A new energy vehicle personalized recommendation method based on a knowledge graph double-view fusion

Abstract

The application discloses a new energy vehicle personalized recommendation method based on a knowledge graph double-view fusion, and comprises the following steps: generating a new energy vehicle multi-source data warehouse to obtain a new energy vehicle knowledge graph; constructing a user-vehicle interaction graph; performing knowledge graph embedding initialization on an entity set and a relation set, performing weighted aggregation on multi-hop neighbor information of a vehicle entity, performing double decision modeling on a user initial interaction representation and a vehicle initial interaction representation, performing double-view collaborative alignment on representations of the same vehicle in a knowledge graph view and a user-vehicle interaction graph view, and calculating a preference score of a user to a candidate vehicle; sorting the candidate vehicle, filtering vehicles that do not meet the conditions, and obtaining a new energy vehicle recommendation list. The application can fuse structured semantic information in the new energy vehicle knowledge graph and behavior preference information in the user-vehicle interaction graph, and improve the accuracy, interpretability and personalization degree of the recommendation result.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and intelligent transportation application technology, and relates to a personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion. It is particularly suitable for intelligent recommendation scenarios of new energy vehicles that require the fusion of vehicle semantics and user interaction information. Background Technology

[0002] With the rapid development of the new energy vehicle market and online configuration platforms, users need to consider multiple constraints when choosing a vehicle, including driving range, charging and battery swapping convenience, policy subsidies, and price budget. Traditional recommendation systems often rely solely on sales volume, tags, or simple collaborative filtering to output results, making it difficult to meet personalized needs. Existing methods have the following shortcomings: First, they cannot fully integrate multi-source heterogeneous data such as vehicle parameters, power batteries, user profiles, charging and battery swapping facilities, and policies, resulting in a lack of deep semantic support for recommendation results; second, they lack high-level modeling of user-vehicle interaction behavior, making it impossible to accurately capture user preferences in different scenarios; third, the recommendation process lacks interpretability, making it difficult for users to understand the reasons for recommendations, thus affecting trust. Summary of the Invention

[0003] To address the aforementioned deficiencies in existing technologies, the present invention aims to provide a personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion. By constructing a knowledge graph covering vehicles, power batteries, facilities, policies, and scenarios, and combining it with a user-vehicle interaction graph, the method employs graph neural networks and a gating fusion strategy to achieve unified expression and adaptive fusion of multi-source information, significantly improving the accuracy and interpretability of new energy vehicle recommendations.

[0004] The present invention is achieved through the following technical solution.

[0005] This invention provides a personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion, including: Acquire multi-source data on new energy vehicles and user behavior data to generate a multi-source data warehouse; Based on the multi-source data warehouse, entity extraction, attribute normalization, disambiguation of entities with the same name, and generation of entity relationships are performed to obtain a knowledge graph of new energy vehicles. Construct a user-vehicle interaction graph based on user behavior data; Based on the set of triples in the knowledge graph of new energy vehicles, the entity set and the relation set are initialized with knowledge graph embedding to obtain the initial embedding matrix of entities and the initial embedding matrix of relations. Based on the initial embedding matrix of entities and the initial embedding matrix of relations, relation-aware graph attention propagation is performed on the knowledge graph of new energy vehicles to perform weighted aggregation of multi-hop neighbor information of vehicle entities and obtain semantic representation of vehicle knowledge. Based on the user-vehicle interaction graph, we learn the initial user interaction representation and the initial vehicle interaction representation, and perform dual-decision modeling on the initial user interaction representation and the initial vehicle interaction representation to obtain the enhanced user representation and the enhanced vehicle interaction representation. Based on the vehicle knowledge semantic representation and the vehicle interaction enhanced representation, the representations of the same vehicle in the knowledge graph view and the user-vehicle interaction graph view are collaboratively aligned in two views, and a fused vehicle representation is generated based on the aligned vehicle knowledge semantic representation and the vehicle interaction enhanced representation. The user-enhanced representation and the fused vehicle representation are input into the preference prediction module to calculate the user's preference score for candidate vehicles; Candidate vehicles are sorted based on preference scores, and vehicles that do not meet the criteria are filtered out to obtain a recommended list of new energy vehicles.

[0006] Preferably, the multi-source data for new energy vehicles includes vehicle technical parameter data, power battery information, charging and swapping infrastructure data, regional policy subsidy data, and usage scenario data; Vehicle technical parameters include model name, brand, body size, drive type, driving range, charging power, and price range; Power battery information includes battery type, battery capacity, cycle life, thermal management method, and energy density; Data on charging and battery swapping infrastructure includes the geographical location of charging stations, charging pile type, rated power, availability, and service radius. Regional policy subsidy data includes applicable regions, applicable vehicle types, subsidy amount, effective date, and expiration date; Usage scenario data includes at least one scenario label from urban commuting, intercity long-distance travel, family car use, and logistics delivery; The multi-source data of new energy vehicles are linked according to vehicle identifier, region identifier, user identifier and time identifier to generate a multi-source data warehouse.

[0007] Preferably, user behavior data includes at least one of the following: user browsing, favorites, test drive appointments, purchases, and usage feedback.

[0008] Preferably, the knowledge graph for new energy vehicles includes a set of entities, a set of relations, and a set of triples; The entity set includes at least vehicle entities, brand entities, power battery entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities; The set of relationships includes at least the relationships of carrying, belonging, adapting, being located, covering, and preference; A set of triples includes a head entity, a relation, and a tail entity.

[0009] Preferably, a vehicle-equipped-power battery ternary array is generated based on the configuration relationship between the vehicle entity and the power battery entity; Generate a vehicle-belongs-brand triplet based on the attribution relationship between the vehicle entity and the brand entity; Generate a vehicle-adaptation-use scenario triplet based on the matching relationship between vehicle entities and use scenario entities; Generate a charging / swapping infrastructure-located-region triplet based on the positional relationship between the charging / swapping infrastructure entity and the regional entity; Generate a policy-coverage-vehicle-region triplet based on the applicable relationship between the policy entity and the vehicle entity or region entity.

[0010] Preferably, the entity set and relation set are initialized with knowledge graph embedding to obtain the initial entity embedding matrix and the initial relation embedding matrix, including: Knowledge graph embedding initialization is performed using an embedding model based on translation distance; Training makes the combined results of head entity embedding and relation embedding approximate those of tail entity embedding; Using real triples in the knowledge graph of new energy vehicles as positive samples, negative samples are generated by replacing head or tail entities. The initial entity embedding matrix and the initial relation embedding matrix are obtained by minimizing the positive sample score and increasing the negative sample score, thus widening the score interval between positive and negative samples.

[0011] Preferably, relationship-aware graph attention propagation is performed on the new energy vehicle knowledge graph, including: For a target vehicle entity, determine the first-order neighbors and multi-hop neighbors of the target vehicle entity from the new energy vehicle knowledge graph Gk; Based on the relationship type between the target vehicle entity and its neighbor entities, a relationship-specific transformation is performed on the neighbor entity representation to obtain a relationship-aware neighbor representation. Based on the target vehicle entity representation and the neighbor entity representation after relation-specific transformation, calculate the attention weight of the neighbor entities to the target vehicle entity. The neighbor entity representations are weighted and aggregated according to attention weights, and the target vehicle entity representation is updated by combining residual connections. After at least one layer of attention propagation through a relation-aware graph, a semantic representation of vehicle knowledge is obtained.

[0012] Preferably, the dual-decision modeling includes: Establish independent decision-making space and conformist decision-making space; Multiple learnable independent decision prototypes are set up in the independent decision space to represent the individualized preferences formed by users based on budget, range, charging conditions and configuration requirements; Multiple learnable conformity decision-making prototypes are set up in the conformity decision-making space to represent the social preferences formed by users under the influence of brand popularity, market sales and social reputation. The correlation weights between the user's initial interaction representation and the independent decision prototype and the conformity decision prototype are calculated respectively to obtain the user's independent decision representation and the user's conformity decision representation. The correlation weights between the initial interaction representation of the vehicle and the independent decision-making prototype and the conformity decision-making prototype are calculated respectively to obtain the vehicle independent decision-making representation and the vehicle conformity decision-making representation. By fusing the user's initial interaction representation, the user's independent decision representation, and the user's conformity decision representation, an enhanced user representation is obtained; By fusing the vehicle's initial interaction representation, vehicle independence decision representation, and vehicle conformity decision representation, an enhanced vehicle interaction representation is obtained.

[0013] Preferably, dual-view collaborative alignment includes: The semantic representation of vehicle knowledge in the knowledge graph view and the enhanced representation of vehicle interaction in the user-vehicle interaction graph view are used as positive sample pairs. The representations of different vehicles in the two views are used as negative sample pairs; By increasing the similarity between positive sample pairs and decreasing the similarity between negative sample pairs through a contrastive learning loss function, we obtain aligned semantic representations of vehicle knowledge and enhanced representations of vehicle interaction. The aligned vehicle knowledge semantic representation and the vehicle interaction enhancement representation are weighted and fused or gated to obtain the fused vehicle representation.

[0014] Preferably, the user's preference score for the candidate vehicle is calculated by the vector inner product or cosine similarity between the user's enhanced representation and the fused vehicle representation; Candidate vehicles are ranked based on preference scores, and those that do not meet the criteria are filtered out. Based on the relationship paths between vehicle entities and brand entities, technical parameter entities, power battery entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities in the new energy vehicle knowledge graph, recommendation explanation information is generated to obtain a new energy vehicle recommendation list.

[0015] The present invention, by adopting the above technical solution, has the following beneficial effects: 1. A knowledge graph of new energy vehicles is constructed based on multi-source data of new energy vehicles, and a user-vehicle interaction graph is constructed based on user behavior data. This makes the semantic information of vehicles and the behavioral preference information of users have clear data sources, which solves the problem that traditional recommendation methods are difficult to integrate multi-source heterogeneous data.

[0016] 2. By embedding knowledge graphs into initialization and using relation-aware graph attention propagation, multi-hop neighbor information of vehicle entities is aggregated in a differentiated weighted manner, enabling vehicle representation to absorb high-level semantic information such as brand, technical parameters, power battery, charging and swapping facilities, policies, and scenarios.

[0017] 3. By using dual-decision modeling, users' car purchase decisions are divided into independent decisions and conformity decisions, which can simultaneously depict users' personalized preferences based on their personal needs as well as social preferences influenced by market popularity and brand reputation.

[0018] 4. By using a dual-view collaborative alignment mechanism, the semantic representation of vehicle knowledge in the knowledge graph view and the enhanced representation of vehicle interaction in the user-vehicle interaction graph view are aligned and fused, which improves the recommendation model's ability to fuse heterogeneous view information.

[0019] 5. It can generate recommendation explanation information based on entity relationship paths in the knowledge graph, so that the recommendation results not only have high accuracy, but also good interpretability. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart of the personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to the present invention; Figure 2(a) and Figure 2(b) are schematic diagrams of knowledge graph embedding initialization in an embodiment of the present invention; Figure 2(a) shows the translation representation of the head entity, relation and tail entity in the vector space in the knowledge graph triple, and Figure 2(b) shows the projection representation of the entity in the relation specific space; Figure 3 This is a schematic diagram of the attention propagation process of the relationship-aware graph in an embodiment of the present invention; Figure 4 This is a schematic diagram of the dual-decision modeling module structure in an embodiment of the present invention; Figure 5 This is a schematic diagram of the dual-view collaborative alignment and fusion process in an embodiment of the present invention. Detailed Implementation

[0021] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions of the present invention are used to explain the present invention, but are not intended to limit the present invention.

[0022] Reference Figure 1 This invention provides a personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion. The method includes the following steps: Step S1: Obtain multi-source data on new energy vehicles and user behavior data to generate a multi-source data warehouse.

[0023] Acquire multi-source data on new energy vehicles and user behavior data. Multi-source data on new energy vehicles includes vehicle technical parameters, power battery information, charging and swapping infrastructure data, regional policy subsidy data, and usage scenario data.

[0024] User behavior data includes at least one of the following: browsing, saving, booking test drives, purchasing, and user feedback.

[0025] In this embodiment, vehicle technical parameter data includes vehicle model name, brand, body size, drive type, driving range, charging power, and price range; power battery information includes battery type, battery capacity, cycle life, thermal management method, and energy density; charging and swapping infrastructure data includes charging station geographical location, charging pile type, rated power, availability status, and service radius; regional policy subsidy data includes applicable areas, applicable vehicle models, subsidy amount, effective time, and expiration time; and usage scenario data includes scenario tags such as urban commuting, intercity long-distance travel, family car use, and logistics delivery.

[0026] The above data is linked according to vehicle identifier, region identifier, user identifier, and time identifier to obtain a multi-source data warehouse. This multi-source data warehouse provides the data foundation for subsequent construction of a new energy vehicle knowledge graph and recommendation constraint filtering.

[0027] Step S2: Construct a knowledge graph for new energy vehicles.

[0028] Based on the multi-source data warehouse obtained in step S1, entity extraction, attribute normalization, disambiguation of entities with the same name, and entity relation generation are performed to obtain the new energy vehicle knowledge graph Gk=(E,R,T). Where E represents the entity set, R represents the relation set, and T represents the triple set.

[0029] The entity set E includes at least vehicle entities, brand entities, power battery entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities. The relation set R includes at least the relations of "equipped with", "belongs to", "adapted to", "located in", "covered", and "preference". Each triple in the triple set T is represented as (h, r, t), where h represents the head entity, r represents the relation, and t represents the tail entity.

[0030] The generation of entity relationships in the knowledge graph of new energy vehicles includes: The following relationships are generated: Installation relationship is generated based on the configuration relationship between the vehicle entity and the power battery entity; Belonging relationship is generated based on the affiliation relationship between the vehicle entity and the brand entity; Adaptation relationship is generated based on the matching relationship between the vehicle entity and the usage scenario entity; Location relationship is generated based on the location relationship between the charging and swapping infrastructure entity and the regional entity; Coverage relationship is generated based on the applicability relationship between the regional policy subsidy entity and the vehicle entity or regional entity; Preference relationship is generated based on the association relationship between the user profile or user behavior tag and the vehicle entity or vehicle attribute entity.

[0031] Specifically, a "vehicle-equipped-power battery" triplet is generated based on the configuration relationship between the vehicle entity and the power battery entity; a "vehicle-belongs to-brand" triplet is generated based on the affiliation relationship between the vehicle entity and the brand entity; a "vehicle-adapted-usage scenario" triplet is generated based on the matching relationship between the vehicle entity and the usage scenario entity; a "charging and swapping infrastructure-located-region" triplet is generated based on the location relationship between the charging and swapping infrastructure entity and the regional entity; and a "policy-coverage-vehicle / -region" triplet is generated based on the applicability relationship between the policy entity and the vehicle entity or the regional entity.

[0032] After the above processing, the knowledge graph of new energy vehicles will be stored in a graph database for subsequent graph embedding learning and recommended explanation path query.

[0033] Step S3: Construct a user-vehicle interaction graph based on user behavior data.

[0034] Construct a user-vehicle interaction graph Gu=(U,I,Y) based on the user behavior data obtained in step S1. Here, U represents the set of users, I represents the set of vehicles, and Y represents the set of interaction edges between users and vehicles.

[0035] The interaction weights between users and vehicles are assigned based on the type, duration, and intensity of user behavior. For example, browsing behavior receives a low weight, favorites and test drive appointments receive a medium weight, purchases and high-rated feedback receive a high weight, and negative feedback receives a negative weight or a reduced weight. Repeated actions by the same user on the same vehicle within a preset time window are merged, and the interaction weights are updated based on a time decay coefficient.

[0036] The user-vehicle interaction graph Gu obtained in step S3 is used for subsequent learning of user and vehicle behavior preference representations in the interaction graph view.

[0037] Step S4: Initialize the knowledge graph embedding.

[0038] Based on the set of triples T in the new energy vehicle knowledge graph Gk obtained in step S2, the entity set E and the relation set R are initialized with knowledge graph embedding to obtain the initial entity embedding matrix and the initial relation embedding matrix.

[0039] Initialization. Referring to Figures 2(a) and 2(b), for the triple (h,r,t), the head entity embedding, relation embedding, and tail entity embedding are trained to satisfy the constraint that h+r is close to t, so that the combined result of the head entity embedding and relation embedding is close to the tail entity embedding.

[0040] During training, real triples in the knowledge graph Gk of new energy vehicles are used as positive samples, and negative samples are constructed by replacing the head or tail entities. By minimizing the score of positive samples and increasing the score of negative samples, the score interval between positive and negative samples is widened, enabling the model to distinguish between correct and incorrect triples.

[0041] The initial entity embedding matrix and the initial relation embedding matrix obtained in step S4 are used as inputs for the relation-aware graph attention propagation in step S5.

[0042] Step S5: Generate a semantic representation of vehicle knowledge.

[0043] Using the initial entity embedding matrix and initial relation embedding matrix obtained in step S4 as input, relation-aware graph attention propagation is performed on the new energy vehicle knowledge graph Gk obtained in step S2 to perform weighted aggregation of the multi-hop neighbor information of vehicle entities, thereby obtaining the semantic representation of vehicle knowledge.

[0044] Specifically, refer to Figure 3 This illustrates the propagation relationships between the target entity, neighboring entities, relation-specific transformations, attention weights, residual connections, and updated entity representations.

[0045] For the target vehicle entity, firstly, the first-order neighbors and multi-hop neighbors of the target vehicle entity are determined from the new energy vehicle knowledge graph Gk. Neighbor entities can include brand entities, power battery entities, technical parameter entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities.

[0046] Secondly, for each neighboring entity, a relationship-specific transformation is performed based on the relationship type between the target vehicle entity and that neighboring entity to obtain a relationship-aware neighbor representation.

[0047] Then, attention weights of neighbor entities to the target vehicle entity are calculated based on the target vehicle entity representation and the relation-aware neighbor representation after relation-specific transformation. These attention weights characterize the importance of different neighbor entities to the target vehicle entity.

[0048] The neighbor entity representations are then weighted and aggregated according to attention weights, and the target vehicle entity representation is updated by combining residual connections.

[0049] Finally, after at least one layer of attention propagation through a relation-aware graph, a semantic representation of vehicle knowledge is obtained.

[0050] The vehicle knowledge semantic representation obtained in step S5 is used to characterize the high-order semantic features of vehicles in the new energy vehicle knowledge graph view.

[0051] The relation-aware graph attention propagation adopts a multi-head attention mechanism, in which multiple attention heads calculate the attention weights of neighboring entities to the target vehicle entity, and the outputs of multiple attention heads are averaged or concatenated and fused to obtain a high-order knowledge semantic representation of the target vehicle entity.

[0052] Step S6: Based on the user-vehicle interaction graph, learn the initial user interaction representation and the initial vehicle interaction representation, and perform dual-decision modeling on the initial user interaction representation and the initial vehicle interaction representation to obtain the enhanced user representation and the enhanced vehicle interaction representation.

[0053] Using the user-vehicle interaction graph Gu obtained in step S3 as input, the initial user interaction representation and the initial vehicle interaction representation are learned. The initial user interaction representation is used to characterize the user's historical interaction preferences, and the initial vehicle interaction representation is used to characterize the vehicle's behavioral feedback features in the user interaction network.

[0054] Based on this, independent decision-making space and conformity decision-making space are established.

[0055] Reference Figure 4 The diagram shows that the user embedding and vehicle embedding in the user-vehicle interaction graph enter the independence decision space and the conformity decision space, respectively, and are then weighted by Softmax to generate enhanced user representations and vehicle representations.

[0056] Multiple learnable independent decision-making prototypes are set up in the independent decision-making space to represent the individualized preferences formed by users based on personal factors such as budget, battery life, charging conditions, and configuration requirements.

[0057] Multiple learnable conformity decision-making prototypes are set up in the conformity decision-making space to represent the social preferences formed by users under the influence of factors such as brand popularity, sales ranking, and social reputation.

[0058] The relevance weights between the initial user interaction representation and the independent decision prototype and the conformity decision prototype are calculated separately to obtain the user independent decision representation and the user conformity decision representation. The initial user interaction representation, the user independent decision representation, and the user conformity decision representation are then fused to obtain the user augmented representation.

[0059] Similarly, the correlation weights between the initial vehicle interaction representation and the independent decision-making prototype and the conformist decision-making prototype are calculated separately to obtain the vehicle independent decision-making representation and the vehicle conformist decision-making representation. The initial vehicle interaction representation, the vehicle independent decision-making representation, and the vehicle conformist decision-making representation are then fused to obtain the vehicle interaction enhancement representation.

[0060] The vehicle interaction augmentation representation obtained in step S6 is used to characterize the vehicle's behavioral preference features in the user-vehicle interaction graph view, and the user augmentation representation is used for subsequent preference prediction.

[0061] Step S7: Perform dual-view collaborative alignment and fusion.

[0062] Using the vehicle knowledge semantic representation obtained in step S5 and the vehicle interaction enhancement representation obtained in step S6 as input, the representations of the same vehicle in the knowledge graph view and the user-vehicle interaction graph view are aligned in a dual-view collaborative manner, and the aligned vehicle knowledge semantic representation and vehicle interaction enhancement representation are used as input.

[0063] Specifically, refer to Figure 5 The paper demonstrates how user-vehicle interaction graph view and new energy vehicle knowledge graph view generate vehicle interaction view representation and vehicle knowledge graph view representation respectively, and how cross-view comparison learning and feature fusion are performed through dual-view collaborative alignment module to obtain fused vehicle representation.

[0064] The semantic representation of vehicle knowledge in the knowledge graph view and the enhanced representation of vehicle interaction in the user-vehicle interaction graph view of the same vehicle are taken as positive sample pairs, and the representations of different vehicles in the two views are taken as negative sample pairs. By contrastive learning loss, the similarity between positive sample pairs is increased and the similarity between negative sample pairs is decreased, so that the representation of the same vehicle in the two views is mapped to a more consistent representation space.

[0065] After completing the dual-view collaborative alignment, the aligned vehicle knowledge semantic representation and the vehicle interaction enhancement representation are fused to obtain a fused vehicle representation. The fusion method can be either weighted fusion or gated fusion. The fused vehicle representation simultaneously includes both vehicle structured semantic information and user behavior preference information.

[0066] Step S8: Calculate preference scores.

[0067] The user-enhanced representation obtained in step S6 and the fused vehicle representation obtained in step S7 are input into the preference prediction module to calculate the user's preference score for the candidate vehicles.

[0068] In this embodiment, the preference score can be calculated either by the vector inner product between the user-enhanced representation and the fused vehicle representation, or by cosine similarity.

[0069] Let user enhancement be represented as:

[0070] The fusion vehicle is represented as:

[0071] Where d represents the dimension of the embedding vector. Preference scores can be calculated either through the vector inner product between the user's augmented representation and the fused vehicle representation, or through cosine similarity. In this embodiment, the user... For candidate vehicles The preference rating is defined as:

[0072] in, This represents the user's predicted preference rating for the vehicle. This represents the transpose of the user-enhanced representation. This indicates the vehicle being represented. A higher preference score indicates a greater level of interest from the user in the candidate vehicles.

[0073] A higher preference score indicates a greater level of interest in the candidate vehicles.

[0074] Step S9: Sort the candidate vehicles according to the preference score, filter out vehicles that do not meet the conditions, and obtain a new energy vehicle recommendation list.

[0075] Candidate vehicles are ranked based on the preference scores obtained in step S8. Simultaneously, based on the multi-source data warehouse obtained in step S1 and the new energy vehicle knowledge graph obtained in step S2, and considering user budget, charging conditions, regional policies, and usage scenario constraints, candidate vehicles that do not meet the hard criteria are filtered out.

[0076] After filtering, a Top-N recommended list of new energy vehicles is output. Further, based on the relationship paths between vehicle entities and brand entities, technical parameter entities, power battery entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities in the new energy vehicle knowledge graph, recommendation explanations are generated. For example, when a vehicle is recommended to a user, explanations such as "This model's range meets the needs of intercity travel," "This model is eligible for subsidies in the user's area," and "This model has high charging facility coverage nearby" can be output.

[0077] The advantages of this invention can be further illustrated by the following simulation experiments: This experiment is used to verify the effectiveness of the new energy vehicle personalized recommendation method based on knowledge graph dual-view fusion proposed in this invention. The experiment focuses on verifying the following: First, verify whether a knowledge graph of new energy vehicles constructed from multi-source data of new energy vehicles can provide vehicle semantic information for recommendation tasks; Second, verify whether the user-vehicle interaction graph can provide user behavior preference information for recommendation tasks; Third, we will verify whether attention propagation of relation-aware graphs can improve the expressive power of semantic representation of vehicle knowledge; Fourth, verify whether dual-decision modeling can characterize users' independent and herd-like decisions in the process of purchasing new energy vehicles; Fifth, verify whether dual-view collaborative alignment can integrate information from the knowledge graph view and the user-vehicle interaction graph view, thereby improving recommendation accuracy and ranking quality.

[0078] The experiment was conducted based on a self-constructed new energy vehicle recommendation dataset. The new energy vehicle knowledge graph includes multiple entity types, such as vehicle entities, brand entities, technical parameter entities, power battery entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities. The knowledge graph size is as follows: Table 1. Statistics on the Scale of the Knowledge Graph for New Energy Vehicles

[0079] The user-vehicle interaction graph consists of a set of users, a set of vehicles, and a set of interaction edges between users and vehicles. Interaction edges are generated by behaviors such as browsing, saving, scheduling test drives, purchasing, and providing feedback, and interaction weights are set according to the type, duration, and intensity of the behavior.

[0080] The experimental procedure was strictly performed in accordance with steps S1 to S9 of claim 1.

[0081] Step S1: Obtain vehicle technical parameter data, power battery information, charging and swapping infrastructure data, regional policy subsidy data, usage scenario data, and user behavior data to generate a multi-source data warehouse.

[0082] Step S2: Based on the multi-source data warehouse, perform entity extraction, attribute normalization, disambiguation of entities with the same name, and generation of entity relationships to construct a knowledge graph Gk=(E,R,T) for new energy vehicles.

[0083] Step S3: Construct a user-vehicle interaction graph Gu=(U,I,Y) based on user behavior data, and set interaction weights for the interaction edges between users and vehicles.

[0084] Step S4: Based on the set of triples T in the knowledge graph Gk of new energy vehicles, perform knowledge graph embedding initialization on the entity set E and the relation set R to obtain the initial entity embedding matrix and the initial relation embedding matrix.

[0085] Step S5: Using the initial entity embedding matrix and the initial relation embedding matrix as input, perform relation-aware graph attention propagation on the new energy vehicle knowledge graph Gk to obtain the semantic representation of vehicle knowledge.

[0086] Step S6: Using the user-vehicle interaction graph Gu as input, learn the initial user interaction representation and the initial vehicle interaction representation, and perform dual-decision modeling through independent decision prototype and conformity decision prototype to obtain the enhanced user representation and the enhanced vehicle interaction representation.

[0087] Step S7: Using vehicle knowledge semantic representation and vehicle interaction enhanced representation as input, perform dual-view collaborative alignment on the representations of the same vehicle in the knowledge graph view and the user-vehicle interaction graph view, and fuse them to obtain a fused vehicle representation.

[0088] Step S8: Input the user enhanced representation and the fused vehicle representation into the preference prediction module to calculate the user's preference score for the candidate vehicles.

[0089] Step S9: Sort the candidate vehicles according to preference scores, and filter them in combination with user budget, charging conditions, regional policies and usage scenario constraints to output a Top-N recommended list of new energy vehicles.

[0090] The experiment used common settings for graph neural network recommendation models. The embedding dimension was set to 64; the number of relationship-aware graph attention propagation layers was set to 2; the number of attention heads was set to 8; the number of independent decision prototypes and conformity decision prototypes were both set to 8; the evaluation metrics were Recall@K and NDCG@K, where K was set to 20, 50, and 100.

[0091] The evaluation indicators include: (1) Recall rate measures the ability of the recommendation list to cover vehicles that users are genuinely interested in. The higher the Recall@K, the stronger the recommendation model's ability to recall vehicles that the user has actually interacted with from the first K recommendations.

[0092]

[0093] Where U represents the set of test users. Represented as user Generated Top- Recommendation result set, Indicates users in the test set A collection of real-world interactive vehicles. Used to measure the proportion of vehicles of genuine interest in the test set that are covered by the recommended results.

[0094] (2) Normalized depreciation cumulative gain, used to evaluate the ranking quality of the recommendation list:

[0095]

[0096] in, Indicates user The recommended list is at the top. Accumulated gain at bit loss This represents the cumulative gain due to loss under ideal sorting. Indicates the number in the recommended list Relevance tags for items at a location: When a vehicle at that location belongs to a user When taking the real interaction set in the test set, take Otherwise take NDCG@K is used to measure the number of recommendations before the list is published. The sorting quality of a bit; a larger value indicates a better sorting result.

[0097] To verify the recommendation effectiveness of the method of the present invention, it was compared with various recommendation models, including MF, CKE, LightGCN, KGAT, CKAN, KGIN, KGCL, and KGCN. Table 2 Performance comparison of each model on the new energy vehicle recommendation dataset Table4.2 Model Performance on the New Energy Vehicle RecommendationDataset

[0098] As shown in Table 2, the method of the present invention achieves superior results in terms of Recall@20, Recall@100, NDCG@20, NDCG@50, and NDCG@100. Specifically, the Recall@20, Recall@50, and Recall@100 of the method of the present invention are 0.1268, 0.1503, and 0.1805, respectively, and the NDCG@20, NDCG@50, and NDCG@100 are 0.0854, 0.0938, and 0.1396, respectively.

[0099] Compared with the best external baseline model, the method of this invention improves recall@20, recall@50, recall@100, NDCG@20, NDCG@50, and NDCG@100 by 10.65%, 3.73%, 4.88%, 6.48%, 2.40%, and 7.72%, respectively. Experimental results show that the method of this invention can effectively integrate semantic information from knowledge graphs and user-vehicle interaction information, and has good overall performance in terms of candidate vehicle recall capability and recommendation ranking quality.

[0100] The above experiments demonstrate that the method of this invention can complete personalized recommendations for new energy vehicles following the process of "multi-source data acquisition—construction of a new energy vehicle knowledge graph—construction of a user-vehicle interaction graph—knowledge graph embedding initialization—relationship-aware graph attention propagation—dual-decision modeling—dual-view collaborative alignment—preference score calculation—recommendation list output." Through relationship-aware graph attention propagation, this method can mine high-order semantic information of vehicle entities; through dual-decision modeling, it can characterize users' individualized preferences and conformity preferences; and through dual-view collaborative alignment, it can integrate complementary information from the knowledge graph view and the interaction graph view, thereby improving the accuracy and ranking quality of the recommendation results.

[0101] This invention is not limited to the above embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.

Claims

1. A personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion, characterized in that, include: Acquire multi-source data on new energy vehicles and user behavior data to generate a multi-source data warehouse; Based on the multi-source data warehouse, entity extraction, attribute normalization, disambiguation of entities with the same name, and generation of entity relationships are performed to obtain a knowledge graph of new energy vehicles. Construct a user-vehicle interaction graph based on user behavior data; Based on the set of triples in the knowledge graph of new energy vehicles, the entity set and the relation set are initialized with knowledge graph embedding to obtain the initial embedding matrix of entities and the initial embedding matrix of relations. Based on the initial embedding matrix of entities and the initial embedding matrix of relations, relation-aware graph attention propagation is performed on the knowledge graph of new energy vehicles to perform weighted aggregation of multi-hop neighbor information of vehicle entities and obtain semantic representation of vehicle knowledge. Based on the user-vehicle interaction graph, we learn the initial user interaction representation and the initial vehicle interaction representation, and perform dual-decision modeling on the initial user interaction representation and the initial vehicle interaction representation to obtain the enhanced user representation and the enhanced vehicle interaction representation. Based on the vehicle knowledge semantic representation and the vehicle interaction enhanced representation, the representations of the same vehicle in the knowledge graph view and the user-vehicle interaction graph view are collaboratively aligned in two views, and a fused vehicle representation is generated based on the aligned vehicle knowledge semantic representation and the vehicle interaction enhanced representation. The user-enhanced representation and the fused vehicle representation are input into the preference prediction module to calculate the user's preference score for candidate vehicles; Candidate vehicles are sorted based on preference scores, and vehicles that do not meet the criteria are filtered out to obtain a recommended list of new energy vehicles.

2. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, Multi-source data for new energy vehicles includes vehicle technical parameter data, power battery information, charging and swapping infrastructure data, regional policy subsidy data, and usage scenario data; Vehicle technical parameters include model name, brand, body size, drive type, driving range, charging power, and price range; Power battery information includes battery type, battery capacity, cycle life, thermal management method, and energy density; Data on charging and battery swapping infrastructure includes the geographical location of charging stations, charging pile type, rated power, availability, and service radius. Regional policy subsidy data includes applicable regions, applicable vehicle types, subsidy amount, effective date, and expiration date; Usage scenario data includes at least one scenario label from urban commuting, intercity long-distance travel, family car use, and logistics delivery; The multi-source data of new energy vehicles are linked according to vehicle identifier, region identifier, user identifier and time identifier to generate a multi-source data warehouse.

3. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, User behavior data includes at least one of the following: browsing, saving, booking test drives, purchasing, and user feedback.

4. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, The knowledge graph for new energy vehicles includes sets of entities, sets of relations, and sets of triples. The entity set includes at least vehicle entities, brand entities, power battery entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities; The set of relationships includes at least the relationships of carrying, belonging, adapting, being located, covering, and preference; A set of triples includes a head entity, a relation, and a tail entity.

5. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 4, characterized in that, A vehicle-equipped-power battery ternary array is generated based on the configuration relationship between the vehicle entity and the power battery entity. Generate a vehicle-belongs-brand triplet based on the attribution relationship between the vehicle entity and the brand entity; Generate a vehicle-adaptation-use scenario triplet based on the matching relationship between vehicle entities and use scenario entities; Generate a charging / swapping infrastructure-located-region triplet based on the positional relationship between the charging / swapping infrastructure entity and the regional entity; Generate a policy-coverage-vehicle-region triplet based on the applicable relationship between the policy entity and the vehicle entity or region entity.

6. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, Knowledge graph embedding initialization is performed on the entity set and relation set to obtain the initial entity embedding matrix and the initial relation embedding matrix, including: Knowledge graph embedding initialization is performed using an embedding model based on translation distance; Training makes the combined results of head entity embedding and relation embedding approximate those of tail entity embedding; Using real triples in the knowledge graph of new energy vehicles as positive samples, negative samples are generated by replacing head or tail entities. The initial entity embedding matrix and the initial relation embedding matrix are obtained by minimizing the positive sample score and increasing the negative sample score, thus widening the score interval between positive and negative samples.

7. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, Perform relationship-aware graph attention propagation on the new energy vehicle knowledge graph, including: For a target vehicle entity, determine the first-order neighbors and multi-hop neighbors of the target vehicle entity from the new energy vehicle knowledge graph Gk; Based on the relationship type between the target vehicle entity and its neighbor entities, a relationship-specific transformation is performed on the neighbor entity representation to obtain a relationship-aware neighbor representation. Based on the target vehicle entity representation and the neighbor entity representation after relation-specific transformation, calculate the attention weight of the neighbor entities to the target vehicle entity. The neighbor entity representations are weighted and aggregated according to attention weights, and the target vehicle entity representation is updated by combining residual connections. After at least one layer of attention propagation through a relation-aware graph, a semantic representation of vehicle knowledge is obtained.

8. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, The dual-decision modeling includes: Establish independent decision-making space and conformist decision-making space; Multiple learnable independent decision prototypes are set up in the independent decision space to represent the individualized preferences formed by users based on budget, range, charging conditions and configuration requirements; Multiple learnable conformity decision-making prototypes are set up in the conformity decision-making space to represent the social preferences formed by users under the influence of brand popularity, market sales and social reputation. The correlation weights between the user's initial interaction representation and the independent decision prototype and the conformity decision prototype are calculated respectively to obtain the user's independent decision representation and the user's conformity decision representation. The correlation weights between the initial interaction representation of the vehicle and the independent decision-making prototype and the conformity decision-making prototype are calculated respectively to obtain the vehicle independent decision-making representation and the vehicle conformity decision-making representation. By fusing the user's initial interaction representation, the user's independent decision representation, and the user's conformity decision representation, an enhanced user representation is obtained; By fusing the vehicle's initial interaction representation, vehicle independence decision representation, and vehicle conformity decision representation, an enhanced vehicle interaction representation is obtained.

9. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, Perform collaborative alignment of two views, including: The semantic representation of vehicle knowledge in the knowledge graph view and the enhanced representation of vehicle interaction in the user-vehicle interaction graph view are used as positive sample pairs. The representations of different vehicles in the two views are used as negative sample pairs; By increasing the similarity between positive sample pairs and decreasing the similarity between negative sample pairs through a contrastive learning loss function, we obtain aligned semantic representations of vehicle knowledge and enhanced representations of vehicle interaction. The aligned vehicle knowledge semantic representation and the vehicle interaction enhancement representation are weighted and fused or gated to obtain the fused vehicle representation.

10. The personalized recommendation method for new energy vehicles based on knowledge graph dual-view fusion according to claim 1, characterized in that, The user's preference rating for candidate vehicles is calculated using the vector inner product or cosine similarity between the user's enhanced representation and the fused vehicle representation. Candidate vehicles are ranked based on preference scores, and those that do not meet the criteria are filtered out. Based on the relationship paths between vehicle entities and brand entities, technical parameter entities, power battery entities, charging and swapping infrastructure entities, policy entities, and usage scenario entities in the new energy vehicle knowledge graph, recommendation explanation information is generated to obtain a new energy vehicle recommendation list.

Images