Automobile parts multi-view recommendation method and system based on user interest drift identification

Abstract

The application discloses a multi-view recommendation method and system for automobile parts based on user interest drift recognition, and belongs to the technical field of intelligent recommendation, comprising: obtaining user purchase behavior data and constructing a user purchase behavior sequence; based on the user purchase behavior sequence, calculating the degree of deviation between the current purchase window representation and the historical stable preference representation to identify candidate drift events of user purchase; judging the interest drift evolution state of the candidate drift events and generating a structured drift state vector; constructing a multi-view state space; fusing each view state in the multi-view state space to obtain a joint state; under the joint state, evaluating the comprehensive recommendation value of each candidate part in the candidate part action space to generate a comprehensive value score of each candidate part; and sorting the candidate parts according to the comprehensive value score to generate an automobile part recommendation result; the application can improve the accuracy of supply and demand matching, long-term benefits and adaptability to complex business constraints in the automobile part purchase scene.

Description

Technical Field

[0001] This invention relates to the field of intelligent recommendation technology, and in particular to a multi-perspective recommendation method and system for automotive parts based on user interest drift identification. Background Technology

[0002] In the context of a collaborative digital ecosystem across the industry chain, auto parts procurement, as a crucial link connecting the supply and demand sides, is characterized by strong task-driven factors, strong business constraints, and highly dynamic changes. Auto parts procurement decisions differ from those for conventional consumer goods recommendations; they are influenced by a complex interplay of multiple business factors, including vehicle compatibility, the progress of repair tasks, inventory fluctuations, procurement cycles, and urgent replenishment needs.

[0003] In real-world procurement scenarios, user behavior sequences typically exhibit high non-stationarity. Traditional recommendation methods are usually modeled based on static user profiles or assumptions of smooth interest evolution. This approach struggles to accurately distinguish between random local disturbances, periodic deviations, or long-term goal shifts embedded in the behavior sequences. It is highly susceptible to misjudging temporary replenishment behavior as long-term stable demand, or failing to capture the true shift in procurement focus in a timely manner. Consequently, the supply-demand match between the recommended parts and the company's current actual procurement needs is low, resulting in poor circulation efficiency.

[0004] The current status and shortcomings of existing technologies are as follows: The definition of drift triggers is ambiguous: In the procurement of auto parts, the occurrence of unusual purchases within a certain time window does not necessarily mean that the actual procurement orientation has shifted. It may just be a local anomaly caused by emergency repairs, temporary replenishment, or one-off projects. Methods used in e-commerce scenarios often lack specific judgment on the boundary between normal fluctuations and true drift trigger events, making them susceptible to the amplification of short-term shock signals. This can lead to misjudging events that are actually event-driven disturbances as sustained changes in interest.

[0005] Lack of Evolutionary Stage Differentiation: In the automotive parts procurement scenario, the key to behavioral shifts lies not only in whether a change has occurred, but also in what evolutionary stage it is in. A single short-term emergency purchase may show a significant shift, but it does not necessarily mean that long-term preferences have shifted; conversely, some relatively mild, continuous shifts may be early signals of a shift from short-term demand to long-term procurement tendencies. If only a single drift score or coarse-grained level is output, it is difficult to effectively distinguish between short-term disturbances, transitional evolution, and long-term migration, thus leading to distorted state assessments.

[0006] Insufficient expression of detection results: Existing methods often only provide a drift score or a single label in their output format. While this can achieve a preliminary judgment, it is difficult to retain hierarchical information during the drift process. For the automotive parts procurement scenario, an effective drift analysis needs to characterize not only the degree of behavioral deviation but also reflect its evolutionary stage and business semantics. If the result expression is too compressed, it is neither conducive to state interpretation nor to forming a reusable structured description.

[0007] Recommendation decision-making suffers from a single perspective: Traditional collaborative filtering, sequence recommendation, or general multi-perspective recommendation methods typically use user history interactions, item similarity, or content features as primary inputs, making it difficult to simultaneously integrate the user's current drift state, user-part collaboration relationships, and part business attributes. In the automotive parts procurement process, recommendation decisions not only need to determine whether a user currently has a procurement need, but also need to consider the relationships between parts, vehicle model compatibility features, business attribute constraints, and supply fulfillment conditions.

[0008] Distortion caused by the superposition of local rewards: Automotive parts recommendation involves multiple feedback signals, such as matching user needs, the accessibility of collaborative relationships, and the feasibility of parts business. These signals are not independent but rather mutually influential and have a stage-based transmission relationship. If traditional methods are still used to simply accumulate local rewards from different perspectives, it will not only fail to reveal the structural dependence between benefits but also easily accumulate essentially similar effective signals repeatedly. This will amplify local preferences, introduce benefit biases, and weaken the stability of the recommendation strategy in complex business environments.

[0009] Therefore, it is necessary to propose a multi-perspective recommendation method and system for automotive parts based on user interest drift identification, in order to solve the problems of inaccurate identification of interest drift behavior, difficulty in reusing drift results, insufficient expression of recommendation state, and distortion of multi-perspective benefit aggregation under complex purchasing behavior. Summary of the Invention

[0010] To address the aforementioned issues, this invention provides a multi-perspective recommendation method and system for automotive parts based on user interest drift identification. It aims to solve technical problems in existing technologies such as ambiguous drift trigger definitions, lack of evolutionary stage discrimination, difficulty in reusing drift identification results, a single perspective in recommendation decisions, and distortion due to local reward superposition. This method can improve the accuracy of supply and demand matching in automotive parts procurement scenarios, enhance long-term benefits, and improve adaptability to complex business constraints.

[0011] This invention provides a multi-view recommendation method for automotive parts based on user interest drift identification, including: Acquire user purchasing behavior data and construct user purchasing behavior sequences; Based on the user's purchasing behavior sequence, the degree of offset between the current purchasing window representation and the historical stable preference representation is calculated to identify candidate drift events in user purchasing. Determine the interest drift evolution state of candidate drift events and generate structured drift state vectors; A multi-perspective state space is constructed, comprising: a user drift demand perspective state determined based on the structured drift state vector; a collaborative relationship perspective state determined based on the historical interaction relationship between the user and the accessory; and an accessory business perspective state determined based on the accessory business attributes. The perspective states in the multi-perspective state space are then fused to obtain a joint state. In the combined state, the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories is evaluated, and a comprehensive value score for each candidate accessory is generated. Candidate accessories are ranked based on the comprehensive value score to generate recommended automotive accessories.

[0012] In some embodiments, calculating the degree of offset between the current purchase window representation and the historical stable preference representation based on the user's purchase behavior sequence to identify candidate drift events in user purchases includes: The user's purchasing behavior sequence is divided into multiple purchasing windows according to a preset time window. The weighted average of the feature vector of each accessory and the purchasing quantity in the current purchasing window is calculated to obtain the representation of the current purchasing window. Select several historical windows before the current window, calculate the mean value represented by each historical window, and obtain the historical stable preference representation; Calculate the cosine similarity between the current procurement window representation and the historical stable preference representation, and calculate the semantic offset based on the cosine similarity; Calculate the changes in accessory categories and accessory business attributes between the current procurement window and the historical window; The semantic offset, the change in accessory category, and the change in accessory business attribute are assigned weight coefficients and then weighted and summed to obtain the basic drift intensity. When the basic drift intensity exceeds a preset threshold, the corresponding continuous time window is determined as a candidate drift event for user purchase.

[0013] In some embodiments, determining the interest drift evolution state of candidate drift events and generating a structured drift state vector includes: Extract the persistence, repetition consistency, backtracking tendency, and event-level drift intensity features of the candidate drift events, input the features into the state discrimination network, and output the evolution state confidence distribution of the candidate drift events as instantaneous perturbation type, transitional evolution type, and stable migration type; Obtain the business attribute features of the accessories corresponding to the candidate drift events, perform weighted mapping on the business attribute features, and generate business sensitivity weights; A comprehensive drift score is generated based on the basic drift intensity, the evolution state confidence distribution, and the business sensitivity weight. The structured drift state vector is constructed based on the basic drift intensity, the evolution state confidence distribution, the business sensitivity weight, and the comprehensive drift score.

[0014] In some embodiments, constructing the multi-view state space includes: The user base representation is concatenated with the structured drift state vector, and the concatenated vector is input into a mapping function. After mapping by the mapping function, the user drift demand perspective state is obtained. The historical interaction relationship between the user and the accessory is constructed into a user-accessory collaboration graph. The user-accessory collaboration graph is input into a first graph neural network. The first graph neural network performs neighborhood propagation and high-order relationship aggregation. Then, the output of the first graph neural network is pooled to obtain the perspective state of the collaboration relationship. The business attribute relationships between parts are constructed into a parts relationship graph. The parts relationship graph is input into a second graph neural network. The second graph neural network performs relationship propagation and neighborhood aggregation. Then, the output of the second graph neural network is pooled to obtain the business perspective state of the parts.

[0015] In some embodiments, including: User nodes and accessory nodes are treated as two types of graph nodes. Interaction edges are established between user nodes and accessory nodes based on the interaction behavior between users and accessories. The weight of the interaction edges is determined based on factors such as purchase quantity, interaction frequency, feedback behavior intensity, and time decay, resulting in a user-accessory collaboration graph.

[0016] In some embodiments, including: By treating the parts nodes as graph nodes, relationship edges are established between the parts nodes based on the relationship between parts categories, vehicle models, substitutes, brand manufacturers, or similar business attributes. The weights of the relationship edges are determined based on the relationship between business rules and attribute similarity, thus obtaining the parts relationship graph.

[0017] In some embodiments, the step of evaluating the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories and generating a comprehensive value score for each candidate accessory in the joint state includes: Based on the aforementioned joint state, obtain the candidate component action space, and select candidate component actions from the candidate component action space; For the candidate accessory actions, a multi-perspective real-time reward vector is constructed, which includes user demand matching reward, collaborative relationship reward and accessory business reward; By performing relation enhancement and temporal accumulation on the multi-view real-time reward vector, the aggregated revenue representation corresponding to each candidate accessory action is obtained; The joint state, the candidate accessory action vector, and the aggregated benefit representation are input into the value assessment network to output the comprehensive value score of the candidate accessory.

[0018] In some embodiments, the step of performing relation enhancement and temporal accumulation on the multi-view real-time reward vector to obtain the aggregated reward representation corresponding to each candidate accessory action includes: By using a reward propagation graph neural network, message propagation and node updates are performed on the multi-view instantaneous reward vector based on the dependency matrix between reward perspectives to obtain an enhanced reward representation. The enhanced reward representation is input into the gated loop unit for time-series state accumulation to obtain the aggregated reward representation.

[0019] In some embodiments, the step of ranking candidate accessories based on the comprehensive value score and generating automotive accessory recommendation results includes: Candidate accessories are sorted in descending order of their comprehensive value scores; Select a predetermined number of candidate accessories that are ranked at the top to generate a set of recommended results; Output the accessory name, accessory category, brand manufacturer, compatibility information, recommendation order, comprehensive score, and reason for recommendation for each candidate accessory in the recommended result set.

[0020] This invention provides a multi-view recommendation system for automotive parts based on user interest drift identification, including: The sequence construction module is used to acquire user purchasing behavior data and construct user purchasing behavior sequences; The identification module is used to calculate the degree of offset between the current purchase window representation and the historical stable preference representation based on the user's purchase behavior sequence, so as to identify candidate drift events in user purchases; The evolution state recognition module is used to determine the interest drift evolution state of candidate drift events and generate structured drift state vectors; A state space construction module is used to construct a multi-view state space, which includes: a user drift demand perspective state determined based on the structured drift state vector, a collaborative relationship perspective state determined based on the historical interaction relationship between users and accessories, and an accessory business perspective state determined based on accessory business attributes; the joint state is obtained by fusing the perspective states in the multi-view state space. The comprehensive scoring module is used to evaluate the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories in the joint state and generate a comprehensive value score for each candidate accessory. The recommendation result generation module is used to sort the candidate accessories according to the comprehensive value score and generate automotive accessory recommendation results.

[0021] Compared with the prior art, the present invention has the following beneficial effects: By identifying interest drift and determining the evolutionary state of user purchasing behavior sequences, a structured drift state vector is generated. Furthermore, a multi-perspective state space is constructed by integrating user drift needs, collaborative relationships, and accessory business attributes. Based on this, a comprehensive value assessment and ranking recommendation of candidate accessories is performed. This can improve the accuracy of supply and demand matching, long-term benefits, and adaptability to complex business constraints in the automotive parts procurement scenario. Attached Figure Description

[0022] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the implementation process of the multi-view recommendation method for automotive parts based on user interest drift identification provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the user interest drift identification model of the present invention; Figure 3 This is a schematic diagram of the multi-view recommendation model of the present invention; Figure 4 This is a schematic diagram of the internal structure of the GNN-GRU module of the present invention; Figure 5 This is a schematic diagram of the structure of a multi-view recommendation system for automotive parts based on user interest drift identification, provided in an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0025] If the application documents contain similar descriptions such as "first, second, third", the following explanation shall be added: In the following description, the terms "first, second, third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first, second, third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of the present invention described herein can be implemented in an order other than that illustrated or described herein.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein is for the purpose of describing embodiments of the invention only and is not intended to limit the invention.

[0027] To address the problems existing in related technologies, this invention provides a multi-perspective recommendation method for automotive parts based on user interest drift identification. The execution entity of this recommendation method can be a computer device. The computer device can be a terminal such as a personal computer, laptop, tablet, in-vehicle computing device, overhead contact line status monitoring terminal, or edge computing device, or it can be implemented as a server. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0028] In some embodiments, the functions implemented by the recommended method provided in this invention can be achieved by the processor of a computer device calling program code, wherein the program code can be stored in a computer storage medium.

[0029] This invention provides a multi-view recommendation method for automotive parts based on user interest drift identification. Figure 1 This is a schematic diagram illustrating the implementation process of the multi-view recommendation method for automotive parts based on user interest drift identification provided in an embodiment of the present invention, as shown below. Figure 1 As shown, it includes: Step S1: Obtain user purchasing behavior data and construct a user purchasing behavior sequence; In this embodiment of the invention, user purchasing behavior data from an auto parts procurement platform is obtained. This purchasing behavior data includes user identifiers, parts identifiers, purchasing time, purchasing quantity, parts category, parts business attributes, user historical purchasing records, and user-parts interaction relationship data. The purchasing behavior data is cleaned, deduplicated, sorted by time, and vectorized, and a user purchasing behavior sequence is constructed according to a preset time window. For users... The procurement behavior sequence is represented as follows: ; in, Indicates user The sequence of procurement activities; Indicates user In the The corresponding auto parts in this procurement transaction; Indicates the corresponding purchase quantity; Indicates the corresponding procurement time; Indicates accessories The feature vector includes at least one of the following: accessory category, compatible vehicle model, brand manufacturer, procurement cycle, turnover attribute, repurchase attribute, or supply stability.

[0030] Step S2: Based on the user's purchasing behavior sequence, calculate the degree of offset between the current purchasing window representation and the historical stable preference representation to identify candidate drift events in user purchasing; In some embodiments, step S2 includes: Step S21: Divide the user's purchasing behavior sequence into multiple purchasing windows according to a preset time window, calculate the weighted average of the feature vector of each accessory and the purchasing quantity in the current purchasing window, and obtain the representation of the current purchasing window; Step S22: Select several historical windows before the current window, calculate the mean value represented by each historical window, and obtain the historical stable preference representation; Step S23: Calculate the cosine similarity between the current procurement window representation and the historical stable preference representation, and calculate the semantic offset based on the cosine similarity; Step S24: Calculate the change in accessory category and the change in accessory business attributes between the current procurement window and the historical window; Step S25: Assign weight coefficients to the semantic offset, the change in accessory category, and the change in accessory business attributes, and perform a weighted summation to obtain the basic drift intensity; Step S26: When the basic drift intensity exceeds a preset threshold, the corresponding continuous time window is determined as a candidate drift event purchased by the user.

[0031] In embodiments of the present invention, such as Figure 2 As shown, the user's purchasing behavior sequence is divided into multiple purchasing windows according to a preset time granularity. Let the user... In the The set of purchasing behaviors in each purchasing window is The current procurement window can be defined as: ; in, Indicates user In the The current procurement focus is displayed in each window; This window indicates the accessories that were purchased within this window; Indicates accessories eigenvectors; Indicates user For accessories The formula represents the quantity of goods purchased. It is a weighted average of the accessory feature vectors within the window and their corresponding purchase quantities, used to indicate the concentration direction of user purchasing demand within the current time window.

[0032] To obtain a comparable historical baseline, select the period before the current window. A historical window is used to construct a stable representation of user historical preferences: ; in, Indicates user In the The historical stable preference representation formed before each window; Indicates the number of historical windows; Indicates the first The procurement focus is represented by a historical window. This formula is an aggregation of window-level representations of multiple historical windows, used as a stable reference baseline for subsequent drift detection.

[0033] In obtaining historically stable preference representations And the current procurement window indicates Then, the degree of deviation of the current purchasing window relative to historical stable preferences is calculated to identify interest drift events in user purchasing behavior, and the basic drift strength is obtained, as shown below: ; in, Indicates user In the The base drift strength in each window; The cosine similarity between the historical stable preference representation and the current procurement window representation; This indicates the degree of semantic shift in current purchasing demand relative to historical stable preferences; This indicates the change in accessory categories between the current procurement window and historical windows; This indicates the change in the attributes of parts and components between the current procurement window and the historical windows; , which is a weighting coefficient used to adjust the contribution ratio of semantic shift, category change, and business attribute change to the drift intensity.

[0034] Candidate drift events are filtered based on the baseline drift intensity and a preset drift threshold: ; in, Represents the set of candidate drift event windows; Indicates the first One candidate drift event; Indicates the drift trigger threshold; This represents the range of the continuous time window corresponding to the candidate drift event. This formula is used to combine windows that meet the significant shift condition into candidate drift events, providing input for subsequent evolutionary state determination.

[0035] Step S3: Determine the interest drift evolution state of candidate drift events and generate a structured drift state vector; In some embodiments, step S3 includes: Step S31: Extract the persistence, repetition consistency, backtracking tendency and event-level drift intensity features of the candidate drift events, input the features into the state discrimination network, and output the evolution state confidence distribution of the candidate drift events as instantaneous perturbation type, transitional evolution type and stable migration type; Step S32: Obtain the business attribute features of the accessories corresponding to the candidate drift events, perform weighted mapping on the business attribute features, and generate business sensitivity weights; Step S33: Generate a comprehensive drift score based on the basic drift intensity, the evolution state confidence distribution, and the business sensitivity weight; Step S34: Construct the structured drift state vector based on the basic drift intensity, the evolution state confidence distribution, the business sensitivity weight, and the comprehensive drift score.

[0036] In this embodiment of the invention, for candidate drift events Further analysis is conducted to extract features such as persistence, repeatability, regression tendency, and event-level drift intensity. Based on these features, the candidate drift event is determined to be either a transient disturbance, a transitional evolution, or a stable migration. Candidate Drift Events The confidence distribution of the evolutionary state is represented as: ; in, Indicates candidate drift events The confidence distribution of the evolutionary state; Indicates the confidence level of a transient perturbation state; Indicates the confidence level of the transitional evolutionary state; The confidence level represents a stable transition state; Indicates candidate drift events The event feature vector consists of event-level drift intensity, persistence feature, repetitive consistency feature, and return tendency feature; and These represent the learnable weight matrix and bias term in the state discrimination process, respectively; It is used to normalize the state discrimination results into an evolution state confidence distribution.

[0037] Furthermore, business sensitivity weights are constructed based on the attributes of the parts business: ; in, Indicates candidate drift events The corresponding business sensitivity weight; This indicates the repurchase characteristics of related accessories; This indicates the turnover characteristics of the relevant parts; Indicates the urgency of the relevant procurement needs; This indicates the stability of the supply of related components; These are the weight parameters corresponding to each business attribute; For bias terms; This represents the business sensitivity weight mapping function, which maps repurchase characteristics, turnover characteristics, urgency characteristics, and supply stability characteristics into business weight values ​​that can participate in drift correction.

[0038] After obtaining the basic drift intensity, evolution state confidence distribution, and business sensitivity weights, a comprehensive drift score is generated: ; in, This represents the overall drift score after business calibration; Indicates candidate drift events The event-level base drift intensity; Indicates the weight of business sensitivity; , , These represent the state adjustment coefficients corresponding to the instantaneous disturbance type, the transitional evolution type, and the stable migration type, respectively; , , These represent the confidence levels of the three evolutionary states. The formula, composed of behavioral shift strength, business importance, and state evolution weights, characterizes the overall impact of the current drift event on procurement recommendation decisions. Among these, the event-level basic drift strength... It can be determined based on the maximum value, average value, etc. of the basic drift intensity.

[0039] Finally, a structured drift state vector is generated: .

[0040] Step S4: Construct a multi-view state space, which includes: user drift demand perspective state determined based on the structured drift state vector, collaborative relationship perspective state determined based on the historical interaction relationship between users and accessories, and accessory business perspective state determined based on accessory business attributes; fuse the perspective states in the multi-view state space to obtain a joint state. In some embodiments, constructing the multi-view state space includes: The user base representation is concatenated with the structured drift state vector, and the concatenated vector is input into a mapping function. After mapping by the mapping function, the user drift demand perspective state is obtained. The historical interaction relationship between the user and the accessory is constructed into a user-accessory collaboration graph. The user-accessory collaboration graph is input into a first graph neural network. The first graph neural network performs neighborhood propagation and high-order relationship aggregation. Then, the output of the first graph neural network is pooled to obtain the perspective state of the collaboration relationship. The business attribute relationships between parts are constructed into a parts relationship graph. The parts relationship graph is input into a second graph neural network. The second graph neural network performs relationship propagation and neighborhood aggregation. Then, the output of the second graph neural network is pooled to obtain the business perspective state of the parts.

[0041] In embodiments of the present invention, such as Figure 3 As shown, this multi-perspective state space consists of the user drift demand perspective, the collaborative relationship perspective, and the accessory business perspective.

[0042] From the perspective of user drift demand: The generated structured drift state vector As part of the dynamic user demand state, this perspective is composed of a user-based representation and a structured drift state vector. The user-based representation preserves long-term user purchasing habits, historical purchasing categories, and stable demand characteristics, while the structured drift state vector injects the drift intensity, evolutionary state confidence, business sensitivity weight, and overall drift score within the current window. The user demand perspective is defined as follows: ; in, This indicates the user's perspective on drifting needs; Represents the user base; Represents the structured drift state vector; This represents a vector concatenation operation; This represents the user-perspective state mapping function, used to fuse the user's basic representation with the drift state prior.

[0043] From the perspective of collaborative relationships: The collaborative relationship perspective describes the degree of association between target users and candidate parts in the historical interaction network. Based on user-part interaction data, this perspective treats purchasing users and auto parts as two types of graph nodes, and combines the user-item bipartite graph construction concept from graph collaborative filtering. Relationship edges are established based on user-part interactions such as purchasing, browsing, adoption, adding to purchase orders, or other interactive behaviors, thus forming a user-part collaborative graph. ; in, Represents a set of user nodes. Represents a set of component nodes. This represents the set of interaction edges between the user and the accessories. This represents the adjacency matrix of the user-part collaboration graph. For any user node... and parts nodes If there are interactive behaviors between the two, such as purchasing, browsing, adopting, adding to a purchase order, or providing feedback, then at the user node... With parts nodes Establish interaction edges between The interaction edge weight between users and accessories can be determined by the purchase quantity, interaction frequency, feedback behavior intensity, and time decay factor, and is expressed as: ; in, Representing users from the perspective of collaborative relationships Accessories Interaction edge weights between them; Indicates user For accessories The cumulative purchase quantity; Indicates user Accessories Frequency of interaction between them; Indicates the intensity of user feedback behavior; This indicates the time interval between the most recent interaction and the current recommendation time. This represents the weighting coefficients corresponding to different interaction factors; This represents the time decay coefficient. The user-accessory interaction matrix is ​​composed of the aforementioned interaction edge weights. And construct the following bipartite graph adjacency matrix: ; in, This represents the user-accessory interaction matrix. Denotes the transpose of the interaction matrix. This represents a zero matrix indicating that there are no direct interaction edges between user nodes or accessory nodes. The initial representations of user nodes and accessory nodes are denoted as follows: and Based on the user-part collaboration graph and its adjacency matrix By using a graph neural network for neighborhood propagation and high-order relation aggregation, the collaborative relationship perspective state is obtained: ; in, Indicates the first The collaborative relationship perspective under each decision-making step; The initial representation matrix of user nodes from the perspective of collaborative relationships; This represents the initial representation matrix of component nodes from the perspective of collaborative relationships; This represents a graph neural network used for propagating cooperative relationships. This represents a pooling aggregation operation, used to aggregate the node-level or neighborhood-level representations output by a graph neural network into a collaborative structural state in the current recommendation context.

[0044] Parts business perspective: The parts business perspective describes the attribute constraints and supply-demand matching characteristics of candidate parts in the automotive parts procurement process. Based on standardized parts business data, this perspective treats automotive parts as graph nodes and combines item-item relationship graphs and k-nearest neighbor graphs based on attribute similarity. According to factors such as part category, vehicle model compatibility, substitution relationships, brand / manufacturer relationships, similarity in procurement cycles, similarity in turnover attributes, similarity in repeat purchase attributes, and similarity in supply stability, relationship edges are established between different parts nodes, thus forming the parts relationship graph. ; in, Represents a set of component nodes. This represents the set of edges representing the relationships between component nodes. This represents the adjacency matrix of the component relationship graph. For any two component nodes... and If the two parts meet the preset association conditions in terms of part category, compatible vehicle model, substitution relationship, brand manufacturer or business attribute similarity, then a relationship edge is established between the two parts. The edge weights between component nodes can be determined by both business rule relationships and attribute similarity, expressed as: ; in, Indicates accessories Accessories The weights of the edges in relation to each other; Function indicating the relationship between accessory categories; Indicates a function that indicates vehicle model compatibility. Indicates a substitution relationship indicator function; This indicates that cosine similarity is used to calculate the similarity between feature vectors of parts and components businesses; This represents the weight coefficients corresponding to different relationship factors. The adjacency matrix of the component relationship graph is composed of the edge weights mentioned above. and the parts business feature matrix Mapped to an initial node representation matrix from the perspective of parts and components business: ; in, This represents a feature matrix for the parts business, consisting of characteristics such as parts category, vehicle model compatibility, brand manufacturer, procurement cycle, turnover attribute, repurchase attribute, and supply stability. This represents the feature mapping function for parts and accessories business. This represents the initial representation matrix of component nodes from the perspective of component business. Based on the component relationship graph. Initial representation matrix of component nodes By using graph neural networks for relationship propagation and neighborhood aggregation, the state from the perspective of parts business is obtained: ; in, Indicates the first The status of the parts business from the perspective of each decision-making step; A graph neural network representing the propagation of business relationships in parts and accessories; This represents a pooling aggregation operation used to form a representation of the accessory business status in the current recommendation context.

[0045] After obtaining the above three perspective states, a unified recommendation state is constructed: ; in, Indicates the recommendation model in the th The joint state of each decision step; , , These represent the user's drift demand perspective, the collaborative relationship perspective, and the accessory business perspective, respectively. and These represent the learnable weight matrix and the bias term, respectively. This represents the multi-perspective state fusion mapping function. This formula is used to compress information from the demand side, relationship side, and business side into the same recommendation decision state space.

[0046] Step S5: In the joint state, evaluate the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories, and generate a comprehensive value score for each candidate accessory; In some embodiments, step S5 includes: Step S51: Based on the joint state, obtain the candidate component action space, and select candidate component actions from the candidate component action space; Step S52: For the candidate accessory actions, construct a multi-perspective real-time reward vector, which includes user demand matching reward, collaborative relationship reward and accessory business reward; Step S53: Perform relation enhancement and temporal accumulation on the multi-view real-time reward vector to obtain the aggregated revenue representation corresponding to each candidate accessory action; Step S54: Input the joint state, the candidate accessory action vector, and the aggregated benefit representation into the value evaluation network, and output the comprehensive value score of the candidate accessory.

[0047] In some embodiments, step S53 includes: Step S531: Using a reward propagation graph neural network, message propagation and node updates are performed on the multi-view instantaneous reward vector based on the dependency matrix between reward perspectives to obtain an enhanced reward representation; Step S532: Input the enhanced reward representation into the gated loop unit for time-series state accumulation to obtain the aggregated revenue representation.

[0048] In embodiments of the present invention, such as Figure 4 As shown, in the joint state Next, actions regarding candidate parts. Construct a multi-view instantaneous reward vector. Among them, Indicates the first The action space consisting of a set of candidate components for each decision step. This represents the candidate accessory action selected from the action space. The multi-view instant reward vector includes user demand matching reward, collaborative structure reward, and accessory business reward, which are used to characterize the degree of matching between the recommended accessory and the user's current drifting demand, the reachability of the recommended action in the user-accessory relationship structure, and the feasibility of the recommended accessory in terms of business attributes, respectively.

[0049] The multi-view instant reward vector is represented as follows: ; in, Indicates the first Heterogeneous local reward vectors for each decision step; This indicates that rewards are matched to user needs. Indicates rewards for collaborative relationships; This indicates a reward for the parts and accessories business.

[0050] To avoid value distortion caused by directly linearly superimposing local rewards, a reward propagation graph neural network is introduced to enhance the relationship between multi-view rewards: ; in, This represents the enhanced reward representation after propagation through relationships; A dependency matrix representing reward perspectives, used to describe the associations, conflicts, or compensation relationships between different reward perspectives; Represents the original multi-view instantaneous reward vector; This represents the reward propagation weight matrix; Indicates the bias term; This represents the reward propagation calibration function, used for nonlinear mapping and structured calibration of the relationship propagation results of multi-source local rewards. The formula comprises the reward dependency structure, the original local rewards, and learnable propagation parameters, used to achieve structured calibration between multi-source rewards.

[0051] Subsequently, the enhanced reward representation is input into the gated loop unit for long-term benefit accumulation: ; in, Indicates the first The hidden state of each decision step; This represents an enhanced reward representation for the current decision step; This indicates the hidden state of the previous decision step; This represents a gated loop unit, used to aggregate historical reward information and current reward information over time.

[0052] Furthermore, the hidden state is mapped to an aggregated revenue representation of candidate accessories: ; in, This represents the aggregated revenue; Represents the linear mapping weight matrix; Indicates the bias term; This represents the hidden state output by the gated recurrent unit. This formula is used to transform high-dimensional temporal hidden states into revenue features that can be computed using the value function.

[0053] The overall value score of the final candidate accessory actions is expressed as follows: ; in, Indicates a state of union Select candidate accessories below The overall value score; Indicates the first The action of selecting candidate parts in each decision-making step; Represents the action vector of the candidate accessory; This represents the aggregated revenue; This represents the mapping function of a multilayer perceptron.

[0054] Step S6: Sort the candidate parts according to the comprehensive value score and generate the recommended results of the car parts.

[0055] In some embodiments, step S6 includes: Step S61: Sort the candidate accessories in descending order of their comprehensive value scores; Step S62: Select a preset number of candidate accessories that are ranked at the top and generate a set of recommended results; Step S63: Output the accessory name, accessory category, brand manufacturer, compatibility information, recommendation order, comprehensive score and recommendation reason for each candidate accessory in the recommendation result set.

[0056] In this embodiment of the invention, candidate parts are sorted according to their comprehensive recommendation value, and a number of parts with the highest scores are selected to form a set of recommended automotive parts results. The output includes the name of the recommended part, part category, brand manufacturer, compatibility information, recommendation order, comprehensive score, and explanation of the reasons for recommendation.

[0057] In some embodiments, the method further includes: recording user feedback behavior on recommendation results and updating the user's purchasing behavior sequence as new data to trigger interest drift behavior identification and recommendation strategy optimization in the next cycle.

[0058] In this embodiment of the invention, user feedback behaviors such as browsing, clicking, adding to purchase order, purchasing, ignoring or returning the recommendation results are recorded and written into the user behavior record; the feedback data is updated to the user purchase behavior sequence and recommendation result record for subsequent interest drift behavior identification, status update and recommendation strategy optimization.

[0059] Based on the foregoing embodiments, this invention provides a multi-view recommendation system for automotive parts based on user interest drift identification. The various modules and units included in the system can be implemented by a processor in a computer device; of course, they can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0060] This invention provides a multi-view recommendation system for automotive parts based on user interest drift identification. Figure 5This is a schematic diagram of the structure of a multi-view recommendation system for automotive parts based on user interest drift identification provided in an embodiment of the present invention, as shown below. Figure 5 As shown, it includes: The sequence construction module is used to acquire user purchasing behavior data and construct user purchasing behavior sequences; The identification module is used to calculate the degree of offset between the current purchase window representation and the historical stable preference representation based on the user's purchase behavior sequence, so as to identify candidate drift events in user purchases; The evolution state recognition module is used to determine the interest drift evolution state of candidate drift events and generate structured drift state vectors; A state space construction module is used to construct a multi-view state space, which includes: a user drift demand perspective state determined based on the structured drift state vector, a collaborative relationship perspective state determined based on the historical interaction relationship between users and accessories, and an accessory business perspective state determined based on accessory business attributes; the joint state is obtained by fusing the perspective states in the multi-view state space. The comprehensive scoring module is used to evaluate the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories in the joint state and generate a comprehensive value score for each candidate accessory. The recommendation result generation module is used to sort the candidate accessories according to the comprehensive value score and generate automotive accessory recommendation results.

[0061] It should be noted that, in the embodiments of the present invention, if the above-mentioned recommended methods are implemented in the form of software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of the present invention are not limited to any specific hardware and software combination.

[0062] Accordingly, embodiments of the present invention provide a storage medium storing a computer program thereon, characterized in that the computer program, when executed by a processor, implements the steps in the recommended method provided in the above embodiments.

[0063] This embodiment provides a computer device; the computer device includes: a processor, at least one communication bus, a user interface, at least one external communication interface, and a memory. The communication bus is configured to enable communication between these components. The user interface may include a display screen, and the external communication interface may include standard wired and wireless interfaces. The processor is configured to execute a program of a recommended method stored in the memory to implement the steps of the recommended method provided in the above embodiment.

[0064] It should be noted that the descriptions of the storage media and computer device embodiments above are similar to the descriptions of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage media and device embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding.

[0065] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0066] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, object, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, object, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, object, or apparatus that includes that element.

[0067] In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0068] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0069] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0070] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0071] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a controller to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0072] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-view recommendation method for automotive parts based on user interest drift identification, characterized in that, include: Acquire user purchasing behavior data and construct user purchasing behavior sequences; Based on the user's purchasing behavior sequence, the degree of offset between the current purchasing window representation and the historical stable preference representation is calculated to identify candidate drift events in user purchasing. Determine the interest drift evolution state of candidate drift events and generate structured drift state vectors; A multi-perspective state space is constructed, comprising: a user drift demand perspective state determined based on the structured drift state vector; a collaborative relationship perspective state determined based on the historical interaction relationship between the user and the accessory; and an accessory business perspective state determined based on the accessory business attributes. The perspective states in the multi-perspective state space are then fused to obtain a joint state. In the combined state, the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories is evaluated, and a comprehensive value score for each candidate accessory is generated. Candidate accessories are ranked based on the comprehensive value score to generate recommended automotive accessories.

2. The method according to claim 1, characterized in that, The step of calculating the degree of offset between the current purchase window representation and the historical stable preference representation based on the user's purchase behavior sequence, in order to identify candidate drift events in user purchases, includes: The user's purchasing behavior sequence is divided into multiple purchasing windows according to a preset time window. The weighted average of the feature vector of each accessory and the purchasing quantity in the current purchasing window is calculated to obtain the representation of the current purchasing window. Select several historical windows before the current window, calculate the mean value represented by each historical window, and obtain the historical stable preference representation; Calculate the cosine similarity between the current procurement window representation and the historical stable preference representation, and calculate the semantic offset based on the cosine similarity; Calculate the changes in accessory categories and accessory business attributes between the current procurement window and the historical window; The semantic offset, the change in accessory category, and the change in accessory business attribute are assigned weight coefficients and then weighted and summed to obtain the basic drift intensity. When the basic drift intensity exceeds a preset threshold, the corresponding continuous time window is determined as a candidate drift event for user purchase.

3. The method according to claim 2, characterized in that, The process of determining the interest drift evolution state of candidate drift events and generating a structured drift state vector includes: Extract the persistence, repetition consistency, backtracking tendency, and event-level drift intensity features of the candidate drift events, input the features into the state discrimination network, and output the evolution state confidence distribution of the candidate drift events as instantaneous perturbation type, transitional evolution type, and stable migration type; Obtain the business attribute features of the accessories corresponding to the candidate drift events, perform weighted mapping on the business attribute features, and generate business sensitivity weights; A comprehensive drift score is generated based on the basic drift intensity, the evolution state confidence distribution, and the business sensitivity weight. The structured drift state vector is constructed based on the basic drift intensity, the evolution state confidence distribution, the business sensitivity weight, and the comprehensive drift score.

4. The method according to claim 1, characterized in that, The construction of the multi-view state space includes: The user base representation is concatenated with the structured drift state vector, and the concatenated vector is input into a mapping function. After mapping by the mapping function, the user drift demand perspective state is obtained. The historical interaction relationship between the user and the accessory is constructed into a user-accessory collaboration graph. The user-accessory collaboration graph is input into a first graph neural network. The first graph neural network performs neighborhood propagation and high-order relationship aggregation. Then, the output of the first graph neural network is pooled to obtain the perspective state of the collaboration relationship. The business attribute relationships between parts are constructed into a parts relationship graph. The parts relationship graph is input into a second graph neural network. The second graph neural network performs relationship propagation and neighborhood aggregation. Then, the output of the second graph neural network is pooled to obtain the business perspective state of the parts.

5. The method according to claim 4, characterized in that, include: User nodes and accessory nodes are treated as two types of graph nodes. Interaction edges are established between user nodes and accessory nodes based on the interaction behavior between users and accessories. The weight of the interaction edges is determined based on factors such as purchase quantity, interaction frequency, feedback behavior intensity, and time decay, resulting in a user-accessory collaboration graph.

6. The method according to claim 4, characterized in that, include: By treating the parts nodes as graph nodes, relationship edges are established between the parts nodes based on the relationship between parts categories, vehicle models, substitutes, brand manufacturers, or similar business attributes. The weights of the relationship edges are determined based on the relationship between business rules and attribute similarity, thus obtaining the parts relationship graph.

7. The method according to claim 1, characterized in that, In the aforementioned joint state, the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories is evaluated, and a comprehensive value score for each candidate accessory is generated, including: Based on the aforementioned joint state, obtain the candidate component action space, and select candidate component actions from the candidate component action space; For the candidate accessory actions, a multi-perspective real-time reward vector is constructed, which includes user demand matching reward, collaborative relationship reward and accessory business reward; By performing relation enhancement and temporal accumulation on the multi-view real-time reward vector, the aggregated revenue representation corresponding to each candidate accessory action is obtained; The joint state, the candidate accessory action vector, and the aggregated benefit representation are input into the value assessment network to output the comprehensive value score of the candidate accessory.

8. The method according to claim 7, characterized in that, The process of enhancing the relationships and accumulating the temporal sequence of the multi-view real-time reward vectors to obtain the aggregated reward representation corresponding to each candidate accessory action includes: By using a reward propagation graph neural network, message propagation and node updates are performed on the multi-view real-time reward vector based on the dependency matrix between reward perspectives to obtain an enhanced reward representation. The enhanced reward representation is input into the gated loop unit for time-series state accumulation to obtain the aggregated reward representation.

9. The method according to claim 1, characterized in that, The step of ranking candidate accessories based on the comprehensive value score and generating recommended automotive accessories results includes: Candidate accessories are sorted in descending order of their comprehensive value scores; Select a predetermined number of candidate accessories that are ranked at the top to generate a set of recommended results; Output the accessory name, accessory category, brand manufacturer, compatibility information, recommendation order, comprehensive score, and reason for recommendation for each candidate accessory in the recommended result set.

10. A multi-view recommendation system for automotive parts based on user interest drift identification, characterized in that, include: The sequence construction module is used to acquire user purchasing behavior data and construct user purchasing behavior sequences; The identification module is used to calculate the degree of offset between the current purchase window representation and the historical stable preference representation based on the user's purchase behavior sequence, so as to identify candidate drift events in user purchases; The evolution state recognition module is used to determine the interest drift evolution state of candidate drift events and generate structured drift state vectors; A state space construction module is used to construct a multi-view state space, which includes: a user drift demand perspective state determined based on the structured drift state vector, a collaborative relationship perspective state determined based on the historical interaction relationship between users and accessories, and an accessory business perspective state determined based on accessory business attributes; the joint state is obtained by fusing the perspective states in the multi-view state space. The comprehensive scoring module is used to evaluate the comprehensive recommendation value of each candidate accessory in the action space of the candidate accessories in the joint state and generate a comprehensive value score for each candidate accessory. The recommendation result generation module is used to sort the candidate accessories according to the comprehensive value score and generate automotive accessory recommendation results.

Images