Trusted O2O Service Recommendation Method in Heterogeneous Environments Based on Deep Learning

Abstract

The present invention discloses a method for recommending trustworthy O2O services in a heterogeneous environment based on deep learning. S1. Collect O2O service data in a heterogeneous environment from multiple channels to form a unified data matrix model; S2. Build a graph structure data model based on the unified data matrix model to generate an initial relationship graph; S3. Obtain graph convolutional node embedding representations; S4. Build a recommendation model according to the graph convolutional node embedding representations and form a recommendation candidate set; S5. Use the whale optimization algorithm to globally search and optimize the key hyperparameters of the recommendation model; S6. Generate personalized O2O service recommendation results based on real-time status. The present invention can adaptively adjust the information propagation range between different nodes, making the model more accurate in dealing with long-distance dependence relationships, and still maintaining a good recommendation effect in the case of data sparsity.

Description

Technical Field

[0001] The present invention relates to the technical field of O2O, and in particular, to a trusted O2O service recommendation method based on deep learning in a heterogeneous environment. Background Art

[0002] With the popularization of the Internet and intelligent mobile devices, the O2O service model (Online To Offline) has developed rapidly, covering multiple fields such as takeaways, travel, housekeeping, medical care, and education. The core of the O2O service lies in accurately matching user needs with offline services through an online platform to achieve efficient resource allocation and service experience. However, due to the complexity of the O2O service environment, including the personalization and diversification of user behaviors, as well as the dynamic changes in geographical location and time factors, traditional recommendation systems have many deficiencies in dealing with O2O service recommendation scenarios.

[0003] Currently, the mainstream O2O service recommendation methods mainly perform personalized recommendations based on collaborative filtering, matrix factorization, and deep learning technologies. Among them, the collaborative filtering method relies on historical user interaction data and makes recommendations by calculating the similarity between users or the implicit relationship between users and services. However, traditional methods perform poorly in data sparsity and cold start problems, especially in O2O environments with a large number of new users or new services, where the recommendation effect significantly decreases. In addition, traditional collaborative filtering methods are difficult to make full use of heterogeneous data of O2O services, resulting in limited recommendation accuracy.

[0004] The recommendation method based on matrix factorization constructs a user-service rating matrix and performs matrix factorization to learn latent feature representations, and has good performance in processing rating data. However, when facing data with multi-dimensional features such as time and space in the O2O scenario, its modeling ability is insufficient and it cannot effectively capture the dynamic changes of user preferences. In addition, since the matrix factorization method usually relies on existing rating data for training, the recommendation effect for unrated services or new users is not good.

[0005] In recent years, deep learning technologies have gradually been applied to recommendation systems. In particular, recommendation methods based on deep neural networks, convolutional neural networks, and graph neural networks have shown great potential in improving recommendation effects. However, existing deep learning recommendation methods still face the following challenges: First, most methods only model a single data source and cannot make full use of the heterogeneous data characteristics in the O2O environment, such as multi-modal information of user behaviors, service information, and geographical locations; Second, traditional deep learning methods are difficult to establish effective multi-level associations among users, services, and geographical locations when dealing with complex relationship data, thus affecting the accuracy of recommendations; Finally, existing methods often rely on empirical hyperparameter tuning in model optimization and lack an effective global search optimization mechanism, resulting in poor generalization ability and robustness of the recommendation model.

[0006] In summary, the existing O2O service recommendation technologies have many deficiencies in data fusion, heterogeneous information modeling, user interest prediction, and model optimization, and it is difficult to meet the requirements of efficient, accurate, and personalized recommendations. Therefore, there is an urgent need for a new recommendation method that can fully integrate various heterogeneous data, establish an efficient relationship modeling mechanism, and improve the accuracy and stability of recommendations through intelligent optimization means to better meet the needs of the O2O service scenario. Summary of the Invention

[0007] An object of the present invention is to propose a trustworthy O2O service recommendation method in a heterogeneous environment based on deep learning. The present invention can adaptively adjust the information propagation range between different nodes, making the model more accurate when dealing with long-distance dependence relationships, and still maintaining a good recommendation effect in the case of data sparsity.

[0008] A trustworthy O2O service recommendation method in a heterogeneous environment based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect O2O service data in a heterogeneous environment from multiple channels, and preprocess the O2O service data in the heterogeneous environment to form a unified data matrix model;

[0010] S2. Based on the unified data matrix model, construct a graph structure data model to generate an initial relationship graph;

[0011] S3. On the basis of the initial relationship graph, construct a multi-perspective adaptive graph diffusion convolutional network to obtain graph convolutional node embedding representations;

[0012] S4. According to the graph convolutional node embedding representations, construct a recommendation model, integrate and generate user interest distributions and service feature representations, and form a recommendation candidate set;

[0013] S5. Use the whale optimization algorithm to globally search and optimize the key hyperparameters of the recommendation model, and feedback the optimized parameters into the multi-perspective adaptive graph diffusion convolutional network to update and optimize the recommendation model;

[0014] S6. Based on the optimized recommendation model, calculate the matching degree and interest degree of the user for each service in the recommendation candidate set, and generate a personalized O2O service recommendation result based on the real-time state.

[0015] Optionally, the S1 includes the following steps:

[0016] S11. Collect O2O service data in a heterogeneous environment from multiple channels. The O2O service data in the heterogeneous environment includes user behavior data, service information data, and geographical location information. The user behavior data includes user browsing records, historical orders, service evaluations, and click behaviors. The service information data includes service categories, merchant characteristics, service prices, and service evaluation scores. The geographical location information includes the user's current location, historical location information, and service-providing geographical locations;

[0017] S12. According to the sources and data formats of the O2O service data in the heterogeneous environment, construct a data preprocessing model, and perform time format normalization, categorical feature encoding, and numerical feature normalization on the O2O service data in the heterogeneous environment to make the heterogeneous data from different sources conform to a unified representation form, and obtain the O2O service data in the heterogeneous environment after format conversion;

[0018] S13. Perform data cleaning on the O2O service data in the heterogeneous environment after format conversion, including mean filling, median filling, and nearest neighbor interpolation;

[0019] S14. Construct a unified data matrix model based on the O2O service data in the heterogeneous environment after format conversion and data cleaning The unified data matrix model includes a user behavior data matrix a service information data matrix and a geographical location information data matrix :

[0020]

[0021] Among them, is the user behavior data matrix, which contains behavior feature vectors of users is the service information data matrix, which contains feature vectors of services is the geographical location information data matrix, which contains feature vectors of geographical locations

[0022] S15. According to the unified data matrix model perform optimization processing on the high-dimensional data, so that the unified data matrix model retains key information while reducing the computational complexity, and store the optimized unified data matrix model in the data storage module and establish an index structure.

[0023] Optionally, the S2 includes the following steps:

[0024] S21. Define the user node set according to the user behavior data matrix Define the user node set , where the user node set includes the unique identifiers of all users;

[0025] S22. Define the service node set according to the service information data matrix Define the service node set , where the service node set includes the unique identifiers of all services;

[0026] S23. Define the geographical location node set according to the geographical location information data matrix Define the geographical location node set , where the geographical location node set includes the unique identifiers of all geographical locations;

[0027] S24. Construct the interaction relationship edge set between users and services according to the user behavior data matrix and the service information data matrix ; ;

[0028] S25. Construct the relationship edge set between services and geographical locations according to the service information data matrix and the geographical location information data matrix ; S26. Construct the relationship edge set between users and geographical locations according to the user behavior data matrix and the geographical location information data matrix ; ; , the edge weight matrix is composed of edge weights : Among them, and are the feature vectors of nodes extracted from the unified data matrix model and node respectively, reflecting the comprehensive characteristics of each node in user behavior, service information, and geographical location data, is the Euclidean distance between node feature vectors, is the scaling parameter that controls the sensitivity of the logic function, is the normalization factor of the feature distance, is the threshold parameter of the logic function, used to adjust the translation of the similarity measurement between nodes, represents the number of interactions between nodes recorded in the user behavior data matrix and node , is the maximum number of interactions for all node pairs and is used to normalize the number of interactions , represents the service information data matrix in the node and the node The corresponding average service evaluation score for the interaction between them, is the highest service evaluation score in the dataset and is used to normalize the average service evaluation score , is the geographical location information data matrix The geographical distance between the recorded nodes and the node in it, is a parameter for adjusting the influence degree of geographical distance, , , are the weight coefficients assigned to the normalized terms of interaction frequency, service evaluation, and geographical distance respectively; S28. Combine the user, service, and geographical location node sets constructed in steps S21 to S27 , the edge set and the edge weight matrix to construct a complete initial relationship graph .

[0029] Optionally, the S3 includes the following steps: S31. Based on the initial relationship graph G, construct a multi-perspective adaptive graph diffusion convolutional network. The multi-perspective adaptive graph diffusion convolutional network adopts a multi-perspective diffusion strategy of local diffusion module and global diffusion module to capture the information propagation characteristics of user, service, and geographical location nodes at different scales; S32. In the local diffusion module, define the local diffusion matrix according to the node type and the edge weight matrix :

[0030] Among them, is the set of adjacent nodes of the node in the local neighborhood, is the local diffusion adjustment factor, which is used to regulate the sensitivity of different types of nodes in local information propagation; S33. In the global diffusion module, define the global diffusion matrix according to the same edge weight matrix :

[0031]

[0032] Among them, is the set of nodes that expand the neighborhood of the node globally, It is a global diffusion regulator, which is used to reflect the influence of nodes in long-distance information transmission; S34. Design a gating fusion mechanism to dynamically fuse the diffusion information of the local diffusion matrix and the diffusion information of the global diffusion matrix, and set the gating coefficient :

[0033]

[0034] where is the feature representation of node in the th layer, is a trainable weight matrix, is a bias vector, is the Sigmoid activation function;

[0035] S35. Use multi-view adaptive graph diffusion convolution to calculate the graph convolution node embedding representation, and perform a fusion operation on the node feature representations and in the th layer:

[0036] where represents the feature matrix of the local diffusion module in the th layer, represents the feature matrix of the global diffusion module in the th layer, is the fused multi-view feature representation matrix, and are the trainable parameter matrices of the local and global diffusion modules in the th layer respectively, represents the element-wise product operation, is a non-linear activation function; S36. Adopt a skip diffusion mechanism to perform weighted fusion on information with different diffusion step lengths, optimize the multi-scale fusion effect of information transmission, introduce a skip diffusion mechanism in the multi-view adaptive graph diffusion convolution network, and obtain the final graph convolution node embedding representation:

[0037]

[0038] where is the final node feature matrix after skip diffusion, is the maximum diffusion step length, is the fusion weight of the information diffused at the th step.

[0039] Optionally, S4 includes the following steps: S41. Based on the final graph convolutional node embedding representation of the multi-view adaptive graph diffusion convolutional network, construct a recommendation model composed of a user interest representation model and a service feature representation model, and respectively apply an interest mapping function and a service node to obtain the final representation vectors: and a feature mapping function where,

[0040]

[0041] Among them, is the final interest representation vector of user , is the final feature representation vector of service , is the final fused node embedding representation matrix; S42. Based on the user interest representation and service feature representation, calculate the recommendation score between user and candidate service :

[0042]

[0043] wherein, represents the vector inner product operation, is the category score of service , is the service category regulation factor; S43. Sort the candidate services according to the recommendation score , and select the highest-scoring service sets to form a personalized O2O service recommendation candidate set.

[0044] Optionally, S5 includes the following steps:

[0045] S51. Determine the key hyperparameter set that affects the recommendation effect according to the recommendation model. The hyperparameter set includes the maximum diffusion step , the local diffusion adjustment factor , the global diffusion adjustment factor and the regularization coefficient :

[0046] S52. Use the whale optimization algorithm to optimize the hyperparameter set , initialize the population size , and set the initial hyperparameter candidate set: wherein, is the A group of hyperparameter individuals is generated based on random initialization or a preset range; S53. Based on the recommendation model, with the objective function Calculate the fitness of each group of hyperparameters .

[0047] S54. The whale optimization algorithm optimizes hyperparameters using the mechanisms of surrounding prey, spiral update, and random search. The mechanism of surrounding prey:

[0048]

[0049] where, is the current optimal hyperparameter, is the contraction coefficient, is the distance between the current position and the optimal solution;

[0050] The spiral update mechanism:

[0051]

[0052] where, , is the spiral shape control parameter, is a random number;

[0053] The random search mechanism:

[0054]

[0055] where, is a random individual in the population; S55. During the optimization process, set the convergence determination criterion, select the optimal hyperparameter:

[0056]

[0057] where, is the convergence threshold. When is less than continuously rounds, the algorithm terminates and outputs the optimal hyperparameter ;

[0058] S56. Feed the optimized hyperparameter back to the multi-view adaptive graph diffusion convolutional network to update the optimized recommendation model.

[0059] Optionally, the S53 includes the following steps: S531. Calculate the fitness evaluation index of the hyperparameter set based on the recommendation model. The fitness evaluation index consists of the hit rate , the normalized discounted cumulative gain and the loss function , Measure the proportion of correctly hitting the user's actual selected service among the top recommended results:

[0060] Among them, is the total number of user sets, represents the user 's actually selected service, is the top recommended service list for the user , is an indicator function, which takes the value of 1 if , otherwise it takes the value of 0;

[0061] Measure the sorting quality of the recommended list:

[0062]

[0063] Among them, is the discounted cumulative gain, indicating the relevance of the top recommended results:

[0064]

[0065] is the ideal discounted cumulative gain, indicating the optimal recommended sorting for the user :

[0066] Among them, is the true relevance of the user to the th service in the recommended results, is the true relevance of the user under the optimal sorting;

[0067] S532. Define the loss function of the recommendation model under the action of the hyperparameter : :

[0068]

[0069] Among them, is the regularization coefficient, is the regularization term, is the reconstruction loss of the recommendation model, calculating the error of the model in the recommendation task:

[0070] ;

[0071] Among them, is the regularization term, is the user The true score of the service is the predicted score of the recommendation model;

[0072] S533. Define the hyperparameter set of the target optimization function according to the fitness evaluation index:

[0073] ;

[0074] Among them, is the weight coefficient of the fitness function. The beneficial effects of the present invention are:

[0075] By constructing an initial relationship graph, the present invention unifies the modeling of user behavior data, service information data, and geographical location information, and uses a multi-view adaptive graph diffusion convolutional network to enhance the modeling ability of complex relationship data. Specifically, MV-ADGCN adopts a local diffusion module and a global diffusion module to capture the near-neighbor interaction relationship and the long-range potential connection respectively, and then forms a multi-scale information fusion mechanism. Traditional graph convolutional networks usually rely on a single diffusion strategy and are difficult to balance local and global information propagation. The multi-view diffusion strategy of the present invention can adaptively adjust the information propagation path, so that user interest features, service features, and geographical location information can be more accurately expressed at different levels.

[0076] In the optimization stage of the recommendation model, the present invention introduces the whale optimization algorithm to globally search and optimize the key hyperparameters affecting the recommendation effect. The whale optimization algorithm alternates three strategies: surrounding the prey, spiral updating, and random search, so as to effectively avoid the local optimal trap of traditional optimization methods and improve the hyperparameter search efficiency and the generalization ability of the recommendation model.

[0077] Based on the graph neural network, the present invention adopts a multi-layer jump diffusion mechanism, allowing information with different diffusion steps to be weighted and fused, thereby enhancing the multi-scale transmission ability of information. Compared with the diffusion method with a fixed step size, the present invention can adaptively adjust the information propagation range between different nodes, making the model more accurate in dealing with long-distance dependence relationships and still maintaining a good recommendation effect in the case of data sparsity. Description of the Drawings

[0078] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation to the present invention. In the drawings:

[0079] Figure 1 is a flowchart of a trustworthy O2O service recommendation method based on deep learning proposed by the present invention. Detailed Embodiments

[0080] Now, the present invention will be further described in detail with reference to the accompanying drawings. These drawings are all simplified schematic diagrams, only schematically showing the basic structure of the present invention, so they only show the components related to the present invention.

[0081] Reference Figure 1 , a trusted O2O service recommendation method in a heterogeneous environment based on deep learning, includes the following steps:

[0082] S1. Collect O2O service data in a heterogeneous environment from multiple channels, and preprocess the O2O service data in a heterogeneous environment to form a unified data matrix model;

[0083] S2. Construct a graph structure data model based on the unified data matrix model to generate an initial relationship graph;

[0084] S3. Construct a multi-perspective adaptive graph diffusion convolutional network based on the initial relationship graph to obtain graph convolutional node embedding representations;

[0085] S4. Construct a recommendation model according to the graph convolutional node embedding representations, integrate and generate user interest distributions and service feature representations, and form a recommendation candidate set;

[0086] S5. Use the whale optimization algorithm to globally search and optimize the key hyperparameters of the recommendation model, and feedback the optimized parameters to the multi-perspective adaptive graph diffusion convolutional network to update and optimize the recommendation model;

[0087] S6. Calculate the matching degree and interest degree of the user for each service in the recommendation candidate set based on the optimized recommendation model, and generate personalized O2O service recommendation results based on the real-time status.

[0088] In this embodiment, S1 includes the following steps:

[0089] S11. Collect O2O service data in a heterogeneous environment from multiple channels. The O2O service data in a heterogeneous environment includes user behavior data, service information data, and geographical location information. The user behavior data includes user browsing records, historical orders, service evaluations, and click behaviors. The service information data includes service categories, merchant characteristics, service prices, and service evaluation scores. The geographical location information includes the user's current location, historical location information, and service-providing geographical location;

[0090] S12. According to the sources and data formats of the O2O service data in a heterogeneous environment, construct a data preprocessing model, perform time format normalization, categorical feature encoding, and numerical feature normalization on the O2O service data in a heterogeneous environment, so that heterogeneous data from different sources conforms to a unified representation form, and obtain the O2O service data in a heterogeneous environment after format conversion;

[0091] S13. Clean the O2O service data in the heterogeneous environment after format conversion, including mean filling, median filling, and nearest neighbor interpolation;

[0092] S14. Construct a unified data matrix model based on the O2O service data in the heterogeneous environment after format conversion and data cleaning , and the unified data matrix model includes a user behavior data matrix , a service information data matrix , and a geographical location information data matrix :

[0093]

[0094] Among them, is the user behavior data matrix, which contains behavioral feature vectors of users, is the service information data matrix, which contains feature vectors of services, is the geographical location information data matrix, which contains feature vectors of geographical locations;

[0095] S15. Based on the unified data matrix model , optimize the high-dimensional data so that the unified data matrix model retains key information while reducing the computational complexity, and store the optimized unified data matrix model in the data storage module and establish an index structure.

[0096] In this embodiment, S2 includes the following steps:

[0097] S21. Define a user node set based on the user behavior data matrix , and the user node set includes the unique identifiers of all users;

[0098] S22. Define a service node set based on the service information data matrix , and the service node set includes the unique identifiers of all services;

[0099] S23. Define a geographical location node set based on the geographical location information data matrix , and the geographical location node set includes the unique identifiers of all geographical locations;

[0100] S24. Based on the user behavior data matrix and the service information data matrix Construct the set of interaction relationship edges between users and services ;

[0101] S25. Based on the service information data matrix and the geographical location information data matrix Construct the set of relationship edges between services and geographical locations ;

[0102] S26. Based on the user behavior data matrix and the geographical location information data matrix Construct the set of relationship edges between users and geographical locations ;

[0103] S27. Calculate the edge weight matrix by integrating the nodes and edge sets in steps S21 - S26 , the edge weight matrix consists of edge weights : Among them, and are the eigenvectors of nodes extracted from the unified data matrix model and node respectively, reflecting the comprehensive characteristics of each node in user behavior, service information, and geographical location data. is the Euclidean distance between node eigenvectors, is the scaling parameter that controls the sensitivity of the logic function, is the normalization factor of the feature distance, is the threshold parameter of the logic function, used to adjust the translation of the similarity measure between nodes, represents the number of interactions between nodes recorded in the user behavior data matrix and node , is the maximum number of interactions among all node pairs, used to normalize the number of interactions , represents the average service evaluation score corresponding to the interaction between nodes in the service information data matrix and node , is the highest service evaluation score in the dataset, used to normalize the average service evaluation score , is the geographical distance between nodes recorded in the geographical location information data matrix and node , is the parameter that adjusts the influence degree of the geographical distance, , , are the weight coefficients for the normalized terms of interaction frequency, service evaluation, and geographical distance, respectively;

[0104] S28. Combine the user, service, and geographical location node sets , edge set and edge weight matrix constructed in steps S21 to S27 to build a complete initial relationship graph .

[0105] In this embodiment, S3 includes the following steps: S31. Based on the initial relationship graph G, construct a multi-perspective adaptive graph diffusion convolutional network. The multi-perspective adaptive graph diffusion convolutional network adopts a multi-perspective diffusion strategy of local diffusion module and global diffusion module to capture the information propagation characteristics of user, service, and geographical location nodes at different scales;

[0106] S32. In the local diffusion module, define the local diffusion matrix :

[0107] where is the set of adjacent nodes of node in the local neighborhood, is the local diffusion adjustment factor used to regulate the sensitivity of different types of nodes in local information propagation;

[0108] S33. In the global diffusion module, define the global diffusion matrix :

[0109]

[0110] where is the set of nodes that expand the neighborhood of node globally, is the global diffusion adjustment factor used to reflect the influence of nodes in long-distance information transmission;

[0111] S34. Design a gating fusion mechanism to dynamically fuse the diffusion information of the local diffusion matrix and the diffusion information of the global diffusion matrix, and set the gating coefficient :

[0112] where is the feature representation of node in the th layer, is the trainable weight matrix, is the bias vector, is the Sigmoid activation function; S35. Use multi-view adaptive graph diffusion convolution to calculate the graph convolution node embedding representation, and perform a fusion operation on the node feature representation of the layer and :

[0113]

[0114] Among them, represents the feature matrix of the local diffusion module of the layer, represents the feature matrix of the global diffusion module of the layer, is the fused multi-view feature representation matrix, and are the trainable parameter matrices of the local and global diffusion modules in the layer respectively, represents the element-wise product operation, is the non-linear activation function;

[0115] S36. Adopt the skip diffusion mechanism to perform weighted fusion on the information of different diffusion steps, optimize the multi-scale fusion effect of information transmission, introduce the skip diffusion mechanism in the multi-view adaptive graph diffusion convolution network, and obtain the final graph convolution node embedding representation:

[0116]

[0117] Among them, is the final node feature matrix after skip diffusion, is the maximum diffusion step, is the step diffusion information fusion weight. In this embodiment, S4 includes the following steps:

[0118] S41. Based on the final graph convolution node embedding representation of the multi-view adaptive graph diffusion convolution network, construct a recommendation model composed of a user interest representation model and a service feature representation model, and use the interest mapping function and the feature mapping function for the user node and the service node to obtain the final representation vector:

[0119]

[0120]

[0121] Among them, is the final interest representation vector of the user , is the final feature representation vector of the service ; and is the node embedding representation matrix of the final fusion

[0122] S42. Calculate the recommendation score between the user and the candidate service based on the user interest representation and the service feature representation: wherein, represents the inner product operation of vectors, is the category score of the service ; and is the service category regulation factor

[0123] S43. Sort the candidate services according to the recommendation score and select the service set with the highest score to form a personalized O2O service recommendation candidate set

[0124] In this embodiment, S5 includes the following steps: S51. Determine the key hyperparameter set that affects the recommendation effect according to the recommendation model. The hyperparameter set includes the maximum diffusion step , the local diffusion adjustment factor , the global diffusion adjustment factor and the regularization coefficient :

[0125]

[0126] S52. Optimize the hyperparameter set using the whale optimization algorithm. Initialize the population size and set the initial hyperparameter candidate set:

[0127]

[0128] wherein, is the th group of hyperparameter individuals, which are generated according to random initialization or a preset range

[0129] S53. Based on the recommendation model, calculate the fitness of each group of hyperparameters using the objective function

[0130] S54. The whale optimization algorithm uses the mechanisms of surrounding prey, spiral update and random search to optimize the hyperparameters. The mechanism of surrounding prey:

[0131]

[0132] wherein,​ is the current optimal hyperparameter, is the contraction coefficient, is the distance between the current position and the optimal solution;

[0133] Spiral update mechanism:

[0134]

[0135] Among them, , is the spiral shape control parameter, is a random number;

[0136] Random search mechanism:

[0137]

[0138] Among them, is a random individual in the population;

[0139] S55. During the optimization process, set the convergence judgment criterion, and select the optimal hyperparameter:

[0140] Among them, is the convergence threshold. When is less than continuously rounds, the algorithm terminates and outputs the optimal hyperparameter ; S56. Feed back the optimized hyperparameter to the multi-view adaptive graph diffusion convolutional network to update the optimized recommendation model.

[0141] In this embodiment, S53 includes the following steps:

[0142] S531. Calculate the fitness evaluation index of the hyperparameter set according to the recommendation model. The fitness evaluation index consists of the hit rate , the normalized discounted cumulative gain and the loss function . measures the proportion of correctly hitting the user's actual selected service in the first recommended results:

[0143] Among them, is the total number of user sets, represents the service actually selected by user , is the first services list recommended to user , is an indicator function, which takes the value 1 if and 0 otherwise;

[0144] measures the sorting quality of the recommendation list:

[0145] where, is the discounted cumulative gain, indicating the relevance of the first recommended results:

[0146]

[0147] is the ideal discounted cumulative gain, indicating the optimal recommendation sorting for user ;

[0148]

[0149] where, is the true relevance of user to the th service in the recommendation result, is the true relevance of user under the optimal sorting;

[0150] S532. Define the loss function of the recommendation model under the action of the hyperparameter : :

[0151]

[0152] where, is the regularization coefficient, is the regularization term, is the reconstruction loss of the recommendation model, calculating the error of the model in the recommendation task:

[0153] ; where, is the regularization term, is the true score of user for service , is the predicted score of the recommendation model;

[0154] S533. Define the objective optimization function of the hyperparameter set according to the fitness evaluation index:

[0155] ;

[0156] where, is the weight coefficient of the fitness function. Example

[0157] At 23:15 PM on March 10, 2024, inside an Internet company in District C, City B, programmer Mr. Zhang was working overtime at the company. When he felt hungry while writing code, he opened a well-known food delivery platform (hereinafter referred to as "Platform A") hoping to quickly find late-night meals that suited his taste. However, since it was relatively late, many merchants had closed. Traditional recommendation methods usually directly recommended a few still-open merchants, lacking pertinence and unable to match the real needs of users, resulting in recommended results that were often options users didn't like or outdated popular recommendations.

[0158] Previously, Mr. Zhang ordered food more often between 19:00 - 20:30 in the evening and preferred Japanese cuisine, barbecue, and Chinese fast food. However, he had never ordered food after 23:00, which was a typical "cold start period problem". If the traditional collaborative filtering recommendation method was used, the system would:

[0159] 1. Based on his past order history, recommend the Japanese cuisine or barbecue he often ordered, but most of these merchants had closed at this time;

[0160] 2. Based on popular merchants, recommend the stores with the highest night-time sales, but these stores might not match his taste.

[0161] The limitations of traditional methods are:

[0162] Ignoring the changes in user behavior in the time dimension and unable to identify the special needs of users in the late-night scenario;

[0163] Not making full use of geographical location information and failing to screen merchants that are closer and have fast delivery speeds;

[0164] Lacking the ability of dynamic optimization and unable to adjust the recommendation strategy according to the user's current state;

[0165] To improve the recommendation quality, Platform A adopted the method of the present invention, analyzed Mr. Zhang's needs more accurately based on the multi-perspective adaptive graph diffusion convolutional network and the whale optimization algorithm, and provided recommendations that better met his actual needs;

[0166] After Mr. Zhang opened the APP, the method of the present invention first collected his real-time status data:

[0167] Time factor: The current time is 23:15, which is an atypical meal ordering period;

[0168] Geographical location information: The user is located in a technology company in District C (longitude 116.3217, latitude 40.0512);

[0169] Historical behavior analysis:

[0170] The past dinner order times were concentrated between 19:00 and 20:30, with a preference for Japanese cuisine, barbecue, and fast food;

[0171] Among recent orders, he is more inclined to merchants with "fast delivery speed", and the average delivery time requirement is within 25 minutes;

[0172] Comparing the user rating data of recent orders, he gives lower ratings to merchants with "low food temperature" or "delivery timeout";

[0173] The method of the present invention combines multi-modal data to optimize the recommendation strategy personalizedly:

[0174] Modeling using a multi-view adaptive graph diffusion convolutional network:

[0175] In the interaction relationship graph of user-merchant-geographical location, automatically identify merchants that are "still open at night, have fast delivery speed, and have high user ratings";

[0176] Combined with historical data, calculate the "common interest characteristics of late-night ordering users", and extract preference patterns suitable for night consumption (such as higher ratings for merchants in the "late-night fast food" category);

[0177] Since Mr. Zhang has never ordered food after 23:00, the parameters of traditional recommendation methods cannot adapt to this new scenario. Therefore, this method uses the whale optimization algorithm to automatically adjust the hyperparameters:

[0178] Increase the weight of merchants with "higher late-night order volume" to enhance the matching degree of recommendations;

[0179] Reduce the weight of merchants with "too long delivery time" to ensure that the takeaway can be delivered as soon as possible;

[0180] Appropriately increase the priority of "fast food, bento, and late-night snack merchants" to match the late-night consumption habits;

[0181] After this method is optimized, the following appear in the recommendation list:

[0182] "A certain bento house" 1.2 km away (rating 4.8, estimated delivery time 18 minutes);

[0183] "A certain barbecue takeaway" 1.5 km away (rating 4.7, estimated delivery time 22 minutes);

[0184] "A certain fried chicken and hamburger store" 2.0 km away (rating 4.5, estimated delivery time 20 minutes);

[0185] In contrast, the list of traditional recommendation methods includes:

[0186] "A certain online celebrity Japanese cuisine store" 3.5 km away (already closed, mis-recommended);

[0187] "A certain fast food chain" 2.8 km away (rating 3.8, delivery takes 35 minutes);

[0188] "A certain spicy hot pot restaurant" 1.9 km away (rating 4.2, delivery time 28 minutes);

[0189] Mr. Zhang finally chose "A certain bento house". He received the takeaway 18 minutes later. The food temperature was appropriate, the experience was good, and he gave a 5-star positive review.

[0190] To further verify the superiority of the method of the present invention, in the late-night meal ordering scenario, the recommendation effects of the method of the present invention and the traditional method were compared. The experimental results show that:

[0191] 1. The recommendation matching degree is increased by 21.4%. The method of the present invention can more accurately match the meal ordering needs of late-night users;

[0192] 2. The user order placement time is reduced by 63%, indicating that the recommendation results are more in line with user expectations and reduce the difficulty of selection;

[0193] 3. The order conversion rate is increased by 22.1%, proving that the method of the present invention is more attractive to both merchants and users;

[0194] 4. The user rating is increased by 0.6 points, indicating that the recommendation quality is improved and the user experience is better;

[0195] 5. The average delivery time is shortened by 8 minutes, indicating that the method of the present invention effectively screens merchants with faster delivery and improves the delivery efficiency;

[0196] This embodiment demonstrates the practical application of the method of the present invention in the late-night meal ordering scenario. Through the multi-perspective adaptive graph diffusion convolution network and the whale optimization algorithm, the method of the present invention can identify the real-time status of users (time, geographical location, historical behavior); dynamically adjust the recommendation parameters to adapt to the cold start period; provide more accurate, fast, and personalized O2O recommendation results;

[0197] The experimental data show that the method of the present invention significantly improves the recommendation matching degree, order conversion rate, and user satisfaction, reduces the user's decision-making time, and optimizes the takeaway delivery efficiency, providing a more intelligent and accurate solution for the O2O service recommendation system.

[0198] The present invention constructs an initial relationship graph, unifies the modeling of user behavior data, service information data, and geographical location information, and uses a multi-perspective adaptive graph diffusion convolutional network to enhance the modeling ability of complex relationship data. Specifically, MV-ADGCN adopts a local diffusion module and a global diffusion module to capture the neighbor interaction relationship and long-range potential connection respectively, and then forms a multi-scale information fusion mechanism. Traditional graph convolutional networks usually rely on a single diffusion strategy and it is difficult to balance local and global information propagation. However, the multi-perspective diffusion strategy of the present invention can adaptively adjust the information propagation path, enabling more accurate expression of user interest features, service features, and geographical location information at different levels.

[0199] In the optimization stage of the recommendation model, the present invention introduces the whale optimization algorithm to globally search for and optimize the key hyperparameters that affect the recommendation effect. The whale optimization algorithm alternately uses three strategies: surrounding the prey, spiral updating, and random search, thus effectively avoiding the local optimal trap of traditional optimization methods and improving the hyperparameter search efficiency and the generalization ability of the recommendation model.

[0200] Based on the graph neural network, the present invention adopts a multi-layer jump diffusion mechanism, allowing weighted fusion of information with different diffusion step lengths, thereby enhancing the multi-scale information transmission ability. Compared with the diffusion method with a fixed step length, the present invention can adaptively adjust the information propagation range between different nodes, making the model more accurate in dealing with long-range dependence relationships and still maintaining a good recommendation effect in the case of data sparsity.

[0201] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A method for recommending trusted O2O services in a heterogeneous environment based on deep learning, characterized in that: The steps include: S1. Collect O2O service data in a heterogeneous environment from multiple channels, and pre-process the O2O service data in the heterogeneous environment to form a unified data matrix model; S2. Build a graph structure data model based on the unified data matrix model to generate an initial relationship graph. The unified data matrix model includes a user behavior data matrix D u , service information data matrix D s And the geographic location information data matrix D g : D=D u ∪D s ∪D g ; Among them, D u is the user behavior data matrix, containing the behavior feature vector u of user i i , D s is the service information data matrix, containing the feature vectors s of j services j , D g is the geographic location information data matrix, containing the feature vectors g of k geographic locations k ; S3. Based on the initial relationship graph, a multi-view adaptive graph diffusion convolution network is constructed to obtain a graph convolution node embedding representation. The multi-view adaptive graph diffusion convolution network adopts a multi-view diffusion strategy of a local diffusion module and a global diffusion module to capture the information propagation characteristics of users, services and geographic location nodes at different scales; S4. constructing a recommendation model based on the graph convolution node embedding representation, integrating and generating user interest distribution and service feature representation, and forming a recommendation candidate set; S5. Use the whale optimization algorithm to globally search and optimize the key hyperparameters of the recommendation model, and feed the optimized parameters back to the multi-view adaptive graph diffusion convolutional network to update the optimized recommendation model; S6. Calculate the user's matching degree and interest in each service in the recommendation candidate set based on the optimized recommendation model, and generate personalized O2O service recommendation results based on real-time status.

2. According to claim 1, a method for recommending trusted O2O services in a heterogeneous environment based on deep learning is characterized in that: The S1 comprises the following steps: S11. Collect O2O service data in a heterogeneous environment from multiple channels, wherein the O2O service data in the heterogeneous environment includes user behavior data, service information data and geographic location information, wherein the user behavior data includes user browsing history, historical orders, service evaluation and click behavior, the service information data includes service category, merchant characteristics, service price and service evaluation score, and the geographic location information includes the user's current location, historical location information and the geographic location of service provision; S12. According to the source and data format of the O2O service data in the heterogeneous environment, a data preprocessing model is constructed to normalize the time format, classify the feature encoding and normalize the numerical features of the O2O service data in the heterogeneous environment, so that the heterogeneous data from different sources conform to a unified representation form, and the O2O service data in the heterogeneous environment after format conversion is obtained; S13. Perform data cleaning on the O2O service data in the heterogeneous environment after format conversion, including mean filling, median filling and nearest neighbor interpolation; S14. Constructing a unified data matrix model D based on the O2O service data in a heterogeneous environment after the format conversion and data cleaning; S15. Based on the unified data matrix model D, the high-dimensional data is optimized so that the unified data matrix model can reduce the computational complexity while retaining key information, and the optimized unified data matrix model is stored in the data storage module to establish an index structure.

3. According to claim 1, a method for recommending trusted O2O services in a heterogeneous environment based on deep learning is characterized in that: The S2 comprises the following steps: S21. Based on user behavior data matrix D u Define user node set V u , the user node set includes unique identifiers of all users; S22. Based on service information data matrix D s Define the service node set V s , the service node set includes unique identifiers of all services; S23. Based on the geographic location information data matrix D g Define the geographic location node set V g , the geographical location node set includes unique identifiers of all geographical locations; S24. Based on user behavior data matrix D u and service information data matrix D s Construct the interaction relationship edge set E between users and services us ; S25. Based on service information data matrix D s and geographic location information data matrix D g Construct the relationship edge set E between services and geographic locations sg ; S26. Based on user behavior data matrix D u and geographic location information data matrix D g Construct the relationship edge set E between users and geographic locations ug ; S27. Calculate the edge weight matrix W by combining the nodes and edge sets in steps S21-S26. The edge weight matrix W is composed of the edge weights w ij composition: Among them, x i With x j are nodes v extracted from the unified data matrix model D. i With node v j The feature vector reflects the comprehensive characteristics of each node in user behavior, service information and geographic location data, ‖x i -x j ‖ is the Euclidean distance between node feature vectors, α is the scaling parameter that controls the sensitivity of the logic function, σ1 is the normalization factor of the feature distance, θ is the threshold parameter of the logic function, which is used to adjust the translation of the similarity measure between nodes, and n ij Represents the user behavior data matrix D u The node v recorded in i With node v j The number of interactions between max is the maximum number of interactions among all node pairs, used to normalize the number of interactions n ij , r ij Represents the service information data matrix D s Middle node v i With node v j The average service evaluation score corresponding to the interaction between max is the highest service evaluation score in the data set, used to normalize the average service evaluation score r ij , d ij is the geographic location information data matrix D g The node v recorded in i With node v j The geographical distance between them, κ is the parameter for adjusting the influence of geographical distance, λ1, λ2, λ3 are the weight coefficients given to the interaction frequency, service evaluation and geographical distance normalization terms respectively; S28. Integrate the user, service and geographic location node set V and edge set E constructed in steps S21 to S27 = E us ∪E sg ∪E ug And the edge weight matrix W, construct the complete initial relationship graph G = (V, E, W).

4. According to claim 1, a method for recommending trusted O2O services in a heterogeneous environment based on deep learning is characterized in that: The S3 comprises the following steps: S31. Constructing a multi-view adaptive graph diffusion convolution network based on the initial relationship graph G; S32. In the local diffusion module, the local diffusion matrix P is defined according to the node type and edge weight matrix (L) : Among them, N L (i) is node v i The set of neighboring nodes in the local neighborhood, is the local diffusion regulation factor, which is used to regulate the sensitivity of different types of nodes in local information propagation; S33. In the global diffusion module, the global diffusion matrix P is defined based on the same edge weight matrix (G) : Among them, N G (i) is node v i Expand the set of nodes in the neighborhood globally, is the global diffusion adjustment factor, which is used to reflect the influence of nodes in long-distance information transmission; S34. Design a gated fusion mechanism to dynamically fuse the diffusion information of the local diffusion matrix with the diffusion information of the global diffusion matrix, and set the gating coefficient ζ i : in, is the node v in the lth layer i The feature representation of , U is the trainable weight matrix, b is the bias vector, and σ(·) is the Sigmoid activation function; S35. Use multi-view adaptive graph diffusion convolution to calculate the graph convolution node embedding representation, and represent the l-th layer node feature and Perform fusion operation: in, represents the feature matrix of the local diffusion module at layer l, represents the feature matrix of the global diffusion module at layer l, is the fused multi-view feature representation matrix, and are the trainable parameter matrices of the local and global diffusion modules at layer l, ⊙ represents the element-wise product operation, and σ(·) is the nonlinear activation function; S36. Using the jump diffusion mechanism, we weightedly fuse the information with different diffusion steps, optimize the multi-scale fusion effect of information transmission, introduce the jump diffusion mechanism into the multi-view adaptive graph diffusion convolution network, and obtain the final graph convolution node embedding representation: in, is the final node feature matrix after jump diffusion, T is the maximum diffusion step length, β t is the fusion weight of the diffusion information in the tth step.

5. According to claim 1, a method for recommending trusted O2O services in a heterogeneous environment based on deep learning is characterized in that: The S4 comprises the following steps: S41. Based on the final graph convolution node embedding representation of the multi-view adaptive graph diffusion convolution network, a recommendation model consisting of a user interest representation model and a service feature representation model is constructed, and the user node v i With service node v j Using the interest mapping function f u (·) and the feature mapping function f s (·) to get the final representation vector: in, For user u i The final interest representation vector of For services j The final feature representation vector of Embedding representation matrix for the final fused node; S42. Calculate the user u based on the user interest representation and service feature representation i With candidate services j Recommended rating between: Among them, <·,·> represents the vector inner product operation, g(s j ) for services j The category score of , η is the service category regulation factor; S43. Based on the recommended score r ij Sort the candidate services and select the k services with the highest scores. i Form a personalized O2O service recommendation candidate set.

6. According to claim 1, a method for recommending trusted O2O services in a heterogeneous environment based on deep learning is characterized in that: The S5 comprises the following steps: S51. Determine a key hyperparameter set Θ that affects the recommendation effect based on the recommendation model, wherein the hyperparameter set includes a maximum diffusion step length T, a local diffusion adjustment factor Global diffusion regulator And the regularization coefficient λ1: S52. Use the whale optimization algorithm to optimize the hyperparameter set Θ, initialize the population size N, and set the initial hyperparameter candidate set: in, is the i-th group of hyperparameter individuals, generated according to random initialization or preset range; S53. Based on the recommendation model, calculate each set of hyperparameters Θ using the objective function F(Θ) i Adaptability; S54. The whale optimization algorithm uses the mechanism of surrounding prey, spiral update and random search to optimize hyperparameters. The mechanism of surrounding prey: in, is the current optimal hyperparameter, A is the shrinkage coefficient, D opt is the distance between the current position and the optimal solution; Spiral update mechanism: in, b is the spiral shape control parameter, l is a random number; Random search mechanism: in, is a random individual in the population; S55. During the optimization process, set the convergence criteria and select the optimal hyperparameters: Where ∈ is the convergence threshold. When ΔF is less than ∈ for k consecutive rounds, the algorithm terminates and outputs the optimal hyperparameter Θ opt ; S56. The optimized hyperparameter Θ opt Feedback is given to the multi-view adaptive graph diffusion convolutional network to update the optimized recommendation model.

7. According to claim 6, a method for recommending trusted O2O services in a heterogeneous environment based on deep learning is characterized in that: The S53 comprises the following steps: S531. Calculate the fitness evaluation index of the hyperparameter set Θ according to the recommended model. The fitness evaluation index is composed of the hit rate HR@K, the normalized discounted cumulative gain NDCG@K and the loss function L(Θ i ), HR@K measures the proportion of services that are correctly selected by users among the first K recommended results in the recommendation list: Among them, |U| is the total number of user sets, s i Represents user u i The actual selected service, S i To recommend to user u i The top K service list of is the indicator function, if s i ∈S i If yes, it takes the value 1, otherwise it takes the value 0; NDCG@K measures the ranking quality of the recommendation list: Among them, DCG@K i is the discounted cumulative gain, which indicates the relevance of the first K recommended results: IDCG@K i is the ideal loss cumulative gain, indicating that user u i The best recommended sorting: Among them, rel i,j For user u i The true relevance of the jth service in the recommendation result, For user u i True correlation under optimal sorting; S532. Define hyperparameter Θ i The loss function of the recommendation model L(Θ i ): L(Θ i )=L recon +λ1·L reg ; Among them, λ1 is the regularization coefficient, L reg is the regularization term, L recon The reconstruction loss of the recommendation model is used to calculate the error of the model in the recommendation task: Among them, L reg is the regularization term, r ij For user u i About Services j The real score, Scoring the recommendation model predictions; S533. Based on the fitness evaluation index, define the hyperparameter set Θ i The objective optimization function is: F(Θ i )=ω1·HR@K+ω2·NDCG@K-ω3·L(Θ i ) Among them, ω1, ω2, ω3 are the weight coefficients of the fitness function.

Images