Intelligent information retrieval and storage method and system based on distributed network

Abstract

The present application relates to the technical field of distributed network information retrieval and storage, and particularly relates to an intelligent information retrieval and storage method and system based on a distributed network. The method comprises the following steps: obtaining original information objects of member consumption behaviors, and performing semantic analysis and structured extraction to obtain knowledge units; distributing the knowledge units to multiple storage nodes according to member characteristic semantic tags and business correlation attributes; generating a global semantic index network according to the member consumption preference correlation relationship of the knowledge units among the nodes; receiving a recommendation retrieval request of a point marketing activity, and determining a target storage node set based on member group characteristics and the correlation path in the global semantic index network; generating a retrieval execution plan in combination with the node load state; generating a target retrieval result containing a target member group and a marketing strategy after parallel retrieval, so as to improve the accuracy and response speed of marketing information retrieval in a distributed environment, and support precise point marketing.

Description

Technical Field

[0001] This invention relates to the field of distributed network information retrieval and storage technology, and in particular to intelligent information retrieval and storage methods and systems based on distributed networks. Background Technology

[0002] In the field of membership points marketing, the storage and management of consumer behavior data typically relies on centralized databases or simple distributed architectures. The conventional approach is to store member transaction records, points change logs, and basic attribute information in structured tables in a relational database, and then perform offline aggregation and analysis through a data warehouse. To ensure query response speed, some systems introduce caching layers or shard frequently accessed data; however, the semantic relationships between data still depend on manually predefined inter-table join rules, and marketing retrieval is often achieved through fixed-dimensional condition matching.

[0003] A significant drawback of existing practices is the severe lack of semantic depth in data storage. Consumer behavior is broken down into discrete transaction fields, severing the inherent connections between member characteristics and business attributes, making it difficult to form a knowledge-based system that reflects genuine consumer preferences. When point-based marketing scenarios targeting specific member groups emerge, search requests can only remain at the surface attribute matching level, failing to effectively utilize the consumer preference chain across storage locations, resulting in unsatisfactory targeting and recall rates for search results.

[0004] Another drawback is the low efficiency of cross-storage node retrieval. In multi-node deployment environments, conventional strategies typically distribute data statically according to data hashing rules or round-robin strategies. Retrieval requests often need to traverse multiple candidate nodes one by one or perform a large amount of data merging on the coordinating node side. This process lacks awareness and scheduling of the real-time load status of each node, which can easily lead to a resource imbalance where some nodes are overloaded while others are idle. Response latency increases significantly with the scale of nodes, failing to meet the timeliness requirements of marketing activities in scenarios with massive membership groups. Summary of the Invention

[0005] The embodiments of the present invention provide an intelligent information retrieval and storage method and system based on distributed networks, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an intelligent information retrieval and storage method based on a distributed network, comprising:

[0007] Obtain the original information object representing the member's consumption behavior, and perform semantic parsing and structured extraction on the original information object to obtain knowledge units;

[0008] Based on the member feature semantic tags and business association attributes of the knowledge units, the knowledge units are allocated to multiple storage nodes in a distributed network. A global semantic index network is generated based on the member consumption preference associations between the knowledge units on different storage nodes.

[0009] Receive a request for a points-based marketing activity recommendation and retrieval for a specific member group, and determine the target storage node set for storing relevant knowledge units based on the member group characteristics in the points-based marketing activity recommendation and retrieval request and the member consumption preference association path in the global semantic index network.

[0010] Based on the association path of member consumption preferences and the node load status of each target storage node, a retrieval execution plan is generated. Based on the retrieval execution plan, a parallel retrieval request is initiated to the set of target storage nodes to generate target retrieval results containing the target member group and corresponding marketing strategies.

[0011] Obtain the original information object representing member consumption behavior, perform semantic parsing and structured extraction on the original information object to obtain knowledge units, including:

[0012] Obtain raw information objects containing member consumption behavior records from the retail enterprise's transaction system. The raw information objects include basic member attribute information, product purchase record information, and points change record information.

[0013] Cross-business domain semantic parsing is performed on the original information object to determine the temporal association pattern between the member's basic attribute information and the product purchase record information, as well as the causal dependency relationship between the product purchase record information and the points change record information.

[0014] Based on the temporal association pattern and the causal dependency, determine the cross-domain semantic features that characterize the semantic association between member consumption preferences and points usage tendency;

[0015] Based on the cross-domain semantic features, multi-level structured extraction is performed. The part of the member's basic attribute information that reflects the member's identity attributes is converted into member feature semantic tags. The part of the cross-domain semantic features that reflects the semantic relationship between the member's consumption preferences and points usage tendency is converted into business association attributes. A knowledge unit containing the member feature semantic tags and the business association attributes is generated.

[0016] Based on the member characteristic semantic tags and business association attributes of the knowledge units, the knowledge units are allocated to multiple storage nodes in a distributed network. A global semantic index network is generated based on the member consumption preference associations between knowledge units on different storage nodes, including:

[0017] A hash mapping function is constructed based on the member feature semantic tags of the knowledge unit. The location identifier of the knowledge unit on the distributed network logical ring is calculated through the hash mapping function. The candidate storage node corresponding to the knowledge unit is determined based on the location identifier.

[0018] Based on the business association attributes of the knowledge unit, a load balancing assessment is performed on the candidate storage nodes. According to the load balancing assessment results, a target storage node is selected from the candidate storage nodes, and the knowledge unit is allocated to the target storage node for storage.

[0019] Cross-node semantic association analysis is performed on knowledge units on multiple storage nodes in a distributed network to identify knowledge unit pairs that represent the same type of member consumption preferences, and to establish the member consumption preference association relationship between the knowledge unit pairs.

[0020] Based on the association of member consumption preferences, a cross-node semantic association path is constructed. The knowledge unit identifiers in the semantic association path, the semantic association strength of the association of member consumption preferences, and the storage node identifiers where the knowledge units are located are structured and organized to generate the global semantic index network.

[0021] Cross-node semantic association analysis is performed on knowledge units across multiple storage nodes in a distributed network to identify pairs of knowledge units representing the same type of member consumption preferences, and to establish member consumption preference associations between these knowledge unit pairs, including:

[0022] Obtain the set of business-related attributes of the knowledge units stored on each storage node from multiple storage nodes in the distributed network;

[0023] The set of business-related attributes is aggregated across nodes to generate a global attribute view that includes the business-related attributes of knowledge units across all storage nodes.

[0024] Based on the global attribute view, member consumption preference types are clustered. By analyzing the semantic similarity features between the business association attributes of knowledge units on different storage nodes, knowledge units representing the same member consumption preference type are grouped into the same preference type cluster, and a preference type mapping table containing the mapping relationship between knowledge unit identifiers and preference type cluster identifiers is generated.

[0025] Based on the preference type mapping table, knowledge unit pairs belonging to the same preference type cluster are identified. For each knowledge unit pair, the business association attributes of the knowledge unit pair are extracted for semantic association measurement, and the semantic association strength between the knowledge unit pairs is calculated.

[0026] Establish the association relationship of member consumption preferences between the knowledge unit pairs based on the semantic association strength.

[0027] The system receives a request for a points-based marketing campaign recommendation, and based on the member group characteristics in the request and the member consumption preference association path in the global semantic index network, determines a target set of storage nodes for storing relevant knowledge units, including:

[0028] Receive the points marketing activity recommendation retrieval request, extract query semantic features representing the characteristics of the member group from the points marketing activity recommendation retrieval request, and convert the query semantic features into query semantic tags that match the member characteristic semantic tags of the knowledge unit;

[0029] Based on the query semantic tags, semantic matching retrieval is performed in the global semantic index network to determine the semantically related knowledge units corresponding to the query semantic tags, and the member consumption preference association path associated with the semantically related knowledge units in the global semantic index network is obtained.

[0030] A path expansion analysis is performed on the association path of the member consumption preference to identify the associated knowledge units that have a relationship with the semantically related knowledge units. The relevance of the associated knowledge units is evaluated based on the semantic association strength in the association relationship of the member consumption preference, and an extended knowledge unit set is generated.

[0031] Extract the storage node identifiers corresponding to the knowledge units from the extended knowledge unit set, perform deduplication and aggregation processing on the storage node identifiers, and generate a target storage node set containing the knowledge units and the associated knowledge units.

[0032] The path expansion analysis is performed on the association path of the member consumption preferences to identify the associated knowledge units that have a relationship with the semantically related knowledge units. The relevance of the associated knowledge units is evaluated based on the semantic association strength in the association relationship of member consumption preferences, generating an extended knowledge unit set, including:

[0033] Extract all member consumption preference associations involving the semantically related knowledge units from the member consumption preference association path, and identify candidate associated knowledge units that form knowledge unit pairs with the semantically related knowledge units based on the knowledge unit pair identifiers recorded in the member consumption preference associations.

[0034] For the candidate associated knowledge unit, the semantic association strength between the semantically related knowledge unit and the candidate associated knowledge unit is obtained from the member consumption preference association relationship;

[0035] Based on the semantic association strength, the semantic matching degree is calculated by matching the member feature semantic tags of the candidate associated knowledge unit with the query semantic tags in the points marketing activity recommendation retrieval request.

[0036] Based on the semantic matching degree, the candidate related knowledge units are screened for relevance. The candidate related knowledge units whose comprehensive relevance scores meet the preset relevance conditions are determined as related knowledge units. The semantically related knowledge units are merged with the screened related knowledge units to generate an extended knowledge unit set.

[0037] Based on the member consumption preference association path and the node load status of each target storage node, a retrieval execution plan is generated. Parallel retrieval requests are then initiated to the target storage node set according to the retrieval execution plan, generating target retrieval results containing the target member group and corresponding marketing strategies, including:

[0038] Obtain the node load status of each target storage node in the target storage node set, evaluate the load capacity of each target storage node based on the node load status, and generate a node load evaluation table.

[0039] Based on the member consumption preference association path, identify the target storage node combination that needs to be jointly retrieved across nodes, and obtain the load capacity evaluation value of each target storage node in the target storage node combination from the node load evaluation table;

[0040] A retrieval execution plan is generated based on the load capacity assessment value and the semantic association strength in the member consumption preference association path;

[0041] According to the retrieval execution plan, a parallel retrieval request is initiated to the target storage node set, and the member consumption behavior data and points usage behavior data obtained from the knowledge units of different target storage nodes are cross-node association and integration.

[0042] Based on the integrated member consumption behavior data and points usage behavior data, target member groups are identified, marketing activity information matching the target member groups is extracted to generate corresponding marketing strategies, and target search results containing the target member groups and corresponding marketing strategies are generated.

[0043] A second aspect of this invention provides an intelligent information retrieval and storage system based on a distributed network, comprising:

[0044] The parsing and extraction unit is used to obtain the original information object representing the member's consumption behavior, and to perform semantic parsing and structured extraction on the original information object to obtain the knowledge unit;

[0045] The semantic indexing unit is used to allocate the knowledge unit to multiple storage nodes in a distributed network based on the member feature semantic tags and business association attributes of the knowledge unit, and to generate a global semantic indexing network based on the member consumption preference association between knowledge units on different storage nodes.

[0046] The retrieval and positioning unit is used to receive a retrieval request for a points marketing activity recommendation targeting a specific member group, and to determine the target storage node set for storing relevant knowledge units based on the member group characteristics in the points marketing activity recommendation retrieval request and the member consumption preference association path in the global semantic index network.

[0047] The parallel retrieval unit is used to generate a retrieval execution plan based on the association path of the member's consumption preferences and the node load status of each target storage node, and to initiate a parallel retrieval request to the set of target storage nodes according to the retrieval execution plan, thereby generating target retrieval results containing the target member group and corresponding marketing strategies.

[0048] A third aspect of the present invention provides an electronic device, comprising:

[0049] processor;

[0050] Memory used to store processor-executable instructions;

[0051] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0053] This invention transforms unstructured consumer behavior data into semantically computable knowledge granularity by semantically parsing and structurally extracting raw information objects, thereby significantly improving the fineness of data organization. Each knowledge unit carries semantic tags representing member characteristics and business-related attributes. When distributed across multiple storage nodes, semantically related data is distributed locally, reducing cross-node association costs at the physical level and laying a data foundation for subsequent retrieval.

[0054] This invention maps the associations of member consumption preferences between knowledge units through a global semantic index network. Member group characteristics in a search request can quickly locate the target storage node set along the association path, avoiding a full network scan. The association path also provides the semantic direction of consumption preferences, narrowing the search scope from all nodes to a subset of relevant nodes, significantly reducing access to invalid data and improving the targeting and response speed of the search.

[0055] This invention generates a retrieval execution plan by combining node load status, enabling parallel retrieval requests to be allocated to each target storage node according to its load capacity, thus avoiding overload of hot nodes. Multiple target nodes execute retrievals simultaneously, fully utilizing the computing and bandwidth resources of the distributed network and significantly shortening the overall retrieval time. The load-aware scheduling method enhances stability in high-concurrency scenarios.

[0056] This invention addresses the recommendation requests for points-based marketing campaigns, generating search results that directly include the target member group and corresponding marketing strategies, saving the additional computational overhead of matching marketing strategies. Semantic storage and distributed layout of knowledge units support horizontal data expansion; as member data dynamically grows, only additional storage nodes are needed, resulting in a near-linear improvement in overall system performance, enhanced robustness, and maintainability. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the intelligent information retrieval and storage method based on a distributed network, according to an embodiment of the present invention.

[0058] Figure 2 This is a flowchart illustrating the process of determining knowledge units according to an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram illustrating the process of determining the load and semantic strength according to an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the process for determining the target retrieval results according to an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0062] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0063] Figure 1 This is a flowchart illustrating the intelligent information retrieval and storage method based on a distributed network, according to an embodiment of the present invention. Figure 1 As shown, intelligent information retrieval and storage methods based on distributed networks include:

[0064] Obtain the original information object representing the member's consumption behavior, and perform semantic parsing and structured extraction on the original information object to obtain knowledge units;

[0065] Based on the member feature semantic tags and business association attributes of the knowledge units, the knowledge units are allocated to multiple storage nodes in a distributed network. A global semantic index network is generated based on the member consumption preference associations between the knowledge units on different storage nodes.

[0066] Receive a request for a points-based marketing activity recommendation and retrieval for a specific member group, and determine the target storage node set for storing relevant knowledge units based on the member group characteristics in the points-based marketing activity recommendation and retrieval request and the member consumption preference association path in the global semantic index network.

[0067] Based on the association path of member consumption preferences and the node load status of each target storage node, a retrieval execution plan is generated. Based on the retrieval execution plan, a parallel retrieval request is initiated to the set of target storage nodes to generate target retrieval results containing the target member group and corresponding marketing strategies.

[0068] Figure 2 This is a flowchart illustrating the process of determining knowledge units according to an embodiment of the present invention. Figure 2 As shown, the original information object representing member consumption behavior is obtained, and semantic parsing and structured extraction are performed on the original information object to obtain knowledge units, including:

[0069] Obtain raw information objects containing member consumption behavior records from the retail enterprise's transaction system. The raw information objects include basic member attribute information, product purchase record information, and points change record information.

[0070] Cross-business domain semantic parsing is performed on the original information object to determine the temporal association pattern between the member's basic attribute information and the product purchase record information, as well as the causal dependency relationship between the product purchase record information and the points change record information.

[0071] Based on the temporal association pattern and the causal dependency, determine the cross-domain semantic features that characterize the semantic association between member consumption preferences and points usage tendency;

[0072] Based on the cross-domain semantic features, multi-level structured extraction is performed. The part of the member's basic attribute information that reflects the member's identity attributes is converted into member feature semantic tags. The part of the cross-domain semantic features that reflects the semantic relationship between the member's consumption preferences and points usage tendency is converted into business association attributes. A knowledge unit containing the member feature semantic tags and the business association attributes is generated.

[0073] When retrieving raw information objects from a retail enterprise's transaction system, it is necessary to cover data sources from multiple business subsystems, including the membership management system, the POS system, and the points management platform. Basic member attribute information includes identity-identifying fields such as member ID, registration time, membership level, geographic location, and age range; product purchase record information includes shopping behavior fields such as transaction timestamp, product category code, purchase quantity, consumption amount, and promotional activity identifier; and points change record information records fields related to the points lifecycle, such as points acquisition time, points source type, points consumption scenario, and points balance change. These three types of information are initially stored separately in different business database tables, linked together by member ID and transaction serial number, but have not yet formed a unified semantic expression structure. During the retrieval phase, the raw information objects need to be validated for completeness, filtering out records with missing fields due to system anomalies, and the timestamp field needs to be standardized and aligned to ensure the data quality for subsequent cross-business domain semantic parsing.

[0074] When performing cross-business domain semantic parsing on raw information objects, the focus is on identifying the temporal association patterns between basic member attribute information and product purchase record information. Determining these temporal association patterns relies on the sequential analysis of multiple purchase behaviors of the same member within a given time window. By using a sliding time window to extract the distribution of member category preferences across different time periods, we can identify changes in the consumption rhythm of specific member groups before and after holidays, promotional seasons, and other key time points. For example, if a certain type of member experiences a significant shift in frequently purchased product categories within 30 days of upgrading their membership, this shift constitutes a typical temporal association pattern, reflecting the driving relationship between changes in member identity attributes and consumption behavior.

[0075] Determining the causal relationship between purchase records and points change records requires establishing a precise event matching mechanism at the transaction event level. Each consumption transaction triggers the increase or consumption of points. By matching the transaction serial number in the purchase record with the associated serial number in the points change record, the complete causal chain from consumption behavior to points change can be reconstructed. Furthermore, by analyzing the correspondence between points consumption scenarios (such as redeeming goods, deducting cash, and participating in activities) and the categories of preceding purchase behaviors, it is possible to identify which points usage methods members prefer in specific consumption scenarios, thereby characterizing the causal dependency structure between points usage preferences and consumption behavior.

[0076] When determining cross-domain semantic features based on temporal correlation patterns and causal dependencies, it is necessary to semantically fuse the results of the two types of analyses mentioned above. The core of cross-domain semantic features lies in expressing the semantic relationship between members' consumption preferences and points usage tendencies, rather than describing a single dimension. Specifically, for the same member group, their consumption preference vector in the category preference dimension and their tendency vector in the points usage scenario dimension are jointly modeled to identify the collaborative patterns between the two dimensions. For example, members who prefer to buy high-priced electrical appliances tend to use points more for cash discounts, while members who prefer to buy fast-moving consumer goods tend to use points more for participating in limited-time redemption activities. Such cross-domain semantic features cannot be obtained independently from data in a single business domain; they must rely on joint analysis across business domains to be effectively identified.

[0077] When performing multi-level structured extraction based on cross-domain semantic features, the extraction process is divided into two levels. The first level focuses on generating semantic tags for member features. This involves extracting combinations of fields reflecting member identity attributes from basic member attribute information and mapping members to several semantic tags according to a predefined tagging system, such as high-value members, new members, and dormant members. Each semantic tag corresponds to specific attribute judgment conditions, ensuring the interpretability and reproducibility of the tags. The design of the tagging system must be closely integrated with the retail business scenario, avoiding the use of purely technical classifications that are detached from business meaning.

[0078] The second layer focuses on generating business-related attributes. It extracts the semantic associations between member consumption preferences and points redemption tendencies from cross-domain semantic features and transforms them into structured business-related attribute fields. These attributes are stored as key-value pairs, where the key describes the semantic association type (e.g., "Category Preference - Points Redemption Association Strength"), and the value is a quantified association metric. During quantification, the association strength is calculated by counting the co-occurrence frequency of members simultaneously meeting specific consumption preference and points redemption tendency conditions within a time window. This frequency is then normalized using the total number of transactions for that member group as the base, resulting in an association strength value ranging from 0 to 1, denoted as [value missing]. ,in The closer the value is to 1, the stronger the semantic association between the two types of behaviors.

[0079] By combining member characteristic semantic tags and business-related attributes, a knowledge unit containing complete semantic information is generated. The data structure design of the knowledge unit must simultaneously meet the needs of subsequent distributed storage allocation and the construction of a global semantic index network. Therefore, during the structured extraction stage, the fields in the knowledge unit are standardized and encoded to ensure consistency in field naming, data type, and value range among knowledge units from different sources. For multiple knowledge units generated for the same member within different time periods, a version link relationship is established through the member ID field to support subsequent tracking and analysis of the evolution trend of member consumption behavior. The entire transformation process from raw information objects to knowledge units constitutes the data foundation for subsequent distributed storage and retrieval operations, and its quality directly determines the accuracy of the global semantic index network and the relevance of the search results recommended by the points marketing campaign.

[0080] Based on the member characteristic semantic tags and business association attributes of the knowledge units, the knowledge units are allocated to multiple storage nodes in a distributed network. A global semantic index network is generated based on the member consumption preference associations between knowledge units on different storage nodes, including:

[0081] A hash mapping function is constructed based on the member feature semantic tags of the knowledge unit. The location identifier of the knowledge unit on the distributed network logical ring is calculated through the hash mapping function. The candidate storage node corresponding to the knowledge unit is determined based on the location identifier.

[0082] Based on the business association attributes of the knowledge unit, a load balancing assessment is performed on the candidate storage nodes. According to the load balancing assessment results, a target storage node is selected from the candidate storage nodes, and the knowledge unit is allocated to the target storage node for storage.

[0083] Cross-node semantic association analysis is performed on knowledge units on multiple storage nodes in a distributed network to identify knowledge unit pairs that represent the same type of member consumption preferences, and to establish the member consumption preference association relationship between the knowledge unit pairs.

[0084] Based on the association of member consumption preferences, a cross-node semantic association path is constructed. The knowledge unit identifiers in the semantic association path, the semantic association strength of the association of member consumption preferences, and the storage node identifiers where the knowledge units are located are structured and organized to generate the global semantic index network.

[0085] Before allocating knowledge units to storage nodes in the distributed network, a hash mapping function needs to be constructed for the member feature semantic tags of each knowledge unit. These member feature semantic tags typically include multi-dimensional attributes such as membership level, product category preference, active time periods, and points sensitivity. These attributes collectively describe the consumption behavior characteristics of a specific member group. The construction process of the hash mapping function involves vectorizing and encoding the aforementioned semantic tags, and then mapping the encoding results to specific positions on the logical ring of the distributed network using a consistent hashing algorithm. Specifically, the logical ring consists of several evenly distributed virtual node positions. Each storage node occupies a continuous interval on the logical ring. The position identifier calculated by the hash mapping function determines which interval it falls into, thus identifying the candidate storage node to which that interval belongs. Let the feature vector of the member feature semantic tags of the knowledge unit after vectorization and encoding be... The hash mapping function will Location identifier mapped onto the logical ring ,but ,in For hash functions, The total length of the logic ring. The range of values ​​is By location identifier By searching clockwise along the logical ring for the starting point of the nearest storage node interval, the candidate storage node corresponding to the knowledge unit can be determined.

[0086] After identifying candidate storage nodes, a load balancing assessment needs to be performed based on the business association attributes of the knowledge units. This is to avoid hotspot issues caused by a large number of knowledge units with similar semantic tags being mapped to the same candidate node. Business association attributes include structured information such as the type of marketing activity associated with the knowledge unit, the points rule encoding, and the effective time window. These attributes reflect the expected access frequency and storage priority of the knowledge unit at the business level. The load balancing assessment comprehensively considers three indicators: the current storage capacity utilization rate of the candidate storage node, the historical retrieval request throughput, and the network transmission latency, constructing a comprehensive score. Let the storage capacity utilization rate of the candidate storage node be... The normalized value of historical retrieval throughput is The normalized value of network transmission delay is Then the overall load score of the candidate node for ,in , , These are the weighting coefficients of the three indicators, satisfying... .score The lower the score, the lighter the current load of the candidate node, and the more suitable it is to receive new knowledge units. When the score of a candidate node exceeds a preset threshold, a virtual node migration mechanism will be triggered to remap the knowledge unit to the next candidate node on the logical ring, until a target storage node with a score that meets the requirements is found. Finally, the knowledge unit, along with its member feature semantic tags and business association attributes, is written into the local index table of the target storage node, completing the storage allocation operation.

[0087] After completing the distributed storage of knowledge units, cross-node semantic association analysis is needed to identify knowledge unit pairs representing the same type of member consumption preferences. The core of cross-node semantic association analysis lies in measuring the semantic similarity between knowledge units on different nodes. In practice, each storage node periodically reports the semantic embedding vectors of all knowledge units within its node to the global coordinator. The global coordinator then aggregates the semantic embedding vectors from each node and uses cosine similarity to calculate the semantic distance between any two knowledge units. Let the semantic embedding vectors of two knowledge units from different storage nodes be... and Then the semantic similarity between the two for ,in Let represent the Euclidean norm of a vector. When Exceeding the preset semantic similarity threshold At that time, knowledge units With knowledge unit Knowledge units that represent the same type of member consumption preferences are identified, and a relationship between them is established. This relationship records the storage node identifiers of the two knowledge units, their respective knowledge unit identifiers, and their corresponding semantic similarity values, serving as the foundational data for constructing the global semantic index network.

[0088] After identifying all knowledge unit pairs and their relationships with member consumption preferences, a cross-node semantic association path is constructed based on these relationships. Essentially, a semantic association path is a directed link connecting multiple knowledge units. Each edge on the link corresponds to a pair of knowledge units with a relationship based on member consumption preferences, and the weight of the edge represents the semantic similarity between the knowledge unit pairs. The path construction employs a breadth-first traversal strategy, starting with a knowledge unit corresponding to a specific high-frequency consumption preference type and expanding layer by layer along the association relationships. Adjacent knowledge units that meet the semantic similarity threshold are sequentially included in the path until the path covers all knowledge units related to that consumption preference type. For path segments spanning different storage nodes, the storage node identifier of each knowledge unit is recorded in the path to accurately locate the physical storage location of the target data during subsequent retrieval execution.

[0089] By structuring all semantic association paths, a global semantic index network can be generated. This network is represented by a graph structure, where each node corresponds to a knowledge unit. Node attributes include the knowledge unit identifier, the identifier of the storage node it resides in, and a summary encoding of the member's feature semantic tags. Each edge in the graph corresponds to a member consumption preference association between a pair of knowledge units, and its attributes include a semantic association strength value and an association type label. The semantic association strength value is directly derived from semantic similarity. After normalization, the values ​​range from 0 to 1, with larger values ​​indicating a stronger correlation between the two knowledge units in terms of member consumption preferences. The global semantic index network is stored locally on each storage node in a sharded manner. Each node only stores the index subgraph directly related to the knowledge units of its own node, while maintaining complete index graph metadata at the global coordinator, including shard boundary information for each subgraph and routing information for cross-node edges. This sharded storage strategy effectively reduces the index maintenance overhead of a single node while ensuring global retrieval capabilities. When a new knowledge unit is written or an existing knowledge unit is updated, an incremental update mechanism is triggered, recalculating and synchronizing only the affected local index subgraphs and related cross-node edges, without requiring a full reconstruction of the global semantic index network. This ensures the real-time performance and consistency of the index network in dynamic data scenarios.

[0090] Cross-node semantic association analysis is performed on knowledge units across multiple storage nodes in a distributed network to identify pairs of knowledge units representing the same type of member consumption preferences, and to establish member consumption preference associations between these knowledge unit pairs, including:

[0091] Obtain the set of business-related attributes of the knowledge units stored on each storage node from multiple storage nodes in the distributed network;

[0092] The set of business-related attributes is aggregated across nodes to generate a global attribute view that includes the business-related attributes of knowledge units across all storage nodes.

[0093] Based on the global attribute view, member consumption preference types are clustered. By analyzing the semantic similarity features between the business association attributes of knowledge units on different storage nodes, knowledge units representing the same member consumption preference type are grouped into the same preference type cluster, and a preference type mapping table containing the mapping relationship between knowledge unit identifiers and preference type cluster identifiers is generated.

[0094] Based on the preference type mapping table, knowledge unit pairs belonging to the same preference type cluster are identified. For each knowledge unit pair, the business association attributes of the knowledge unit pair are extracted for semantic association measurement, and the semantic association strength between the knowledge unit pairs is calculated.

[0095] Establish the association relationship of member consumption preferences between the knowledge unit pairs based on the semantic association strength.

[0096] In a distributed network environment, a large number of knowledge units are stored scattered across different storage nodes. These knowledge units have deep-seated relationships with each other based on member consumption preferences. However, due to their physical distribution across different nodes, it is difficult to perform a comprehensive semantic association analysis directly. Therefore, it is necessary to systematically identify pairs of knowledge units representing the same type of member consumption preferences through a cross-node semantic association analysis mechanism, and establish clear relationships between them based on member consumption preferences. This will provide a foundation for the subsequent construction of a global semantic indexing network.

[0097] When retrieving the set of business-related attributes for knowledge units stored by multiple storage nodes in a distributed network, each storage node reports the business-related attribute information for all knowledge units within its node to the coordinating node. This set of business-related attributes includes, but is not limited to, multi-dimensional attribute fields such as member consumer product category tags, points usage scenario identifiers, consumption frequency range labels, marketing activity participation history markers, and membership level association identifiers. When reporting, each storage node uses the unique identifier of the knowledge unit as the key and the corresponding set of business-related attributes as the value to construct a node-level attribute mapping structure. This ensures that the coordinating node can accurately track the source node information for each attribute record, providing a complete data traceability link for subsequent cross-node aggregation processing.

[0098] When performing cross-node aggregation on the set of business-related attributes, the node-level attribute mapping structures from all storage nodes are merged to generate a global attribute view containing the business-related attributes of knowledge units from all storage nodes. This global attribute view uses a unified attribute model to standardize and align heterogeneous attribute fields reported by each node, addressing issues such as inconsistent field naming and inconsistent attribute granularity caused by differences in local storage strategies across different nodes. Specifically, for synonymous but differently named attribute fields, a unified mapping is performed based on a predefined attribute thesaurus; for attribute values ​​with inconsistent granularity, roll-up or drill-down processing is performed according to a unified attribute granularity specification, ensuring the comparability and consistency of business-related attributes for each knowledge unit in the global attribute view, laying a data quality foundation for subsequent clustering analysis.

[0099] When clustering member consumption preferences based on a global attribute view, the set of business-related attributes for each knowledge unit in the global attribute view is transformed into a structured attribute feature representation. For discrete attribute fields, one-hot encoding is used for numerical processing; for textual attribute fields, a pre-trained language model maps the attribute text into a dense semantic vector, which is then concatenated with the numerical results of the discrete attributes to form a comprehensive attribute feature representation for each knowledge unit. Based on this, by analyzing the semantic similarity features between the business-related attributes of knowledge units on different storage nodes, a hierarchical clustering method is used to cluster all knowledge units by preference type. During hierarchical clustering, the cosine similarity between the comprehensive attribute feature representations of knowledge units is used as the clustering similarity metric. The knowledge units or groups of knowledge units with the highest similarity are gradually merged through a bottom-up agglomeration approach until a preset clustering termination condition is reached. The clustering termination condition can be set as follows: when the average link similarity between two clusters to be merged is lower than a preset preference type discrimination threshold, the merging operation is stopped, and the current clustering result is taken as the final preference type cluster division. Each preference type cluster corresponds to a specific type of member consumption preference, such as high-frequency dining consumption preference, travel points redemption preference, and holiday gift consumption preference. After clustering, a preference type mapping table is generated, which contains the mapping relationship between knowledge unit identifiers and preference type cluster identifiers. This mapping table uses the unique identifier of the knowledge unit as an index, records the preference type cluster identifier to which it belongs, and records the similarity score from the knowledge unit to its cluster center, as auxiliary reference information for subsequent semantic association strength calculation.

[0100] When identifying knowledge unit pairs belonging to the same preference type cluster based on a preference type mapping table, the preference type mapping table is grouped according to the preference type cluster identifier. A list of all knowledge unit identifiers contained within each preference type cluster is extracted. Within the same preference type cluster, these identifier lists are paired to generate a candidate knowledge unit pair set. To control the size of the candidate pairs, for larger preference type clusters, they can be sorted according to the similarity score from the knowledge units to the cluster center. Knowledge units with higher similarity scores are prioritized for pairing, ensuring the quality of the candidate knowledge unit pairs and reducing subsequent computational overhead.

[0101] For each pair of knowledge units in the candidate knowledge unit pair set, extract the business association attributes of the knowledge unit pair for semantic association measurement, and calculate the semantic association strength between the knowledge unit pairs. Let the two knowledge units in the knowledge unit pair be... and The corresponding comprehensive attribute features are represented as follows: and The semantic association strength between the two Calculated in the following way: This calculation method uses cosine similarity to measure the directional consistency of two knowledge units in the business-related attribute space. The value ranges from -1 to 1; the closer the value is to 1, the stronger the semantic association between the two knowledge units in the dimension of member consumption preferences. In practical applications, since the features of business-related attributes are all non-negative, The actual value range is from 0 to 1. (This refers to the semantic association strength.) Knowledge unit pairs that are below the preset effective threshold for association are filtered out from the candidate knowledge unit pair set, and only knowledge unit pairs whose semantic association strength reaches the effective threshold are retained to participate in the establishment of subsequent association relationships.

[0102] When establishing the association between member consumption preferences among knowledge unit pairs based on semantic association strength, the semantic association strength is used as the basis. As a weight attribute for the association relationship, a weighted directed edge is created for each pair of valid knowledge units in the global semantic indexing network. The weight value of the edge is the corresponding semantic association strength value. Simultaneously, the association relationship record also includes the preference type cluster identifiers, source storage node identifiers, and key dimension information for attribute semantic matching of the two knowledge units, forming a structured description of member consumption preference association relationships. This structured association relationship description not only supports rapid retrieval of association paths based on preference type clusters but also provides interpretable association basis for marketing strategy recommendations. For example, when retrieving points marketing activities targeting members with high-frequency dining consumption preferences, relevant knowledge units can be quickly located along high-weight association paths within preference type clusters, improving retrieval accuracy and response efficiency.

[0103] The system receives a request for a points-based marketing campaign recommendation, and based on the member group characteristics in the request and the member consumption preference association path in the global semantic index network, determines a target set of storage nodes for storing relevant knowledge units, including:

[0104] Receive the points marketing activity recommendation retrieval request, extract query semantic features representing the characteristics of the member group from the points marketing activity recommendation retrieval request, and convert the query semantic features into query semantic tags that match the member characteristic semantic tags of the knowledge unit;

[0105] Based on the query semantic tags, semantic matching retrieval is performed in the global semantic index network to determine the semantically related knowledge units corresponding to the query semantic tags, and the member consumption preference association path associated with the semantically related knowledge units in the global semantic index network is obtained.

[0106] A path expansion analysis is performed on the association path of the member consumption preference to identify the associated knowledge units that have a relationship with the semantically related knowledge units. The relevance of the associated knowledge units is evaluated based on the semantic association strength in the association relationship of the member consumption preference, and an extended knowledge unit set is generated.

[0107] Extract the storage node identifiers corresponding to the knowledge units from the extended knowledge unit set, perform deduplication and aggregation processing on the storage node identifiers, and generate a target storage node set containing the knowledge units and the associated knowledge units.

[0108] Upon receiving a request for a points-based marketing campaign recommendation, the system first performs structured parsing of the request content, extracting query semantic features that characterize the member group. These features encompass descriptive information across multiple dimensions, including member spending frequency distribution, points accumulation period, category preferences, spending range, and historical activity participation records. These features exist in the retrieval request in the form of natural language descriptions, structured fields, or semi-structured tags, thus requiring unified semantic normalization. Specifically, a pre-trained semantic encoding model is used to vectorize the query semantic features, obtaining semantic embedding representations. These embedding representations are then converted into query semantic tags that match the member feature semantic tag system of the knowledge unit using a tag mapping dictionary. During the generation of query semantic tags, synonym tags are merged, and higher-level concept tags are hierarchically expanded to ensure that the semantic granularity of the query semantic tags remains consistent with existing tags in the global semantic index network, thus laying an accurate semantic foundation for subsequent semantic matching retrieval.

[0109] After obtaining the query semantic tags, semantic matching retrieval is performed in the global semantic index network. This network uses knowledge units as nodes and member consumption preference relationships as edges, with each knowledge unit node bearing its corresponding member feature semantic tag. The core of semantic matching retrieval is calculating the semantic similarity between the query semantic tags and the member feature semantic tags of each knowledge unit. Knowledge units with similarity exceeding a preset matching threshold are identified as semantically relevant. Semantic similarity is calculated using cosine similarity, based on the cosine of the angle between the embedding vector of the query semantic tag and the embedding vector of the member feature semantic tag of the knowledge unit. For composite query semantic tags containing multiple sub-tags, a weighted aggregation strategy is used to fuse the matching scores of each sub-tag, with weights dynamically allocated based on the semantic importance of each sub-tag in the retrieval request. After initial semantic matching, member consumption preference association paths associated with each semantically relevant knowledge unit are extracted from the global semantic index network. These paths record the co-occurrence relationships of consumption behavior, the sequence relationships of points usage, and the relationships of marketing activity responses between this knowledge unit and other knowledge units, forming the starting basis for subsequent path expansion analysis.

[0110] After obtaining the member consumption preference association paths of semantically related knowledge units, path expansion analysis is performed on these paths to discover extended knowledge units that are indirectly related to the semantically related knowledge units. The path expansion analysis employs a bounded breadth-first traversal strategy, starting from the semantically related knowledge units and expanding outwards layer by layer along the association edges in the global semantic index network. The number of expansion layers is determined by a preset maximum expansion depth parameter. Control measures are implemented to prevent excessive expansion of the path range and the introduction of a large number of low-relevance, noisy knowledge units. During each layer of expansion, candidate related knowledge units in the current layer are filtered based on the semantic correlation strength between them and semantically related knowledge units in terms of member consumption preferences. Only knowledge units with a semantic correlation strength not lower than the expansion filtering threshold are retained. The related knowledge units will be included in the next round of candidate set expansion. The semantic association strength reflects the similarity in consumption behavior patterns and common response tendency of the member groups corresponding to the two knowledge units in points marketing activities. The higher the value, the closer the association between the consumption preferences of the two, and the stronger the potential relevance to the current search request.

[0111] The relevance of the associated knowledge units obtained through path expansion analysis is evaluated to generate an expanded knowledge unit set. The relevance evaluation comprehensively considers two dimensions: first, the degree of direct semantic matching between the associated knowledge units and the query semantic tags, denoted as... The first method is to calculate the cosine similarity between the embedded vectors of member feature semantic tags and the embedded vectors of query semantic tags for the associated knowledge unit; the second method is to calculate the path association strength between the associated knowledge unit and semantically related knowledge units via the member consumption preference association path, denoted as... The score is calculated by multiplying the semantic association strength of each edge along the path, reflecting the attenuation characteristic that the longer the path, the weaker the association. This is the comprehensive relevance score of the associated knowledge units. It is obtained by weighted fusion of the above two dimensions, that is ,in For direct semantic matching weights, For path association strength weights, satisfying All candidate related knowledge units are sorted in descending order of their comprehensive relevance score, and those with a score not lower than the relevance screening threshold are selected. The associated knowledge units constitute an extended knowledge unit set, which, together with the initial semantically related knowledge units, constitutes the complete candidate knowledge unit set for the current retrieval request.

[0112] The storage node identifiers for each knowledge unit are extracted from the extended knowledge unit set and semantically related knowledge units. Each knowledge unit is stored on a specific storage node in the distributed network, and its storage node identifier is recorded and maintained in the knowledge unit's metadata information. Since multiple knowledge units are stored on the same storage node, directly aggregating the storage node identifiers of all knowledge units would result in a large number of duplicates. Therefore, it is necessary to perform deduplication and aggregation processing on the extracted storage node identifiers. The deduplication and aggregation processing performs aggregation operations based on the uniqueness of the storage node identifiers, merging identical storage node identifiers into a single entry. During the merging process, the list of all candidate knowledge units contained on that storage node and their corresponding comprehensive relevance scores are recorded simultaneously. This allows the subsequent retrieval execution stage to accurately locate the range of knowledge units to be retrieved on each storage node, avoiding unnecessary resource consumption caused by full node scanning. After deduplication and aggregation processing, a target storage node set containing semantically related knowledge units and associated knowledge units is generated. This set clearly identifies all storage nodes in the distributed network that need to participate in the recommended retrieval of this points marketing activity, as well as the range of knowledge units to be retrieved on each storage node, providing accurate node location basis for initiating subsequent parallel retrieval requests and generating retrieval execution plans.

[0113] In practical applications, for scenarios with large-scale member data, path expansion analysis involves a huge number of candidate related knowledge units. To ensure the timeliness of retrieval response, an upper limit can be set on the size of the expanded knowledge unit set. When the number of candidate related knowledge units exceeds When doing so, priority will be given to retaining the one with the highest overall relevance score. Each related knowledge unit is added to the extended knowledge unit set to achieve a reasonable balance between retrieval accuracy and retrieval efficiency.

[0114] The path expansion analysis is performed on the association path of the member consumption preferences to identify the associated knowledge units that have a relationship with the semantically related knowledge units. The relevance of the associated knowledge units is evaluated based on the semantic association strength in the association relationship of member consumption preferences, generating an extended knowledge unit set, including:

[0115] Extract all member consumption preference associations involving the semantically related knowledge units from the member consumption preference association path, and identify candidate associated knowledge units that form knowledge unit pairs with the semantically related knowledge units based on the knowledge unit pair identifiers recorded in the member consumption preference associations.

[0116] For the candidate associated knowledge unit, the semantic association strength between the semantically related knowledge unit and the candidate associated knowledge unit is obtained from the member consumption preference association relationship;

[0117] Based on the semantic association strength, the semantic matching degree is calculated by matching the member feature semantic tags of the candidate associated knowledge unit with the query semantic tags in the points marketing activity recommendation retrieval request.

[0118] Based on the semantic matching degree, the candidate related knowledge units are screened for relevance. The candidate related knowledge units whose comprehensive relevance scores meet the preset relevance conditions are determined as related knowledge units. The semantically related knowledge units are merged with the screened related knowledge units to generate an extended knowledge unit set.

[0119] Before performing path expansion analysis on the association paths of member consumption preferences, it is necessary to first clarify all member consumption preference associations in which semantically related knowledge units participate in the global semantic index network. Specifically, each association in the global semantic index network is stored in the form of knowledge unit pairs. Each knowledge unit pair records the identifiers of the two knowledge units and the semantic association strength value between them. By traversing all association records in the global semantic index network, all association entries containing the identifier of the target semantically related knowledge unit are extracted, thereby obtaining the complete set of associations in which that semantically related knowledge unit participates. The core of this step is to perform exact matching queries in the association storage structure using the knowledge unit identifier as the index key, avoiding traversal efficiency issues caused by the large number of associations.

[0120] The process involves retrieving the identifiers of knowledge units recorded in each relationship from the extracted set of relationships. The identifier of the other party, which differs from the target semantically related knowledge unit, is extracted; this identifier is the candidate related knowledge unit's identifier. Since candidate related knowledge units are distributed across different storage nodes in a distributed network, it is necessary to locate the storage node of each candidate related knowledge unit based on the node affiliation information stored in the global semantic index network. Then, the complete attribute information of the candidate related knowledge unit is retrieved from the corresponding node, including its member characteristic semantic tags, business-related attributes, and historical consumption behavior statistical features. This process can be executed in parallel with subsequent semantic matching calculation steps to reduce overall processing latency.

[0121] For each candidate related knowledge unit, the semantic association strength value between the candidate related knowledge unit and the target semantically related knowledge unit is obtained from the member consumption preference association record to which it belongs. The semantic association strength value reflects the degree of association between the two knowledge units in the member consumption preference dimension. The higher the value, the greater the overlap between the two in member behavior patterns and consumption preference features. This strength value has been normalized during the construction stage of the global semantic indexing network and can be directly used for subsequent weighted matching calculations without additional numerical transformation.

[0122] After obtaining the semantic association strength, a weighted matching calculation is performed on the member feature semantic tags of candidate related knowledge units and the query semantic tags in the points marketing activity recommendation retrieval request based on this strength value. The matching calculation process is divided into two levels: the first level is to directly compare the member feature semantic tags of candidate related knowledge units with the query semantic tags, calculate the closeness between the two in the semantic space, and obtain a basic semantic matching score; the second level is to introduce the semantic association strength as a moderating factor to adjust the basic semantic matching score, so that candidate related knowledge units that are more closely associated with the target semantic related knowledge units can obtain a higher comprehensive semantic matching degree in the matching calculation. This two-level matching mechanism can effectively distinguish those candidate knowledge units that have a certain textual similarity with the query semantic tags but have a weak association with the current retrieval context, thereby improving the accuracy of the expanded knowledge unit set.

[0123] Set up candidate related knowledge units The basic semantic matching score is The semantic association strength between it and the target semantically related knowledge units is Then the candidate related knowledge unit Comprehensive semantic matching degree By and The result is obtained by weighted fusion calculation, i.e. .in The similarity is obtained by calculating the cosine similarity between the member feature semantic tag vector of the candidate related knowledge unit and the query semantic tag vector. Read directly from the associated records. Consider the overall semantic matching degree. The value ranges from 0 to 1. A higher value indicates that the candidate related knowledge unit is more relevant in the current retrieval context.

[0124] After completing the comprehensive semantic matching degree calculation, all candidate related knowledge units are processed according to... The data are sorted in descending order of size, and candidate related knowledge units are filtered based on preset relevance conditions. These preset relevance conditions include two constraint dimensions: one is a comprehensive semantic matching degree threshold constraint, which requires... The first constraint is that the number of candidate related knowledge units must meet the minimum matching degree threshold set in advance. Candidate related knowledge units that are below the threshold are considered to be insufficiently relevant to the current retrieval request and are removed from the candidate set. The second constraint is that when the number of candidate related knowledge units that meet the matching degree threshold exceeds the preset size limit, the candidate related knowledge units with the highest comprehensive semantic matching degree are retained first to control the overall size of the expanded knowledge unit set and prevent the parallelism of the retrieval execution plan from getting out of control due to the excessively large target node set in the subsequent retrieval stage.

[0125] After the above two layers of screening, the finally determined related knowledge units are merged with the original semantically related knowledge units to form an extended knowledge unit set. During the merging process, duplicate knowledge units need to be removed; that is, if the identifier of a related knowledge unit is the same as the identifier of a semantically related knowledge unit already included in the set, it is not added again. The extended knowledge unit set is stored in an ordered structure, with semantically related knowledge units arranged first, followed by related knowledge units arranged in descending order of comprehensive semantic matching degree. This facilitates the allocation of retrieval resources according to priority during the subsequent retrieval execution plan generation stage.

[0126] After the extended knowledge unit set is generated, each knowledge unit carries the location information of its storage node and its relevance score in the context of the current retrieval request. This information serves as input for generating the retrieval execution plan, determining the scope of the target storage node set, and rationally allocating parallel retrieval tasks based on the load status of each node. The entire path expansion analysis and extended knowledge unit set generation process fully utilizes the semantic association strength information pre-calculated and stored in the global semantic indexing network, avoiding the computational overhead of real-time calculation of semantic similarity between large-scale knowledge units during the retrieval phase. This effectively controls retrieval response latency while ensuring the quality of the expanded results.

[0127] Based on the member consumption preference association path and the node load status of each target storage node, a retrieval execution plan is generated. Parallel retrieval requests are then initiated to the target storage node set according to the retrieval execution plan, generating target retrieval results containing the target member group and corresponding marketing strategies, including:

[0128] Obtain the node load status of each target storage node in the target storage node set, evaluate the load capacity of each target storage node based on the node load status, and generate a node load evaluation table.

[0129] Based on the member consumption preference association path, identify the target storage node combination that needs to be jointly retrieved across nodes, and obtain the load capacity evaluation value of each target storage node in the target storage node combination from the node load evaluation table;

[0130] A retrieval execution plan is generated based on the load capacity assessment value and the semantic association strength in the member consumption preference association path;

[0131] According to the retrieval execution plan, a parallel retrieval request is initiated to the target storage node set, and the member consumption behavior data and points usage behavior data obtained from the knowledge units of different target storage nodes are cross-node association and integration.

[0132] Based on the integrated member consumption behavior data and points usage behavior data, target member groups are identified, marketing activity information matching the target member groups is extracted to generate corresponding marketing strategies, and target search results containing the target member groups and corresponding marketing strategies are generated.

[0133] After determining the target storage node set, it is necessary to comprehensively collect the current operating status of each target storage node to support the generation of subsequent retrieval execution plans. For each target storage node in the target storage node set, its current CPU utilization, memory usage ratio, disk I / O throughput, network bandwidth utilization, and pending request queue length are collected, among other multi-dimensional status indicators. These indicators are converted into evaluation components with uniform dimensions according to a preset normalization mapping rule. Corresponding weight coefficients are applied to each component and summed to obtain the node load capacity evaluation value for each target storage node. The node number, node identifier, and corresponding load capacity evaluation values ​​of all target storage nodes are summarized and written into a node load evaluation table. The lower the node load capacity evaluation value, the more additional retrieval tasks the node can currently handle; the higher the node load capacity evaluation value, the more strained the node's resources are, and it should be avoided or have its allocation weight reduced during retrieval task allocation.

[0134] Based on the member consumption preference association paths stored in the global semantic index network, the target storage node combinations requiring cross-node joint retrieval are identified. These association paths describe the semantic relationships between different knowledge units. When the knowledge units involved in a certain association path are distributed across two or more target storage nodes, these nodes constitute a cross-node joint retrieval combination. For each identified cross-node joint retrieval combination, the load capacity assessment values ​​of each target storage node within the combination are read from the node load assessment table, serving as the basis for subsequent retrieval task allocation and execution order arrangement. The identification process for cross-node joint retrieval combinations requires traversing all association paths in the global semantic index network that semantically match the current retrieval request, and sorting the paths according to the strength of semantic association within the paths. Node combinations corresponding to paths with higher semantic association strength are processed first to ensure the semantic coverage quality of the retrieval results.

[0135] The retrieval execution plan is generated by comprehensively considering two dimensions: load capacity assessment value and semantic association strength in the member consumption preference association path. For node combinations with high semantic association strength and low load capacity assessment values, parallel retrieval tasks are prioritized and more concurrent retrieval thread resources are allocated. For node combinations with low semantic association strength or high load capacity assessment values, serial or delayed scheduling strategies are adopted to avoid response timeouts on high-load nodes due to excessive concurrent requests. The retrieval execution plan is organized in the form of a directed task graph, where each node corresponds to a retrieval subtask on a target storage node, and directed edges in the graph represent dependencies or data flow between subtasks. For retrieval subtasks without dependencies, parallel retrieval requests are allowed to be triggered simultaneously; for subtasks with data dependencies, they are triggered sequentially according to the topological order in the directed task graph.

[0136] Based on the generated retrieval execution plan, parallel retrieval requests are initiated to each target storage node in the target storage node set. Each parallel retrieval request carries the semantic query conditions, member group feature filtering conditions, and the list of knowledge unit fields to be returned for the current retrieval subtask. After receiving the retrieval request, each target storage node performs the retrieval operation according to the knowledge unit index structure stored locally, and encapsulates the member consumption behavior data and points usage behavior data of the knowledge units that meet the conditions into a retrieval response message and returns it. During the parallel retrieval process, the response latency of each target storage node is monitored in real time. If the response latency of a target storage node exceeds the preset timeout threshold, a timeout retry mechanism is triggered, and the retrieval subtask on that node is rerouted to a backup storage node where the same data replica is located for execution, so as to ensure the overall reliability of the retrieval process.

[0137] After the retrieval response messages from different target storage nodes arrive, the member consumption behavior data and points usage behavior data returned by each node are integrated across nodes according to the semantic relationships described by the member consumption preference association path. The core of cross-node integration lies in establishing a member identity association key between the data returned by different nodes. Using the member's unique identifier as the primary association key, consumption behavior records, points accumulation records, points redemption records, and points validity information from different nodes are concatenated into a complete member behavior sequence in timestamp order. For cases where there are data version conflicts for the same member on different nodes, a timestamp priority strategy is adopted, retaining the data version with the newer timestamp, and the conflict is recorded in the integration log for subsequent auditing. After integration, each member corresponds to a comprehensive record containing their complete consumption behavior trajectory and points usage behavior trajectory.

[0138] Based on integrated member consumption behavior data and points usage behavior data, and according to the target member group screening criteria specified in the points marketing activity recommendation retrieval request, the comprehensive records are subjected to multi-dimensional filtering and cluster analysis to identify target member groups that meet the criteria. Screening criteria can include multiple dimensions such as consumption frequency range, consumption amount range, points balance range, points usage rate range, and recent active time window. For the identified target member groups, features such as their preferred product category distribution, points redemption preferred product category distribution, and historical response rate to marketing activities are further extracted. These features are matched with marketing activity information stored in the knowledge unit to select marketing activities that highly match the target member group's consumption preferences and points usage habits as a recommendation candidate set. The marketing activities in the recommendation candidate set are sorted according to expected response rate and points consumption matching degree, and the top-ranked marketing activities are extracted to generate corresponding marketing strategies. Finally, the member list of the target member group, the group characteristic summary, and the detailed content of the corresponding marketing strategy are packaged into the target retrieval results and returned for the business layer to use for subsequent marketing activity deployment and effect tracking.

[0139] Figure 3 This is a schematic diagram illustrating the process of determining load and semantic strength according to an embodiment of the present invention. Figure 3 As shown, this invention sorts paths based on semantic association strength and generates a directed task graph by considering node load capacity. Node combinations with high association strength and low load (such as node-01 and node-03) are prioritized for parallel retrieval, while high-load nodes (such as node-02) are marked for delayed scheduling. This ensures semantic coverage quality while avoiding system overload.

[0140] Figure 4 This is a schematic diagram illustrating the process of determining the target retrieval results according to an embodiment of the present invention. Figure 4As shown, this demonstrates the final output of the search analysis. Based on the integrated member behavior data, the system filters and clusters data according to multiple dimensions such as purchase frequency, points balance, and active time to identify high-value target groups. Subsequently, by matching their consumption preferences with points redemption habits, the system selects the most relevant strategies (such as points doubling promotions) from the marketing campaign library and packages them into search results containing member lists and strategy details for direct use by the business side.

[0141] A second aspect of this invention provides an intelligent information retrieval and storage system based on a distributed network, comprising:

[0142] The parsing and extraction unit is used to obtain the original information object representing the member's consumption behavior, and to perform semantic parsing and structured extraction on the original information object to obtain the knowledge unit;

[0143] The semantic indexing unit is used to allocate the knowledge unit to multiple storage nodes in a distributed network based on the member feature semantic tags and business association attributes of the knowledge unit, and to generate a global semantic indexing network based on the member consumption preference association between knowledge units on different storage nodes.

[0144] The retrieval and positioning unit is used to receive a retrieval request for a points marketing activity recommendation targeting a specific member group, and to determine the target storage node set for storing relevant knowledge units based on the member group characteristics in the points marketing activity recommendation retrieval request and the member consumption preference association path in the global semantic index network.

[0145] The parallel retrieval unit is used to generate a retrieval execution plan based on the association path of the member's consumption preferences and the node load status of each target storage node, and to initiate a parallel retrieval request to the set of target storage nodes according to the retrieval execution plan, thereby generating target retrieval results containing the target member group and corresponding marketing strategies.

[0146] A third aspect of the present invention provides an electronic device, comprising:

[0147] processor;

[0148] Memory used to store processor-executable instructions;

[0149] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0150] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0151] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0152] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent information retrieval and storage based on distributed networks, characterized in that, include: Obtain the original information object representing the member's consumption behavior, and perform semantic parsing and structured extraction on the original information object to obtain knowledge units; Based on the member feature semantic tags and business association attributes of the knowledge units, the knowledge units are allocated to multiple storage nodes in a distributed network. A global semantic index network is generated based on the member consumption preference associations between the knowledge units on different storage nodes. Receive a request for a points-based marketing activity recommendation and retrieval for a specific member group, and determine the target storage node set for storing relevant knowledge units based on the member group characteristics in the points-based marketing activity recommendation and retrieval request and the member consumption preference association path in the global semantic index network. Based on the association path of member consumption preferences and the node load status of each target storage node, a retrieval execution plan is generated. Based on the retrieval execution plan, a parallel retrieval request is initiated to the set of target storage nodes to generate target retrieval results containing the target member group and corresponding marketing strategies.

2. The method according to claim 1, characterized in that, Obtain the original information object representing member consumption behavior, perform semantic parsing and structured extraction on the original information object to obtain knowledge units, including: Obtain raw information objects containing member consumption behavior records from the retail enterprise's transaction system. The raw information objects include basic member attribute information, product purchase record information, and points change record information. Cross-business domain semantic parsing is performed on the original information object to determine the temporal association pattern between the member's basic attribute information and the product purchase record information, as well as the causal dependency relationship between the product purchase record information and the points change record information. Based on the temporal association pattern and the causal dependency, determine the cross-domain semantic features that characterize the semantic association between member consumption preferences and points usage tendency; Based on the cross-domain semantic features, multi-level structured extraction is performed. The part of the member's basic attribute information that reflects the member's identity attributes is converted into member feature semantic tags. The part of the cross-domain semantic features that reflects the semantic relationship between the member's consumption preferences and points usage tendency is converted into business association attributes. A knowledge unit containing the member feature semantic tags and the business association attributes is generated.

3. The method according to claim 1, characterized in that, Based on the member characteristic semantic tags and business association attributes of the knowledge units, the knowledge units are allocated to multiple storage nodes in a distributed network. A global semantic index network is generated based on the member consumption preference associations between knowledge units on different storage nodes, including: A hash mapping function is constructed based on the member feature semantic tags of the knowledge unit. The location identifier of the knowledge unit on the distributed network logical ring is calculated through the hash mapping function. The candidate storage node corresponding to the knowledge unit is determined based on the location identifier. Based on the business association attributes of the knowledge unit, a load balancing assessment is performed on the candidate storage nodes. According to the load balancing assessment results, a target storage node is selected from the candidate storage nodes, and the knowledge unit is allocated to the target storage node for storage. Cross-node semantic association analysis is performed on knowledge units on multiple storage nodes in a distributed network to identify knowledge unit pairs that represent the same type of member consumption preferences, and to establish the member consumption preference association relationship between the knowledge unit pairs. Based on the association of member consumption preferences, a cross-node semantic association path is constructed. The knowledge unit identifiers in the semantic association path, the semantic association strength of the association of member consumption preferences, and the storage node identifiers where the knowledge units are located are structured and organized to generate the global semantic index network.

4. The method according to claim 3, characterized in that, Cross-node semantic association analysis is performed on knowledge units across multiple storage nodes in a distributed network to identify pairs of knowledge units representing the same type of member consumption preferences, and to establish member consumption preference associations between these knowledge unit pairs, including: Obtain the set of business-related attributes of the knowledge units stored on each storage node from multiple storage nodes in the distributed network; The set of business-related attributes is aggregated across nodes to generate a global attribute view that includes the business-related attributes of knowledge units across all storage nodes. Based on the global attribute view, member consumption preference types are clustered. By analyzing the semantic similarity features between the business association attributes of knowledge units on different storage nodes, knowledge units representing the same member consumption preference type are grouped into the same preference type cluster, and a preference type mapping table containing the mapping relationship between knowledge unit identifiers and preference type cluster identifiers is generated. Based on the preference type mapping table, knowledge unit pairs belonging to the same preference type cluster are identified. For each knowledge unit pair, the business association attributes of the knowledge unit pair are extracted for semantic association measurement, and the semantic association strength between the knowledge unit pairs is calculated. Establish the association relationship of member consumption preferences between the knowledge unit pairs based on the semantic association strength.

5. The method according to claim 1, characterized in that, The system receives a request for a points-based marketing campaign recommendation, and based on the member group characteristics in the request and the member consumption preference association path in the global semantic index network, determines a target set of storage nodes for storing relevant knowledge units, including: Receive the points marketing activity recommendation retrieval request, extract query semantic features representing the characteristics of the member group from the points marketing activity recommendation retrieval request, and convert the query semantic features into query semantic tags that match the member characteristic semantic tags of the knowledge unit; Based on the query semantic tags, semantic matching retrieval is performed in the global semantic index network to determine the semantically related knowledge units corresponding to the query semantic tags, and the member consumption preference association path associated with the semantically related knowledge units in the global semantic index network is obtained. A path expansion analysis is performed on the member consumption preference association path to identify the associated knowledge units that have a member consumption preference association relationship with the semantically related knowledge units. The relevance of the associated knowledge units is evaluated based on the semantic association strength in the member consumption preference association relationship, and an extended knowledge unit set is generated. Extract the storage node identifiers corresponding to the knowledge units from the extended knowledge unit set, perform deduplication and aggregation processing on the storage node identifiers, and generate a target storage node set containing the knowledge units and the associated knowledge units.

6. The method according to claim 5, characterized in that, The path expansion analysis is performed on the association path of the member consumption preferences to identify the associated knowledge units that have a relationship with the semantically related knowledge units. The relevance of the associated knowledge units is evaluated based on the semantic correlation strength in the association relationship of member consumption preferences, generating an extended knowledge unit set, including: Extract all member consumption preference associations involving the semantically related knowledge units from the member consumption preference association path, and identify candidate associated knowledge units that form knowledge unit pairs with the semantically related knowledge units based on the knowledge unit pair identifiers recorded in the member consumption preference associations. For the candidate associated knowledge unit, the semantic association strength between the semantically related knowledge unit and the candidate associated knowledge unit is obtained from the member consumption preference association relationship; Based on the semantic association strength, the semantic matching degree is calculated by matching the member feature semantic tags of the candidate associated knowledge unit with the query semantic tags in the points marketing activity recommendation retrieval request. Based on the semantic matching degree, the candidate related knowledge units are screened for relevance. The candidate related knowledge units whose comprehensive relevance scores meet the preset relevance conditions are determined as related knowledge units. The semantically related knowledge units are merged and organized with the screened related knowledge units to generate an extended knowledge unit set.

7. The method according to claim 1, characterized in that, Based on the member consumption preference association path and the node load status of each target storage node, a retrieval execution plan is generated. Parallel retrieval requests are then initiated to the target storage node set according to the retrieval execution plan, generating target retrieval results containing the target member group and corresponding marketing strategies, including: Obtain the node load status of each target storage node in the target storage node set, evaluate the load capacity of each target storage node based on the node load status, and generate a node load evaluation table. Based on the member consumption preference association path identification, the target storage node combination that needs to be jointly retrieved across nodes is identified, and the load capacity evaluation value of each target storage node in the target storage node combination is obtained from the node load evaluation table. A retrieval execution plan is generated based on the load capacity assessment value and the semantic association strength in the member consumption preference association path; According to the retrieval execution plan, a parallel retrieval request is initiated to the target storage node set, and the member consumption behavior data and points usage behavior data obtained from the knowledge units of different target storage nodes are cross-node association and integration. Based on the integrated member consumption behavior data and points usage behavior data, target member groups are identified, marketing activity information matching the target member groups is extracted to generate corresponding marketing strategies, and target search results containing the target member groups and corresponding marketing strategies are generated.

8. An intelligent information retrieval and storage system based on a distributed network, used to implement the method as described in any one of claims 1-7, characterized in that, include: The parsing and extraction unit is used to obtain the original information object representing the member's consumption behavior, and to perform semantic parsing and structured extraction on the original information object to obtain the knowledge unit; The semantic indexing unit is used to allocate the knowledge unit to multiple storage nodes in a distributed network based on the member feature semantic tags and business association attributes of the knowledge unit, and to generate a global semantic indexing network based on the member consumption preference association between knowledge units on different storage nodes. The retrieval and positioning unit is used to receive a retrieval request for a points marketing activity recommendation targeting a specific member group, and to determine the target storage node set for storing relevant knowledge units based on the member group characteristics in the points marketing activity recommendation retrieval request and the member consumption preference association path in the global semantic index network. The parallel retrieval unit is used to generate a retrieval execution plan based on the association path of the member's consumption preferences and the node load status of each target storage node, and to initiate a parallel retrieval request to the set of target storage nodes according to the retrieval execution plan, thereby generating target retrieval results containing the target member group and corresponding marketing strategies.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images