtain a first label includes:

[0027] Based on a geographic information database, the business district type, housing price tier, and community functional attributes of the current geographic location are obtained; the geographic information database integrates massive amounts of Points of Interest (POI) data and Areas of Interest (AOI) data.

[0028] The first tag corresponding to the current geographical location data is obtained by integrating the business district type, housing price tier, and community functional attributes to which the current geographical location belongs.

[0029] In one possible implementation, the step of using a large language model to transform the interaction data into a natural language description of user consumption motivation to obtain a first semantic note includes:

[0030] The interactive data is filtered based on filtering rules to obtain filtered data; the filtering rules are to remove user IDs, specific product IDs, and transaction serial numbers from the interactive data.

[0031] Generate semantic conversion instructions based on preset prompts and the filtered data;

[0032] The semantic transformation instruction is input into the large language model for behavioral narrative processing to obtain the first semantic note;

[0033] The prompt words are used to guide the large language model to delete sensitive information in the interaction data and to transform the discrete interaction data into a natural language description that includes consumption motivation.

[0034] In one possible implementation, the consumption intention prediction results include the target user's purchase probability for various types of goods and personalized recommendation suggestions.

[0035] A user consumption behavior prediction device with cross-tenant privacy protection, the device comprising:

[0036] The first acquisition unit, in response to a consumption prediction request for a target user issued by any tenant, is used to acquire the target user's interaction data and the target user's current geographical location data; the interaction data includes product click records and product browsing records;

[0037] The data transformation unit is used to transform the interactive data into a natural language description of the user's consumption motivation using a large language model, thereby obtaining a first semantic note.

[0038] A data abstraction unit is used to abstract the current geographic location data into regional commercial attribute tags that do not contain specific street information, thus obtaining a first tag;

[0039] The first vectorization unit is used to vectorize and concatenate the first semantic note and the first tag to obtain the target feature vector.

[0040] The retrieval unit is used to retrieve at least one reference vector in a pre-built public vector database that has a similarity to the target feature vector greater than or equal to a similarity threshold.

[0041] The prediction unit is used to perform inference analysis using the large language model based on the at least one reference vector and the group features associated with the reference vector, as well as the target feature vector, to obtain the consumption intention prediction result of the target user.

[0042] A user consumption behavior prediction device with cross-tenant privacy protection includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the user consumption behavior prediction method with cross-tenant privacy protection as described above.

[0043] A computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the user consumption behavior prediction method under cross-tenant privacy protection as described above.

[0044] Compared with the prior art, this application has the following beneficial effects:

[0045] This application provides a method and related products for predicting user consumption behavior under cross-tenant privacy protection. Specifically, when executing the method for predicting user consumption behavior under cross-tenant privacy protection provided in this application embodiment, upon receiving a consumption prediction request from any tenant, the recent interaction data of the target user and the user's current geographical location data are obtained. This interaction data includes the user's product click records and product browsing records, thereby providing a comprehensive view of user activity. Next, a large language model is used to transform this interaction data into natural language descriptions to extract the user's potential consumption motives and generate a first semantic note. At the same time, to protect user privacy, the user's actual geographical location information is abstracted into regional commercial attribute tags that do not contain specific street information, thereby forming a first tag. Subsequently, these two are vectorized and concatenated to construct a target feature vector. This feature vector is then compared with reference vectors in a pre-built public vector database, and at least one reference vector with a similarity greater than or equal to a similarity threshold is retrieved. Finally, based on the retrieved reference vectors and their associated group characteristics, combined with the target feature vector, inference analysis is performed through a large language model to finally obtain the prediction result of the target user's consumption intention. This application utilizes a large language model to transform users' raw interaction data into natural language notes describing their consumption motivations, and abstracts geographical location into regional business attributes (rather than specific addresses). This process removes sensitive information locally, thus mitigating the risk of leakage of users' personal information and addresses. Simultaneously, by retrieving similar group characteristics from a vector database for inference, brands can leverage cross-domain common consumption trends to help predict new user intentions. This effectively breaks down data silos and improves the accuracy of new user predictions without infringing on user privacy. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in this embodiment or the prior art, the drawings used in the description of the embodiment or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 A flowchart illustrating a method for predicting user consumption behavior under cross-tenant privacy protection, provided in this application embodiment;

[0048] Figure 2 A flowchart illustrating a method for constructing a public vector database, as provided in this application embodiment;

[0049] Figure 3 This is a schematic diagram of a user consumption behavior prediction device under cross-tenant privacy protection provided in an embodiment of this application. Detailed Implementation

[0050] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0051] When new users first encounter a brand, the brand often struggles to accurately determine their purchase intentions due to a lack of data on their past consumer behavior. While these users may have already established mature consumer behavior patterns across other brands—patterns that could serve as crucial for predicting their needs—this cross-brand behavioral information cannot be effectively utilized due to data silos between brands. To overcome this limitation, different brands typically need to use "hard identifiers" such as phone numbers and IDFA (IDF) for matching, thereby obtaining cross-brand user behavior data and building a more complete user profile based on this data.

[0052] However, this data exchange method relying on "hard identifiers" essentially falls under the category of personal information processing. It not only requires explicit user authorization, increasing the complexity and cost of business implementation, but also poses significant privacy and security risks. Once this core identity data is leaked, it could directly link to specific individuals, leading to the exposure of their identity information. At the same time, traditional user profiles often include detailed geographic location information such as street names and house numbers. This type of information has strong deanonymization capabilities, allowing for the reverse deduction of specific individuals from addresses, which can also lead to the leakage of user identities and violate the basic principles of personal information protection.

[0053] To address this issue, this application provides a method and related products for predicting user consumption behavior under cross-tenant privacy protection. When any tenant issues a consumption prediction request for a target user, the method acquires the target user's interaction data and current geographic location data. Then, a large language model is used to transform the interaction data into a natural language description of the user's consumption motivation, generating a first semantic note. Simultaneously, the current geographic location data is abstracted into a regional commercial attribute tag without specific street information, resulting in a first tag. The first semantic note and the first tag are then vectorized and concatenated to generate a target feature vector. Further, at least one reference vector with a similarity greater than or equal to a similarity threshold to the target feature vector is retrieved from a pre-built public vector database. Based on the retrieved reference vector and its associated group features, as well as the target feature vector, a large language model is used for inference analysis to ultimately derive the target user's consumption intention prediction result. This application not only accurately predicts user consumption intentions without historical user data but also effectively protects user privacy, avoiding the direct use of sensitive personal information and reducing the risk of privacy leakage. Furthermore, by utilizing a public vector database to retrieve similar features, the accuracy and efficiency of the prediction are improved.

[0054] 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 this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0055] See Figure 1 The figure is a flowchart of a method for predicting user consumption behavior under cross-tenant privacy protection provided in an embodiment of this application. Figure 1 As shown, the method for predicting user consumption behavior under cross-tenant privacy protection may include steps S101-S105:

[0056] S101: In response to a consumption prediction request for a target user issued by any tenant, obtain the target user's interaction data and the target user's current geographical location data.

[0057] To break down data barriers between brands and accurately predict new users' purchasing intentions while protecting user privacy, the system first acquires target user interaction data and current geographic location data upon receiving a prediction request from a tenant. This interaction data includes product click records and browsing history, thus providing a comprehensive view of user activity.

[0058] S102: Use a large language model to transform the interaction data into a natural language description of user consumption motivation to obtain a first semantic note, and abstract the current geographical location data into a regional commercial attribute label that does not contain specific street information to obtain a first label.

[0059] To eliminate privacy-sensitive information from raw data at the source and transform interactive data into high-value semantic features that can be used for cross-brand matching, the deep understanding capabilities of large language models can be leveraged to guide the model through "semantic desensitization" by constructing specific prompts.

[0060] Specifically, for interactive data, the prompt word instruction model ignores sensitive information such as specific user identifiers (IDs), specific product IDs, transaction serial numbers, and precise timestamps, focusing instead on extracting the underlying consumption motivations of users. For geographic location data, the prompt word instruction model masks specific latitude, longitude, and house numbers, retaining only regional commercial attributes. This prompt word-based processing method ensures that the output only contains abstract descriptions related to consumption intentions, thus preserving the business value of the data without compromising privacy.

[0061] For example, suppose the target user's interaction data shows: "User A clicked on product ID-89757 (a certain brand of baby diapers) at 22:30 on October 1, 2023", and the GPS location shows that it is located at "501, Building 3, Xingfu Garden, XX District, XX City".

[0062] Prompt word design based on behavioral data:

[0063] Input prompt: "Please analyze the interaction data 'User A clicked on product ID-89757 (a certain brand of baby diapers) at 22:30 on October 1, 2023,' and the GPS location shows that the user is located at 'Building 3, Room 501, Xingfu Garden, XX District, XX City.' To infer what stage of life the user might be in? What are their core purchasing concerns? Please describe this in one sentence, without including sensitive information such as user ID, specific product ID, transaction serial number, or precise timestamp."

[0064] Model output (first semantic note): "The user is suspected of being in the process of preparing for a second pregnancy and is choosing diapers for the newborn, with the core concern being size fit."

[0065] Prompt word design based on geographic location data:

[0066] Input prompt: "Based on the following latitude and longitude or address, please determine the commercial attributes and consumption level of this area. Please output 'the type of business district + housing price tier + community function attributes'. Specific street names or house numbers are strictly prohibited."

[0067] Model output (first label): "Mid-to-high-end residential cluster, high housing price tiers, mother-and-baby friendly living community".

[0068] Through the above process, the specific "User A" and "Building 3, Unit 501, Happiness Garden" were transformed into the abstract "User preparing for a second child + Resident of a mid-to-high-end mother-and-baby friendly community", effectively blocking the path of identity leakage.

[0069] This approach conceals the user's specific identity and whereabouts while preserving their core consumption needs and contextual information, enabling subsequent cross-brand searches to accurately identify user groups with similar characteristics while protecting privacy.

[0070] S103: Vectorize and concatenate the first semantic note and the first tag to obtain the target feature vector.

[0071] To transform the anonymized text semantics and contextual attributes into computer-computable mathematical features and achieve deep semantic alignment across brands, the first semantic note and the first tag can be further mapped to a high-dimensional vector space. Subsequently, vector concatenation is performed to generate a target feature vector that uniquely represents the user's current consumption intention and environment. This process merges discrete text descriptions and category labels into a continuous vector representation, preserving not only the user's deep consumption motivations and contextual information but also eliminating specific identifiers. This allows for secure similarity comparison of user data from different brands at the vector level, breaking down data silos and achieving cross-domain collaboration where "data is usable but not visible."

[0072] For example, suppose the target user's first semantic note is "suspected of being in the second pregnancy preparation period, focusing on product comfort and size fit," and the first tag is "top-tier business district in a first-tier city, high-net-worth business area, luxury goods cluster." Through vectorization, the semantic note is transformed into an embedding vector [0.12,-0.35,0.89,...], and the tag is transformed into a vector [0.91,0.05,-0.42,...]. Concatenating these two vectors yields the target feature vector [0.12,-0.35,0.89,...,0.91,0.05,-0.42,...]. This vector contains both the user's consumption motivation characteristics and the commercial attributes of their geographical location, without carrying any personally identifiable information. By comparing this vector with vectors in a public vector database, user groups with similar consumption motivations and scenario characteristics can be found, thus providing accurate consumption intention predictions for new users.

[0073] S104: Retrieve at least one reference vector in a pre-built public vector database that has a similarity to the target feature vector that is greater than or equal to a similarity threshold.

[0074] To aid predictions by leveraging common consumption patterns across brands without directly sharing raw data, the generated target feature vectors can be input into a pre-built public vector database. Vector similarity retrieval algorithms (such as cosine similarity) are then used to find reference vectors that are closest to the target user's features (i.e., similarity greater than or equal to a similarity threshold, such as 0.85). This process achieves "vector-level" cross-brand data collaboration, enabling brands to utilize the massive user behavior patterns (existing in vector form) accumulated by other brands to provide reference data for new or cold-start users. This effectively breaks down data silos and improves prediction accuracy while protecting data privacy.

[0075] S105: Based on the at least one reference vector and the group features associated with the reference vector, as well as the target feature vector, the large language model is used for inference analysis to obtain the consumption intention prediction result of the target user.

[0076] To deeply integrate retrieved group patterns with the individual characteristics of target users, thereby generating interpretable and accurate predictions, the powerful contextual understanding and reasoning capabilities of large language models can be leveraged. This involves integrating "semantic reconstruction of target feature vectors," "group behavior patterns associated with reference vectors," and "specific prediction instructions" into structured prompts, which are then input into the model for comprehensive reasoning. This prompt-based reasoning approach allows the model to analyze the current user's immediate behavior by incorporating historical experience from similar groups (such as purchase conversion rates and preference tendencies), thus outputting decision-making criteria that include not only prediction results but also recommendation justifications.

[0077] For example, suppose the target user's feature vector, after restoration, is "browsing diapers late at night and inquiring about sizes, located in a top-tier business district in a first-tier city, and belonging to a high-net-worth business area and luxury goods cluster," while the retrieved reference vector's associated group characteristics are "this type of user typically purchases XL size within 3 days, with a high repurchase rate." The system will construct the following prompt word input large language model:

[0078] Please analyze the user intent based on the following information:

[0079] 1. Current user behavior: {semantic description of the target user} (e.g., suspected of preparing for a second pregnancy, focusing on comfort, seeking advice late at night).

[0080] 2. Characteristics of similar groups: {Group patterns of reference vectors} (e.g., 70% of users of the same type will buy size XL, prefer ultra-thin styles, and have an average order value of 300 yuan).

[0081] 3. Task Requirements: Predict the probability of this user making a purchase and recommend products within the next 72 hours.

[0082] The large language model analyzes the prompt word and outputs a prediction: "This user has a strong purchase intention (90% confidence), it is recommended to prioritize pushing coupons for XL size ultra-thin diapers." This process uses the prompt word as a bridge to realize the transformation of vector data into human-understandable language and intelligent decision-making.

[0083] Based on the content of S101-S105, upon receiving a consumption prediction request for a target user from any tenant, the system first acquires the target user's interaction data and their current geographical location data. Next, a large language model is used to transform the interaction data into a natural language description of the user's consumption motivation, resulting in a first semantic note. The current geographical location data is then abstracted into a regional commercial attribute label without specific street information, resulting in a first tag. Then, the first semantic note and the first tag are vectorized and concatenated to obtain a target feature vector. At least one reference vector with a similarity greater than or equal to a similarity threshold with the target feature vector is retrieved from a pre-built public vector database. Finally, based on at least one reference vector and the associated group features, as well as the target feature vector, a large language model is used for inference analysis to obtain the target user's consumption intention prediction result. This application achieves effective prediction of cross-brand user purchase intentions while ensuring user privacy and security.

[0084] In one possible implementation, such as Figure 2 As shown, Figure 2 A flowchart of a method for constructing a public vector database provided in this application embodiment can be implemented through steps S201-S204:

[0085] S201: Obtain the second semantic notes of multiple historical users uploaded by any tenant and the second tags corresponding to the second semantic notes.

[0086] To break down data silos between brands and aggregate anonymized data from multiple sources to uncover universal group consumption patterns, the first step is to acquire multiple sets of related data uploaded by any tenant. Specifically, this includes second semantic notes and corresponding second tags for each historical user. The second semantic notes are generated based on the tenant's anonymized consumer behavior logs: the tenant first anonymizes the original consumer behavior data, including product click records, product browsing records, product order records, and customer service records, removing sensitive information such as user IDs, phone numbers, and specific product IDs. Then, a large language model is used to transform these structured behavior logs into natural language descriptions focusing on consumption motivations. The second tags are the result of privacy-enhancing the historical geographic location data of each user. Sensitive location information such as specific streets and house numbers is removed through abstraction and refinement, retaining only feature tags reflecting the regional commercial attributes. Each second tag is associated with a corresponding second semantic note, ensuring data consistency and usability.

[0087] S202: Vectorize and concatenate the second semantic notes and second tags of each historical user to obtain the behavioral feature vector of each historical user, and cluster the behavioral feature vectors of each historical user to obtain multiple vector clusters.

[0088] To transform anonymized discrete semantic information into computer-computable continuous features and uncover potential common consumption patterns across brand user groups, the second semantic note and second tag of each historical user can be mapped to a high-dimensional vector space, generating corresponding semantic vectors and attribute vectors. These are then concatenated to form a behavioral feature vector that comprehensively represents the user's consumption motivation and environment. Based on this, clustering algorithms (such as K-Means) are used to group the massive user behavioral feature vectors, aggregating users with similar features to form multiple tight vector clusters. This process not only achieves data dimensionality reduction and structuring but, more importantly, breaks down brand boundaries at the vector level, allowing users with similar consumption preferences (such as a focus on high-end maternal and infant products) to be grouped together, thus laying the foundation for extracting universal group patterns.

[0089] For example, assuming that after preliminary processing, the behavioral feature vector generated by tenant A's historical user 1 is V1 (representing "focus on DHA content + top-tier business districts in first-tier cities, high-net-worth business areas"), and the behavioral feature vector generated by tenant B's historical user 2 is V2 (representing "purchasing high-end wet wipes + top-tier business districts in first-tier cities, luxury goods gathering areas"), although these two users come from different brands, in the vector space, because their consumption motivation both point to high-end maternal and infant needs, and their geographical locations are both abstracted as top-tier business districts where high-net-worth individuals gather, V1 and V2 are very close. Through clustering algorithms, the system will cluster V1, V2, and other vectors with similar characteristics into a cluster (e.g., named "High-Net-Worth Maternal and Infant Care Cluster"). Another user from tenant C, with a feature vector of V3 (representing "focus on low-price promotions + ordinary communities, cost-effective business districts"), will be clustered into a completely different "price-sensitive cluster".

[0090] S203: Extract the cluster center vector or typical sample vector from each vector cluster, and perform statistical analysis based on multiple behavioral feature vectors in each vector cluster to obtain the group characteristics of each vector cluster.

[0091] To transform the clustered vectors into structured knowledge that can be understood and used for reasoning by a large language model, two core vector extraction methods can be employed: First, extract the cluster center vector for each vector cluster and condense the core semantic features of the group by calculating the spatial mean of all behavioral feature vectors within the cluster, serving as the "common representative" of the group. Second, select typical sample vectors within the cluster, i.e., select key vectors that have high similarity to other vectors within the cluster and accurately reflect the core behavioral patterns of the group, serving as the "concrete representative" of the group. After obtaining representative vectors of the group through either method, further statistical analysis is performed on the original semantic information (including second semantic notes and second tags) corresponding to all behavioral feature vectors within the cluster, summarizing group characteristics that reflect the group's common consumption trends, preferences, and conversion patterns. These group characteristics are presented in natural language form, including users' life stages, spending power, focus areas, and typical purchase paths, providing interpretable and quantifiable references for subsequent consumption intention prediction.

[0092] For example, suppose a vector cluster consists of a large number of user vectors from different tenants. The semantic features of these vectors often involve "concern about DHA content," "preference for imported infant formula," and "location in a top-tier business district in a first-tier city." Through statistical analysis, the system finds that 70% of users in this cluster, after focusing on ingredient safety, will purchase high-end imported infant formula, and 20% will also purchase high-end baby care products, mostly from high-net-worth business districts. Based on this, the system extracts the cluster center vector and generates a corresponding group characteristic description: "High-net-worth individuals, valuing product quality, in the pre-pregnancy / parenting stage; after focusing on ingredient safety, 70% will purchase high-end imported infant formula, and 20% will also purchase high-end baby care products." This group characteristic retains the core consumption pattern of this type of user while completely eliminating any personally identifiable information, making it safe for subsequent cross-brand inference.

[0093] S204: Associate and store each cluster center vector or typical sample vector with the corresponding group features to obtain the public vector database.

[0094] To solidify the anonymized, clustered, and analyzed group knowledge into a structured resource for subsequent retrieval and reasoning, the cluster center vector or typical sample vector of each vector cluster can be associated and stored with its corresponding group characteristics, thus constructing a public vector database. Each record in this database consists of two parts: a "vector index" and "semantic knowledge." The cluster center vector serves as the index key for fast retrieval, used to locate similar users in high-dimensional space; while the group characteristics serve as a behavioral specification for that type of user, providing specific consumption preferences and conversion patterns. This storage structure achieves an organic combination of "vector retrieval efficiency" and "semantic interpretation capability," enabling the rapid acquisition of the most valuable cross-brand group insights through a single vector similarity comparison when processing new user requests.

[0095] For example, assuming that after preliminary processing, the system generates a vector group named "High-Net-Worth Maternal and Infant Care Cluster". The system uses the cluster center vector of this cluster (e.g., [0.85, 0.12, -0.34, ...]) as the key and the corresponding group characteristic description (e.g., "high-net-worth individuals, valuing product quality, in the pre-pregnancy / childcare stage, after paying attention to ingredient safety, 70% will purchase high-end imported maternal and infant formula, and 20% will also purchase high-end infant care products") as the value, storing them in a public vector database in key-value pairs. After the database is built, its internal structure may look like the following:

[0096] Index Vector 1 → Group Feature 1: "High-Net-Worth Maternal and Infant Care Cluster... (as described above)";

[0097] Index Vector 2 → Group Feature 2: "Price-sensitive family clusters... (Describes preference for low-price promotions)".

[0098] Through the construction process of the public vector database in steps S201-S204, the public vector database retains only the anonymized group-level features and cluster center vectors, without containing any information that can identify individuals. This enables secure sharing of cross-brand data at the vector layer, providing reliable support for subsequent consumer intention prediction based on group patterns.

[0099] In one possible implementation, the extraction process of cluster center vectors in each vector cluster includes:

[0100] Using a pre-defined vector clustering algorithm (such as the K-Means algorithm), the spatial mean vector of all behavioral feature vectors within each vector cluster is calculated, and this mean vector is determined as the cluster center vector of the corresponding vector cluster.

[0101] The cluster center vector is used to represent the core semantic features of all behavioral feature vectors within the vector cluster. It can serve as a "typical representative" of the group in the subsequent retrieval process, ensuring that when retrieving similar vectors based on the cluster center vector, user feature vectors with similar consumption needs and behavioral patterns can be accurately matched.

[0102] In this way, the cluster center vector not only condenses the common features of the entire cluster, but also forms a stable and reusable retrieval anchor in the vector space, enabling the feature vector of a new user to quickly find the most similar group, thereby realizing cross-brand consumption intention prediction.

[0103] In one possible implementation, step S204 involves performing statistical analysis based on multiple behavioral feature vectors in each vector cluster to obtain the group characteristics of each vector cluster, including A1-A3:

[0104] A1: For each vector cluster, perform text semantic mining on the second semantic note corresponding to each behavioral feature vector in the vector cluster to extract common features of user consumption preferences and obtain common consumption preferences. Also, perform attribute summarization on the second label corresponding to each behavioral feature vector in the vector cluster to extract common features of the user's consumption environment and obtain common consumption environment.

[0105] To extract business-guided group profiles from clustered vector clusters, this method performs deep semantic aggregation and attribute summarization on the anonymized data of all users within the cluster. Specifically, for each vector cluster, textual semantic mining is performed on the second semantic notes corresponding to each behavioral feature vector in the vector cluster to extract high-frequency keywords and core demands of users in terms of consumption motivation, product preferences, etc., forming the common consumption preferences of the group. At the same time, attribute summarization is performed on the second tags corresponding to each behavioral feature vector in the vector cluster to extract the common features of the user's consumption environment, thereby transforming the specific description of the business district into abstract consumption capacity and life scenario characteristics, forming the common consumption environment of the group.

[0106] For example, suppose a vector cluster contains 1,000 user vectors from different tenants.

[0107] Extracting common consumption preferences: The system analyzed the second semantic notes corresponding to these 1,000 users and found that descriptions such as "focus on the safety of imported ingredients", "pursuit of organic certification", and "not sensitive to price" appeared most frequently. The system integrated these high-frequency features and concluded that the common consumption preferences of this cluster are: "Highly educated parents who attach great importance to the safety and natural properties of maternal and infant products and belong to the high-end consumer group."

[0108] Extracting Common Consumption Environment: The system summarized the second tags corresponding to these 1,000 users and found that 90% of the tags included "top-tier business districts in first-tier cities" or "high-net-worth business districts". Based on the preset mapping rules, the system extracted these geographical attributes into the common consumption environment of this cluster as: "living in high-value locations in core cities, possessing extremely strong immediate purchasing power and high-end consumption capabilities".

[0109] A2: Statistically determine the conversion probability of consumption behavior for each vector cluster under the corresponding common consumption preferences and common consumption environment.

[0110] To transform abstract group profiles into quantifiable and predictable decision-making criteria, in-depth statistical analysis can be performed on historical behavioral data within each vector cluster. This involves statistically analyzing the frequency of user purchases, add-to-cart actions, or repeat purchases under specific combinations of "common consumption preferences + common consumption environment," thereby obtaining the conversion probability of the group's consumption behavior.

[0111] For example, assuming a vector cluster of "highly educated parents + high-value locations in core cities," the system performs retrospective statistics on 1000 historical user records within the cluster. The results show that among users who prioritize "ingredient safety" and reside in "top-tier business districts in first-tier cities," 70% completed the purchase of high-end imported infant formula within 3 days of browsing, and 20% also purchased organic cotton diapers. Based on this, the system calculates the conversion probability of this cluster's consumption behavior as: "70% conversion rate for high-end infant formula purchase, 20% conversion rate for high-end diapers purchase."

[0112] A3: Integrate the common consumption preferences, common consumption environment, and consumption behavior conversion probabilities of each vector cluster to obtain the group characteristics of each vector cluster.

[0113] To create a comprehensive, accurate, and interpretable group profile for each vector cluster, this method structurally integrates the common consumption preferences, common consumption environments, and consumption behavior conversion probabilities obtained in the preceding steps, ultimately generating group characteristics for each vector cluster. These group characteristics not only include what users "pay attention to" and "what environment they are in," but also "what behaviors they might engage in," thus providing complete, quantifiable, and reasonable group-level knowledge support for subsequent consumption intention prediction.

[0114] Through steps A1-A3, the group characteristics include not only what users "want" (consumption preferences) and "what environment they are in" (consumption environment), but also "what they might do" (conversion probability), thus providing interpretable and quantifiable group-level decision-making basis for subsequent consumption intention prediction.

[0115] In one possible implementation, the step of abstracting the current geographic location data into a regional commercial attribute label that does not contain specific street information to obtain a first label includes:

[0116] Based on a pre-built geographic information database, the business district type, housing price range, and community functional attributes of the current geographical location are obtained. This geographic information database integrates massive amounts of Point of Interest (POI) and Area of ​​Interest (AOI) data, providing a multi-level, structured description of the geographical location. Subsequently, the business district type, housing price range, and community functional attributes of the current geographical location are structurally integrated to obtain the first tag corresponding to the current geographical location data.

[0117] In this way, the first tag can accurately reflect the commercial value and demographic characteristics of the user's area without containing any sensitive information such as specific streets or house numbers, thus providing effective contextual information for subsequent consumption intention prediction while protecting privacy.

[0118] In one possible implementation, step S102 utilizes a large language model to transform the interaction data into a natural language description of the user's consumption motivation, resulting in a first semantic note, including:

[0119] The interactive data is filtered using filtering rules to remove sensitive information such as user IDs, specific product IDs, transaction serial numbers, and precise timestamps. Then, semantic transformation instructions are generated based on preset prompts and the interactive data. These prompts guide the large language model to automatically remove sensitive information (such as user IDs, specific product IDs, transaction serial numbers, and precise timestamps) from the interactive data during processing, transforming the discrete, structured interactive data into a natural language description containing the user's consumption motivation, demand preferences, and life stage. Subsequently, the semantic transformation instructions are input into the large language model for behavioral narrative processing, and the model outputs a first semantic note that accurately reflects the user's current consumption intention.

[0120] In this way, First Semantic Notes not only achieves semantic extraction of raw behavioral data, but also completes automatic filtering of privacy information during the generation process, ensuring that the user characteristics used for cross-brand matching do not contain any personally identifiable content, thereby improving data usability while protecting user privacy and security.

[0121] It should be noted that the transformation process of the second semantic note is similar to that of the first semantic note, and will not be repeated here.

[0122] In one possible implementation, the consumption intention prediction result may include, but is not limited to, two key components: first, the probability of target users purchasing various types of goods, which is quantified as a percentage to show the likelihood of users making a purchase, providing an intuitive basis for screening high-value demand goods; second, the accompanying personalized recommendation suggestions, which are generated based on users' consumption motives, their consumption environment, and common consumption patterns of the group, to guide brands to efficiently connect with user needs.

[0123] For example, taking "pregnant women who are concerned about the safety of ingredients in maternal and infant products and live in top-tier business districts in first-tier cities" as an example, their consumption intention prediction results are as follows:

[0124] 1. Purchase probability: High-end imported DHA-containing milk powder 82%, organic cotton newborn diapers 75%, natural ingredient baby diaper rash cream 53%, ordinary domestic milk powder 28%, affordable diapers 21%;

[0125] 2. Personalized Recommendation Suggestions: ① For "Premium Imported DHA-Containing Infant Formula" (82% High Purchase Probability): As a core recommended product, emphasize the three core selling points during communication: "import traceability, additive-free ingredients, and DHA to support brain development," and use purchase cases from users in the same demographic to enhance trust; ② For "Organic Cotton Newborn Diapers" (75% Relatively High Purchase Probability): As a recommended complementary product, design a "formula + diapers" pre-pregnancy essential package, offer exclusive discounts to increase conversion rates, and highlight the gentle and skin-friendly characteristics of organic cotton, suitable for newborns' sensitive skin; ③ For "Baby Natural Ingredient Diaper Rash Cream" (53% Moderate Purchase Probability): Not a proactively recommended product, but can be mentioned casually when users inquire about diapers, emphasizing the selling points of "natural ingredients and non-irritating" to meet users' potential baby care needs; ④ For ordinary domestic infant formula and affordable diapers with a purchase probability of less than 30%: Not recommended at this time to avoid contradicting users' high-end consumption preferences and reducing user experience.

[0126] See Figure 3 , Figure 3 This is a schematic diagram of a user consumption behavior prediction device under cross-tenant privacy protection, provided as an embodiment of this application. Figure 3 As shown, the user consumption behavior prediction device under cross-tenant privacy protection includes:

[0127] The first acquisition unit 301, in response to a consumption prediction request for a target user issued by any tenant, is used to acquire the target user's interaction data and the target user's current geographical location data; the interaction data includes product click records and product browsing records;

[0128] Data conversion unit 302 is used to convert the interaction data into a natural language description of user consumption motivation using a large language model, thereby obtaining a first semantic note;

[0129] Data abstraction unit 303 is used to abstract the current geographic location data into regional commercial attribute tags that do not contain specific street information, thereby obtaining a first tag;

[0130] The first vectorization unit 304 is used to vectorize and concatenate the first semantic note and the first tag to obtain the target feature vector.

[0131] The retrieval unit 305 is used to retrieve at least one reference vector in a pre-built public vector database that has a similarity to the target feature vector that is greater than or equal to a similarity threshold.

[0132] The prediction unit 306 is used to perform inference analysis using the large language model based on the at least one reference vector and the group features associated with the reference vector, as well as the target feature vector, to obtain the consumption intention prediction result of the target user.

[0133] In one possible implementation, the device further includes:

[0134] The second acquisition unit is used to acquire the second semantic notes and the second tags corresponding to the second semantic notes of multiple historical users uploaded by any tenant; the second semantic note is the user's consumption motivation obtained by converting the consumption behavior data of each historical user into natural language description through the large language model; the second tag corresponding to the second semantic note is the regional commercial attribute tag that does not contain specific street information after abstracting the historical geographical location data of each historical user; the consumption behavior data is log data describing the user's consumption behavior after the tenant has desensitized it, and the consumption behavior data includes product click records, product browsing records, product order records and customer service records;

[0135] The second vectorization unit is used to vectorize and concatenate the second semantic notes and second tags of each historical user to obtain the behavioral feature vector of each historical user.

[0136] Clustering units are used to cluster the behavioral feature vectors of each historical user to obtain multiple vector clusters;

[0137] The extraction unit is used to extract the cluster center vector or typical sample vector from each vector cluster.

[0138] The statistical unit is used to perform statistical analysis based on multiple behavioral feature vectors in each vector cluster to obtain the group characteristics of each vector cluster; the group characteristics are used to describe the common and distinctive consumption trends and preferences of users in the vector cluster.

[0139] The associated storage unit is used to associate and store each cluster center vector or typical sample vector with the corresponding group features to obtain the public vector database.

[0140] In one possible implementation, the extraction unit is specifically used for:

[0141] Using a pre-defined vector clustering algorithm, the spatial mean vector of all behavioral feature vectors within each vector cluster is calculated, and this mean vector is determined as the cluster center vector of the corresponding vector cluster.

[0142] The cluster center vector is used to characterize the core semantic features of all behavioral feature vectors within the corresponding vector cluster.

[0143] In one possible implementation, the statistical unit is specifically used for:

[0144] For each vector cluster, text semantic mining is performed on the second semantic note corresponding to each behavioral feature vector in the vector cluster to extract common features of user consumption preferences and obtain common consumption preferences. Furthermore, attribute summarization is performed on the second label corresponding to each behavioral feature vector in the vector cluster to extract common features of the user's consumption environment and obtain common consumption environment.

[0145] Statistically analyze the conversion probability of consumption behavior for each vector cluster under the corresponding common consumption preferences and common consumption environment;

[0146] By integrating the common consumption preferences, common consumption environments, and consumption behavior conversion probabilities of each vector cluster, the group characteristics of each vector cluster are obtained.

[0147] In one possible implementation, the data abstraction unit 303 is specifically used for:

[0148] Based on a geographic information database, the business district type, housing price range, and community functional attributes of the current geographical location are obtained; the geographic information database integrates massive amounts of POI location data and AOI area data;

[0149] The first tag corresponding to the current geographical location data is obtained by integrating the business district type, housing price tier, and community functional attributes of the current geographical location.

[0150] In one possible implementation, the data conversion unit 302 is specifically used for:

[0151] The interactive data is filtered based on filtering rules to obtain filtered data; the filtering rules are to remove user ID, specific product ID and transaction serial number from the consumer behavior data.

[0152] Generate semantic conversion instructions based on preset prompts and the filtered data;

[0153] The semantic transformation instruction is input into the large language model for behavioral narrative processing to obtain the first semantic note;

[0154] The prompt words are used to guide the large language model to delete sensitive information in the interaction data and to transform the discrete interaction data into a natural language description that includes consumption motivation.

[0155] In one possible implementation, the consumption intention prediction results include the target user's purchase probability for various types of goods and personalized recommendation suggestions.

[0156] In addition, this application embodiment also provides a user consumption behavior prediction device under cross-tenant privacy protection, including: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the user consumption behavior prediction method under cross-tenant privacy protection as described above.

[0157] In addition, this application embodiment also provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the user consumption behavior prediction method under cross-tenant privacy protection as described above.

[0158] This application utilizes a large language model to transform users' raw behavioral data into natural language notes describing their consumption motivations, and abstracts geographical location into regional business attributes (rather than specific addresses). This process removes sensitive information locally, effectively mitigating the risk of leakage of users' personal information and addresses. Simultaneously, by retrieving similar group characteristics from a vector database and performing inference analysis, it becomes possible to leverage cross-domain common consumption trends to help predict the purchase intentions of new users. This method not only ensures user privacy but also breaks down data silos between different brands without requiring explicit authorization, significantly improving the accuracy of new user predictions.

[0159] The foregoing provides a detailed description of a user consumption behavior prediction method and related products under cross-tenant privacy protection provided in this application. The various embodiments in the specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

[0160] 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.

[0161] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A method for predicting user consumption behavior under cross-tenant privacy protection, characterized in that, The method includes: In response to a consumption prediction request for a target user issued by any tenant, the system acquires the target user's interaction data and the target user's current geographical location data; the interaction data includes product click records and product browsing records. The interaction data is transformed into a natural language description of user consumption motivation using a large language model to obtain a first semantic note, and the current geographical location data is abstracted into a regional commercial attribute label that does not contain specific street information to obtain a first label. The first semantic note and the first tag are vectorized and concatenated to obtain the target feature vector; Retrieve at least one reference vector from a pre-built public vector database that has a similarity to the target feature vector greater than or equal to a similarity threshold; Based on the at least one reference vector and the group features associated with the reference vector, as well as the target feature vector, the large language model is used for inference analysis to obtain the consumption intention prediction result of the target user.

2. The method according to claim 1, characterized in that, The construction process of the public vector database includes: The system retrieves second semantic notes and corresponding second tags from multiple historical users uploaded by any tenant. The second semantic note represents the user's consumption motivation, obtained by converting the consumption behavior data of each historical user into natural language descriptions using the large language model. The second tag corresponding to the second semantic note is a regional commercial attribute tag that abstracts the historical geographical location data of each historical user without containing specific street information. The consumption behavior data is log data describing user consumption behavior after the tenant has anonymized the data, including product click records, product browsing records, product order records, and customer service records. The second semantic notes and second tags of each historical user are vectorized and concatenated to obtain the behavioral feature vectors of each historical user. The behavioral feature vectors of each historical user are then clustered to obtain multiple vector clusters. Cluster center vectors or typical sample vectors are extracted from each vector cluster, and statistical analysis is performed based on multiple behavioral feature vectors in each vector cluster to obtain the group characteristics of each vector cluster; the group characteristics are used to describe the common and distinctive consumption trends and preferences of users in the vector cluster. The common vector database is obtained by associating and storing each cluster center vector or typical sample vector with the corresponding group features.

3. The method according to claim 2, characterized in that, The extraction process of cluster center vectors in each vector cluster includes: Using a pre-defined vector clustering algorithm, the spatial mean vector of all behavioral feature vectors within each vector cluster is calculated, and this mean vector is determined as the cluster center vector of the corresponding vector cluster. The cluster center vector is used to characterize the core semantic features of all behavioral feature vectors within the corresponding vector cluster.

4. The method according to claim 2, characterized in that, The statistical analysis based on multiple behavioral feature vectors in each vector cluster yields the group characteristics of each vector cluster, including: For each vector cluster, text semantic mining is performed on the second semantic note corresponding to each behavioral feature vector in the vector cluster to extract common features of user consumption preferences and obtain common consumption preferences. Furthermore, attribute summarization is performed on the second label corresponding to each behavioral feature vector in the vector cluster to extract common features of the user's consumption environment and obtain common consumption environment. The conversion probability of consumption behavior for each vector cluster under the corresponding common consumption preferences and common consumption environment is statistically analyzed. By integrating the common consumption preferences, common consumption environments, and consumption behavior conversion probabilities of each vector cluster, the group characteristics of each vector cluster are obtained.

5. The method according to claim 1, characterized in that, The step of abstracting the current geographic location data into regional commercial attribute tags that do not contain specific street information to obtain a first tag includes: Based on a geographic information database, the business district type, housing price tier, and community functional attributes of the current geographic location are obtained; the geographic information database integrates massive amounts of Points of Interest (POI) data and Areas of Interest (AOI) data. The first tag corresponding to the current geographical location data is obtained by integrating the business district type, housing price tier, and community functional attributes to which the current geographical location belongs.

6. The method according to claim 1, characterized in that, The process of using a large language model to transform the interaction data into a natural language description of user consumption motivation, resulting in a first semantic note, includes: The interactive data is filtered based on filtering rules to obtain filtered data; the filtering rules are to remove user IDs, specific product IDs, and transaction serial numbers from the interactive data. Generate semantic conversion instructions based on preset prompts and the filtered data; The semantic transformation instruction is input into the large language model for behavioral narrative processing to obtain the first semantic note; The prompt words are used to guide the large language model to delete sensitive information in the interaction data and to transform the discrete interaction data into a natural language description that includes consumption motivation.

7. The method according to claim 1, characterized in that, The consumption intention prediction results include the target user's purchase probability for various types of goods and personalized recommendation suggestions.

8. A device for predicting user consumption behavior under cross-tenant privacy protection, characterized in that, The device includes: The first acquisition unit, in response to a consumption prediction request for a target user issued by any tenant, is used to acquire the target user's interaction data and the target user's current geographical location data; the interaction data includes product click records and product browsing records; The data transformation unit is used to transform the interactive data into a natural language description of the user's consumption motivation using a large language model, thereby obtaining a first semantic note. A data abstraction unit is used to abstract the current geographic location data into regional commercial attribute tags that do not contain specific street information, thus obtaining a first tag; The first vectorization unit is used to vectorize and concatenate the first semantic note and the first tag to obtain the target feature vector. The retrieval unit is used to retrieve at least one reference vector in a pre-built public vector database that has a similarity to the target feature vector greater than or equal to a similarity threshold. The prediction unit is used to perform inference analysis using the large language model based on the at least one reference vector and the group features associated with the reference vector, as well as the target feature vector, to obtain the consumption intention prediction result of the target user.

9. A user consumption behavior prediction device with cross-tenant privacy protection, characterized in that, include: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor, when executing the computer program, implements the user consumption behavior prediction method under cross-tenant privacy protection as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a terminal device, cause the terminal device to perform the user consumption behavior prediction method under cross-tenant privacy protection as described in any one of claims 1-7.

Images