A knowledge graph-based personalized recommendation method and system

Abstract

The application provides a kind of personalized recommendation method and system based on knowledge graph, it is related to knowledge graph technical field.The application first collects the basic information of product source, the delivery cycle of product, financial constraint, policy constraint and user behavior and other multi-source heterogeneous data, and writes it into shopping knowledge graph by intelligent cleaning and entity-relation extraction;Subsequently, combined with user budget, preference and purchase qualification, user portrait is constructed, and the candidate set of compliant product source is reasoned and screened in the knowledge graph with the portrait as the retrieval condition;Then, according to the preference matching degree, price adaptation degree, complete degree and appreciation potential of each product source in the candidate set, a comprehensive score is generated and sorted;Finally, output the highest comprehensive score of several product sources and the corresponding three-section recommendation explanation, realize the depth fusion of multi-source data, ensure that the recommendation result is accurate and compliant, and has interpretability.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology in general artificial intelligence technology, and in particular to a personalized recommendation method and system based on knowledge graph. Background Technology

[0002] With the opening up of various product information platforms, big data government interfaces, and fintech, shoppers can access massive amounts of product sources, interest rates, and policy information online. Knowledge graphs, due to their excellent semantic representation, are widely used in recommendation systems for products, tourism, and film / television. Collaborative filtering remains the mainstream interest modeling method, and the integration of these two approaches is becoming an important trend for improving recommendation accuracy.

[0003] Currently, the industry is transitioning from "interest-driven" to a "constraint-interest dual-driven" model: on the one hand, it needs to capture shoppers' aesthetic preferences and usage needs; on the other hand, it must comply with relevant financial policies, government consumer voucher issuance policies, purchase restrictions, regional product launch cycles, and other rigid constraints. With the increasing abundance of multimodal data (images, text, geographic information), graph neural networks-large language models collaborative reasoning and "explainable recommendations" are gradually becoming research hotspots.

[0004] Existing solutions often follow the consumer product recommendation framework, ignoring the value decision-making characteristics of shopping scenarios, and mainly suffer from the following technical problems: ① lack of cross-domain knowledge integration covering policy, finance and market supply and demand data; ② insufficient cold start mitigation for new product sources and new users; ③ weak ability to decouple and personalize the trade-offs between multiple objectives (price-quality-appreciation); ④ difficulty in responding to policy changes in real time and outputting compliance explanations. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a personalized recommendation method and system based on knowledge graphs. By integrating multi-source data to construct a knowledge graph, and combining user profiles with multi-dimensional constraints, it achieves accurate, compliant, and interpretable personalized shopping recommendations.

[0006] Based on a first key aspect of the present invention, a knowledge graph-based personalized recommendation method is provided, comprising the following steps performed by a computer system:

[0007] Collect heterogeneous data from multiple sources, including basic information about the product source, the product launch cycle, financial constraints, policy constraints, and user behavior.

[0008] The multi-source heterogeneous data is cleaned and entity-relation extraction is performed to obtain extracted information, which is then written into the shopping knowledge graph.

[0009] Based on user budget, preferences, and purchase eligibility information, construct user demand profiles;

[0010] Using the user demand profile as the retrieval criteria, a candidate set of product sources that meet policy and financial constraints is obtained by reasoning in the shopping knowledge graph.

[0011] The attribute information of each product source in the product source candidate set is compared with the user demand profile to calculate the matching degree score of each product source relative to the user demand, and a comprehensive score is generated by combining the target factors of price level, supporting services, and appreciation potential.

[0012] The candidate product sources are sorted according to the comprehensive score, and the top-scoring product sources and their corresponding recommendation explanations are output.

[0013] Preferably, the multi-source heterogeneous data includes basic information data of product sources, market supply and demand data, financial constraint data, policy constraint data, and user behavior data; the product source information data includes supply volume, grade, type, total price, unit price, production date, and geographical location (place of origin); the market supply and demand data includes total market supply, product shelf life, total amount of replaceable products, and total market demand; the financial constraint data includes interest rates, credit payment periods, installment payment ratios, and user credit scores; the policy constraint data includes purchase restrictions, loan restrictions, purchase eligibility requirements, and tax standards; and the user behavior data includes user browsing history, collection history, consultation history, trial history, and proactive feedback content.

[0014] As a further preferred embodiment, in the aforementioned method, the multi-source heterogeneous data is cleaned and entity-relation extraction is performed to obtain extracted information, and the extracted information is written into a shopping knowledge graph, including:

[0015] The multi-source heterogeneous data is subjected to format standardization, missing value completion, and outlier detection to obtain a cleaned structured dataset;

[0016] Based on natural language processing technology, named entity recognition is performed on text information in structured datasets to extract core entities such as product source, location, supporting services, policy terms, financial products and user behavior.

[0017] Combining a pre-defined domain ontology dictionary and semantic rules, attribute relationships and associations are mined from the core entities to generate extracted information in units of "entity-relationship-entity" triples;

[0018] The extracted information is written into a graph database to construct the node set and edge set of the shopping knowledge graph.

[0019] Preferably, a user needs profile is constructed based on user budget, preferences, and purchase eligibility information, including:

[0020] The system obtains basic information provided by the user, estimates the user's shopping budget based on the basic information, and extracts demographic features; the basic information includes age, marital status, occupation, annual income, family structure, and city / region of residence.

[0021] The system analyzes users' historical behavior data and uses a time-weighted modeling method to generate a distribution of users' shopping behavior preferences; the historical behavior data includes browsing history, collection history, consultation history, and trial history.

[0022] The demographic characteristics, shopping budget capacity, and shopping behavior preference distribution are fused to obtain the preliminary user profile features.

[0023] Based on the aforementioned basic information and current policy constraint data, the preset purchase eligibility verification module is invoked to evaluate the user's shopping area restrictions, loan limit, and eligibility status, and the evaluation results are used as profile constraint attributes.

[0024] A user demand profile vector is constructed by combining the preliminary profile features with the constraint attributes.

[0025] Preferably, the expression for the shopping behavior preference distribution is:

[0026]

[0027] in, For users The vector of the shopping behavior preference distribution; The total number of historical behavior records accumulated for the user; For the first The attribute vectors of the product sources associated with each behavior record are generated after being uniformly represented by a semantic embedding model; For the first The interaction depth factor corresponding to each behavior record is naturally obtained based on the behavior category mapping: browsing is 1, collecting is 2, consulting is 3, and trying is 4. This is the current timestamp. For the first The timestamp of the occurrence of the action; The effective response weight function represents the behavior; The normalization factor is defined as the sum of the weights of all behavioral responses, i.e. , used to Normalized into a probabilistic preference vector.

[0028] Preferably, based on the basic information and current policy constraint data, a preset purchase eligibility verification module is invoked to evaluate the user's shopping area restrictions, loan limit, and eligibility status, and the evaluation result is used as a profile constraint attribute, including:

[0029] Obtain the user's basic identity information; the basic identity information includes household registration type, marital status, number of family members, tax or social security records, and fixed asset holdings.

[0030] Based on the market regulation policies corresponding to the current region, retrieve the purchase restriction rules, loan restriction regulations, and purchase restriction identification standards for the current region;

[0031] The basic identity information is matched with the market regulation policy to determine whether the user is eligible to purchase in each target area, and the corresponding list of area restrictions is output.

[0032] The system calls a pre-defined purchase eligibility verification module to automatically calculate the maximum loan amount a user can obtain based on their income level, debt situation, credit history, and current interest rate environment.

[0033] Based on the regional restriction list and whether the maximum loan amount meets the eligibility requirements, a purchase eligibility status identifier is generated in the preliminary profile features, and the purchase eligibility status identifier is used as the profile constraint attribute.

[0034] Preferably, using the user demand profile as the retrieval criteria, a candidate set of product sources that meet policy and financial constraints is obtained by reasoning within the shopping knowledge graph, including:

[0035] Using the budget range, applicable scenarios, special needs, and type preferences in the user demand profile as the starting point for retrieval, semantic subgraph construction is performed in the shopping knowledge graph to filter a preliminary set of product sources that meet the conditions of spatial location, product quality, and price.

[0036] Based on the purchase eligibility status and financial constraints in the user demand profile, product source nodes that the user does not have the purchase eligibility or payment ability are excluded from the preliminary product source set to form a compliant product source set.

[0037] Based on the multi-hop path relationship between user entities and product source entities in the shopping knowledge graph, the graph association confidence of each compliant product source entity in the compliant product source set is calculated; the graph association confidence reflects the degree of semantic fit between the product source and user needs.

[0038] Product source entities with a graph association confidence level not lower than a preset threshold in the compliant product source set are selected as the product source candidate set.

[0039] Preferably, the expression for the confidence level of the map association is:

[0040]

[0041] in, Calculate the confidence level of the graph association between user entities and product source entities; The number of different semantic path types connecting the user entity and the product source entity is automatically obtained by traversing and counting the shopping knowledge graph. The shortest path length between the user entity and the product source entity in the shopping knowledge graph reflects semantic distance and is calculated using a standard graph search algorithm. Graph embedding vectors for user entities. For the graph embedding vector of the product source entity, and Generated from the same unsupervised embedding model; is the Euclidean distance between vectors, used to measure differences in semantic similarity.

[0042] Preferably, the attribute information of each product source in the candidate product source set is compared with the user demand profile to calculate the matching degree score of each product source relative to the user demand. A comprehensive score is then generated by combining target factors such as price level, supporting services, and appreciation potential, including:

[0043] For each product source candidate set, construct a product source attribute vector by assigning attributes such as supply quantity, type, grade, geographical location (origin), applicable scenarios, and transportation accessibility.

[0044] The cosine similarity between the product source attribute vector and the user demand profile vector is calculated to obtain the product source preference matching score; the formula for calculating the product source preference matching score is as follows: ;in, The product source preference matching score is represented by x; the product source attribute vector is represented by y; and the user demand profile vector is represented by y. Represents the dot product of vectors; The second norm of a vector;

[0045] Based on the user's budget range and the listed price of the product, a price fit score is calculated. This price fit score is obtained by normalizing the difference between the product's listed price and the upper and lower limits of the budget using a normalization function. The formula for calculating the price fit score is as follows: ;in, The price fit score; The listed price for the product source; and These are the user's lower and upper budget limits, respectively.

[0046] The coverage ratio of supporting services for statistical product sources is calculated, and the completeness score of supporting services is determined. The formula for calculating the completeness score of supporting services is as follows: ;in, The completeness score of the matching system; These represent the coverage ratios of the four designated service items within a preset service radius, with values ​​ranging from 0 to 1.

[0047] An automated valuation model, trained based on historical price sequences and urban product supply cycle variables, is invoked to predict the price growth rate of the product source within a specified future period. This prediction is then compared with the benchmark growth rate in the same city to obtain an appreciation potential score. The formula for calculating the appreciation potential score is as follows: ;in, The appreciation potential score; Forecast the growth rate of the product source over a specified future period; This is the benchmark growth rate for the same city during the same period; adding 1 to the denominator is to prevent... Divide-by-zero risk at that time;

[0048] The product source preference matching score, price suitability score, accessory completeness score, and appreciation potential score are each normalized to 0-1, and then the geometric mean of the normalized results is calculated to obtain the comprehensive score; the formula for calculating the comprehensive score is as follows: ;in, The overall score is as described above; These are the normalized results of the product source preference matching score, the price suitability score, the completeness of supporting facilities score, and the appreciation potential score, respectively.

[0049] Based on a second key aspect of the present invention, a knowledge graph-based personalized shopping recommendation system is provided, comprising:

[0050] The multi-source data acquisition unit is used to collect multi-source heterogeneous data such as basic information of product sources, product launch cycle, financial constraints, policy constraints, and user behavior.

[0051] The graph construction and update unit is used to perform cleaning and entity-relation extraction on the multi-source heterogeneous data to obtain extracted information, and write the extracted information into the shopping knowledge graph;

[0052] The user profile generation unit is used to build user demand profiles based on user budget, preferences and purchase eligibility information;

[0053] The candidate product source reasoning and filtering unit is used to reason in the shopping knowledge graph to obtain a set of candidate product sources that meet policy and financial constraints, based on the user demand profile as the retrieval condition.

[0054] The matching degree calculation and scoring generation unit is used to compare the attribute information of each product source in the product source candidate set with the user demand profile, calculate the matching degree score of each product source relative to the user demand, and generate a comprehensive score by combining the target factors of price level, supporting services, and appreciation potential.

[0055] The sorting and recommendation explanation output unit is used to sort the product source candidate set according to the comprehensive score and output the top-scoring product sources and their corresponding recommendation explanation information.

[0056] According to a third key aspect of the present invention, an electronic device is provided, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus;

[0057] The memory stores a computer program that, when executed by the processor, causes the processor to perform the aforementioned knowledge graph-based personalized recommendation method.

[0058] Based on a fourth key aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon that, when executed, implements the knowledge graph-based personalized recommendation method as described above.

[0059] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0060] This invention collects heterogeneous data from multiple sources, including basic information about product sources, product launch cycles, financial constraints, policy constraints, and user behavior, and constructs a structured shopping knowledge graph to achieve deep semantic connections between user needs and product source information. Based on this, it combines user budget, preferences, and purchase eligibility information to build user demand profiles, and uses graph reasoning to filter candidate sets of product sources that meet policy and financial constraints, ensuring the feasibility and compliance of the recommendation results.

[0061] Meanwhile, by calculating the matching degree between product sources and user profiles, and integrating target factors such as price, accessories, and appreciation potential to form a comprehensive score, the system achieves a multi-dimensional value ranking, and finally outputs the preferred product sources and recommendation explanations, which improves the accuracy, practicality, and user trust of recommendations, significantly outperforming traditional recommendation methods based solely on behavior or tags. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, obtaining other drawings based on these drawings without creative effort still falls within the scope of the present invention.

[0063] Figure 1 The following is an execution flowchart of a knowledge graph-based personalized recommendation method provided as an embodiment of the present invention;

[0064] Figure 2 This is a schematic diagram of a system architecture for implementing a knowledge graph-based personalized recommendation method, provided as an embodiment of the present invention. Detailed Implementation

[0065] The preferred embodiments of the present invention will be described in detail below to provide a clearer understanding of the purpose, features, and advantages of the invention. It should be understood that the following embodiments are not intended to limit the scope of the invention, but are merely illustrative of the essential spirit of the technical solution of the invention.

[0066] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known techniques associated with the invention may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0067] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0068] The purpose of this invention is to provide a personalized recommendation method and system based on knowledge graphs. By integrating multi-source data to construct a knowledge graph, and combining user profiles with multi-dimensional constraints, it achieves accurate, compliant, and interpretable personalized shopping recommendations.

[0069] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0070] Figure 1 The method flowchart provided in the embodiments of the present invention is as follows: Figure 1 As shown, this invention provides a personalized shopping recommendation method based on knowledge graphs, including the following steps 100-600 executed by a computer system:

[0071] Step 100: Collect multi-source heterogeneous data, including basic information about the product source, product launch cycle, financial constraints, policy constraints, and user behavior.

[0072] Step 200: Perform cleaning and entity-relation extraction on the multi-source heterogeneous data to obtain extracted information, and write the extracted information into the shopping knowledge graph;

[0073] Step 300: Based on user budget, preferences, and purchase eligibility information, construct a user needs profile;

[0074] Step 400: Using the user demand profile as the retrieval criteria, reason in the shopping knowledge graph to obtain a candidate set of product sources that meet policy and financial constraints;

[0075] Step 500: Compare the attribute information of each product source in the product source candidate set with the user demand profile, calculate the matching degree score of each product source relative to the user demand, and generate a comprehensive score by combining the target factors of price level, supporting services, and appreciation potential.

[0076] Step 600: Sort the candidate product sources according to the comprehensive score, and output the top-scoring product sources and their corresponding recommendation explanations.

[0077] In this invention, the multi-source heterogeneous data collection in step 100 is not limited to simply pulling data from multiple channels, but rather uses a "heterogeneous data labeling and semantic mapping" mechanism to structure, align fields, and fuse semantics of various types of raw data, ensuring that the precise entity attributes and relational edges required for subsequent graph construction can be accurately obtained.

[0078] The solution of the present invention can be widely applied to various shopping scenarios such as buying a house, buying a car, buying home appliances, and purchasing electromechanical equipment. The following embodiment uses buying a house as an example to illustrate the specific implementation process of the present invention.

[0079] Example 1: Implementation of the method of the present invention in personalized home purchase recommendations

[0080] Specifically, for product information data, a unified geocoding is first used as the primary key to standardize fields such as area, floor, unit type, total price, unit price, and year of construction to the same coordinate system and unit of measurement. Unmapped supplementary attributes are retained in the original interface return, and missing geographical location dimensions are automatically filled in through a mapping dictionary. For urban planning data, GIS semantic services are called to overlay subway planning lines, school district boundaries, and commercial land use patches onto the property coordinates. Geographic buffer analysis is used to automatically calculate the spatial proximity index with the property. This process is the key technology for constructing the "space-function" composite edge in this invention and cannot be omitted.

[0081] In the data processing stage related to financial and policy constraints, this invention employs "timeliness context synchronization verification" technology. By accessing the real-time interest rate and loan product catalog interfaces of financial institutions and aligning the date version number with the property entry time, it ensures that the collected mortgage interest rate, loan term, and down payment ratio always match the user's query period. Simultaneously, regarding purchase restrictions, loan restrictions, and tax policies, this invention constructs a "regional policy matrix," which bidirectionally maps the user's intended region code to a policy clause index table. When policies are updated, only the index table needs to be updated, without changing the mapping logic, significantly improving system maintenance efficiency and data consistency.

[0082] The collection and preprocessing of user behavior data adopts the "multi-source log fusion and temporal sparse filling" method: the system's self-developed behavior record aggregation engine can subscribe in real time to the event streams such as house viewing, consultation, and collection returned by the front-end tracking points and third-party brokerage platforms, and encrypt and map the user identifiers of different channels to a unified ID system; for sparse or discontinuous behavior sequences, this invention designs a temporal interpolation algorithm based on a sliding window, which intelligently fills only short-term gaps without relying on manual rules, thereby preserving the true temporal characteristics of user behavior and providing unbiased original input for subsequent preference modeling.

[0083] In step 200 of this invention, multi-source heterogeneous data cleaning is centered on "adaptive attribute distribution reconstruction": After receiving the original housing source interface and log stream, the system first constructs "attribute quantile distribution maps" for numerical fields (such as total price, unit price, and year of construction) and enumeration fields (such as apartment type and decoration status). The system continuously calculates the latest 1%–99% quantiles using an online-updating sliding window statistics module to detect outliers and generate "field distribution thresholds". When a new record exceeds the threshold, the system calls the asynchronous callback interface to trigger verification on the supply side or user side, instead of directly discarding or hard replacing it.

[0084] For missing value completion, this invention employs Conditional Random Field (CRF) combined with contextual features: using the geographical code of the property location, the average price of adjacent properties, the construction year segmentation, and the combination of unit types within the same community as node features, a multi-source co-occurrence map is constructed. By training a time-annotated CRF model, missing values ​​are inferred simultaneously along the spatial-attribute axis. This mechanism ensures that the completion results conform to the true data distribution at both the micro (single unit) and macro (same area) scales, without the need for manually setting interpolation rules.

[0085] To address the diversity of terminology in the fixed asset field, this invention innovatively employs a "two-stage distillation and dynamic dictionary expansion" strategy: In the first stage, based on a large-scale general pre-trained language model, a real estate professional lexicon (including building names, subway lines, and policy terms) is injected into a lightweight Transformer encoder using "knowledge distillation" technology to obtain a basic NER model. In the second stage, a "real-time feedback fine-tuning" process is run on this model: the system monitors tens of thousands of new entities in user inquiries and external policy announcements after the system goes live, automatically constructing incremental labeled samples. After several rounds of online fine-tuning, the NER model adaptively expands to include entity categories. The entity disambiguation module combines "character-level edit distance" and "geographical context overlap" algorithms to map multiple aliases of the same entity (such as "Line 10 Metro" and "Metro Line 10") to the same ID, achieving low redundancy under high recall.

[0086] In the relation extraction and graph writing stages, this invention is based on the "multi-level ontology mapping and priority path ranking" technology: First, predefined "urban planning ontology," "financial policy ontology," and "user behavior ontology" are loaded, and entity pairs are automatically identified by referring to the dependency syntax tree. For each entity pair, the system attempts to map all possible relations in the ontology in parallel, and uses a semantic slot weight solver to calculate the confidence score of candidate relations. This solver considers syntactic dependency distance, co-occurrence frequency, and ontology priority together, eliminating the need for manual scoring. Subsequently, the "multi-hop minimum semantic path" algorithm calculates the optimal connection path of entity pairs in the entity-relation network, and generates triples with the highest confidence relation. Batch triples are stored in the database through the "versioned writing" module, supporting historical snapshot backtracking and incremental updates, ensuring that the knowledge graph is both scalable and can accurately express deep semantic relationships in multi-source heterogeneous data.

[0087] Preferably, the expression for the distribution of home-buying behavior preferences is:

[0088]

[0089] in, For users The vector of the distribution of home-buying behavior preferences; The total number of historical behavior records accumulated for the user; For the first The attribute vectors of the properties associated with each behavioral record are generated after being uniformly represented by a semantic embedding model; For the first The interaction depth factor corresponding to each behavior record is naturally obtained based on the behavior category mapping: browsing is 1, collecting is 2, consulting is 3, and viewing the property on-site is 4. This is the current timestamp. For the first The timestamp of the occurrence of the action; The effective response weight function represents the behavior; The normalization factor is defined as the sum of the weights of all behavioral responses, i.e. , used to Normalized into a probabilistic preference vector.

[0090] In this invention, the generation of behavioral preference distribution is achieved through a "semantic embedding and adaptive time decay fusion" mechanism: The system first automatically converts each user behavior (such as browsing, saving, consulting, or viewing a property) into a multi-dimensional embedded representation of the corresponding property attribute. Then, it combines the inherent depth level of the behavior type (browsing is the lowest, saving is the next highest, consulting is the highest, and viewing is the highest) and the time interval between the behavior and the current time point to calculate the timeliness weight of each record. The weight integrates the intensity and freshness of the behavior, making recent deep interactions have a greater impact on user preferences, while ensuring that long-term preferences can also be gradually accumulated. Subsequently, the system sums all weighted embedded representations according to their weights and normalizes the resulting vector using the sum of weights, ultimately outputting a probabilistic preference distribution vector that reflects both the user's deep interaction tendency and dynamically responds to the latest behavior. This process strictly avoids any manual weight setting or empirical parameter tuning, ensuring that preference modeling is both interpretable and capable of end-to-end automated updates.

[0091] Specifically, this embodiment uses a "unified identity attribute mapping" mechanism to map a user's household registration type, marital status, number of family members, social security and tax records, and existing property ownership into a structured attribute set. Using the administrative code of the user's intended home purchase area as the key, it retrieves the corresponding area's purchase restrictions, loan restrictions, and first / second home purchase criteria from government data interfaces or local policy databases in real time. This embodiment emphasizes the "semantic rule-based" approach in this process—using a custom domain-specific language (DSL) to convert natural language policy clauses into executable rules, eliminating the need for manual parsing. Furthermore, when policies are updated, only the rule mapping needs to be updated, without affecting the core verification logic.

[0092] This embodiment automatically assesses user eligibility through an "attribute-rule mapping matrix": it compares the aforementioned structured identity attributes with multiple policy rules item by item to generate eligibility judgment results segmented by region (e.g., a first-time homebuyer in a certain region is limited to one property, and a second property requires meeting incremental social security contribution standards, etc.). This matrix-based assessment process ensures that complex and ever-changing regulatory policies can be implemented dynamically and item by item, without arbitrary hard coding, and generates a traceable list of regional restrictions upon output, clearly identifying eligible and temporarily unavailable purchase areas.

[0093] This embodiment utilizes the "Credit Boundary Calculator" module to automatically calculate the maximum loan amount based on the user's income level, debt ratio, and credit score, combined with the current floating mortgage rate and loan product terms. This embodiment integrates real-time interest rate query and product term parsing functions, enabling loan amount constraint estimation within milliseconds. It embeds the regional restriction list and the maximum loan amount result into the user's initial profile, generating three constraint attributes: "Purchasable Area," "Maximum Loan Capacity," and "Eligibility Status," providing strict compliance boundaries for subsequent graph reasoning and property selection.

[0094] In step 400, this embodiment introduces a "spatial-semantic subgraph construction" mechanism in the primary stage of graph reasoning: the system maps budget range, commuting location, school district requirements, and housing type preferences from the user's demand profile into multi-dimensional search tags, and calls spatial index and attribute index services in parallel in the housing knowledge graph to constrain the matching relationship between housing entities and geographical coordinates, school district tags, and price ranges; subsequently, based on the entity set fused with the above search results, a graph query language (such as Cypher) and a customized subgraph pattern template are used to generate a subgraph containing all housing entities that meet the basic spatial, structural, and price conditions in one go, without having to perform multiple simple filters separately, thereby significantly reducing the number of graph traversals and ensuring the integrity of the subgraph. Next, this embodiment uses "compliance secondary screening" to directly inject the regional eligibility status and maximum loan amount generated in the user profile into the subgraph node attributes, and batch-removes housing entities that do not meet any compliance attributes, forming a set of compliant entities containing only housing that the user can purchase and pay for.

[0095] Based on the set of compliant entities, this embodiment innovatively employs a "multi-hop path diversity assessment" method to calculate the confidence level of graph association: For each compliant property entity, this embodiment starts with the user entity and uses a restricted breadth-first search algorithm to enumerate up to three-hop semantic paths. Each path is automatically scored according to "reciprocal of hop count × relationship type entropy," with the entropy value automatically updated by the system's online statistics of the probability of occurrence of each relationship type in the entire graph, without the need for manual weight setting. The scores of all paths for the same property are summed and divided by the number of paths to achieve standardization, yielding the semantic fit value between the property and the user. Finally, this embodiment outputs all property entities with semantic fit values ​​not lower than a preset confidence threshold as the final candidate set, ensuring that both path diversity and semantic distance and relationship structure depth are considered.

[0096] Preferably, the expression for the confidence level of the map association is:

[0097]

[0098] in, Calculate the confidence level of the graph association between user entities and property entities; The number of different semantic path types connecting the user entity and the property entity is automatically obtained by traversing and counting the housing knowledge graph. The shortest path length between the user entity and the property entity in the home purchase knowledge graph reflects the semantic distance and is calculated using a standard graph search algorithm. Graph embedding vectors for user entities. The graph embedding vector for the property entity. and Generated from the same unsupervised embedding model; is the Euclidean distance between vectors, used to measure differences in semantic similarity.

[0099] In this embodiment, when calculating the confidence of the graph association, the knowledge graph of home purchase is first traversed in parallel to automatically count the number of different semantic paths and the shortest path length between user entities and housing entities. Then, the entity embedding vectors generated by the same unsupervised embedding model are used to calculate the Euclidean distance between the two entities. Finally, the path diversity, semantic distance and vector similarity are seamlessly integrated into a unified confidence value to ensure that both the complexity of the relational structure and the deep semantic fit are taken into account.

[0100] Preferably, the attribute information of each property in the candidate property set is compared with the user demand profile to calculate the matching degree score of each property relative to the user demand. A comprehensive score is then generated by combining target factors such as property price level, supporting facilities, and appreciation potential, including:

[0101] For each property in the candidate property set, extract the attributes of area, unit type, floor, geographical location, school district coverage, and transportation accessibility to construct a property attribute vector;

[0102] The cosine similarity between the property attribute vector and the user demand profile vector is calculated to obtain the property preference matching score; the formula for calculating the property preference matching score is as follows: ;in, y is the property preference matching score; x is the property attribute vector; y is the user demand profile vector; Represents the dot product of vectors; The second norm of a vector;

[0103] Based on the user's budget range and the property listing price, a price fit score is calculated. This price fit score is obtained by normalizing the difference between the property price and the upper and lower limits of the budget using a normalization function. The formula for calculating the price fit score is as follows: ;in, The price fit score; The listed price of the property; and These are the user's lower and upper budget limits, respectively.

[0104] The coverage ratio of surrounding educational, medical, commercial, and rail transit facilities for properties is statistically analyzed, and a completeness score for these facilities is calculated. The formula for calculating the completeness score is as follows: ;in, The completeness score of the matching system; These represent the coverage ratios of education, healthcare, commerce, and rail transit facilities within a preset service radius, with values ​​ranging from 0 to 1.

[0105] An automated valuation model trained based on historical housing price sequences and urban planning variables is invoked to predict the price growth rate of the property within a specified future period, and the result is compared with the benchmark growth rate in the same city to obtain an appreciation potential score; the formula for calculating the appreciation potential score is as follows: ;in, The appreciation potential score; Predict the growth rate of properties over a specified future period; This is the benchmark growth rate for the same city during the same period; adding 1 to the denominator is to prevent... Divide-by-zero risk at that time;

[0106] The scores for property preference matching, price suitability, amenities completeness, and appreciation potential are each normalized to 0-1. The geometric mean of the normalized results is then calculated to obtain the comprehensive score. The formula for calculating the comprehensive score is as follows: ;in, The overall score is as described above; These are the normalized results of the property preference matching score, the price suitability score, the completeness of supporting facilities score, and the appreciation potential score, respectively.

[0107] In constructing the property attribute vector, this embodiment automatically acquires and quantifies the property's area, unit type, floor, geographical coordinates, school district coverage, and real-time traffic accessibility by uniformly calling spatial index and multi-dimensional attribute storage service, ensuring that the attribute vector can accurately reflect the spatial and functional characteristics of the property.

[0108] In the calculation of preference matching degree and sub-item scores, this embodiment maps cosine similarity, price deviation, supporting facility coverage ratio and appreciation potential score to 0–1 through built-in normalization service, and automatically completes geometric mean calculation using a dedicated "multi-dimensional quality fusion engine", avoiding external parameter tuning and realizing end-to-end automation of the whole process.

[0109] Ultimately, the comprehensive score output by this embodiment reflects the deep semantic alignment between user demand profiles and property attributes, while also taking into account financial feasibility, completeness of supporting facilities, and investment value, providing a comprehensive and balanced quantitative basis for subsequent ranking and recommendation explanation modules.

[0110] In this embodiment, step 600 completes the final recommendation output through a "dynamic threshold sorting and interpretable result packaging" mechanism: The system first stores the comprehensive scores of all candidate properties in descending order into the memory cache, and adaptively adjusts the extraction threshold based on user historical clicks and feedback data to dynamically determine the number of outputs instead of a fixed N; after sorting, the "path-driven explanation generator" is called to automatically extract key related nodes and relationships along the optimal semantic subgraph path between each recommended property and user entity, and constructs a multi-dimensional explanation template by combining property attribute vectors and user profile features, and then outputs a three-part recommendation description of "reasons for eligibility + highlights of preference matching + overview of investment value" through a lightweight text synthesis module; finally, in this embodiment, the sorting results with explanation information are written back to the front-end interface according to priority index to ensure a one-to-one correspondence between the sorting logic and the explanation content and end-to-end real-time delivery.

[0111] Corresponding to the above methods, such as Figure 2 As shown, this embodiment also provides a knowledge graph-based personalized recommendation system for home purchases, including:

[0112] The multi-source data acquisition unit is used to collect multi-source heterogeneous data on housing information, planning, financial constraints, policy constraints, and user behavior.

[0113] The graph construction and update unit is used to perform cleaning and entity-relation extraction on the multi-source heterogeneous data to obtain extracted information, and write the extracted information into the home purchase knowledge graph;

[0114] The user profile generation unit is used to build user demand profiles based on user budget, preferences, and home purchase eligibility information.

[0115] The candidate housing listing reasoning and filtering unit is used to reason in the housing purchase knowledge graph to obtain a set of candidate housing listings that meet policy and financial constraints, based on the user demand profile as the retrieval condition.

[0116] The matching degree calculation and score generation unit is used to compare the attribute information of each property in the property candidate set with the user demand profile, calculate the matching degree score of each property relative to the user demand, and generate a comprehensive score by combining the target factors of housing price level, supporting facilities, and appreciation potential.

[0117] The sorting and recommendation explanation output unit is used to sort the candidate housing units according to the comprehensive score and output the highest-scoring housing units and their corresponding recommendation explanation information.

[0118] The beneficial effects of this embodiment are as follows:

[0119] (1) This embodiment achieves semantic-level accurate matching of user needs and housing information through deep integration of multi-source heterogeneous data and (housing) knowledge graph, which significantly improves the accuracy of recommendations.

[0120] (2) This embodiment introduces a real-time policy and financial constraint verification mechanism, which can automatically remove non-compliant or non-payable housing listings to ensure the legality and feasibility of the recommendation results.

[0121] (3) This embodiment adopts a comprehensive scoring method that integrates multiple objectives such as preference matching, price suitability, completeness of supporting facilities and appreciation potential, effectively taking into account user preferences, economic affordability and long-term value.

[0122] (4) The recommendation explanations generated by the graph path and the large language model in this embodiment make the recommendation decision transparent and traceable, thereby enhancing user trust and system acceptability.

[0123] The technical terms, principles, or means related to the technical solutions of the present invention mentioned in the above embodiments, which are not described in detail above, are all well-known technologies or common practices that are known to those skilled in the art.

[0124] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0125] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A personalized recommendation method based on knowledge graphs, characterized in that, This includes the following steps performed by the computer system: Collect multi-source heterogeneous data from product sources; the multi-source heterogeneous data includes basic information data of product sources, market supply and demand data, policy constraint data, financial constraint data, and user behavior data; The multi-source heterogeneous data is cleaned and entity-relation extraction is performed to obtain extracted information, which is then written into the shopping knowledge graph. Based on user budget, preferences, and purchase eligibility information, construct user demand profiles; Using the user demand profile as the retrieval criteria, a candidate set of product sources that meet policy and financial constraints is obtained by reasoning in the shopping knowledge graph. The attribute information of each product source in the product source candidate set is compared with the user demand profile to calculate the matching degree score of each product source relative to the user demand, and a comprehensive score is generated by combining the target factors of price level, supporting services, and appreciation potential. The product source candidate set is sorted according to the comprehensive score, and the top-scoring product sources and their corresponding recommendation explanations are output. Using the user demand profile as the retrieval criteria, a candidate set of product sources that meet policy and financial constraints is obtained through reasoning within the shopping knowledge graph, including: Using the budget range, applicable scenarios, special needs, and type preferences in the user demand profile as the starting point for retrieval, semantic subgraph construction is performed in the shopping knowledge graph to filter a preliminary set of product sources that meet the conditions of spatial location, product quality, and price. Based on the purchase eligibility status and financial constraints in the user demand profile, product source nodes that the user does not have the purchase eligibility or payment ability are excluded from the preliminary product source set to form a compliant product source set. Based on the multi-hop path relationship between user entities and product source entities in the shopping knowledge graph, the graph association confidence of each compliant product source entity in the compliant product source set is calculated; the graph association confidence reflects the degree of semantic fit between the product source and user needs. Entities with a graph association confidence level not lower than a preset threshold in the compliant product source set are selected as the product source candidate set. The expression for the confidence level of the graph association is: in, Calculate the confidence level of the graph association between user entities and product source entities; The number of different semantic path types connecting the user entity and the product source entity is automatically obtained by traversing and counting the shopping knowledge graph. The shortest path length between the user entity and the product source entity in the shopping knowledge graph reflects semantic distance and is calculated using a standard graph search algorithm. Graph embedding vectors for user entities. For the graph embedding vector of the product source entity, and Generated from the same unsupervised embedding model; is the Euclidean distance between vectors, used to measure differences in semantic similarity.

2. The knowledge graph-based personalized recommendation method according to claim 1, characterized in that, The basic information data of the product source includes supply quantity, grade, type, total price, unit price, production date, and geographical location of the product source; The market supply and demand data include total market supply, product shelf life, total amount of replaceable products, and total market demand. The policy constraint data includes purchase restriction policies, loan restriction policies, purchase eligibility requirements, and tax standards; The financial constraint data includes interest rates, credit payment terms, installment payment ratios, and user credit scores. The user behavior data includes the user's browsing history, favorites history, consultation history, trial history, and proactive feedback.

3. The knowledge graph-based personalized recommendation method according to claim 1, characterized in that, The multi-source heterogeneous data is cleaned and entity-relation extraction is performed to obtain extracted information, which is then written into a shopping knowledge graph, including: The multi-source heterogeneous data is subjected to format standardization, missing value completion, and outlier detection to obtain a cleaned structured dataset; Based on natural language processing technology, named entity recognition is performed on text information in structured datasets to extract core entities such as product source, location, supporting services, policy terms, financial products and user behavior. Combining a pre-defined domain ontology dictionary and semantic rules, attribute relationships and associations are mined from the core entities to generate extracted information in units of "entity-relationship-entity" triples; The extracted information is written into a graph database to construct the node set and edge set of the shopping knowledge graph.

4. The knowledge graph-based personalized recommendation method according to claim 1, characterized in that, Based on user budget, preferences, and purchase eligibility information, construct a user needs profile, including: The system obtains basic information provided by the user, estimates the user's shopping budget based on the basic information, and extracts demographic features; the basic information includes age, marital status, occupation, annual income, family structure, and city / region of residence. The system analyzes users' historical behavior data and uses a time-weighted modeling method to generate a distribution of users' shopping behavior preferences; the historical behavior data includes browsing history, collection history, consultation history, and trial history. The demographic characteristics, shopping budget capacity, and shopping behavior preference distribution are fused to obtain the preliminary user profile features. Based on the aforementioned basic information and current policy constraint data, the preset purchase qualification verification module is invoked to evaluate the user's shopping area restrictions, maximum loan amount, and qualification status, and the evaluation results are used as profile constraint attributes. A user demand profile vector is constructed by combining the preliminary profile features with the constraint attributes.

5. The knowledge graph-based personalized recommendation method according to claim 4, characterized in that, The expression for the shopping behavior preference distribution is: in, For users The vector of the shopping behavior preference distribution; The total number of historical behavior records accumulated for the user; For the first The attribute vectors of the product sources associated with each behavior record are generated after being uniformly represented by a semantic embedding model; For the first The interaction depth factor corresponding to each behavior record is naturally obtained based on the behavior category mapping: browsing is 1, collecting is 2, consulting is 3, and trying is 4. This is the current timestamp. For the first The timestamp of the occurrence of the action; The effective response weight function represents the behavior; The normalization factor is defined as the sum of the weights of all behavioral responses, i.e. , used to Normalized into a probabilistic preference vector.

6. The knowledge graph-based personalized recommendation method according to claim 4, characterized in that, Based on the aforementioned basic information and current policy constraints, a preset purchase eligibility verification module is invoked to assess the user's shopping area restrictions, maximum loan amount, and eligibility status. The assessment results are then used as profile constraint attributes, including: Obtain the user's basic identity information; the basic identity information includes household registration type, marital status, number of family members, tax or social security records, and fixed asset holdings. Based on the market regulation policies corresponding to the current region, retrieve the purchase restriction rules, loan restriction regulations, and purchase restriction identification standards for the current region; The basic identity information is matched with the market regulation policy to determine whether the user is eligible to purchase in each target area, and the corresponding list of area restrictions is output. The system calls a pre-defined purchase eligibility verification module to automatically calculate the maximum loan amount a user can obtain based on their income level, debt situation, credit history, and current interest rate environment. Based on the regional restriction list, the maximum loan amount, and whether the eligibility conditions are met, a purchase eligibility status identifier is generated in the preliminary profile features, and the purchase eligibility status identifier is used as the profile constraint attribute.

7. A knowledge graph-based personalized shopping recommendation system for implementing the method as described in any one of claims 1-6, characterized in that, include: A multi-source data acquisition unit is used to collect multi-source heterogeneous data from product sources; the multi-source heterogeneous data includes basic information data of product sources, market supply and demand data, policy constraint data, financial constraint data, and user behavior data; The graph construction and update unit is used to perform cleaning and entity-relation extraction on the multi-source heterogeneous data to obtain extracted information, and write the extracted information into the shopping knowledge graph; The user profile generation unit is used to build user demand profiles based on user budget, preferences and purchase eligibility information; The candidate product source reasoning and filtering unit is used to reason in the shopping knowledge graph to obtain a set of candidate product sources that meet policy and financial constraints, based on the user demand profile as the retrieval condition. The matching degree calculation and scoring generation unit is used to compare the attribute information of each product source in the product source candidate set with the user demand profile, calculate the matching degree score of each product source relative to the user demand, and generate a comprehensive score by combining the target factors of price level, supporting services, and appreciation potential. The sorting and recommendation explanation output unit is used to sort the product source candidate set according to the comprehensive score and output the top-scoring product sources and their corresponding recommendation explanation information.

8. An electronic device, characterized in that, include: The processor, communication interface, memory, and communication bus are connected, with the processor, communication interface, and memory communicating with each other via the communication bus. The memory stores a computer program that, when executed by the processor, causes the processor to perform the knowledge graph-based personalized recommendation method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed, it implements the knowledge graph-based personalized recommendation method as described in any one of claims 1-6.

Images