Intelligent recommendation method and system for mall commodities based on knowledge graph

Abstract

This invention relates to the field of knowledge graph technology, specifically to a method and system for intelligent product recommendation in e-commerce based on knowledge graphs. The method includes: collecting product data, advertising data, and user interaction data; constructing a product knowledge graph; and establishing a mapping relationship between user advertising interaction behavior and product nodes in the product knowledge graph. Based on the product nodes corresponding to user advertising interaction behavior, the invention performs association path analysis in the product knowledge graph, conducts continuity analysis on the product node sequences corresponding to the association paths, and calculates the path continuity index corresponding to each effective association path. This invention constructs a product knowledge graph and establishes a user advertising interaction mapping, performs association path continuity analysis and constraint field construction based on the interaction behavior sequence, identifies unstable nodes and corrects paths, and dynamically updates the graph association strength, enabling the graph evolution to accurately reflect the user's real-time interests.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, specifically to a knowledge graph-based intelligent product recommendation method and system for e-commerce. Background Technology

[0002] With the increasing sophistication of internet advertising services, intelligent product recommendation based on knowledge graphs has become a key technology for improving the accuracy of ad targeting and user conversion rates. Existing methods typically construct product knowledge graphs and leverage user ad interactions (such as ad clicks, product browsing, and adding items to cart) to facilitate association and propagation within the graph, aiming to uncover recommended products related to users' potential interests. However, in real-world internet advertising scenarios, user ad interaction sequences often contain numerous accidental clicks, short-term fluctuations, and noise interference, leading to discontinuous and unstable propagation paths within the knowledge graph, severely impacting the accuracy of graph updates and the stability of recommendations. Furthermore, existing technologies often employ static statistical indicators such as the number of propagations and paths when evaluating the activity of product node associations, neglecting the artificially inflated activity caused by overlapping coverage and convergence during the association propagation process. This results in insufficient differentiation among candidate recommended products, making it difficult to effectively support accurate recommendation decisions in real-time ad delivery. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in the background technology by proposing a knowledge graph-based intelligent product recommendation method and system for e-commerce.

[0004] The technical solution of this invention: A method for intelligent product recommendation in e-commerce based on knowledge graphs, comprising: S1. Collect product data, advertising data, and user interaction data to construct a product knowledge graph and establish a mapping relationship between user advertising interaction behavior and product nodes in the product knowledge graph; S2. Based on the product nodes corresponding to user advertising interaction behavior, perform association path analysis in the product knowledge graph, perform continuity analysis on the product node sequence corresponding to the association path, calculate the path continuity index corresponding to each effective association path, construct an association path constraint field based on the path continuity index, identify nodes where association propagation is unstable and correct the constraints on the effective association paths, and dynamically update the association strength between each product node in the product knowledge graph based on the corrected effective association paths to generate an updated product knowledge graph that matches the current advertising behavior. S3. Based on the strength of association between product nodes in the updated product knowledge graph, analyze the association propagation characteristics of each product node, calculate the association activity level of product nodes, and generate a set of candidate recommended products based on the association activity level of product nodes. S4. Based on the real-time internet advertising status of the products corresponding to the candidate recommended product set, perform advertising adaptation analysis on the candidate recommended products and generate a product recommendation sequence.

[0005] As a further improvement to this technical solution, in step S1, a product knowledge graph is constructed, and a mapping relationship is established between user advertising interaction behavior and product nodes in the product knowledge graph, including the following steps: S1.1 Collect product data, advertising data, and user advertising interaction data; preprocess the various types of data; and generate basic data for the product knowledge graph. S1.2 Extract product entities, brand entities, and category entities based on the basic data of the product knowledge graph. Establish relationship edges between entities based on the attribution relationship between products and brands and the classification relationship between products and categories. Use each entity as a product node, each relationship as a connection edge, and the initial association strength as the edge weight to construct the product knowledge graph. S1.3 Read the ad identifier in the user ad interaction data, retrieve the corresponding product node in the product knowledge graph, and establish an index mapping relationship between each interaction record in the user ad interaction data and the corresponding product node to generate the user ad interaction behavior mapping result; S1.4 Sort the user advertising interaction behavior mapping results according to user identifier and advertising interaction time, and generate a continuous advertising interaction behavior sequence for each user.

[0006] As a further improvement to this technical solution, step S2, generating an updated product knowledge graph that matches the current advertising behavior, includes the following steps: S2.1 Based on the continuous advertising interaction behavior sequence and corresponding product nodes, each product node is used as the starting node. The association path is retrieved in the product knowledge graph along the product category relationship and brand relationship. All association paths from the starting node to the associated product node are obtained, and association paths with cyclic access are removed to generate a set of valid association paths. S2.2 For each effective association path, form a product node sequence according to the connection order of product nodes in the path, analyze the continuous change of the association relationship between adjacent product nodes, count the path length where the association relationship is maintained continuously and the position where the association relationship changes, introduce the position weighting factor, and calculate the path continuity index corresponding to each effective association path. S2.3 Construct an association path constraint field in the product knowledge graph based on the path continuity index and change location distribution characteristics of each effective association path; S2.4. Analyze the propagation state of each effective associated path based on the associated path constraint field, identify the unstable nodes of associated propagation, and use the unstable nodes of associated propagation as the termination positions to extract the corresponding effective associated paths and generate the corrected effective associated paths. S2.5. Based on the corrected effective association paths, calculate the association propagation between each product node, and dynamically update the association strength between each product node in the product knowledge graph according to the association propagation, generating an updated product knowledge graph that matches the current advertising behavior.

[0007] As a further improvement to this technical solution, in step S2.3, the associated path constraint field is constructed in the product knowledge graph based on the path continuity index and change position distribution characteristics of each effective associated path, including the following steps: S2.31. Read all valid association paths and their corresponding path continuity indices, assign the path continuity index of each valid association path to all product nodes that constitute the valid association path, and count all path continuity indices corresponding to each product node to form a set of path continuity indices corresponding to each product node. S2.32. Based on the set of path continuity indices corresponding to each product node, calculate the node continuity index corresponding to each product node, and write the node continuity index corresponding to each product node into the corresponding product node in the product knowledge graph. S2.33. Read the distribution characteristics of the change positions corresponding to each effective association path, count the frequency of each commodity node being marked as a change position in its effective association path and the relative advance degree of the change position, and calculate the change impact coefficient corresponding to each commodity node. S2.34. Using the node continuity index corresponding to each product node as the node status and the change impact coefficient as the node change sensitivity index, the node is continuously expanded according to the relationship between each product node in the product knowledge graph, forming a node continuity region covering the entire product knowledge graph. S2.35. Based on the continuous index of each node, the continuous regions of each node are marked with state, generating stable propagation regions and propagation offset regions, and the continuous regions of all nodes and their corresponding state marks are combined to form an associated path constraint field.

[0008] As a further improvement to this technical solution, in step S2.4, the propagation state of each effective associated path is analyzed based on the associated path constraint field, and the nodes where the associated propagation becomes unstable are identified, including the following steps: S2.41. Read the associated path constraint field and valid associated paths, map each valid associated path to the associated path constraint field in the order of the connection of the product nodes, and obtain the state flag corresponding to the continuous area of ​​the node where each product node is located, and generate the propagation state sequence corresponding to each valid associated path. S2.42. Along the propagation direction of each effective associated path, perform sequential analysis on the propagation state sequence, determine whether the state markers corresponding to adjacent product nodes have changed, and identify propagation state transition nodes; S2.43. Starting from the propagation state transition node, continue to analyze the state markers of each product node after the current valid association path, and identify the association propagation instability nodes.

[0009] As a further improvement to this technical solution, step S3, generating a candidate recommended product set, includes the following steps: S3.1 Based on the updated product knowledge graph, take the product node corresponding to the current advertising interaction behavior as the starting product node, perform association propagation analysis along the relationship between product nodes in the updated product knowledge graph, count all related product nodes, and extract the association strength and association propagation path between the starting product node and each related product node to generate a product node association propagation dataset. S3.2 Based on the product node association propagation dataset, statistically analyze the association propagation characteristics of each product node in the association propagation process, and calculate the initial node association activity level of each product node based on the association propagation characteristics. S3.3 Based on the changing characteristics of the association propagation coverage of each product node, identify the association propagation convergence domain of the product node, calculate the propagation convergence index of each product node, and correct the initial node association activity level based on the propagation convergence index to generate the corrected product node association activity level. S3.4. Sort the products based on the activity level of the corrected product nodes and generate a set of candidate recommended products.

[0010] As a further improvement to this technical solution, step S3.3, generating the corrected activity level of the product node association, includes the following steps: S3.31. Read the initial node association activity level and the association propagation coverage objects corresponding to each product node. Taking each product node as the center, count the overlap between association propagation coverage objects and generate the association propagation convergence region. S3.32. For each associated propagation convergence domain, statistically analyze the changes in the number of newly propagated objects covered according to the order of associated propagation, and calculate the propagation convergence index of the corresponding associated propagation convergence domain. S3.33. Based on the propagation convergence index corresponding to each product node, the initial node association activity level is corrected to generate the corrected product node association activity level.

[0011] As a further improvement to this technical solution, in S3.32, the change in the number of newly added propagation coverage objects is statistically analyzed according to the order of associated propagation, and the propagation convergence index of the corresponding associated propagation convergence domain is calculated, including the following steps: For each associated propagation convergence domain, the propagation coverage objects corresponding to each propagation stage are read sequentially according to the propagation order. The number of newly added propagation coverage objects in the current propagation stage compared to the previous propagation stage is counted to form a sequence of newly added propagation coverage objects. The changing trend of newly added propagation coverage objects is analyzed based on the sequence of newly added propagation coverage objects, and the propagation convergence index of the corresponding associated propagation convergence domain is calculated.

[0012] As a further improvement to this technical solution, step S4, generating a product recommendation sequence, includes the following steps: S4.1 Read the candidate recommended product set and collect the real-time internet advertising delivery status data corresponding to the candidate recommended products to generate a candidate product advertising status dataset; S4.2 Based on the candidate product advertising status dataset, analyze the advertising availability status and advertising display activity of each candidate recommended product, and calculate the advertising suitability of each candidate recommended product. S4.3. Combining the activity level of the corrected product node association and the degree of ad adaptation of each candidate recommended product, a comprehensive adaptation analysis is performed on the candidate recommended products to generate a comprehensive recommendation priority for each candidate recommended product. S4.4 Sort the candidate recommended products according to their comprehensive recommendation priority to generate a product recommendation sequence.

[0013] On the other hand, the present invention provides a knowledge graph-based intelligent recommendation system for e-commerce products, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the above-mentioned knowledge graph-based intelligent recommendation method for e-commerce products.

[0014] Compared with the prior art, the above-mentioned technical solution of the present invention has the following beneficial technical effects: by constructing a product knowledge graph and establishing a user advertising interaction mapping, performing continuous analysis of association paths and constructing a constraint field based on the interaction behavior sequence, identifying unstable nodes and correcting paths, and dynamically updating the graph association strength, the graph evolution accurately reflects the user's real-time interests; furthermore, by correcting node activity through the propagation convergence index, the artificial increase in activity caused by repeated propagation is weakened, and a candidate recommendation set is generated; then, by integrating the real-time advertising delivery status for adaptation analysis, a comprehensive priority is generated by weighting and a recommendation sequence is output, realizing deep synergy between graph evolution and advertising delivery, and effectively solving the problem of disconnect between recommendation and advertising inventory. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall method of the present invention. Detailed Implementation

[0016] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Example 1: Please refer to Figure 1 As shown, this embodiment provides a knowledge graph-based intelligent product recommendation method for e-commerce, including the following steps: S1. Collect product data, advertising data, and user interaction data to construct a product knowledge graph and establish a mapping relationship between user advertising interaction behavior and product nodes in the product knowledge graph; In this embodiment, a product knowledge graph is constructed, and a mapping relationship is established between user advertising interaction behavior and product nodes in the product knowledge graph, including the following steps: S1.1 Collect product data, advertising data, and user advertising interaction data generated during the Internet advertising service process, preprocess the various types of data (including data cleaning, format unification, and identifier association), and generate basic data for the product knowledge graph. Among them, product data includes product identifiers, product categories, brand information, and product attributes; advertising data includes advertising identifiers and advertising-related product information; and user advertising interaction data includes ad impressions, ad clicks, product browsing, product favorites, adding to cart, and product purchase records. S1.2 Extract product entities, brand entities, and category entities based on the basic data of the product knowledge graph. Establish relationship edges between entities based on the attribution relationship between products and brands (i.e., brand relationship) and the classification relationship between products and categories (i.e., category relationship). Assign an initial association strength value to each relationship edge (the initial default value is uniformly set to 1.0). Use each entity as a product node, each relationship as a connection edge, and the initial association strength as the edge weight to construct the product knowledge graph. S1.3 Read the ad identifier or product identifier in the user ad interaction data, retrieve the corresponding product node in the product knowledge graph, and establish an index mapping relationship between each interaction record in the user ad interaction data and the corresponding product node to generate the user ad interaction behavior mapping result. S1.4. Sort the user advertising interaction behavior mapping results according to user identifier and advertising interaction time to generate a continuous advertising interaction behavior sequence for each user (first, classify all product node mapping results according to user identifier, and merge all interaction records belonging to the same user into the same user set; then sort the product node mapping results in each user set by time from earliest to latest according to advertising interaction time, so that each product node maintains the same chronological order as the actual advertising interaction process; if there are multiple interaction records with the same advertising interaction time, the sorting order is determined according to the preset behavior priority (purchase, add to cart, favorite, product browsing, ad click, ad exposure); finally, connect each product node in sequence according to the sorting result to generate a continuous advertising interaction behavior sequence for each user, and retain the interaction time and interaction behavior type corresponding to each product node as input data for subsequent product knowledge graph association path analysis).

[0018] S2. Based on the product nodes corresponding to user advertising interaction behavior, perform association path analysis in the product knowledge graph. Analyze the continuity of the product node sequences corresponding to the association paths, calculate the path continuity index for each effective association path, construct an association path constraint field based on the path continuity index, identify nodes where association propagation becomes unstable, and correct the constraints on effective association paths. Based on the corrected effective association paths, dynamically update the association strength between each product node in the product knowledge graph, generating an updated product knowledge graph that matches the current advertising behavior (addressing the challenges in internet advertising recommendation scenarios where the product knowledge graph is affected by noise interference from advertising interaction data and cross-propagation of multiple association paths). This invention addresses the specific problems of discontinuous propagation, frequent node instability, and decreased graph update accuracy caused by the lack of association. It introduces a path continuity index to quantify the stability of each effective association path, constructs an association path constraint field to identify unstable propagation nodes, and corrects the constraints on effective association paths, dynamically updating the association strength between product nodes. Furthermore, this invention establishes a state-aware and dynamic correction mechanism based on path continuity and changing position distribution. This mechanism can accurately filter out discontinuous and unstable noise propagation paths, suppress the negative impact of unstable nodes on the global association structure of the graph, and make knowledge graph updates more robust and adaptive, thus providing a more reliable and accurate structured basis for subsequent recommendations. In this embodiment, generating an updated product knowledge graph that matches the current advertising behavior includes the following steps: S2.1 Based on the continuous advertising interaction sequence and corresponding product nodes, using each product node as the starting node, perform association path retrieval in the product knowledge graph along product category and brand relationships to obtain all association paths from the starting node to the associated product nodes, and remove association paths with cyclic access to generate a set of valid association paths; specifically: using the product node corresponding to the current advertising interaction in the continuous advertising interaction sequence as the starting product node, traverse the product knowledge graph layer by layer outward according to product category and brand relationships to retrieve product nodes that are related to the starting product node; each time a new product node is reached... When a product node is selected, it is added to the corresponding associated path. The process continues recursively along the product category and brand relationships connected to the current product node until no new associated product nodes are found. This yields all associated paths from the starting product node to each associated product node. Then, a validity check is performed on all associated paths to determine if the same product node is accessed repeatedly. If a duplicate product node is found, the associated path is considered to have a circular access pattern and is removed. If all product nodes in a path are accessed only once, the associated path is retained as a valid associated path, ultimately generating a set of valid associated paths. S2.2 For each valid association path, form a product node sequence according to the connection order of product nodes in the path, analyze the continuous changes in the association relationship between adjacent product nodes, and count the path length where the association relationship is maintained continuously and the position where the association relationship changes. Introduce a position weighting factor (i.e., The following steps are performed to calculate the path continuity index for each valid association path: For each valid association path, a sequence of product nodes is formed according to their connection order within the path. The association relationship types between adjacent product nodes are then read sequentially (i.e., the specific relationship categories corresponding to the edges connecting two product nodes in the product knowledge graph). The current association relationship type is compared with the previous adjacent product node's association relationship type. When the association relationship types remain consistent, the continuity length is increased by 1, and the analysis continues along the path. When the association relationship type changes, the position number of the current product node is recorded as the change position, along with the relative position ratio of this change position to the total path length. A new continuity length is then calculated. After traversing the entire association path, the maximum continuity length, the number of association relationship changes, and the relative position ratio of each change position are obtained, and the corresponding path continuity index is generated. (In the formula, For the first The length of the longest continuous path (in terms of the number of edges) in which the association relationship is maintained continuously among the valid association paths. For the first The total path length (number of edges) of the valid associated paths. For the first The number of times the association relationship changes in a valid association path. For the first The first of the valid association paths The sequential numbering of each change position in the path (counting edges starting from the starting node); this formula, by introducing a distance penalty for the change position, gives a higher penalty weight to changes occurring closer to the starting point of the path, and is used to quantify the propagation continuity and stability of the associated path, providing a basis for continuous evaluation for the subsequent construction of the associated path constraint field; S2.3 Construct an association path constraint field in the product knowledge graph based on the path continuity index and change location distribution characteristics of each effective association path; The process of constructing a path constraint field in the product knowledge graph based on the path continuity index and change location distribution characteristics of each effective path includes the following steps: S2.31. Read all valid association paths and their corresponding path continuity indices, assign the path continuity index of each valid association path to all product nodes that constitute the valid association path, and count all path continuity indices corresponding to each product node to form a set of path continuity indices corresponding to each product node. S2.32. Based on the path continuity index set corresponding to each product node, calculate the node continuity index corresponding to each product node. (In the formula, For goods passing through nodes The number of valid associated paths, For the first The path continuity index corresponding to each valid association path is calculated, and the node continuity index corresponding to each product node is written into the corresponding product node in the product knowledge graph as a continuous state attribute of the product node, providing node state information for the subsequent construction of the association path constraint field. S2.33. Read the distribution characteristics of the changed positions corresponding to each effective association path, count the frequency of each product node being marked as a changed position in its effective association path and the relative advance degree of the changed position, and calculate the change impact coefficient corresponding to each product node. (In the formula, For indicator functions, when the product node In the The value is 1 if the location is recorded as a change in the valid associated path, and 0 otherwise. For the first The total path length of the valid associated paths, For product nodes In the The position number used as the location of change in the effective associated path, and the change impact coefficient. This is used to quantify the tendency and scope of influence of product nodes in causing relationship changes during the association propagation process. The larger the value, the earlier the node appears as a change position and the more significant the disruption to the stability of the subsequent propagation path. The change influence coefficient corresponding to each product node is written into the corresponding product node in the product knowledge graph as a change state attribute of the product node. S2.34. Using the node continuity index corresponding to each product node as the node state and the change impact coefficient as the node change sensitivity index, the system continuously expands according to the association relationships between product nodes in the product knowledge graph, forming a continuous node region covering the entire product knowledge graph. (Starting from any product node, the system continuously expands along the association relationships directly connected to that product node in the product knowledge graph to adjacent product nodes, comparing the change magnitude of the node continuity index corresponding to adjacent product nodes; when the change magnitude of the node continuity index between adjacent product nodes is less than the first preset change threshold and the change impact coefficients of adjacent product nodes are all lower than the second preset change threshold, the adjacent product nodes are included in the same continuous expansion region, and the system continues to expand outward along the association relationships using that adjacent product node as a new expansion node.) When the change in the continuous index of a node is greater than or equal to the first preset change threshold, or the change influence coefficient of any product node is greater than or equal to the second preset change threshold, the expansion in the current direction is terminated, and the above process is repeated with other product nodes that have not yet completed the expansion, until all product nodes have completed the continuous expansion, forming multiple continuous regions of nodes covering the entire product knowledge graph; wherein, the first preset change threshold is dynamically set based on the standard deviation of the continuous index of all product nodes, and the value is half of the standard deviation; the second preset change threshold is dynamically set based on the quartile of the change influence coefficient of all product nodes, and the value is the upper quartile); the discrete continuous index attributes of each product node are diffused into a structured continuous region covering the entire graph, providing a spatial basis for subsequent stability and offset state marking of each region; S2.35. Based on the continuous index of each node, the continuous regions of each node are marked with state labels to generate stable propagation regions and propagation deviation regions. This provides a global state benchmark for the propagation stability assessment of effective associated paths, enabling subsequent path analysis to quickly determine whether the propagation trend has shifted based on the type of region where the node is located (the node continuous index of all commodity nodes in the continuous region of each node is averaged. If the average value is higher than the preset continuous threshold, it is marked as a stable propagation region; if the average value is lower than the preset continuous threshold, it is marked as a propagation deviation region; where the preset continuous threshold is dynamically set based on the average of the node continuous index of all commodity nodes). The continuous regions of all nodes and the corresponding state labels are combined to form an associated path constraint field, providing a global state benchmark for the propagation stability assessment of effective associated paths, thereby accurately identifying and trunculating unstable propagation nodes and guiding path correction and map updates. S2.4. Based on the constraint field of the associated path, analyze the propagation state of each effective associated path and identify the unstable nodes of the associated propagation. Use the unstable nodes as the termination positions to truncate the corresponding effective associated paths and generate corrected effective associated paths. (Using the unstable nodes as the termination positions, the effective associated paths are truncated, retaining only the product nodes and their relationships between the starting product node and the unstable node, and deleting all product nodes and their corresponding relationships after the unstable node. For effective associated paths where unstable nodes are not identified, the original effective associated paths are retained. After correcting all effective associated paths, a set of corrected effective associated paths is generated, and the product node sequence, path length, and relationship information corresponding to each corrected effective associated path are retained as input data for subsequent statistical analysis of product node association propagation and dynamic updating of product node association strength in the product knowledge graph. By truncating the unstable nodes to terminate unreliable propagation segments, the stable prefixes of the effective associated paths are retained, purifying the propagation data used to update the graph from the source and avoiding noise paths from polluting the calculation of node association strength.) Furthermore, based on the associated path constraint field analysis, the propagation state of each effective associated path is analyzed, and nodes where the associated propagation becomes unstable are identified, including the following steps: S2.41. Read the associated path constraint field and valid associated paths, map each valid associated path to the associated path constraint field in the order of the connection of the product nodes, and obtain the state flag corresponding to the continuous area of ​​the node where each product node is located, and generate the propagation state sequence corresponding to each valid associated path. S2.42. Along the propagation direction of each effective association path, perform sequential analysis on the propagation state sequence, determine whether the state markers corresponding to adjacent product nodes have changed, identify propagation state transition nodes, and use them as the starting reference positions for subsequent determination of unstable propagation nodes. This is used to mark the critical point where an effective association path enters the offset region from the stable region (when the propagation state continuously enters the propagation offset region from the stable propagation region, record the position where the propagation state changes, and determine the product node where the propagation state changes for the first time as the propagation state transition node). S2.43. Starting from the propagation state transition node, continue to analyze the state markers of each product node after the current effective association path, and identify the associated propagation instability node (first check whether there is at least one subsequent product node after the propagation state transition node; if there is no subsequent node, then abandon the node as an unstable node; if there is a subsequent node, continue to analyze the state markers of the subsequent product nodes in sequence along the propagation direction of the current effective association path; if the state markers of all product nodes from the transition node to the end of the current effective association path are all propagation offset areas, and the proportion of these subsequent nodes to the total number of nodes in the path exceeds the preset proportion threshold (default is 10%), then the propagation state transition node is determined as an associated propagation instability node, which is used to effectively locate the key node position in the effective association path that causes the propagation trend to change from stable to offset, and provide accurate termination coordinates for subsequent path truncation correction; otherwise, the propagation state transition node is not determined as an associated propagation instability node, and the propagation state analysis is continued for the subsequent product nodes in the current effective association path). S2.5. Based on the corrected effective association paths, statistically analyze the association propagation between each product node (including the number of effective association propagations, association coverage, and path participation; the number of effective association propagations is the cumulative frequency of product node pairs being propagated through all corrected effective association paths; association coverage is the range of related product nodes that a product node can reach through the corrected effective association paths; and path participation is the number of product nodes that a product node participates in the corrected effective association paths). Dynamically update the association strength between each product node in the product knowledge graph based on the association propagation results (when the number of effective association propagations between product node pairs increases, the increase is accumulated to the corresponding product knowledge graph). The original association strength between product nodes (i.e., the initial edge weight assigned in step S1.2); when the association coverage of a product node changes, the association strength is incrementally updated based on the number of newly added associated product nodes; when the number of times a product node participates in the corrected effective association path changes, the association strength is corrected based on the change in the number of path participations; after completing the association strength calculation for all product nodes, the updated association strength is written into the association relationship between the corresponding product nodes in the product knowledge graph, replacing the original association strength value (initial value or previous update value), thus completing the dynamic update of the product knowledge graph and generating an updated product knowledge graph that matches the current advertising behavior.

[0019] S3. Based on the strength of association between product nodes in the updated product knowledge graph, analyze the association propagation characteristics of each product node, calculate the association activity level of product nodes, and generate a set of candidate recommended products based on the association activity level of product nodes. In this embodiment, generating a candidate recommended product set includes the following steps: S3.1 Based on the updated product knowledge graph, take the product node corresponding to the current advertising interaction behavior as the starting product node, perform association propagation analysis along the association relationship between product nodes in the updated product knowledge graph, count all related product nodes that can be reached during the association propagation process, and extract the association strength and association propagation path between the starting product node and each related product node to generate a product node association propagation dataset. S3.2 Based on the product node association propagation dataset, statistically analyze the association propagation characteristics of each product node during the association propagation process (including the number of association propagations, the number of path participants, and the coverage of association propagation for each product node), and calculate the initial node association activity level corresponding to each product node based on the association propagation characteristics. (In the formula, This represents the normalized number of association propagation attempts, which is the cumulative frequency of a product node being reached through association propagation in all corrected valid association paths. It is a relative value mapped to the [0,1] interval using the min-max normalization method. The normalized path participation count is the relative value of the total number of corrected valid association paths in which a product node participates, after max-min normalization. It is used to characterize its participation density in the path structure. The normalized coverage of the association propagation is the result of max-min normalization of the total number of different product nodes that a product node can reach through the corrected effective association path. This is used to measure the breadth of its influence. For product node indexing, As a weight for the number of propagations, The weights for the number of participants in the path. The weight of the coverage of the associated propagation. Weight , , The initial values ​​were determined by optimizing historical interaction data using a grid search method, with recommendation accuracy as the optimization objective. During system operation, values ​​were adjusted using a sliding window approach based on online feedback data. The weight values ​​are recalculated every time window to achieve adaptive adjustment of the weights. S3.3. Based on the changing characteristics of the association propagation coverage of each product node, identify the association propagation convergence domain corresponding to the product node, calculate the propagation convergence index corresponding to each product node, and correct the initial node association activity level based on the propagation convergence index to generate the corrected product node association activity level (addressing the specific problems of artificially inflated activity and insufficient differentiation of candidate recommended products caused by repeated superposition of association propagation paths and overlapping coverage in the evaluation of product association activity based on static propagation statistics; by identifying the association propagation convergence domain corresponding to each product node and calculating the propagation convergence index, the initial node association activity level is corrected based on the decay trend of newly added propagation coverage objects; this invention introduces a dynamic perception mechanism for propagation convergence, which can effectively identify and suppress the artificially high activity level of nodes that have entered the convergence state due to repeated propagation, so that the corrected node association activity level more realistically reflects its actual propagation driving force and expansion potential, thereby improving the quality differentiation of the candidate recommended product set and the rationality of the recommendation ranking). The process of generating the corrected activity level of product nodes includes the following steps: S3.31. Read the initial node association activity level and the association propagation coverage objects corresponding to each product node. Taking each product node as the center, count the overlap between association propagation coverage objects and generate the association propagation convergence region (using a two-stage determination method: In the first stage, take each product node as the center node and construct the corresponding coverage object set of its association propagation coverage objects; compare the coverage object set of the current product node with the coverage object sets corresponding to its associated product nodes in turn, count the number of the same product nodes in the two coverage object sets, and calculate the proportion of the number of the same product nodes to the total number of product nodes in the union of the two coverage object sets. When this proportion is greater than the preset overlap threshold, mark the connection edge between the current product node and the associated product node as an edge with a propagation convergence relationship; where the preset overlap threshold is 0.5, and supports adaptive adjustment based on the sparsity of the product knowledge graph; the specific quantification standard is: using graph density). To measure, among which, This represents the total number of relation edges in the graph. The total number of product nodes; when When it is determined to be a sparse graph, the threshold is reduced to 0.3; when When it is determined to be a dense spectrum, the threshold is raised to 0.6; when In the first stage, the default threshold of 0.5 is maintained. In the second stage, only edges marked as having propagation convergence relationships are retained in the product knowledge graph, and connected components are searched along these edges. All product nodes contained in each connected component are merged into the same association propagation convergence domain. The above process is repeated until all product nodes have been traversed, generating multiple association propagation convergence domains. These domains are used to identify node groups with highly overlapping coverage objects during the association propagation process, serving as the basic calculation unit for determining whether the propagation is approaching saturation and correcting the artificial increase in node activity accordingly. S3.32. For each associated propagation convergence domain, statistically analyze the changes in the number of newly covered objects according to the associated propagation order, and calculate the propagation convergence index of the corresponding associated propagation convergence domain (wherein, the propagation convergence index corresponding to each associated propagation convergence domain is assigned to all product nodes within that associated propagation convergence domain, serving as the propagation convergence index corresponding to each product node). Furthermore, the changes in the number of newly covered objects are statistically analyzed according to the order of propagation, and the propagation convergence index of the corresponding propagation convergence region is calculated, including the following steps: For each convergence region of the correlation propagation, the propagation coverage objects corresponding to each propagation stage are read sequentially according to the propagation order. The number of newly added propagation coverage objects in the current propagation stage compared to the previous propagation stage is counted to form a sequence of the number of newly added propagation coverage objects. ( For the first The number of newly covered targets in each stage of the transmission process , (This refers to the number of propagation stages); based on the sequence of newly added propagation coverage objects, analyze the changing trend of newly added propagation coverage objects, and calculate the propagation convergence index of the corresponding associated propagation convergence region (calculate the overall decay trend value of the sequence of newly added propagation coverage objects). ,in, To prevent extremely small positive values ​​where the denominator is zero, calculate the monotonically decaying penalty term of the sequence. This term is used to penalize non-monotonic fluctuations in the sequence; the penalty term is 0 when the sequence is strictly monotonically decreasing. The propagation convergence index is calculated by combining the decay trend value and the monotonically decaying penalty term. , When the sequence of newly added propagation coverage targets continues to decrease monotonically, A value approaching 1 indicates that correlation propagation tends to be highly convergent; as sequence fluctuations increase, (If the value approaches 0, it indicates that the association propagation has not converged.) This provides a quantitative convergence basis for correcting the association activity level of nodes. By measuring the degree of decay of the propagation coverage expansion, the degree of artificial increase in activity level can be judged, and differentiated adjustments can be made. S3.33. Based on the propagation convergence index corresponding to each product node, the initial node association activity level is corrected to generate the corrected product node association activity level. (In the formula, For the first The initial node activity level associated with each product node. The attenuation control coefficient is used to control the intensity of the convergence penalty and avoid excessively weakening the activity of core nodes (determined through historical data grid search). It is used to weaken the artificial increase in activity of product nodes caused by repeated propagation during association propagation, so that the association activity of product nodes more realistically reflects their actual association propagation ability. S3.4. Sort the products based on the corrected activity level of the product nodes and generate a set of candidate recommended products (select product nodes with a corrected activity level higher than a preset activity threshold as candidate recommended product nodes, and extract the product information corresponding to the candidate recommended product nodes to generate a set of candidate recommended products; wherein, the preset activity threshold is determined based on the statistical distribution of the corrected activity level of all candidate product nodes, and the value is the median of the corrected activity level of all candidate product nodes).

[0020] S4. Based on the real-time internet advertising status of the products corresponding to the candidate recommended product set, perform advertising adaptation analysis on the candidate recommended products and generate a product recommendation sequence; In this embodiment, generating a product recommendation sequence includes the following steps: S4.1 Read the candidate recommended product set and collect the real-time internet advertising delivery status data corresponding to the candidate recommended products, including at least the advertising delivery status and advertising display frequency, and generate a candidate product advertising status dataset. S4.2. Based on the candidate product ad status dataset, analyze the ad availability status and ad display activity of each candidate recommended product, and calculate the ad fit degree for each candidate recommended product. Specifically: Based on the candidate product ad status dataset, first determine whether the ad for each candidate recommended product is in the running state. Assign a value of 1 to ads in the normal running state and a value of 0 to ads in the paused, ended, or not running state, as the ad availability status. Then, count the number of times the ad for each candidate recommended product is displayed within a preset time window, and process the number of displays using the max-min normalization method to obtain the ad display activity degree. Finally, perform a weighted fusion of the ad availability status and the ad display activity degree to generate the ad fit degree. (In the formula, This indicates the ad's availability status. A value of 1 indicates the ad is ready to run; otherwise, a value of 0 indicates it is ready to run. To normalize the ad display activity level, The weighting of the ad's availability status. Weighting of ad display activity Weight and Personalized settings based on advertisers' historical campaign preferences); S4.3. Combining the corrected activity level of product nodes and the degree of ad adaptation for each candidate recommended product, a comprehensive adaptation analysis is performed on the candidate recommended products to generate a comprehensive recommendation priority for each candidate recommended product. Specifically, this involves: reading the corrected activity level of product nodes and the degree of ad adaptation for each candidate recommended product, and matching them according to the product identifier of the candidate recommended product; then normalizing the corrected activity level of product nodes and the degree of ad adaptation (using the max-min normalization method) to eliminate numerical differences between different indicators; and calculating the comprehensive recommendation priority for each candidate recommended product based on the normalized corrected activity level of product nodes and the degree of ad adaptation. (In the formula, To determine the normalized activity level of product nodes, To determine the degree of ad fit after normalization, The weight assigned to the activity level of the associated product node. The weighting of ad relevance. and Dynamically set according to the business objectives of the recommendation system, for example: when the business objective focuses on improving click-through rate, The value is 0.7. The value is 0.3; when the business objective focuses on increasing advertising revenue, The value is 0.4. The value is 0.6; the default value is... , The revised product node association activity level is used to characterize the degree of association between the product and the current user's advertising interaction interest, and the advertising adaptation level is used to characterize the real-time matching degree of the product's corresponding advertising. S4.4. Sort the candidate recommended products according to their comprehensive recommendation priority and generate a product recommendation sequence (read the comprehensive recommendation priority of each candidate recommended product and sort them from high to low according to the comprehensive recommendation priority; when multiple candidate recommended products have the same comprehensive recommendation priority, the candidate recommended product with higher activity level of the corrected product node association is selected first; if the activity level of the corrected product node association is still the same, the candidate recommended product with higher ad fit is selected first; after sorting all candidate recommended products, generate the product recommendation sequence according to the sorting result, and output the product identifier, recommendation order and corresponding ad information as the final product recommendation result).

[0021] Example 2: This example provides a knowledge graph-based intelligent product recommendation system for e-commerce, including a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor executes the computer program to implement the knowledge graph-based intelligent product recommendation method for e-commerce described in Example 1.

[0022] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A knowledge graph-based intelligent product recommendation method for e-commerce platforms, characterized in that: include: S1. Collect product data, advertising data, and user interaction data to construct a product knowledge graph and establish a mapping relationship between user advertising interaction behavior and product nodes in the product knowledge graph; S2. Based on the product nodes corresponding to user advertising interaction behavior, perform association path analysis in the product knowledge graph, perform continuity analysis on the product node sequence corresponding to the association path, calculate the path continuity index corresponding to each effective association path, construct an association path constraint field based on the path continuity index, identify nodes where association propagation is unstable and correct the constraints on the effective association paths, and dynamically update the association strength between each product node in the product knowledge graph based on the corrected effective association paths to generate an updated product knowledge graph that matches the current advertising behavior. S3. Based on the strength of association between product nodes in the updated product knowledge graph, analyze the association propagation characteristics of each product node, calculate the association activity level of product nodes, and generate a set of candidate recommended products based on the association activity level of product nodes. S4. Based on the real-time internet advertising status of the products corresponding to the candidate recommended product set, perform advertising adaptation analysis on the candidate recommended products and generate a product recommendation sequence.

2. The knowledge graph-based intelligent product recommendation method for e-commerce platforms according to claim 1, characterized in that, In step S1, a product knowledge graph is constructed, and a mapping relationship is established between user advertising interaction behavior and product nodes in the product knowledge graph, including the following steps: S1.1 Collect product data, advertising data, and user advertising interaction data; preprocess the various types of data; and generate basic data for the product knowledge graph. S1.2 Extract product entities, brand entities, and category entities based on the basic data of the product knowledge graph. Establish relationship edges between entities based on the attribution relationship between products and brands and the classification relationship between products and categories. Use each entity as a product node, each relationship as a connection edge, and the initial association strength as the edge weight to construct the product knowledge graph. S1.3 Read the ad identifier in the user ad interaction data, retrieve the corresponding product node in the product knowledge graph, and establish an index mapping relationship between each interaction record in the user ad interaction data and the corresponding product node to generate the user ad interaction behavior mapping result; S1.4 Sort the user advertising interaction behavior mapping results according to user identifier and advertising interaction time, and generate a continuous advertising interaction behavior sequence for each user.

3. The intelligent product recommendation method for e-commerce based on knowledge graphs according to claim 1, characterized in that, In step S2, generating an updated product knowledge graph that matches the current advertising behavior includes the following steps: S2.1 Based on the continuous advertising interaction behavior sequence and corresponding product nodes, each product node is used as the starting node. The association path is retrieved in the product knowledge graph along the product category relationship and brand relationship. All association paths from the starting node to the associated product node are obtained, and association paths with cyclic access are removed to generate a set of valid association paths. S2.2 For each effective association path, form a product node sequence according to the connection order of product nodes in the path, analyze the continuous change of the association relationship between adjacent product nodes, count the path length where the association relationship is maintained continuously and the position where the association relationship changes, introduce the position weighting factor, and calculate the path continuity index corresponding to each effective association path. S2.3 Construct an association path constraint field in the product knowledge graph based on the path continuity index and change location distribution characteristics of each effective association path; S2.

4. Analyze the propagation state of each effective associated path based on the associated path constraint field, identify the unstable nodes of associated propagation, and use the unstable nodes of associated propagation as the termination positions to extract the corresponding effective associated paths and generate the corrected effective associated paths. S2.

5. Based on the corrected effective association paths, calculate the association propagation between each product node, and dynamically update the association strength between each product node in the product knowledge graph according to the association propagation, generating an updated product knowledge graph that matches the current advertising behavior.

4. The knowledge graph-based intelligent product recommendation method for e-commerce platforms according to claim 3, characterized in that, In step S2.3, the association path constraint field is constructed in the product knowledge graph based on the path continuity index and change position distribution characteristics of each effective association path, including the following steps: S2.

31. Read all valid association paths and their corresponding path continuity indices, assign the path continuity index of each valid association path to all product nodes that constitute the valid association path, and count all path continuity indices corresponding to each product node to form a set of path continuity indices corresponding to each product node. S2.

32. Based on the set of path continuity indices corresponding to each product node, calculate the node continuity index corresponding to each product node, and write the node continuity index corresponding to each product node into the corresponding product node in the product knowledge graph. S2.

33. Read the distribution characteristics of the change positions corresponding to each effective association path, count the frequency of each commodity node being marked as a change position in its effective association path and the relative advance degree of the change position, and calculate the change impact coefficient corresponding to each commodity node. S2.

34. Using the node continuity index corresponding to each product node as the node status and the change impact coefficient as the node change sensitivity index, the node is continuously expanded according to the relationship between each product node in the product knowledge graph, forming a node continuity region covering the entire product knowledge graph. S2.

35. Based on the continuous index of each node, the continuous regions of each node are marked with state, generating stable propagation regions and propagation offset regions, and the continuous regions of all nodes and their corresponding state marks are combined to form an associated path constraint field.

5. The intelligent recommendation method for e-commerce products based on knowledge graphs according to claim 3, characterized in that, In step S2.4, the propagation state of each effective associated path is analyzed based on the associated path constraint field, and the nodes where the associated propagation becomes unstable are identified. This includes the following steps: S2.

41. Read the associated path constraint field and valid associated paths, map each valid associated path to the associated path constraint field in the order of the connection of the product nodes, and obtain the state flag corresponding to the continuous area of ​​the node where each product node is located, and generate the propagation state sequence corresponding to each valid associated path. S2.

42. Along the propagation direction of each effective associated path, perform sequential analysis on the propagation state sequence, determine whether the state markers corresponding to adjacent product nodes have changed, and identify propagation state transition nodes; S2.

43. Starting from the propagation state transition node, continue to analyze the state markers of each product node after the current valid association path, and identify the association propagation instability nodes.

6. The intelligent product recommendation method for e-commerce based on knowledge graphs according to claim 1, characterized in that, In step S3, generating a candidate recommended product set includes the following steps: S3.1 Based on the updated product knowledge graph, take the product node corresponding to the current advertising interaction behavior as the starting product node, perform association propagation analysis along the relationship between product nodes in the updated product knowledge graph, count all related product nodes, and extract the association strength and association propagation path between the starting product node and each related product node to generate a product node association propagation dataset. S3.2 Based on the product node association propagation dataset, statistically analyze the association propagation characteristics of each product node in the association propagation process, and calculate the initial node association activity level of each product node based on the association propagation characteristics. S3.3 Based on the changing characteristics of the association propagation coverage of each product node, identify the association propagation convergence domain of the product node, calculate the propagation convergence index of each product node, and correct the initial node association activity level based on the propagation convergence index to generate the corrected product node association activity level. S3.

4. Sort the products based on the activity level of the corrected product nodes and generate a set of candidate recommended products.

7. The intelligent recommendation method for e-commerce products based on knowledge graphs according to claim 6, characterized in that, In step S3.3, generating the corrected activity level of the product node association includes the following steps: S3.

31. Read the initial node association activity level and the association propagation coverage objects corresponding to each product node. Taking each product node as the center, count the overlap between association propagation coverage objects and generate the association propagation convergence region. S3.

32. For each associated propagation convergence domain, statistically analyze the changes in the number of newly propagated objects covered according to the order of associated propagation, and calculate the propagation convergence index of the corresponding associated propagation convergence domain. S3.

33. Based on the propagation convergence index corresponding to each product node, the initial node association activity level is corrected to generate the corrected product node association activity level.

8. The intelligent recommendation method for e-commerce products based on knowledge graphs according to claim 7, characterized in that, In step S3.32, the changes in the number of newly covered objects are statistically analyzed according to the order of associated propagation, and the propagation convergence index of the corresponding associated propagation convergence domain is calculated, including the following steps: For each associated propagation convergence domain, the propagation coverage objects corresponding to each propagation stage are read sequentially according to the propagation order. The number of newly added propagation coverage objects in the current propagation stage compared to the previous propagation stage is counted to form a sequence of newly added propagation coverage objects. The changing trend of newly added propagation coverage objects is analyzed based on the sequence of newly added propagation coverage objects, and the propagation convergence index of the corresponding associated propagation convergence domain is calculated.

9. The intelligent recommendation method for e-commerce products based on knowledge graphs according to claim 1, characterized in that, In step S4, generating the product recommendation sequence includes the following steps: S4.1 Read the candidate recommended product set and collect the real-time internet advertising delivery status data corresponding to the candidate recommended products to generate a candidate product advertising status dataset; S4.2 Based on the candidate product advertising status dataset, analyze the advertising availability status and advertising display activity of each candidate recommended product, and calculate the advertising suitability of each candidate recommended product. S4.

3. Combining the activity level of the corrected product node association and the degree of ad adaptation of each candidate recommended product, a comprehensive adaptation analysis is performed on the candidate recommended products to generate a comprehensive recommendation priority for each candidate recommended product. S4.4 Sort the candidate recommended products according to their comprehensive recommendation priority to generate a product recommendation sequence.

10. A knowledge graph-based intelligent product recommendation system for e-commerce platforms, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: The processor executes a computer program to implement the knowledge graph-based intelligent product recommendation method for e-commerce as described in any one of claims 1-9.

Images