Cold chain logistics vehicle source recommendation method and device based on dynamic interest and knowledge graph

Abstract

The present invention discloses a method and device for recommending cold chain logistics vehicle sources based on dynamic interests and knowledge graphs. First, the cold chain logistics vehicle source information is cleaned and sorted, the cold chain logistics vehicle source knowledge graph framework is defined, and the required knowledge graph is constructed; the historical vehicle source information of the consignor is integrated by the attention mechanism, and the feature vector obtained by integration is used as the consignor's long-term interest representation, and the hierarchical interest representation of the consignor is mined by the interest propagation algorithm, and the multi-layer interest representation is used as the consignor's short-term interest representation, and the consignor's long-term and short-term interest representation is integrated to obtain the consignor's dynamic interest representation; the high-order structural information of the knowledge graph is mined by the knowledge graph convolutional neural network to obtain the vehicle source feature representation; finally, the cold chain logistics vehicle source recommendation is realized according to the inner product of the consignor's dynamic interest representation and the vehicle source feature representation. The present invention makes full use of the high-order structural information and semantic information of the knowledge graph, deeply mines the user's dynamic interests, and realizes personalized recommendation.

Description

Technical Field

[0001] The present invention relates to the field of cold chain logistics recommendation technology, and in particular to a method and device for recommending cold chain logistics vehicle sources based on dynamic interests and knowledge graphs. Background Art

[0002] At present, there is a large amount of vehicle source information on the truckless carrier platform, and it is time-consuming and laborious for shippers to choose the vehicle source that meets their needs.

[0003] In the field of cold chain logistics recommendations, existing literature has made recommendations based on the interactive information between users and items. Although this type of method is simple and practical, it often faces the problem of data sparsity. In addition, most methods in the literature focus on capturing users' long-term interests and lack modeling of users' short-term interests. It is impossible to mine users' dynamic information to learn dynamically changing interest characteristics, resulting in the recommended items not meeting users' preferences. Summary of the invention

[0004] Purpose of the invention: In view of the problems existing in the above methods, the present invention provides a method and device for recommending vehicles for cold chain logistics based on dynamic interests and knowledge graphs, which utilizes structured information in knowledge graphs to mine the fine-grained relationship between cargo owners and vehicle sources, and more accurately obtains cargo owners' interests. At the same time, starting from the cargo owners' long-term and short-term interests, the interest characteristics that change over time are analyzed, which is more reasonable to be applied in actual scenarios.

[0005] Technical solution: The present invention proposes a cold chain logistics vehicle source recommendation method based on dynamic interests and knowledge graphs, comprising the following steps:

[0006] Step 1: Clean and organize the cold chain logistics vehicle source information, define the cold chain logistics vehicle source knowledge graph framework, and construct the required knowledge graph G;

[0007] Step 2: Use the attention mechanism to fuse the historical vehicle source information of the cargo owner, and use the fused feature vector as the cargo owner's long-term interest representation u l The interest propagation algorithm is used to mine the hierarchical interest representation of the consignor, and the multi-layer interest representation is used as the short-term interest representation u of the consignor. s , integrating the short-term interest of shippers s Expressing long-term and short-term interests with cargo owners l Get the cargo owner's dynamic interest expression d ;

[0008] Step 3: Use the knowledge graph convolutional neural network to mine the high-order structural information of the knowledge graph to obtain the vehicle source feature representation v u ;

[0009] Step 4: According to the dynamic interest of the shipper d and vehicle source feature representation v uThe internal area is used to recommend cold chain logistics vehicles.

[0010] Furthermore, the specific method of step 1 is:

[0011] Step 1.1: Analyze the entities in the cold chain logistics vehicle source knowledge graph based on the vehicle source data of the vehicle-free carrier platform, mainly including: license plate number, vehicle model, load capacity, refrigeration capacity and region. There are relationships between different entity classes, and the relationship category is defined as the category of the entity to which the relationship points;

[0012] Step 1.2: Extract vehicle source data from the database of the truckless carrier platform, obtain entity tables and relationship tables according to pre-defined entity classification, and store them in the database;

[0013] Step 1.3: According to the cold chain logistics vehicle source entities and relationships in the database, the cold chain logistics vehicle source knowledge graph G is constructed through the graph database Neo4j. The cold chain logistics vehicle source knowledge graph G consists of triples (h, r, t), where h∈E, t∈E, r∈R represent the head entity, tail entity and relationship between entities of the triple, respectively, and E and R represent the entity and relationship sets in the knowledge graph G, respectively.

[0014] Furthermore, the specific method for obtaining the long-term interest expression of the cargo owner in step 2 is:

[0015] Step 2.1: Get the complete historical vehicle source set H = {v1, v2, ..., v n};

[0016] Step 2.2: Initialize the historical vehicle source vector v i and the candidate vehicle source vector v;

[0017] Step 2.3: Define the loop variable i, assign an initial value of 1, and define the long-term interest of the shipper as u l ;

[0018] Step 2.4: If i≤len(H), go to step 2.5, otherwise jump to step 2.10;

[0019] Step 2.5: Translate the historical vehicle source vector v i It is concatenated with the candidate vehicle source vector v and input into the attention network to obtain the attention weight a′ i , where a′ i =σ(W a [v i , v]+b a ), W a is the training weight, b a is the bias vector;

[0020] Step 2.6: Normalize the fusion weight a through the softmax functioni ,in

[0021] Step 2.7: According to the fusion weight a i and the vehicle source vector v i The product is the current vehicle source fusion vector C i ;

[0022] Step 2.8: l =u l +C i ;

[0023] Step 2.9: i=i+1, jump to step 2.4;

[0024] Step 2.10: End the loop and get the final long-term interest representation u of the consignor l .

[0025] Furthermore, the step of obtaining the cargo owner's short-term interest expression in step 2 includes:

[0026] Step 2.11: Filter the K most recently interacted vehicle sources from the shipper interaction sequence;

[0027] Step 2.12: Define loop variable i2, with an initial value of 1, and define the number of propagation layers, with a value of h;

[0028] Step 2.13: Take K vehicle sources as the knowledge graph seed set

[0029] Step 2.14: If i2≤h, go to step 2.15, otherwise jump to step 2.20;

[0030] Step 2.15: Obtain the i2-layer neighbor node set of the cargo owner u through propagation according to the entity

[0031] Step 2.16: The interest set of the i2 layer of the cargo owner u is Among them, h∈E, t∈E, r∈R represent the head entity, tail entity and relationship between entities of the triple respectively, and E and R represent the entity and relationship sets in the knowledge graph G respectively;

[0032] Step 2.17: By calculating the candidate vehicle source vector v and the interest set Middle head entity h i In the relational space i The correlation below gets the intra-layer fusion weight Where R i ∈R d 、h i ∈R d and v∈R d For the relationshipi 、Head entity h i and the embedding representation of item v;

[0033] Step 2.18: Weight the tail entity vector representation in the interest set to obtain the i2-layer cargo owner interest representation in

[0034] Step 2.19: i2=i2+1, jump to step 2.14;

[0035] Step 2.20: End the loop and get the short-term interest set of the shipper

[0036] Step 2.21: Define loop variable i3, assign an initial value of 1, and define the short-term interest representation u of the cargo owner s ;

[0037] Step 2.22: If i3≤len(O1), go to step 2.23, otherwise jump to step 2.28;

[0038] Step 2.23: Vector representation of the shipper’s short-term interest set O1 And the candidate vehicle source vector representation v is input into the attention network to calculate the attention weight in W h is the training weight, b h is the bias vector;

[0039] Step 2.24: Normalize the fusion weights using the softmax function in

[0040] Step 2.25: According to the fusion weight and the vector representation of the short-term interest set of the shipper The product is used to obtain the current shipper's interest fusion vector

[0041] Step 2.26:

[0042] Step 2.27: i3=i3+1, jump to step 2.22;

[0043] Step 2.28: End the loop and get the final cargo owner's short-term interest representation u s .

[0044] Furthermore, the step 2 integrates the short-term interest representation u of the cargo owner s Expressing long-term and short-term interests with cargo owners l Get the cargo owner's dynamic interest expression dThe specific operations are:

[0045] The long-term interest of the cargo owner is expressed as u l and short-term interest representation u s Input into the fully connected neural network to obtain the dynamic interest representation of the cargo owner o d =W b [u l ,u s ], where W b is the training weight.

[0046] Furthermore, the specific method of step 3 is:

[0047] Step 3.1: Use graph convolutional neural network to extract vehicle source features and input the dynamic interest representation of the cargo owner d ;

[0048] Step 3.2: Define loop variable i4, assign an initial value of 1, define v0 as the initial feature vector of candidate vehicle source v, and set the number of aggregation layers to H;

[0049] Step 3.3: If i4≤H, go to step 3.4, otherwise jump to step 3.12;

[0050] Step 3.4: Define loop variable i5, assign an initial value of 1, define N(v) as the neighbor set of the entity, and the length of the set is M;

[0051] Step 3.5: If i5≤M, go to step 3.6, otherwise jump to step 3.8;

[0052] Step 3.6: Based on the relationship r between entities and dynamic interest o d Calculate the shipper’s preference for the relationship, and the calculation formula is:

[0053] Step 3.7: i5=i5+1, jump to step 3.5;

[0054] Step 3.8: End the loop and obtain the fusion weights of different entities through normalization The calculation formula is

[0055] Step 3.9: Use linear combination of neighbor information to obtain the neighborhood information of the node The calculation formula is

[0056] Step 3.10: Aggregate the current vehicle source embedding representation and neighborhood information to obtain a new vehicle source embedding representation. The calculation formula is: Where W c represents the learnable weight, b crepresents the bias vector, σ represents the activation function;

[0057] Step 3.11: i4=i4+1, jump to step 3.3;

[0058] Step 3.12: End the loop and get the final vehicle source feature representation v u , where v u =v H .

[0059] Furthermore, the specific method of step 4 is:

[0060] Step 4.1: The final cargo owner's dynamic interest representation d And the final vehicle source feature representation v u Perform inner product to predict the interaction probability between cargo owners and vehicle sources

[0061] Step 4.2: Through the loss function

[0062] Update the model parameters and optimize the model. The first item is the cross entropy loss, the second item is the knowledge graph embedding unit loss, λ1 is the loss value parameter, and the third item is the regularization loss, λ2 is the regularization parameter.

[0063] The present invention also discloses a device for recommending cold chain logistics vehicle sources based on dynamic interests and knowledge graphs, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program executes the steps of the method for recommending cold chain logistics vehicle sources based on dynamic interests and knowledge graphs when loaded into the processor.

[0064] Beneficial effects:

[0065] 1. The present invention introduces knowledge graph into the cold chain logistics recommendation method, effectively solving the existing data sparsity and cold start problems, and further improving the accuracy, diversity and interpretability of recommendations. By mining the historical interaction information of the consignor to obtain the long-term and short-term interests of the consignor, integrating the long-term and short-term interests to obtain the dynamic interests of the consignor, better mining the characteristics of the consignor, improving the interpretability of the consignor's interests, and more accurately recommending vehicles to the consignor, it can effectively increase the economic performance of cold chain logistics.

[0066] 2. The present invention filters out the most recent K historical vehicle sources that the consignor interacts with through time series, establishes a knowledge graph centered on the consignor's historical vehicle sources, and obtains the vector representation of the consignor through interest propagation, so as to more effectively utilize the knowledge graph to represent the consignor's short-term interests.

[0067] 3. The present invention uses the attention mechanism to simulate the information attenuation characteristics in the process of knowledge graph information propagation, which can more accurately capture high-order semantic relationships to enrich the cargo owner representation, and extract vehicle source features through graph convolutional neural networks. This method not only takes into account the user's information, but also comprehensively considers the information of the relationships and tail entities in the knowledge graph, and has good interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is the overall flow chart of the present invention;

[0069] Figure 2 The process of constructing a cold chain logistics vehicle source knowledge graph for the present invention;

[0070] Figure 3 The present invention mines the dynamic interest expression process of the cargo owner;

[0071] Figure 4 This is the process of extracting vehicle source features in the present invention. DETAILED DESCRIPTION

[0072] The following is combined with Figure 1-4 To further illustrate the present invention, it should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent modifications of the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0073] Step 1: Clean and organize the cold chain logistics vehicle source information, define the cold chain logistics vehicle source knowledge graph framework, and construct the required knowledge graph G, as follows: Figure 2 As shown:

[0074] Step 1.1: According to the vehicle source data of the vehicle-free carrier platform, the entities in the cold chain logistics vehicle source knowledge graph are mainly: license plate number, vehicle model, load capacity, refrigeration capacity, and region. There are relationships between different entity classes, and the relationship category is defined as the category of the entity to which the relationship points.

[0075] Step 1.2: Extract vehicle source data from the database of the truckless carrier platform, obtain entity tables and relationship tables based on pre-defined entity classifications, and store them in the database.

[0076] Step 1.3: According to the cold chain logistics vehicle source entities and relationships in the database, the cold chain logistics vehicle source knowledge graph G is constructed through the graph database Neo4j. The cold chain logistics vehicle source knowledge graph G consists of triples (h, r, t), where h∈E, t∈E, r∈R represent the head entity, tail entity and relationship between entities of the triple, respectively, and E and R represent the entity and relationship sets in the knowledge graph G, respectively.

[0077] Step 2: Use the attention mechanism to fuse the historical vehicle source information of the cargo owner, and use the fused feature vector as the cargo owner's long-term interest representation u l The interest propagation algorithm is used to mine the hierarchical interest representation of the consignor, and the multi-layer interest representation is used as the short-term interest representation of the consignor u s , integrating the short-term interest of shippers s Expressing long-term and short-term interests with cargo owners l Get the cargo owner's dynamic interest expression d , such as Figure 3 As shown:

[0078] Step 2.1: Get the complete historical vehicle source set H = {v1, v2, ..., v n};

[0079] Step 2.2: Initialize the historical vehicle source vector v i and the candidate vehicle source vector v;

[0080] Step 2.3: Define the loop variable i, assign an initial value of 1, and define the long-term interest of the shipper as u l ;

[0081] Step 2.4: If i≤len(H), go to step 2.5, otherwise jump to step 2.10;

[0082] Step 2.5: Translate the historical vehicle source vector v i It is concatenated with the candidate vehicle source vector v and input into the attention network to obtain the attention weight a′ i , where a′ i =σ(W a [v i , v]+b a ), W a is the training weight, b a is the bias vector;

[0083] Step 2.6: Normalize the fusion weight a through the softmax function i ,in

[0084] Step 2.7: According to the fusion weight a i and the vehicle source vector v i The product is the current vehicle source fusion vector C i ;

[0085] Step 2.8: l =u l +C i ;

[0086] Step 2.9: i=i+1, jump to step 2.4;

[0087] Step 2.10: End the loop and get the final long-term interest representation u of the consignor l ;

[0088] Step 2.11: Filter the K most recently interacted vehicle sources from the shipper interaction sequence;

[0089] Step 2.12: Define loop variable i2, with an initial value of 1, and define the number of propagation layers, with a value of h;

[0090] Step 2.13: Take K vehicle sources as the knowledge graph seed set

[0091] Step 2.14: If i2≤h, go to step 2.15, otherwise jump to step 2.20;

[0092] Step 2.15: Obtain the i2-layer neighbor node set of the cargo owner u through propagation according to the entity

[0093] Step 2.16: The interest set of the i2 layer of the cargo owner u is

[0094] Step 2.17: By calculating the candidate vehicle source vector v and the interest set Middle head entity h i In the relational space i The correlation below gets the intra-layer fusion weight Where R i ∈R d 、h i ∈R d and v∈R d For the relationship i 、Head entity h i and the embedding representation of item v;

[0095] Step 2.18: Weight the tail entity vector representation in the interest set to obtain the i2-layer cargo owner interest representation in

[0096] Step 2.19: i2=i2+1, jump to step 2.14;

[0097] Step 2.20: End the loop and get the short-term interest set of the shipper

[0098] Step 2.21: Define loop variable i3, assign an initial value of 1, and define the short-term interest representation u of the cargo owner s ;

[0099] Step 2.22: If i3≤len(O1), go to step 2.23, otherwise jump to step 2.28;

[0100] Step 2.23: Vector representation of the shipper’s short-term interest set O1 And the candidate vehicle source representation v is input into the attention network to calculate the attention weight in W h is the training weight, b h is the bias vector;

[0101] Step 2.24: Normalize the fusion weights using the softmax function in

[0102] Step 2.25: According to the fusion weight and the vector representation of the short-term interest set of the shipper The product is used to obtain the current shipper's interest fusion vector

[0103] Step 2.26:

[0104] Step 2.27: i3=i3+1, jump to step 2.22;

[0105] Step 2.28: End the loop and get the final cargo owner's short-term interest representation u s ;

[0106] Step 2.29: Denote the long-term interest of the cargo owner as u l and short-term interest representation u s Input into the fully connected neural network to obtain the dynamic interest representation of the cargo owner o d =W b [u l ,u s ], where W b is the training weight.

[0107] Step 3: Use the knowledge graph convolutional neural network to mine the high-order structural information of the knowledge graph to obtain the vehicle source feature representation v u , such as Figure 4 As shown:

[0108] Step 3.1: Use graph convolutional neural network to extract vehicle source features and input the dynamic interest representation of the cargo owner d ;

[0109] Step 3.2: Define loop variable i4, assign an initial value of 1, define v0 as the initial feature vector of candidate vehicle source v, and set the number of aggregation layers to H;

[0110] Step 3.3: If i4≤H, go to step 3.4, otherwise jump to step 3.12;

[0111] Step 3.4: Define loop variable i5, assign an initial value of 1, define N(v) as the neighbor set of the entity, and the length of the set is M;

[0112] Step 3.5: If i5≤M, go to step 3.6, otherwise jump to step 3.8;

[0113] Step 3.6: Based on the relationship r between entities and dynamic interest o d Calculate the shipper’s preference for the relationship, and the calculation formula is:

[0114] Step 3.7: i5=i5+1, jump to step 3.5;

[0115] Step 3.8: End the loop and obtain the fusion weights of different entities through normalization The calculation formula is

[0116] Step 3.9: Use linear combination of neighbor information to obtain the neighborhood information of the node The calculation formula is

[0117] Step 3.10: Aggregate the current vehicle source embedding representation and neighborhood information to obtain a new vehicle source embedding representation. The calculation formula is: Where W c represents the learnable weight, b c represents the bias vector, σ represents the activation function;

[0118] Step 3.11: i4=i4+1, jump to step 3.3;

[0119] Step 3.12: End the loop and get the final vehicle source feature representation v u , where v u =v H .

[0120] Step 4: Dynamic interest expression based on cargo owner d and vehicle source feature representation v u Recommended cold chain logistics vehicles:

[0121] Step 4.1: The final cargo owner's dynamic interest representation d And the final car source representation v u Perform inner product to predict the interaction probability between cargo owners and vehicle sources

[0122] Step 4.2: Through the loss function

[0123] Update the model parameters and optimize the model. The first item is the cross entropy loss, the second item is the knowledge graph embedding unit loss, λ1 is the loss value parameter, and the third item is the regularization loss, λ2 is the regularization parameter.

[0124] The present invention can be combined with a computer system to become a cold chain logistics vehicle source recommendation device based on dynamic interests and knowledge graphs. The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the above-mentioned cold chain logistics vehicle source recommendation method based on dynamic interests and knowledge graphs is implemented.

[0125] The above embodiments are only for illustrating the technical concept and features of the present invention, and their purpose is to enable people familiar with the technology to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent transformation or modification made according to the spirit of the present invention should be included in the protection scope of the present invention.

Claims

1. A cold chain logistics vehicle source recommendation method based on dynamic interests and knowledge graph, characterized in that: The steps include: Step 1: Clean and organize the cold chain logistics vehicle source information, define the cold chain logistics vehicle source knowledge graph framework, and construct the required knowledge graph G; Step 2: Use the attention mechanism to fuse the historical vehicle source information of the cargo owner, and use the fused feature vector as the cargo owner's long-term interest representation u l The interest propagation algorithm is used to mine the hierarchical interest representation of the consignor, and the multi-layer interest representation is used as the short-term interest representation u of the consignor. s , integrating the short-term interest of shippers s Long-term interest expressed with cargo owners l Get the cargo owner's dynamic interest expression d ; The specific method of obtaining the long-term interest expression of the cargo owner is: Step 2.1: Get the complete historical vehicle source set H = {v1,v2,...,v n }; Step 2.2: Initialize the historical vehicle source vector v i and the candidate vehicle source vector v; Step 2.3: Define the loop variable i, assign an initial value of 1, and define the long-term interest of the shipper as u l ; Step 2.4: If i≤len(H), go to step 2.5, otherwise jump to step 2.10; Step 2.5: Translate the historical vehicle source vector v i It is concatenated with the candidate vehicle source vector v and input into the attention network to obtain the attention weight a′ i , where a′ i =σ(W a [v i ,v]+b a ), W a is the training weight, b a is the bias vector; Step 2.6: Normalize the fusion weight a through the softmax function i ,in Step 2.7: According to the fusion weight a i and the historical vehicle source vector v i The product of gets the current vehicle source fusion vector C i ; Step 2.8:u l =u l +C i ; Step 2.9: i=i+1, jump to step 2.4; Step 2.10: End the loop and get the final long-term interest representation u of the consignor l ; The steps to obtain short-term expressions of interest from cargo owners include: Step 2.11: Filter the K most recently interacted vehicle sources from the shipper interaction sequence; Step 2.12: Define loop variable i2, with an initial value of 1, and define the number of propagation layers, with a value of h; Step 2.13: Take K vehicle sources as the knowledge graph seed set Step 2.14: If i2≤h, go to step 2.15, otherwise jump to step 2.20; Step 2.15: Obtain the i2-layer neighbor node set of the shipper u according to the entity propagation method Step 2.16: The interest set of the i2 layer of the cargo owner u is Among them, h∈E, t∈E, r∈R represent the head entity, tail entity and relationship between entities of the triple respectively, and E and R represent the entity and relationship sets in the knowledge graph G respectively; Step 2.17: By calculating the candidate vehicle source vector v and the interest set Middle head entity h i In the relational space i The correlation below gets the intra-layer fusion weight in and For the relationship i 、Head entity h i and the embedding representation of the candidate vehicle source vector v; Step 2.18: Weight the tail entity vector representation in the interest set to obtain the i2-layer cargo owner interest representation in Step 2.19: i2=i2+1, jump to step 2.14; Step 2.20: End the loop and get the short-term interest set of the shipper Step 2.21: Define loop variable i3, assign an initial value of 1, and define the cargo owner's short-term interest representation u s ; Step 2.22: If i3≤len(O1), go to step 2.23, otherwise jump to step 2.28; Step 2.23: Vector representation of the shipper’s short-term interest set O1 And the candidate vehicle source vector v is input into the attention network to calculate the attention weight in W h is the training weight, b h is the bias vector; Step 2.24: Normalize the fusion weights using the softmax function in Step 2.25: According to the fusion weight and the vector representation of the short-term interest set of the shipper The product of gets the current consignor's interest fusion vector Step 2.26: Step 2.27: i3=i3+1, jump to step 2.22; Step 2.28: End the loop and get the final cargo owner's short-term interest representation u s ; Fusion cargo owners short-term interest expression s Long-term interest expressed with cargo owners l Get the cargo owner's dynamic interest expression d The specific operations are: The long-term interest of the cargo owner is expressed as u l and short-term interest representation u s Input into the fully connected neural network to obtain the dynamic interest representation of the cargo owner o d =W b [u l ,u s ], where W b is the training weight; Step 3: Use graph convolutional neural networks to mine high-order structural information of the knowledge graph to obtain vehicle source feature representation v u ; Step 3.1: Use graph convolutional neural network to extract vehicle source features and input the dynamic interest representation of the cargo owner d ; Step 3.2: Define loop variable i4, assign an initial value of 1, define v0 as the initial feature vector of candidate vehicle source vector v, and set the number of aggregation layers to Q layers; Step 3.3: If i4≤Q, go to step 3.4, otherwise jump to step 3.12; Step 3.4: Define loop variable i5, assign an initial value of 1, define N(v) as the neighbor set of the entity, and the length of the set is M; Step 3.5: If i5≤M, go to step 3.6, otherwise jump to step 3.8; Step 3.6: Based on the relationship between entities r and the dynamic interest representation o of the cargo owner d Calculate the shipper’s preference for the relationship, and the calculation formula is: Step 3.7: i5=i5+1, jump to step 3.5; Step 3.8: End the loop and obtain the fusion weights of different entities through normalization The calculation formula is Step 3.9: Use linear combination of neighbor information to obtain the neighborhood information of the node The calculation formula is Step 3.10: Aggregate the current vehicle source embedding representation and neighborhood information to obtain a new vehicle source embedding representation. The calculation formula is: Where W c represents the learnable weight, b c represents the bias vector, σ represents the activation function; Step 3.11: i4=i4+1, jump to step 3.3; Step 3.12: End the loop and get the final vehicle source feature representation v u , where v u =v Q ; Step 4: According to the dynamic interest of the shipper d and vehicle source feature representation v u The internal area is used to recommend cold chain logistics vehicles.

2. The cold chain logistics vehicle source recommendation method based on dynamic interests and knowledge graph according to claim 1 is characterized in that: The specific method of step 1 is: Step 1.1: Analyze the entities in the cold chain logistics vehicle source knowledge graph based on the vehicle source data of the vehicle-free carrier platform, including: license plate number, vehicle model, load capacity, refrigeration capacity and region. There are relationships between different entity classes, and the relationship category is defined as the category of the entity to which the relationship points; Step 1.2: Extract vehicle source data from the database of the truckless carrier platform, obtain entity tables and relationship tables according to pre-defined entity classification, and store them in the database; Step 1.3: According to the cold chain logistics vehicle source entities and relationships in the database, the cold chain logistics vehicle source knowledge graph G is constructed through the graph database Neo4j. The cold chain logistics vehicle source knowledge graph G consists of triples (h, r, t), where h∈E, t∈E, r∈R represent the head entity, tail entity and relationship between entities of the triple, respectively, and E and R represent the entity and relationship sets in the knowledge graph G, respectively.

3. The cold chain logistics vehicle source recommendation method based on dynamic interests and knowledge graph according to claim 1 is characterized in that: The specific method of step 4 is: Step 4.1: The final cargo owner's dynamic interest representation d And the final vehicle source feature representation v u Perform inner product to predict the interaction probability between cargo owners and vehicle sources Step 4.2: Through the loss function Update the model parameters and optimize the model. The first item is the cross entropy loss, the second item is the knowledge graph embedding unit loss, λ1 is the loss value parameter, and the third item is the regularization loss, λ2 is the regularization parameter.

4. A cold chain logistics vehicle source recommendation device based on dynamic interests and knowledge graphs, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is loaded into the processor, the steps of the cold chain logistics vehicle source recommendation method based on dynamic interests and knowledge graphs described in any one of claims 1 to 3 are executed.

Images