Model training method, recommended information generation method, device, equipment and medium

Abstract

Embodiments of the present disclosure disclose a model training method, a recommended information generation method, apparatuses, devices, and media. A specific implementation of the method includes: obtaining a sample set; based on the sample set, performing the following training steps: generating each dynamic user-item modality graph information; generating corresponding integrated user-item modality graph information; generating node feature information; inputting each node feature information into an initial interaction information prediction model to obtain interaction information; comparing the interaction information corresponding to each sample in at least one sample with the corresponding sample interaction information; determining whether the initial interaction information prediction model reaches a preset optimization target according to the comparison result; and in response to determining that the initial interaction information prediction model reaches the optimization target, taking the initial interaction information prediction model as a trained interaction information prediction model. The implementation improves the accuracy of recommended items, and reduces the waste of network resources in recommending items and when a user views an item.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the technical field of computer technology, and in particular, to a model training method, a recommended information generation method, an apparatus, a device, and a medium. BACKGROUND

[0002] A multi-modal recommendation method is a method of extracting modal feature information according to multi-modal data, and then strengthening the feature information of users and items in the recommendation system. At present, when generating recommended information by using multi-modal data, the commonly used way is: first, the modal feature information of the multi-modal data is fused into the user-item interaction graph, and then the graph neural network technology is used to encode and learn the graph information, and then the modal knowledge enhanced high-order feature information of the user and the item is generated.

[0003] However, when the above method is used to generate recommended information, the following technical problems often exist:

[0004] First, when the modal feature information of the modal data is directly used to construct the graph information, the noise data in the modal data has a large interference on the graph information, which leads to low accuracy of the constructed graph information, low reliability of the node feature information and interaction information generated by using the graph information, and further leads to low accuracy of the recommended items when the items are recommended according to the generated interaction information, resulting in waste of network resources for recommending and for users to view the items.

[0005] Second, only one layer of graph information is used to generate the final high-order feature information, which leads to poor model generalization ability and low accuracy of the model prediction result.

[0006] The above information disclosed in the background section is only intended to enhance the understanding of the background of the present inventive concept, and therefore, it can include information that does not form the prior art known to those of ordinary skill in the art in the country. SUMMARY

[0007] The summary section is provided to introduce the concepts briefly in a simplified form, which will be described in detail in the specific embodiments section. The summary section is not intended to identify key or essential features of the claimed technology nor is it intended to be used to limit the scope of the claimed technology.

[0008] Some embodiments of the present disclosure propose an interaction information prediction model training method, a recommended information generation method, an apparatus, an electronic device, and a computer readable medium to solve one or more of the technical problems mentioned in the background section.

[0009] In a first aspect, some embodiments of the present disclosure provide an interactive information prediction model training method, which comprises: obtaining a sample set, wherein the samples in the sample set comprise sample user-item interaction information, the sample user-item interaction information comprises sample user information, sample item information and sample interaction information, the sample user information comprises a sample user modality information group, and the sample item information comprises a sample item modality information group; based on the sample set, performing the following training steps: based on the sample user-item interaction information included in each sample in the sample set, generating each dynamic user-item modality graph information, wherein each dynamic user-item modality graph information in the dynamic user-item modality graph information corresponds to one sample, and each dynamic user-item modality graph information comprises a dynamic user modality graph and a dynamic item modality graph; for each sample in the at least one sample, based on the pre-generated predefined user-item modality graph information corresponding to the sample and the dynamic user-item modality graph information corresponding to the sample in the each dynamic user-item modality graph information, generating integrated user-item modality graph information corresponding to the sample, wherein the integrated user-item modality graph information comprises an integrated user modality graph and an integrated item modality graph; based on the integrated user modality graph and the integrated item modality graph included in the integrated user-item modality graph information corresponding to each sample in the at least one sample, generating node feature information, wherein the node feature information comprises each user node feature information corresponding to the integrated user modality graph and each item node feature information corresponding to the integrated item modality graph; inputting each node feature information corresponding to the at least one sample into an initial interactive information prediction model to obtain the interaction information corresponding to each sample in the at least one sample; comparing the interaction information corresponding to each sample in the at least one sample with the corresponding sample interaction information; determining whether the initial interactive information prediction model reaches a preset optimization target according to the comparison result; and in response to determining that the initial interactive information prediction model reaches the optimization target, taking the initial interactive information prediction model as a trained interactive information prediction model.

[0010] In a second aspect, some embodiments of the present disclosure provide a method for generating recommendation information, the method comprising: for each user information in a set of user information, performing the following steps: obtaining each user-item interaction information corresponding to the user information; for each user-item interaction information in the set of user-item interaction information, determining a preset number of item information in a set of preset item information that satisfy a preset interaction condition as interaction item information corresponding to the user-item interaction information; performing integration processing on the set of user-item interaction information and the determined set of interaction item information to obtain a set of integrated user-item interaction information; generating interaction information corresponding to each integrated user-item interaction information in the set of integrated user-item interaction information based on a pre-trained interaction information prediction model, wherein the interaction information prediction model is trained by the method described in any of the implementations of the first aspect; performing sorting processing on the generated set of interaction information to obtain an interaction information sequence; and generating item recommendation information corresponding to the user information according to the interaction information sequence.

[0011] In a third aspect, some embodiments of the present disclosure provide an interactive information prediction model training apparatus. The apparatus comprises: an obtaining unit configured to obtain a sample set, wherein a sample in the sample set comprises sample user-item interaction information, the sample user-item interaction information comprises sample user information, sample item information, and sample interaction information, the sample user information comprises a set of sample user modal information, and the sample item information comprises a set of sample item modal information; and a training unit configured to perform the following training steps based on the sample set: generating, based on each sample user-item interaction information included in at least one sample in the sample set, a respective dynamic user-item modal graph information, wherein each dynamic user-item modal graph information in the respective dynamic user-item modal graph information corresponds to a sample, and each dynamic user-item modal graph information comprises a dynamic user modal graph and a dynamic item modal graph; for each sample in the at least one sample, generating, based on a predefined user-item modal graph information corresponding to the sample in the respective predefined user-item modal graph information and a dynamic user-item modal graph information corresponding to the sample in the respective dynamic user-item modal graph information, an integrated user-item modal graph information corresponding to the sample, wherein the integrated user-item modal graph information comprises an integrated user modal graph and an integrated item modal graph; generating, based on the integrated user modal graph and the integrated item modal graph included in the integrated user-item modal graph information corresponding to each sample in the at least one sample, node feature information, wherein the node feature information comprises respective user node feature information corresponding to the integrated user modal graph and respective item node feature information corresponding to the integrated item modal graph; inputting the respective node feature information corresponding to the at least one sample into an initial interactive information prediction model to obtain interaction information corresponding to each sample in the at least one sample; comparing the interaction information corresponding to each sample in the at least one sample with the corresponding sample interaction information; determining whether the initial interactive information prediction model reaches a preset optimization target according to a comparison result; and in response to determining that the initial interactive information prediction model reaches the optimization target, taking the initial interactive information prediction model as a trained interactive information prediction model.

[0012] In a fourth aspect, some embodiments of the present disclosure provide a recommendation information generation apparatus, the apparatus comprising: an execution unit configured to, for each user information in a set of user information, perform the following steps: obtaining respective user-item interaction information corresponding to the user information; for each user-item interaction information in the respective user-item interaction information, determining a preset number of item information in a set of preset item information that satisfy a preset interaction condition as interaction item information corresponding to the user-item interaction information; performing integration processing on the respective user-item interaction information and the determined respective interaction item information to obtain an integrated user-item interaction information group; generating interaction information corresponding to each integrated user-item interaction information in the integrated user-item interaction information group based on a pre-trained interaction information prediction model, wherein the interaction information prediction model is trained by the method described in any implementation manner of the first aspect; performing sorting processing on the generated respective interaction information to obtain an interaction information sequence; and generating item recommendation information corresponding to the user information according to the interaction information sequence.

[0013] In a fifth aspect, some embodiments of the present disclosure provide an electronic device, comprising: one or more processors; a storage device having one or more programs stored thereon, when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation manner of the first aspect or the second aspect.

[0014] In a sixth aspect, some embodiments of the present disclosure provide a computer readable medium having a computer program stored thereon, wherein the program is executed by a processor to implement the method described in any implementation manner of the first aspect or the second aspect.

[0015] The above various embodiments of the present disclosure have the following beneficial effects: through the interaction information prediction model training method of some embodiments of the present disclosure, the interference of noise data in modal data on graph information can be effectively reduced, the accuracy of the constructed graph information can be improved, and the reliability of the node feature information and the interaction information generated using the graph information can be improved. Furthermore, the accuracy of the recommended items according to the interaction information can be improved, and the waste of network resources when the user views the items can be reduced. Specifically, the reason why the reliability of the node feature information and the interaction information generated using the graph information is low is that when the modal feature information of the modal data is directly used to construct the graph information, the noise data in the modal data has a large interference on the graph information, resulting in low accuracy of the constructed graph information, low reliability of the node feature information and the interaction information generated using the graph information, and further resulting in poor accuracy of the recommended items when the items are recommended according to the generated interaction information, causing waste of network resources when the user views the items. Based on this, the interaction information prediction model training method of some embodiments of the present disclosure first acquires a sample set, wherein the samples in the sample set include sample user-item interaction information, the sample user-item interaction information includes sample user information, sample item information and sample interaction information, the sample user information includes a sample user modal information group, and the sample item information includes a sample item modal information group. In this way, data for model training can be obtained. Then, the following training steps are performed based on the sample set: first, based on each sample user-item interaction information included in at least one sample in the sample set, each dynamic user-item modal graph information is generated, wherein each dynamic user-item modal graph information corresponds to one sample, and each dynamic user-item modal graph information includes a dynamic user modal graph and a dynamic item modal graph. In this way, a dynamic modal relationship graph can be obtained through data, and the structure of the dynamic modal relationship graph can be optimized through the constantly updated node feature information in the dynamic modal relationship graph, reducing the influence of noise data on the generated graph information. Second, for each sample in the at least one sample, based on the predefined user-item modal graph information corresponding to the sample in each predefined user-item modal graph information and the dynamic user-item modal graph information corresponding to the sample in the each dynamic user-item modal graph information, integrated user-item modal graph information corresponding to the sample is generated, wherein the integrated user-item modal graph information includes an integrated user modal graph and an integrated item modal graph. In this way, a more comprehensive and stable graph information can be generated by combining the predefined user-item modal graph information and the dynamic user-item modal graph information, more effectively reducing the influence of noise data on the graph information and improving the accuracy of the generated interaction information.Then, based on the integrated user modal graph and the integrated item modal graph included in the integrated user-item modal graph information corresponding to each of the at least one sample, node feature information is generated, wherein the node feature information includes user node feature information corresponding to the integrated user modal graph and item node feature information corresponding to the integrated item modal graph. In this way, the node feature information can be more accurately obtained by integrating the user-item modal graph information, and the accuracy of the interaction information generated by using the node feature information can be improved. Then, each node feature information corresponding to the at least one sample is input into the initial interaction information prediction model to obtain the interaction information corresponding to each of the at least one sample. In this way, the model can be trained by the obtained node feature information to obtain more accurate interaction information, and the accuracy of the recommended items can be improved. Then, the interaction information corresponding to each of the at least one sample is compared with the corresponding sample interaction information. In this way, the comparison result can be obtained by comparing the interaction information with the corresponding sample interaction information, and whether the model reaches the preset optimization target can be verified by the comparison result. Then, whether the initial interaction information prediction model reaches the preset optimization target is determined according to the comparison result. In this way, the performance of the model can be verified by the comparison result. Finally, in response to determining that the initial interaction information prediction model reaches the optimization target, the initial interaction information prediction model is taken as the trained interaction information prediction model. In this way, an interaction information prediction model reaching the preset optimization target can be obtained, and more accurate interaction information can be obtained by using this model, and the accuracy of the recommended items can be improved. Because the predefined user-item modal graph information and the dynamic user-item modal graph information can be obtained, and then the integrated user-item modal graph information is generated, the interference of noise data on the graph information is reduced by the multi-layer constructed graph information, the reliability of the node feature information and the interaction information is improved, and the accuracy of the recommended items generated according to the interaction information is improved, and the waste of network resources in recommending and viewing items by users is reduced. Because the performance of the model can be verified by the comparison result, the deviation of the model can be identified, and the model can be adjusted to ensure that the model reaches the optimization target, more reliable interaction information can be generated, and the accuracy of the recommended items generated according to the interaction information can be improved, and the waste of network resources in recommending and viewing items by users is reduced. In summary, by processing the graph information, the interference of noise data on the graph information can be reduced, the reliability of the node feature information and the interaction information generated by using the graph information can be improved, the accuracy of the recommended items generated according to the interaction information can be improved, and the waste of network resources in recommending and viewing items by users can be reduced. BRIEF DESCRIPTION OF DRAWINGS

[0016] The above and other features, aspects and advantages of various embodiments of the present disclosure will become more apparent with reference to the following detailed description when taken in conjunction with the accompanying drawings. Throughout the drawings, similar or common elements of the drawings are denoted by like reference numerals. It is to be understood that the drawings are schematic, and elements and features are not necessarily drawn to scale.

[0017] Figure 1 is a flowchart of some embodiments of an interaction information prediction model training method according to the present disclosure;

[0018] Figure 2 is a flowchart of some embodiments of a recommended information generation method according to the present disclosure;

[0019] Figure 3 is a structural schematic diagram of some embodiments of an interaction information prediction model training apparatus according to the present disclosure;

[0020] Figure 4 is a structural schematic diagram of some embodiments of a recommended information generation apparatus according to the present disclosure;

[0021] Figure 5 is a structural schematic diagram of an electronic device suitable for use to implement some embodiments of the present disclosure. DETAILED DESCRIPTION

[0022] Embodiments of the present disclosure will be described in more detail with reference to the drawings. Although some embodiments of the present disclosure are shown in the drawings, it is understood that the present disclosure can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure can be more thoroughly and completely understood. It should be understood that the drawings and embodiments of the present disclosure are for exemplary purposes only, and are not intended to limit the scope of protection of the present disclosure.

[0023] It should also be noted that, for ease of description, only parts related to the present application are shown in the drawings. The embodiments in the present disclosure and the features in the embodiments can be combined with each other without conflict.

[0024] It should be noted that the terms "first", "second", etc. mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0025] It should be noted that the adjectives "one", "multiple" mentioned in the present disclosure are illustrative and not limiting, and those skilled in the art should understand that, unless otherwise explicitly stated in the context, it should be understood as "one or more".

[0026] The names of messages or information exchanged between the plurality of devices in the embodiments of the present disclosure are only for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0027] The present disclosure will be described in detail below with reference to the accompanying drawings and in conjunction with embodiments.

[0028] Figure 1 A flow 100 of some embodiments of an interaction information prediction model training method according to the present disclosure is shown. The interaction information prediction model training method includes the following steps:

[0029] Step 101, obtaining a sample set.

[0030] In some embodiments, an execution subject (e.g., a computing device) of the interaction information prediction model training method can obtain a sample set. The execution subject can be a terminal or a server that trains the interaction information prediction model. The samples in the sample set can be training data used to train the interaction information prediction model. The sample user-item interaction information can be information corresponding to a user and an item that has interacted or has not interacted. The sample user-item interaction information can include sample user information, sample item information, and sample interaction information. The interaction can be that the user corresponding to the sample user information has performed a click operation on a detail page of the item corresponding to the sample item information. The sample user information can include personal information of the user. For example, the sample user information can include account information of the user, item flow records (e.g., household supplies purchased within a year), and evaluation information of the items in the item flow records. The evaluation information can include, but is not limited to, evaluation text, evaluation scores, and evaluation videos. The sample user information can include a sample user modality information group. The sample user modality information group can be user multi-modal data used to represent the sample user information. The user multi-modal data can be various different types of data used to represent the sample user information. The user multi-modal data can include, but is not limited to, text data used to record the account information, text data used to record the item flow records, and video data used to record the item evaluations. The sample item information can be various information corresponding to any item in the item flow records of the sample user information or various information corresponding to any item not recorded in the item flow records of the sample user information. For example, the sample item information can include image information of the item, item flow volume (e.g., item purchase volume), and unit value information (e.g., item unit price). The sample item information can include a sample item modality information group. The sample item modality information in the sample item modality information group can be item multi-modal data used to represent the sample item information. The item multi-modal data can be various different types of data used to represent the sample item information. The item multi-modal data can include, but is not limited to, image data used to record the image information, text data used to record the item flow volume, and text data used to record the unit value information. The sample interaction information can be label information used to represent whether the sample user information and the sample item information have interacted. For example, the sample interaction information can be “interaction” or “no interaction”.

[0031] At step 102, the following training steps are performed based on the sample set:

[0032] At step 1021, each dynamic user-item modality graph information is generated based on the respective sample user-item interaction information included in at least one sample in the sample set.

[0033] In some embodiments, the execution subject can generate each dynamic user-item modality graph information based on the respective sample user-item interaction information included in at least one sample in the sample set. Each of the dynamic user-item modality graph information corresponds to a sample. Each dynamic user-item modality graph information includes a dynamic user modality graph and a dynamic item modality graph.

[0034] In some optional implementations of some embodiments, the execution subject can generate each dynamic user-item modality graph information based on the respective sample user-item interaction information included in at least one sample in the sample set by:

[0035] In a first step, each sample user-item interaction feature information is generated based on each sample user-item interaction information included in at least one sample in the sample set. Each sample user-item interaction feature information in the sample user-item interaction feature information corresponds to a sample, and each sample user-item interaction feature information includes a sample user modal feature information group and a sample item modal feature information group. The sample user modal feature information in the sample user modal feature information group can be a feature vector corresponding to the sample user modal information in the sample user modal information group. The sample item modal feature information in the sample item modal feature information group can be a feature vector corresponding to the sample item modal information in the sample item modal information group. In practice, first, the execution subject can select different pre-trained models to extract features from each sample user interaction information according to the data type of each sample user interaction information, to obtain each sample user interaction feature information. Then, the execution subject can select different pre-trained models to extract features from each sample item interaction information according to the data type of each sample item interaction information, to obtain each sample item interaction feature information. Finally, the execution subject can combine each sample user interaction feature information and each sample item interaction feature information into the sample user-item interaction feature information. The pre-trained model can be various neural network models capable of feature extraction. For example, the pre-trained model can include but is not limited to a ViT model (Vision Transformer, ViT) and a BERT model (Bidirectional Encoder Representations from Transformers). As an example: when the data type of each sample user-item interaction information is text data, the selected pre-trained model can be a BERT model. When the data type of each sample user-item interaction information is image data, the selected pre-trained model can be a ViT model.

[0036] In a second step, each predefined user-item modal graph information is generated based on each sample user-item interaction feature information. Each predefined user-item modal graph information in the predefined user-item modal graph information corresponds to a sample, and each predefined user-item modal graph information includes a predefined user modal graph and a predefined item modal graph. The predefined user modal graph can be a weighted undirected graph used to represent the connection relationship between each sample user modal feature information in the sample user modal feature information group. The predefined item modal graph can be a weighted undirected graph used to represent the connection relationship between each sample item modal feature information in the sample item modal feature information group.

[0037] In the third step, the execution subject generates the dynamic user-item modal graph information based on the pre-defined user-item modal graph information. Each of the dynamic user-item modal graph information corresponds to a sample, and each of the dynamic user-item modal graph information includes a dynamic user modal graph and a dynamic item modal graph. The dynamic user modal graph can be a weighted undirected graph with nodes and weights changing over time after the pre-defined user modal graph is processed by a pre-trained graph neural network. The dynamic item modal graph can be a weighted undirected graph with nodes and weights changing over time after the pre-defined item modal graph is processed by the graph neural network. The graph neural network can be a neural network taking a weighted undirected graph as input and outputting a plurality of feature vectors generated for each node in the weighted undirected graph in different iterations.

[0038] In some optional implementations of some embodiments, the execution subject can generate the pre-defined user-item modal graph information based on the sample user-item interaction feature information by the following steps:

[0039] In the first step, for each of the sample user-item interaction feature information, the execution subject performs the following steps:

[0040] In the first sub-step, for each of the pair of sample user modal feature information in the sample user modal feature information set included in the sample user-item interaction feature information, the execution subject performs the following steps:

[0041] In sub-step one, the execution subject generates a similarity between the pair of sample user modal feature information as a user modal similarity based on the pair of sample user modal feature information. In practice, the execution subject can determine a cosine similarity between the pair of sample user modal feature information as the user modal similarity.

[0042] In sub-step two, the execution subject generates a user modal probability based on the user modal similarity. The user modal probability can be a cosine similarity satisfying a preset constraint condition. The preset constraint condition can be that the cosine similarity is greater than 0. In practice, the execution subject can input the user modal similarity into an activation function to obtain the user modal probability. The activation function can be a rectified linear unit (ReLU).

[0043] A second sub-step, generating a predefined user modality graph based on the generated respective user modality probabilities. In practice, first, the execution subject can take each of the sample user modality feature information in the sample user modality feature information group as a node, and take the generated respective user modality probabilities as weights. Then, the predefined user modality graph is constructed by the obtained respective nodes and respective weights.

[0044] A third sub-step, for each of the two sample item modality feature information in the sample item modality feature information group included in the sample user-item interaction feature information, the following steps are performed:

[0045] Sub-step one, generating an item modality similarity between the two sample item modality feature information based on the two sample item modality feature information. In practice, the execution subject can determine the cosine similarity between the two sample item modality feature information as the item modality similarity.

[0046] Sub-step two, generating an item modality probability based on the item modality similarity. The item modality probability can be a cosine similarity that meets a preset constraint condition. The preset constraint condition can be that the cosine similarity is greater than 0. In practice, the execution subject can input the item modality similarity into an activation function to obtain the item modality probability. The activation function can be a linear rectifier function.

[0047] A fourth sub-step, generating a predefined item modality graph based on the generated respective item modality probabilities. In practice, first, the execution subject can take each of the sample item modality feature information in the sample item modality feature information group as a node, and take the generated respective item modality probabilities as weights. Then, the predefined item modality graph is constructed according to the obtained respective nodes and respective weights.

[0048] A fifth sub-step, determining the predefined user modality graph and the predefined item modality graph as predefined user-item modality graph information. In practice, the execution subject can combine the predefined user modality graph and the predefined item modality graph into the predefined user-item modality graph information.

[0049] In some optional implementations of some embodiments, the execution subject can generate respective dynamic user-item modality graph information based on respective predefined user-item modality graph information by the following steps:

[0050] A first step, for each of the predefined user-item modality graph information in the respective predefined user-item modality graph information, the following steps are performed:

[0051] In a first sub-step, based on the predefined user-item modal graph information, a set of user-item dynamic node feature information is generated. Each user-item dynamic node feature information in the set of user-item dynamic node feature information includes a set of user dynamic node feature information and a set of item dynamic node feature information. The user dynamic node feature information in the set of user dynamic node feature information can be a feature vector corresponding to the sample user modal information obtained by inputting the sample user modal feature information into the graph neural network and processing the sample user modal feature information by the message passing mechanism in the graph neural network. The item dynamic node feature information in the set of item dynamic node feature information can be a feature vector corresponding to the sample item modal information obtained by inputting the sample item modal feature information into the graph neural network and processing the sample item modal feature information by the message passing mechanism in the graph neural network. In practice, first, the execution subject can traverse each node of the predefined user modal graph included in the predefined user-item modal graph information by the graph neural network, and set an iteration number. Then, each node in the predefined user modal graph is randomly initialized to obtain an initial feature vector corresponding to each node. Then, for each initial feature vector in the obtained set of initial feature vectors, the product of the user normalization matrix, the initial feature vector, and the initial transformation matrix is input into the nonlinear activation function to obtain a first-order feature vector, and the iteration number is reduced by one. Then, for each first-order feature vector in the obtained set of first-order feature vectors, the product of the user normalization matrix, the first-order feature vector, and the first-order transformation matrix is input into the nonlinear activation function to obtain a second-order feature vector, and the iteration number is reduced by one. This process is repeated until the iteration number meets the preset parameter condition. The obtained set of initial feature vectors is taken as a set of user dynamic node feature information, the obtained set of first-order feature vectors is taken as another set of user dynamic node feature information, and this process is repeated until a set of user dynamic node feature information is obtained. The user normalization matrix can be a matrix obtained by the normalization degree matrix and the neighbor matrix corresponding to the predefined user modal graph. In practice, first, the execution subject can generate a diagonal matrix by the predefined user modal graph, in which each row can correspond to a node of the predefined user modal graph, each column can correspond to a node of the predefined user modal graph, the elements on the main diagonal can correspond to the degrees of the nodes in the predefined user modal graph, and the elements other than the elements on the main diagonal are all 0. Then, for each element on the main diagonal of the degree matrix, the element on the main diagonal is replaced by the inverse of the square root of the element on the main diagonal to obtain the normalization degree matrix.Then, the execution subject can generate a matrix in which each row corresponds to a sample user modal feature information and each column corresponds to a sample user modal feature information, and the modal probability between the two sample user modal feature information is used as the corresponding matrix element, as the pre-defined user adjacency matrix corresponding to the pre-defined user modal graph. Then, the execution subject can determine the product of the normalized degree matrix, the adjacency matrix, and the normalized degree matrix as the user normalized matrix. The manner of obtaining each item dynamic node feature information set is similar to this, and will not be repeated here. Finally, the execution subject can combine each user dynamic node feature information set obtained and the item dynamic node feature information set corresponding to the user dynamic node feature information set to obtain each user-item node feature information set. The initial feature vector can be a feature vector obtained by randomly initializing each node in the pre-defined user modal graph by the graph neural network. The first-order feature vector can be a feature vector obtained by processing the initial feature vector of the node by the graph neural network. The second-order feature vector can be a feature vector obtained by processing the first-order feature vector of the node by the graph neural network. The nonlinear activation function can be a linear rectification function. The number of iterations can be a pre-set value. Here, the specific setting of the number of iterations is not limited. The initial conversion matrix can be a matrix in which each row corresponds to a first-order feature vector and each column corresponds to an initial feature vector, and the corresponding element is a parameter value generated by the graph neural network for converting the initial feature vector into the corresponding first-order feature vector. The parameter value can be a numerical value generated by the graph neural network. Here, the specific setting of the parameter value is not limited.

[0052] In the second sub-step, each user-item dynamic graph information is generated based on the user-item dynamic node feature information set. Each user-item dynamic graph information includes user dynamic graph information and item dynamic graph information. In practice, first, the execution subject can generate the cosine similarity between each two user dynamic node feature information in the user dynamic node feature information set. Then, the obtained cosine similarity is input into an activation function to obtain the modal probability between each two user dynamic node feature information in the user dynamic node feature information set. Then, the execution subject can take each user dynamic node feature information in the user dynamic node feature information set as a node and take the modal probability between each two user dynamic node feature information as a corresponding weight to obtain the user dynamic graph information. The way to obtain the item dynamic graph information is similar to the above, and details are not repeated here. Finally, the execution subject can combine the user dynamic graph information and the item dynamic graph information into the user-item dynamic graph information. The user dynamic graph information can be a weighted undirected graph generated by taking each user dynamic node feature information in the user dynamic node feature information set as a node and taking the modal probability between each two user dynamic node feature information as a weight. The item dynamic graph information can be a weighted undirected graph generated by taking each item dynamic node feature information in the item dynamic node feature information set as a node and taking the modal probability between each two item dynamic node feature information as a weight. The activation function can be a linear rectifier function.

[0053] A third sub-step, based on the above-mentioned each user-item dynamic graph information, generating each user-item to-be-fused adjacency matrix information. Wherein, each user-item to-be-fused adjacency matrix information in the above-mentioned each user-item to-be-fused adjacency matrix information includes user to-be-fused adjacency matrix and item to-be-fused adjacency matrix. The above-mentioned user to-be-fused adjacency matrix can be the adjacency matrix corresponding to the above-mentioned user dynamic graph information. The above-mentioned item to-be-fused adjacency matrix can be the adjacency matrix corresponding to the above-mentioned item dynamic graph information. In practice, first, the above-mentioned execution subject can generate a matrix whose each row can correspond to the above-mentioned user dynamic node feature information and each column can correspond to the above-mentioned user dynamic node feature information through the user dynamic graph information included in the above-mentioned user-item dynamic graph information, and the modal probability between two user dynamic node feature information is taken as the matrix corresponding to the element of the matrix, as the user to-be-fused adjacency matrix corresponding to the above-mentioned user dynamic graph information. Then, the above-mentioned execution subject can generate a matrix whose each row can correspond to the above-mentioned item dynamic node feature information and each column can correspond to the above-mentioned item dynamic node feature information through the item dynamic graph information included in the above-mentioned user-item dynamic graph information, and the modal probability between two item dynamic node feature information is taken as the matrix corresponding to the element of the matrix, as the item to-be-fused adjacency matrix corresponding to the above-mentioned item dynamic graph information. Finally, the above-mentioned execution subject can combine the above-mentioned user to-be-fused adjacency matrix and the above-mentioned item to-be-fused adjacency matrix into the above-mentioned user-item to-be-fused adjacency matrix information.

[0054] A fourth sub-step, fusing the above-mentioned each user-item to-be-fused adjacency matrix information to obtain user-item dynamic adjacency matrix information. Wherein, the above-mentioned user-item dynamic adjacency matrix information includes user dynamic adjacency matrix and item dynamic adjacency matrix. In practice, first, the above-mentioned execution subject can sum each user to-be-fused adjacency matrix in the above-mentioned each user-item to-be-fused adjacency matrix information. Then, the user dynamic adjacency matrix can be obtained by dividing the adjacency matrix obtained after summing by the above-mentioned iteration number. Then, the above-mentioned execution subject can sum each item to-be-fused adjacency matrix in the above-mentioned each user-item to-be-fused adjacency matrix information. Then, the item dynamic adjacency matrix can be obtained by dividing the adjacency matrix obtained after summing by the above-mentioned iteration number. Finally, the above-mentioned execution subject can combine the above-mentioned user dynamic adjacency matrix and the above-mentioned item dynamic adjacency matrix into the above-mentioned user-item dynamic adjacency matrix information.

[0055] In a fifth sub-step, based on the user-item dynamic adjacency matrix information, dynamic user-item modal graph information is generated. The dynamic user-item modal graph information includes a dynamic user modal graph and a dynamic item modal graph. The dynamic user modal graph can be a weighted undirected graph corresponding to the user dynamic adjacency matrix. The dynamic item modal graph can be a weighted undirected graph corresponding to the item dynamic adjacency matrix. In practice, first, the execution subject can take the rows or columns in the user dynamic adjacency matrix included in the user-item dynamic adjacency matrix information as nodes, and the elements in the matrix as weights corresponding to each two nodes, to obtain a dynamic user modal graph corresponding to the user dynamic adjacency matrix. Then, the execution subject can take the rows or columns in the item dynamic adjacency matrix included in the user-item dynamic adjacency matrix information as nodes, and the elements in the matrix as weights corresponding to each two nodes, to obtain a dynamic item modal graph corresponding to the item dynamic adjacency matrix. Finally, the execution subject can combine the dynamic user modal graph and the dynamic item modal graph into the dynamic user-item modal graph information.

[0056] The technical solution and related content thereof serve as one of the invention points of the embodiments of the present disclosure, and solve the second technical problem mentioned in the background art, i.e., generating final high-order feature information only according to one layer of constructed graph information, resulting in poor model generalization ability and low accuracy of model prediction results. Factors that result in poor model generalization ability and low accuracy of model prediction results are usually as follows: generating final high-order feature information only according to one layer of constructed graph information, resulting in poor model generalization ability and low accuracy of model prediction results. If the above factors are solved, the generalization ability of the model can be improved, and the accuracy of model prediction can be improved. To achieve this effect, for each sample user item interaction feature information in each sample user item interaction feature information, the following steps are performed: first, for each two sample user modal feature information in the sample user modal feature information group included in the sample user item interaction feature information, the following steps are performed: first, based on the two sample user modal feature information, the similarity between the two sample user modal feature information is generated as user modal similarity. In this way, the similarity between the two sample user modal feature information can be obtained, the user group with similar modal feature information is obtained, and the noise data with low similarity can be filtered out. Based on the user modal similarity, a user modal probability is generated. In this way, the connection relationship between the two sample user modal feature information can be obtained through the user modal probability, which provides a data basis for subsequent graph information construction, and can reduce the interference of noise data on the construction of graph information. Based on the generated each user modal probability, a predefined user modal graph is generated. In this way, a modal relationship graph between all sample user modal feature information can be constructed through the modal probability, and the interference of noise data can be further reduced. For each two sample item modal feature information in the sample item modal feature information group included in the sample user item interaction feature information, the following steps are performed: based on the two sample item modal feature information, the similarity between the two sample item modal feature information is generated as item modal similarity. In this way, the similarity between the two sample item modal feature information can be obtained, the item information with similar modal feature information is obtained, and the interference of noise data on the subsequent construction of graph information can be reduced. Based on the item modal similarity, an item modal probability is generated. In this way, the modal probability between the two sample item modal feature information can be obtained, which provides a data basis for subsequent graph information construction. Based on the generated each item modal probability, a predefined item modal graph is generated. In this way, a modal relationship graph between all sample item modal feature information can be constructed through the modal probability, and the interference of noise data can be further reduced. The predefined user modal graph and the predefined item modal graph are determined as predefined user item modal graph information. In this way, the predefined user modal graph and the predefined item modal graph can be combined as predefined user item modal graph information.For each of the predefined user-item modal graph information, the following steps are performed: first, based on the predefined user-item modal graph information, a set of user-item dynamic node feature information is generated, wherein each user-item node feature information set in the set of user-item dynamic node feature information includes a set of user dynamic node feature information and a set of item dynamic node feature information. Thus, based on the constructed predefined user-item modal graph information, a set of user-item dynamic node feature information can be generated, which can more accurately reflect the current state of users and items, further improve the generalization ability of the model, and improve the accuracy of model prediction. Then, based on the set of user-item dynamic node feature information, a set of user-item dynamic graph information is generated, wherein each user-item dynamic graph information in the set of user-item dynamic graph information includes user dynamic graph information and item dynamic graph information. Thus, through the constructed user dynamic graph information and item dynamic graph information, the changing relationship between users and items can be better reflected, the generalization ability of the model can be improved, and a data basis can be provided for subsequent processing. Then, based on the set of user-item dynamic graph information, a set of user-item to-be-fused adjacency matrix information is generated, wherein each user-item to-be-fused adjacency matrix information in the set of user-item to-be-fused adjacency matrix information includes a user to-be-fused adjacency matrix and an item to-be-fused adjacency matrix. Thus, through the obtained set of user-item to-be-fused adjacency matrix information, the quantization of the graph information can be completed, and a basis for subsequent fusion of each user dynamic graph information and each item dynamic graph information can be provided. Then, the set of user-item to-be-fused adjacency matrix information is fused to obtain user-item dynamic adjacency matrix information, wherein the user-item dynamic adjacency matrix information includes a user dynamic adjacency matrix and an item dynamic adjacency matrix. Thus, through the fusion of the adjacency matrix, the fusion of the user dynamic graph information and the item dynamic graph information can be further realized, the generalization ability of the model can be improved, and the accuracy of model prediction can be further improved. Finally, based on the user-item dynamic adjacency matrix information, dynamic user-item modal graph information is generated, wherein the dynamic user-item modal graph information includes a dynamic user modal graph and a dynamic item modal graph. Thus, through the obtained dynamic user modal graph and dynamic item modal graph, the multi-dimensional feature vectors and dynamic changes between all users and all items can be obtained, further improving the generalization ability of the model and the accuracy of model prediction. Because the dynamic user-item modal graph information is generated based on the predefined user-item modal graph information, the interference of noise data can be reduced through the obtained two-layer graph information, and the accuracy of model prediction can be improved.In addition, the dynamic user-item modal graph information is generated by the multi-layer fused adjacency matrix, and the dynamic user-item modal graph information can optimize the graph information by using the continuously updated node features, further reduce the interference of noise data, and improve the generalization ability of the model and the prediction accuracy of the model.

[0057] At step 1022, for each sample in the at least one sample, the corresponding sample integrated user-item modal graph information is generated based on the pre-defined user-item modal graph information corresponding to the sample in the pre-generated each pre-defined user-item modal graph information and the dynamic user-item modal graph information corresponding to the sample in the each dynamic user-item modal graph information.

[0058] In some embodiments, the execution subject can generate, for each sample in the at least one sample, the corresponding sample integrated user-item modal graph information based on the pre-defined user-item modal graph information corresponding to the sample in the pre-generated each pre-defined user-item modal graph information and the dynamic user-item modal graph information corresponding to the sample in the each dynamic user-item modal graph information. The integrated user-item modal graph information includes an integrated user modal graph and an integrated item modal graph. The integrated user modal graph can be a weighted undirected graph obtained by processing the pre-defined user modal graph included in the pre-defined user-item modal graph information and the dynamic user modal graph included in the dynamic user-item modal graph information. The integrated item modal graph can be a weighted undirected graph obtained by processing the pre-defined item modal graph included in the pre-defined user-item modal graph information and the dynamic item modal graph included in the dynamic user-item modal graph information.

[0059] In some optional implementations of some embodiments, the execution subject can generate, for each sample in the at least one sample, the corresponding sample integrated user-item modal graph information based on the pre-defined user-item modal graph information corresponding to the sample in the pre-generated each pre-defined user-item modal graph information and the dynamic user-item modal graph information corresponding to the sample in the each dynamic user-item modal graph information by the following steps:

[0060] In a first step, based on the predefined user-item modal graph information corresponding to the above-mentioned samples, a predefined user-item adjacency matrix information corresponding to the above-mentioned predefined user-item modal graph information is generated. The predefined user-item adjacency matrix information includes a predefined user adjacency matrix and a predefined item adjacency matrix. In practice, first, the execution subject can generate a predefined user adjacency matrix corresponding to the predefined user modal graph in the predefined user-item modal graph information, in which each row can correspond to the sample user modal feature information, each column can correspond to the sample user modal feature information, and the modal probability between two sample user modal feature information is taken as the matrix element of the corresponding matrix. Then, the execution subject can generate a predefined item adjacency matrix corresponding to the predefined item modal graph in the predefined user-item modal graph information, in which each row can correspond to the sample item modal feature information, each column can correspond to the sample item modal feature information, and the modal probability between two sample item modal feature information is taken as the matrix element of the corresponding matrix. Finally, the execution subject can combine the predefined user adjacency matrix and the predefined item adjacency matrix into the predefined user-item adjacency matrix information. The predefined user adjacency matrix can be the adjacency matrix corresponding to the predefined user modal graph. The predefined item adjacency matrix can be the adjacency matrix corresponding to the predefined item modal graph.

[0061] In a second step, the product of the predefined user adjacency matrix and a first preset coefficient is determined as a predefined user coefficient matrix. The first preset coefficient can be any value within a preset range. Here, the specific setting of the first preset coefficient is not limited. The preset range can be [0, 1].

[0062] In a third step, the product of the predefined item adjacency matrix and the first preset coefficient is determined as a predefined item coefficient matrix.

[0063] In the fourth step, the dynamic user-item adjacency matrix information corresponding to the dynamic user-item modal graph information is generated based on the dynamic user-item modal graph information of the above-mentioned sample. The dynamic user-item adjacency matrix information includes a dynamic user adjacency matrix and a dynamic item adjacency matrix. In practice, first, the execution subject can generate a matrix by using the dynamic user modal graph included in the dynamic user-item modal graph information, where each row corresponds to a node in the dynamic user modal graph, each column corresponds to a node in the dynamic user modal graph, and the weight between two nodes is used as the corresponding matrix element. The matrix is the dynamic user adjacency matrix corresponding to the dynamic user modal graph. Then, the execution subject can generate a matrix by using the dynamic item modal graph included in the dynamic user-item modal graph information, where each row corresponds to a node in the dynamic item modal graph, each column corresponds to a node in the dynamic item modal graph, and the weight between two nodes is used as the corresponding matrix element. The matrix is the dynamic item adjacency matrix corresponding to the dynamic item modal graph. Finally, the execution subject can combine the dynamic user adjacency matrix and the dynamic item adjacency matrix to obtain the dynamic user-item adjacency matrix information. The dynamic user adjacency matrix can be the adjacency matrix corresponding to the dynamic user modal graph. The dynamic item adjacency matrix can be the adjacency matrix corresponding to the dynamic item modal graph.

[0064] In the fifth step, the product of the dynamic user adjacency matrix and a second preset coefficient is determined as a dynamic user coefficient matrix. The second preset coefficient can be the absolute value of the difference between the first preset coefficient and 1. Here, the specific setting of the first preset coefficient is not limited.

[0065] In the sixth step, the product of the dynamic item adjacency matrix and the second preset coefficient is determined as a dynamic item coefficient matrix.

[0066] In the seventh step, the sum of the predefined user coefficient matrix and the dynamic user coefficient matrix is determined as an integrated user adjacency matrix.

[0067] In the eighth step, the sum of the predefined item coefficient matrix and the dynamic item coefficient matrix is determined as an integrated item adjacency matrix.

[0068] In the ninth step, the integrated user adjacency matrix and the integrated item adjacency matrix are determined as integrated user-item adjacency matrix information. In practice, the execution subject can combine the integrated user adjacency matrix and the integrated item adjacency matrix to obtain the integrated user-item adjacency matrix information.

[0069] In the tenth step, based on the integrated user-item adjacency matrix information, integrated user-item modality graph information corresponding to the sample is generated. The integrated user-item modality graph information includes an integrated user modality graph and an integrated item modality graph. The integrated user modality graph can be a weighted undirected graph corresponding to the integrated user adjacency matrix. The integrated item modality graph can be a weighted undirected graph corresponding to the integrated item adjacency matrix. In practice, first, the execution subject can determine the rows or columns of the integrated user adjacency matrix included in the integrated user-item adjacency matrix information as nodes. Then, each element in the integrated user adjacency matrix is determined as the weight corresponding to each two nodes, and the integrated user modality graph corresponding to the integrated user adjacency matrix is obtained. Then, the execution subject can determine the rows or columns of the integrated item adjacency matrix included in the integrated user-item adjacency matrix information as nodes. Then, each element in the integrated item adjacency matrix is determined as the weight corresponding to each two nodes, and the integrated item modality graph corresponding to the integrated item adjacency matrix is obtained. Finally, the execution subject can combine the integrated user modality graph and the integrated item modality graph into the integrated user-item modality graph information.

[0070] The technical solution and related content thereof serve as one inventive point of the embodiments of the present disclosure, and solve the second technical problem mentioned in the background art, i.e., generating final high-order feature information only based on one layer of graph information, resulting in poor model generalization ability and low accuracy of model prediction results. Factors that result in poor model generalization ability and low accuracy of model prediction results are usually as follows: generating final high-order feature information only based on one layer of graph information, resulting in poor model generalization ability and low accuracy of model prediction results. If the above factors are solved, the generalization ability of the model can be improved, and the accuracy of model prediction can be improved. To achieve this effect, the present disclosure generates predefined user item adjacency matrix information corresponding to the predefined user item modal graph information of the corresponding sample based on the predefined user item modal graph information of the corresponding sample, wherein the predefined user item adjacency matrix information includes a predefined user adjacency matrix and a predefined item adjacency matrix. In this way, the quantization of the predefined user item modal graph information can be achieved by generating the predefined user item adjacency matrix information, and a data basis can be provided for subsequent processing. The product of the predefined user adjacency matrix and a first preset coefficient is determined as a predefined user coefficient matrix. In this way, the elements in the predefined user adjacency matrix can be adjusted, important element information can be highlighted, unimportant element information can be reduced, and noise data interference on the subsequent integrated user modal graph can be reduced. The product of the predefined item adjacency matrix and the first preset coefficient is determined as a predefined item coefficient matrix. In this way, the elements in the predefined item adjacency matrix can be adjusted, important element information can be highlighted, unimportant element information can be reduced, and noise data interference on the subsequent integrated item modal graph can be reduced. Dynamic user item adjacency matrix information corresponding to the dynamic user item modal graph information of the corresponding sample is generated based on the dynamic user item modal graph information of the corresponding sample, wherein the dynamic user item adjacency matrix information includes a dynamic user adjacency matrix and a dynamic item adjacency matrix. In this way, the quantization of the dynamic user item modal graph information can be achieved by generating the dynamic user item adjacency matrix information, and a data basis can be provided for subsequent processing. The product of the dynamic user adjacency matrix and a second preset coefficient is determined as a dynamic user coefficient matrix. In this way, the elements in the dynamic user adjacency matrix can be adjusted, important element information can be highlighted, unimportant element information can be reduced, and noise data interference on the subsequent integrated user modal graph can be reduced. The product of the dynamic item adjacency matrix and the second preset coefficient is determined as a dynamic item coefficient matrix. In this way, the elements in the dynamic item adjacency matrix can be adjusted, important element information can be highlighted, unimportant element information can be reduced, and noise data interference on the subsequent integrated item modal graph can be reduced. The sum of the predefined user coefficient matrix and the dynamic user coefficient matrix is determined as an integrated user adjacency matrix.Thus, the multi-layer dimensional sample user modal information can be obtained through the obtained integrated user adjacency matrix, and the generalization ability of the model and the accuracy of the model prediction can be further improved. The sum of the above predefined item coefficient matrix and the above dynamic item coefficient matrix is determined as an integrated item adjacency matrix. Thus, the multi-layer dimensional sample item modal information can be obtained through the obtained integrated user adjacency matrix, and the generalization ability of the model and the accuracy of the model prediction can be further improved. The above integrated user adjacency matrix and the above integrated item adjacency matrix are determined as integrated user item adjacency matrix information. Thus, the integrated user adjacency matrix and the integrated item adjacency matrix can be combined into the integrated user item adjacency matrix information. Based on the above integrated user item adjacency matrix information, integrated user item modal graph information corresponding to the above sample is generated, wherein the above integrated user item modal graph information includes an integrated user modal graph and an integrated item modal graph. Thus, the modal graph information of the user and the item can be updated through the processing of the adjacency matrix information, and the integrated user modal graph and the integrated item modal graph obtained after the processing can comprehensively reflect the connection relationship between all users and all items, which helps to improve the generalization ability of the model and further improve the accuracy of the model prediction. Because the integrated user item modal graph information is obtained on the basis of the predefined user item modal graph information and the dynamic user item modal graph information, the interference of noise data can be effectively reduced through the constructed multi-layer graph information, the generalization ability of the model is improved, and the accuracy of the model prediction is improved.

[0071] In step 1023, node feature information is generated based on the integrated user modal graph and the integrated item modal graph included in the integrated user item modal graph information corresponding to each sample in the at least one sample.

[0072] In some embodiments, the execution subject can generate node feature information based on the integrated user modal graph and the integrated item modal graph included in the integrated user item modal graph information corresponding to each of the at least one sample. The node feature information includes user node feature information corresponding to the integrated user modal graph and item node feature information corresponding to the integrated item modal graph. In practice, first, the execution subject can input the integrated user modal graph into the graph neural network. Then, the execution subject can generate a set of feature vectors corresponding to the integrated user modal graph through the message passing mechanism in the graph neural network. Then, for each set of feature vectors in the set of feature vectors, the execution subject can sum each feature vector in each set of feature vectors. Then, the execution subject can divide the summed feature vector by the iteration number to obtain a user node feature information, obtain each user node feature information corresponding to each set of feature vectors, and further obtain each user node feature information. Then, the execution subject can input the integrated item modal graph into the graph neural network. Then, the execution subject can generate a set of feature vectors corresponding to the integrated item modal graph through the message passing mechanism in the graph neural network. Then, for each set of feature vectors in the set of feature vectors, the execution subject can sum each feature vector in each set of feature vectors. Then, the execution subject can divide the summed feature vector by the iteration number to obtain an item node feature information, obtain each item node feature information corresponding to each set of feature vectors, and further obtain each item node feature information. Finally, the execution subject can combine the user node feature information corresponding to the integrated user modal graph and the item node feature information corresponding to the integrated item modal graph to obtain the node feature information.

[0073] At step 1024, the execution subject inputs the node feature information corresponding to each of the at least one sample into the initial interaction information prediction model to obtain the interaction information corresponding to each of the at least one sample.

[0074] In some embodiments, the execution subject can input the node feature information corresponding to each of the at least one sample into an initial interaction information prediction model to obtain the interaction information corresponding to each of the at least one sample. The initial interaction information prediction model can be a neural network model taking each user node feature information and each item node feature information included in the node feature information as input and taking the interaction information between the sample user information corresponding to each user node feature information and the sample item information corresponding to each item node feature information as output. For example, the initial interaction information prediction model can be a convolutional neural network. The interaction information can be information indicating that the user corresponding to the sample user information interacts with the item corresponding to the sample item information. The interaction can be that the user corresponding to the sample user information clicks the detail page of the item corresponding to the sample item information. The interaction information can be, but is not limited to, label information indicating whether the user corresponding to the sample user information interacts with the item corresponding to the sample item information. For example, the interaction information can be "interaction" or "no interaction".

[0075] In some optional implementations of some embodiments, the execution subject can input the node feature information corresponding to each of the at least one sample into an initial interaction information prediction model to obtain the interaction information corresponding to each of the at least one sample by the following steps:

[0076] In a first step, user-item interaction data feature information is generated based on pre-stored user-item interaction data. The user-item interaction data can be data corresponding to a user and an item that have interacted. The user-item interaction data can include user data and item data. The interaction can be that the user corresponding to the user data has clicked on a detail page of the item corresponding to the item data. The user-item interaction data feature information includes user interaction data feature information and item interaction data feature information. The user interaction data feature information can be a feature vector corresponding to the user data in the user-item interaction data. The item interaction data feature information can be a feature vector corresponding to the item data in the user-item interaction data. The user data can be data corresponding to a user. For example, the user data can be account data of the user, item flow records (for example, all household items purchased within a year), and evaluations of items in the item flow records. The item data can be data corresponding to an item. For example, the item data can be image data of the item, item flow volume (for example, item purchase volume), and unit value information (for example, item unit price). The item in the item data can be an item that has interacted with the user corresponding to the user data. In practice, first, the execution subject can determine each user data and each item data included in the user-item interaction data as each user node and each item node according to the user-item interaction data. Then, the user node and the item node corresponding to the user and the item that have interacted can be connected. Then, the number of times the item data appears in the item flow records of the user data can be determined as the weight between the corresponding user node and the item node, and a user-item interaction bipartite graph is obtained. Then, the user-item interaction bipartite graph can be feature extracted by a pre-trained feature extraction network to obtain the user interaction data feature information and the item interaction data feature information. Finally, the execution subject can combine the user interaction data feature information and the item interaction data feature information to obtain the user-item interaction data feature information. The feature extraction network can be various neural networks that take a bipartite graph as input and output a feature vector of each node in the bipartite graph. For example, the feature extraction network can be a graph convolutional neural network.

[0077] In a second step, for the node feature information corresponding to each sample in at least one sample, the following steps are performed:

[0078] In a first sub-step, the sum of each user node feature information included in the node feature information is determined as accumulated user node feature information.

[0079] A second sub-step, determining a sum of the accumulated user node feature information and user interaction data feature information included in the user item interaction data feature information as the to-be-fused user node feature information.

[0080] A third sub-step, determining a ratio of the to-be-fused user node feature information and the preset quantity as the fused user node feature information.

[0081] A fourth sub-step, determining a sum of each item node feature information included in the node feature information as the accumulated item node feature information.

[0082] A fifth sub-step, determining a sum of the accumulated item node feature information and item interaction data feature information included in the user item interaction data feature information as the to-be-fused item node feature information.

[0083] A sixth sub-step, determining a ratio of the to-be-fused item node feature information and the preset quantity as the fused item node feature information.

[0084] A seventh sub-step, generating interaction information between the fused user node feature information and the fused item node feature information based on the fused user node feature information and the fused item node feature information. In practice, first, the execution subject can perform inner product processing on the fused user node feature information and the fused item node feature information to obtain an inner product value. Then, the interaction information is output according to the inner product value. The inner product value can be a value obtained by performing inner product processing on the fused user node feature information and the fused item node feature information. Specifically, in response to determining that the inner product value meets a preset output condition, the output interaction information is “interaction”. In response to determining that the inner product value does not meet the preset output condition, the output interaction information is “no interaction”. The preset output condition can be that the inner product value is greater than a preset interaction value. The preset interaction value can be a preset value. Here, the specific setting of the preset interaction value is not limited.

[0085] Step 1025, comparing the interaction information corresponding to each sample in the at least one sample with the corresponding sample interaction information.

[0086] In some embodiments, the execution subject can compare the interaction information corresponding to each sample in the at least one sample with the corresponding sample interaction information.

[0087] Step 1026, determining whether the initial interaction information prediction model meets a preset optimization target according to the comparison result.

[0088] In some embodiments, the execution subject can determine whether the initial interaction information prediction model reaches a preset optimization target according to a comparison result. The comparison result can be information used to represent whether the interaction information is the same as or different from the corresponding sample interaction information. For example, the comparison result can be "same". The preset optimization target can be that a proportion of "same" information in the comparison result meets a preset optimization threshold. The preset optimization threshold can be a preset numerical value. Here, the specific setting of the preset optimization threshold is not limited.

[0089] In step 1027, in response to determining that the initial interaction information prediction model reaches the optimization target, the initial interaction information prediction model is taken as a trained interaction information prediction model.

[0090] In some embodiments, the execution subject can take the initial interaction information prediction model as a trained interaction information prediction model in response to determining that the initial interaction information prediction model reaches the optimization target.

[0091] Optionally, after step 1027, in response to determining that the initial interaction information prediction model does not reach the optimization target, the execution subject can further adjust the network parameters of the initial interaction information prediction model, and use a sample set composed of unused samples to use the adjusted initial interaction information prediction model as the initial interaction information prediction model to execute the training step again. In practice, the execution subject can first execute an experiment through a Tensorflow framework. Then, the parameters are adjusted in a gradient descent manner. Specifically, the execution subject can adjust the model parameters through an Adam optimizer (Adaptive Moment Estimation). For example, the execution subject can set the initial learning rate to 0.001. The batch can be set to 256. The vector dimension of the initial modal feature vector of the user and the item in the graph neural network can be set to 32. The vector dimension of the interaction feature vector can be set to 32. The ratio of the regularization technique to eliminate nodes can be set to 0.2. The number of iterations of the message passing mechanism in the graph neural network can be set to 3. The regularization technique can be a dropout technique.

[0092] The above various embodiments of the present disclosure have the following beneficial effects: through the interaction information prediction model training method of some embodiments of the present disclosure, the interference of noise data in modal data on graph information can be effectively reduced, the accuracy of the constructed graph information can be improved, and the reliability of the node feature information and the interaction information generated using the graph information can be improved. Furthermore, the accuracy of the recommended items according to the interaction information can be improved, and the waste of network resources when the user views the items can be reduced. Specifically, the reason why the reliability of the node feature information and the interaction information generated using the graph information is low is that when the modal feature information of the modal data is directly used to construct the graph information, the noise data in the modal data has a large interference on the graph information, resulting in low accuracy of the constructed graph information, low reliability of the node feature information and the interaction information generated using the graph information, and further resulting in poor accuracy of the recommended items when the items are recommended according to the generated interaction information, causing waste of network resources when the user views the items. Based on this, the interaction information prediction model training method of some embodiments of the present disclosure first acquires a sample set, wherein the samples in the sample set include sample user-item interaction information, the sample user-item interaction information includes sample user information, sample item information and sample interaction information, the sample user information includes a sample user modal information group, and the sample item information includes a sample item modal information group. In this way, data for model training can be obtained. Then, the following training steps are performed based on the sample set: first, based on each sample user-item interaction information included in at least one sample in the sample set, each dynamic user-item modal graph information is generated, wherein each dynamic user-item modal graph information corresponds to one sample, and each dynamic user-item modal graph information includes a dynamic user modal graph and a dynamic item modal graph. In this way, a dynamic modal relationship graph can be obtained through data, and the structure of the dynamic modal relationship graph can be optimized through the constantly updated node feature information in the dynamic modal relationship graph, reducing the influence of noise data on the generated graph information. Second, for each sample in the at least one sample, based on the predefined user-item modal graph information corresponding to the sample in each predefined user-item modal graph information and the dynamic user-item modal graph information corresponding to the sample in the each dynamic user-item modal graph information, integrated user-item modal graph information corresponding to the sample is generated, wherein the integrated user-item modal graph information includes an integrated user modal graph and an integrated item modal graph. In this way, a more comprehensive and stable graph information can be generated by combining the predefined user-item modal graph information and the dynamic user-item modal graph information, more effectively reducing the influence of noise data on the graph information and improving the accuracy of the generated interaction information.Then, based on the integrated user modal graph and the integrated item modal graph included in the integrated user-item modal graph information corresponding to each of the at least one sample, node feature information is generated, wherein the node feature information includes user node feature information corresponding to the integrated user modal graph and item node feature information corresponding to the integrated item modal graph. In this way, the node feature information can be more accurately obtained by integrating the user-item modal graph information, and the accuracy of the interaction information generated by using the node feature information can be improved. Then, each of the node feature information corresponding to the at least one sample is input into the initial interaction information prediction model to obtain the interaction information corresponding to each of the at least one sample. In this way, the model can be trained by using the obtained node feature information, and more accurate interaction information can be obtained, and the accuracy of the recommended items can be improved. Then, the interaction information corresponding to each of the at least one sample is compared with the corresponding sample interaction information. In this way, the comparison result can be obtained by comparing the interaction information with the corresponding sample interaction information, and whether the model reaches the preset optimization target can be verified by using the comparison result. Then, whether the initial interaction information prediction model reaches the preset optimization target is determined according to the comparison result. In this way, the performance of the model can be verified by using the comparison result. Finally, in response to determining that the initial interaction information prediction model reaches the optimization target, the initial interaction information prediction model is used as the trained interaction information prediction model. In this way, an interaction information prediction model that reaches the preset optimization target can be obtained, and more accurate interaction information can be obtained by using the model, and the accuracy of the recommended items can be improved. Because the predefined user-item modal graph information and the dynamic user-item modal graph information can be obtained, and then the integrated user-item modal graph information is generated, the interference of the noise data on the graph information is reduced by using the multi-layer constructed graph information, the reliability of the node feature information and the interaction information is improved, and the accuracy of the recommended items generated according to the interaction information is improved, and the waste of network resources in recommending and viewing items by users is reduced. Because the performance of the model can be verified by using the comparison result, the deviation of the model can be identified, and the model can be adjusted to ensure that the model reaches the optimization target, more reliable interaction information can be generated, and the accuracy of the recommended items generated according to the interaction information can be improved, and the waste of network resources in recommending and viewing items by users is reduced. In summary, by processing the graph information, the interference of the noise data on the graph information can be reduced, the reliability of the node feature information and the interaction information generated by using the graph information can be improved, the accuracy of the recommended items generated according to the interaction information can be improved, and the waste of network resources in recommending and viewing items by users can be reduced.

[0093] With continued reference to Figure 2 , a flow 200 illustrating some embodiments of a method of generating recommendation information according to the present disclosure is shown. The method of generating recommendation information includes the following steps:

[0094] Step 201, for each user information in the user information set, the following steps are performed:

[0095] Step 2011, obtaining each user item interaction information corresponding to the user information.

[0096] In some embodiments, the execution subject (for example, a computing device) of the recommendation information generation method can obtain each user item interaction information corresponding to the user information. The execution subject can be a terminal or a server that generates recommendation information. The user information can be information corresponding to a user. For example, the user information can include but is not limited to: account information corresponding to the user, item flow record (for example: daily necessities purchased within a year), and evaluation information of items in the item flow record. The evaluation information can include but is not limited to: evaluation text, evaluation score, and evaluation video. The user item interaction information can be various information corresponding to each item in the item flow record corresponding to the user information. For example, the user item interaction information can include but is not limited to: image information corresponding to each item, item flow volume (for example: item purchase volume), and unit value information (for example: item unit price).

[0097] Step 2012, for each user item interaction information in the user item interaction information, determining a preset number of item information in the preset item information set that satisfies a preset interaction condition as interaction item information corresponding to the user item interaction information.

[0098] In some embodiments, the execution subject can determine, for each user item interaction information in the user item interaction information, a preset number of item information in the preset item information set that satisfies a preset interaction condition as interaction item information corresponding to the user item interaction information. The item information in the item information set can be various information corresponding to an item. The item information can include but is not limited to: image information corresponding to the item, item flow volume (for example: item purchase volume), and unit value information (for example: item unit price). The item information in the item information set can be information corresponding to any item in the item flow record of the user information or information corresponding to any item not recorded in the item flow record of the user information. The preset interaction condition can be that the user corresponding to the user information has not interacted with the item corresponding to the item information. The interaction can be that the user corresponding to the user information has performed a click operation on the detail page of the item corresponding to the item information. The preset number can be a pre-set value. Here, the specific setting of the preset number is not limited.

[0099] Step 2013: Integrate the user item interaction information and the determined interactive item information to obtain an integrated user item interaction information group.

[0100] In some embodiments, the executing entity may integrate the various user item interaction information and the determined information of each interactive item to obtain an integrated user item interaction information group. In practice, the executing entity may combine the various user item interaction information and the determined information of each interactive item into the integrated user item interaction information group.

[0101] Step 2014: Based on the pre-trained interaction information prediction model, generate the interaction information corresponding to each integrated user item interaction information in the integrated user item interaction information group.

[0102] In some embodiments, the execution entity can generate interaction information corresponding to each integrated user item interaction information in the integrated user item interaction information group based on a pre-trained interaction information prediction model. The interaction information prediction model is obtained through... Figure 1 The training steps in the corresponding embodiments are used. The interaction information corresponding to each integrated user-item interaction information in the aforementioned integrated user-item interaction information group can be information about the interaction between the aforementioned user information and the aforementioned item information. The interaction information can be, but is not limited to, the probability of the aforementioned user information interacting with the aforementioned item information. In practice, the executing entity can input each integrated user-item interaction information in the aforementioned integrated user-item interaction information group into the interaction information prediction model to obtain the interaction information corresponding to each integrated user-item interaction information.

[0103] Step 2015: Sort the generated interaction information to obtain the interaction information sequence.

[0104] In some embodiments, the execution entity can sort the generated interaction information to obtain an interaction information sequence. In practice, the execution entity can sort the generated interaction information in descending order to obtain the interaction information sequence.

[0105] Step 2016: Generate item recommendation information corresponding to user information based on the interaction information sequence.

[0106] In some embodiments, the execution subject can generate the item recommendation information corresponding to the user information according to the interaction information sequence. In practice, the execution subject can determine the item information corresponding to a preset number of interaction information in the interaction information sequence satisfying a preset sequence condition as the item recommendation information corresponding to the user information. For example, the execution subject can select the first five interaction information from the interaction information sequence, and determine the item information corresponding to the first five interaction information as the item recommendation information. The item recommendation information can be used to represent a preset number of item information in the preset item information set satisfying the preset sequence condition. The preset sequence condition can be that the interaction information is the first preset number of interaction information in the interaction information sequence. The preset number can be a preset value. Here, the specific setting of the preset number is not limited.

[0107] Optionally, after step 2016, the execution subject can also send the item recommendation information to the user terminal corresponding to the user information. Thus, the execution subject can send the generated item recommendation information with high accuracy to the user, and can reduce the waste of network resources when recommending network resources and the user views the item.

[0108] The above various embodiments of the present disclosure have the following beneficial effects: through the recommendation information generation method of some embodiments of the present disclosure, the accuracy of the recommended items when recommending items according to interaction information can be improved, and the waste of network resources when recommending network resources and users viewing items can be reduced. Specifically, the reason why the accuracy of the model prediction is low and the diversity of the generated recommendation information is insufficient is that when the modal feature information of the modal data is directly used to construct the graph information, the noise data in the modal data has a greater interference on the graph information, resulting in that the accuracy of the constructed graph information is low, the reliability of the node feature information and the interaction information generated by using the graph information is low, and then the accuracy of the recommended items when recommending items according to the generated interaction information is poor, resulting in the waste of network resources when recommending network resources and users viewing items. Based on this, the recommendation information generation method of some embodiments of the present disclosure first, for each user information in the user information set, the following steps are performed: first, obtain each user-item interaction information corresponding to the user information. In this way, the item information actually interacted with the user can be obtained, providing a data basis for recommendation information generation. For each user-item interaction information in the above user-item interaction information, a preset number of item information in the preset item information set that satisfies the preset interaction condition is determined as the interaction item information corresponding to the user-item interaction information. In this way, the diversity of the sample can be improved by introducing item information that has not actually interacted with the user, thereby improving the accuracy of the generated interaction information and reducing the waste of network resources when recommending network resources and users viewing items. The above each user-item interaction information and each interaction item information are integrated to obtain an integrated user-item interaction information group. In this way, by combining the item information actually interacted with the user and the item information not actually interacted with the user, a basis for generating more accurate interaction information can be provided, thereby reducing the waste of network resources when recommending network resources and users viewing items. Based on the pre-trained interaction information prediction model, the interaction information corresponding to each integrated user-item interaction information in the integrated user-item interaction information group is generated, wherein the interaction information prediction model is trained by Figure 1The steps in the corresponding embodiments are trained. Thus, the probability of the user interacting with the item can be obtained. The generated interaction information is sorted to obtain an interaction information sequence. Thus, the item information with the highest probability of interacting with the user can be obtained by sorting. According to the above interaction information sequence, the item recommendation information corresponding to the above user information is generated. Thus, according to the interaction information sequence, the item information with the highest probability of interacting with the user can be preferentially recommended and displayed on the user terminal, reducing the waste of network resources in recommending and the user viewing the item. Also, when the model makes a prediction, not only the item information that has actually interacted with the user is input, but also the item information that has not actually interacted with the user, so the accuracy of the generated interaction information can be improved, and the waste of network resources in recommending and the user viewing the item can be reduced. Also, after obtaining the interaction information, the interaction information is sorted, and the item recommendation information corresponding to the above user information is generated according to the interaction information sequence, so the accuracy of the generated item recommendation information can be improved, and the waste of network resources in recommending and the user viewing the item can be reduced.

[0109] Further reference Figure 3 As an implementation of the method shown in the above figures, the disclosure provides some embodiments of an interaction information prediction model training device, which corresponds to the method embodiments shown in Figure 1 The device can be applied to various electronic devices.

[0110] As Figure 3As shown, the interaction information prediction model training apparatus 300 of some embodiments includes an obtaining unit 301 and a training unit 302. The obtaining unit 301 is configured to obtain a sample set, wherein the samples in the sample set include sample user-item interaction information, the sample user-item interaction information includes sample user information, sample item information and sample interaction information, the sample user information includes a set of sample user modal information, and the sample item information includes a set of sample item modal information. The training unit 302 is configured to perform the following training steps based on the sample set: based on each sample user-item interaction information included in at least one sample in the sample set, generating each dynamic user-item modal graph information, wherein each dynamic user-item modal graph information in the dynamic user-item modal graph information corresponds to one sample, and each dynamic user-item modal graph information includes a dynamic user modal graph and a dynamic item modal graph; for each sample in the at least one sample, based on a predefined user-item modal graph information corresponding to the sample in each predefined user-item modal graph information generated in advance and a dynamic user-item modal graph information corresponding to the sample in the each dynamic user-item modal graph information, generating integrated user-item modal graph information corresponding to the sample, wherein the integrated user-item modal graph information includes an integrated user modal graph and an integrated item modal graph; based on the integrated user modal graph and the integrated item modal graph included in the integrated user-item modal graph information corresponding to each sample in the at least one sample, generating node feature information, wherein the node feature information includes each user node feature information corresponding to the integrated user modal graph and each item node feature information corresponding to the integrated item modal graph; inputting each node feature information corresponding to the at least one sample to an initial interaction information prediction model to obtain interaction information corresponding to each sample in the at least one sample; comparing the interaction information corresponding to each sample in the at least one sample with the sample interaction information corresponding to the sample; determining whether the initial interaction information prediction model reaches a preset optimization target according to the comparison result; and in response to determining that the initial interaction information prediction model reaches the optimization target, taking the initial interaction information prediction model as a trained interaction information prediction model.

[0111] It can be understood that the units described in the apparatus 300 correspond to the steps in the methods described with reference to Figure 1 Thus, the operations, features and advantages described above for the methods also apply to the apparatus 300 and the units included therein, which will not be described here again.

[0112] Further with reference to Figure 4 as an implementation of the methods shown in the above figures, the present disclosure provides some embodiments of a recommendation information generation apparatus, which correspond to the method embodiments shown in Figure 2 and can be applied in various electronic devices.

[0113] As Figure 4 shown, the recommendation information generation apparatus 400 of some embodiments includes an execution unit 401. Wherein the execution unit 401 is configured to, for each user information in the user information set, perform the following steps: obtain each user-item interaction information corresponding to the above-mentioned user information; for each user-item interaction information in the above-mentioned each user-item interaction information, determine a preset number of item information in the preset item information set that satisfies the preset interaction condition as the interaction item information corresponding to the above-mentioned user-item interaction information; perform integration processing on the above-mentioned each user-item interaction information and each determined interaction item information, to obtain an integrated user-item interaction information group; generate the interaction information corresponding to each integrated user-item interaction information in the above-mentioned integrated user-item interaction information group based on a pre-trained interaction information prediction model, wherein the above-mentioned interaction information prediction model is trained by the steps in the corresponding embodiments; perform sorting processing on the generated each interaction information to obtain an interaction information sequence; and generate the item recommendation information corresponding to the above-mentioned user information according to the above-mentioned interaction information sequence. Figure 1

[0114] It can be understood that the units recorded in the apparatus 400 correspond to each step in the method described with reference to Figure 2 . Therefore, the operations, features and beneficial effects described above for the method also apply to the apparatus 400 and the units contained therein, which will not be described here again.

[0115] Reference is made below to Figure 5 , which shows a structural schematic diagram of an electronic device 500 suitable for use to implement some embodiments of the present disclosure. Figure 5 The electronic device shown is merely an example and should not impose any limitation on the function and use range of the embodiments of the present disclosure.

[0116] As Figure 5 shown, the electronic device 500 can include a processing apparatus (such as a central processor, a graphics processor, etc.) 501, which can perform various appropriate actions and processes according to programs stored in a read-only memory (ROM) 502 or loaded into a random access memory (RAM) 503 from a storage apparatus 508. In the RAM 503, various programs and data required for the operation of the electronic device 500 are also stored. The processing apparatus 501, the ROM 502, and the RAM 503 are connected to each other through a bus 504. An input / output (I / O) interface 505 is also connected to the bus 504.

[0117] ​In general, the following devices can be connected to the I / O interface 505: input devices 506 including, for example, a touch screen, a touch pad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, and the like; output devices 507 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, and the like; storage devices 508 including, for example, a magnetic tape, a hard disk, and the like; and communication devices 509. The communication devices 509 can allow the electronic device 500 to communicate wirelessly or wired with other devices to exchange data. Although Figure 5 The electronic device 500 is shown with various devices, but it is understood that all of the shown devices are not required to be implemented or present. More or less devices can alternatively be implemented or present. Figure 5 Each block shown in the flowcharts can represent a device or multiple devices as needed.

[0118] In particular, processes described above with reference to the flowcharts can be implemented as a computer software program according to some embodiments of the present disclosure. For example, some embodiments of the present disclosure include a computer program product including a computer program carried on a computer readable medium, the computer program containing program codes for executing the methods shown in the flowcharts. In some such embodiments, the computer program can be downloaded and installed from a network through the communication devices 509, or installed from the storage devices 508, or installed from the ROM 502. When the computer program is executed by the processing devices 501, the above-mentioned functions defined in the methods of some embodiments of the present disclosure are performed.

[0119] Note that the computer readable medium in some embodiments of the present disclosure can be a computer readable signal medium or a computer readable storage medium or any combination thereof. The computer readable storage medium can be, for example but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any suitable combination of the foregoing. More specific examples of the computer readable storage medium can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In some embodiments of the present disclosure, the computer readable storage medium can be any tangible medium that contains or stores a program used by an instruction execution system, apparatus or device, or that can be used by or in connection with an instruction execution system, apparatus or device. In some embodiments of the present disclosure, the computer readable signal medium can include a computer readable program code propagated in or on a carrier medium, in which the computer readable program code is embodied. Such propagated computer readable program code can take many forms, including but not limited to, an electromagnetic signal, an optical signal or any suitable combination of the foregoing. The computer readable signal medium can also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate or transport a program for use by or in connection with an instruction execution system, apparatus or device. Program code embodied on a computer readable medium can be transmitted using any suitable medium, including but not limited to, wire, cable, wireless, RF, infrared or any suitable combination of the foregoing.

[0120] In some embodiments, the client, server, or both can communicate using any current known or future developed network protocol, such as HTTP (HyperText Transfer Protocol), and can be interconnected with any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include local area networks ("LAN"), wide area networks ("WAN"), the Internet, and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any current known or future developed networks.

[0121] The computer readable medium can be included in the electronic device or exist separately from the electronic device. The computer readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: obtain a sample set, wherein sample user item interaction information included in the sample set includes sample user information, sample item information, and sample interaction information, the sample user information includes a set of sample user modal information, and the sample item information includes a set of sample item modal information; perform the following training steps based on the sample set: generate each dynamic user item modal graph information based on each sample user item interaction information included in at least one sample in the sample set, wherein each dynamic user item modal graph information in the dynamic user item modal graph information corresponds to one sample, and each dynamic user item modal graph information includes a dynamic user modal graph and a dynamic item modal graph; for each sample in the at least one sample, generate integrated user item modal graph information corresponding to the sample based on pre-defined user item modal graph information corresponding to the sample in each pre-defined user item modal graph information and dynamic user item modal graph information corresponding to the sample in the dynamic user item modal graph information, wherein the integrated user item modal graph information includes an integrated user modal graph and an integrated item modal graph; generate node feature information based on the integrated user modal graph and the integrated item modal graph included in the integrated user item modal graph information corresponding to each sample in the at least one sample, wherein the node feature information includes each user node feature information corresponding to the integrated user modal graph and each item node feature information corresponding to the integrated item modal graph; input each node feature information corresponding to the at least one sample to an initial interaction information prediction model to obtain interaction information corresponding to each sample in the at least one sample; compare the interaction information corresponding to each sample in the at least one sample with the corresponding sample interaction information; determine whether the initial interaction information prediction model reaches a preset optimization target according to a comparison result; and in response to determining that the initial interaction information prediction model reaches the optimization target, use the initial interaction information prediction model as a trained interaction information prediction model.

[0122] or the electronic device is caused to: for each user information in the set of user information, perform the following steps: obtain each user-item interaction information corresponding to the user information; for each user-item interaction information in the set of each user-item interaction information, determine a preset number of item information in a set of preset item information that satisfy a preset interaction condition as interaction item information corresponding to the user-item interaction information; perform integration processing on the set of each user-item interaction information and the determined set of each interaction item information to obtain a set of integrated user-item interaction information; generate interaction information corresponding to each integrated user-item interaction information in the set of integrated user-item interaction information based on a pre-trained interaction information prediction model, wherein the interaction information prediction model is trained by the steps in the corresponding embodiments; perform sorting processing on the generated set of each interaction information to obtain an interaction information sequence; and generate item recommendation information corresponding to the user information according to the interaction information sequence. Figure 1

[0123] Computer program code for carrying out operations of some embodiments of the present disclosure can be written in any of one or more programming languages, including object oriented programming languages such as Java, Smalltalk, C++, or conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider).

[0124] ​The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0125] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including an acquisition unit and a training unit. The names of these units do not necessarily limit the specific unit; for example, an acquisition unit may also be described as a "unit for acquiring a sample set".

[0126] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0127] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for training an interaction information prediction model, comprising: obtaining a sample set, wherein a sample in the sample set comprises sample user-item interaction information, the sample user-item interaction information comprising sample user information, sample item information, and sample interaction information, the sample user information comprising a set of sample user modal information, and the sample item information comprising a set of sample item modal information; performing the following training steps based on the sample set: generating, based on each sample user-item interaction information included in at least one sample in the sample set, a respective dynamic user-item modal graph information, wherein each dynamic user-item modal graph information in the respective dynamic user-item modal graph information corresponds to one sample, and each dynamic user-item modal graph information comprises a dynamic user modal graph and a dynamic item modal graph; for each sample in the at least one sample, generating, based on a predefined user-item modal graph information corresponding to the sample in the respective predefined user-item modal graph information and a dynamic user-item modal graph information corresponding to the sample in the respective dynamic user-item modal graph information, an integrated user-item modal graph information corresponding to the sample, wherein the integrated user-item modal graph information comprises an integrated user modal graph and an integrated item modal graph; generating, based on the integrated user modal graph and the integrated item modal graph included in the integrated user-item modal graph information corresponding to each sample in the at least one sample, node feature information, wherein the node feature information comprises respective user node feature information corresponding to the integrated user modal graph and respective item node feature information corresponding to the integrated item modal graph; inputting the respective node feature information corresponding to the at least one sample into an initial interaction information prediction model to obtain interaction information corresponding to each sample in the at least one sample; comparing the interaction information corresponding to each sample in the at least one sample with the corresponding sample interaction information; determining whether the initial interaction information prediction model reaches a preset optimization target according to a comparison result; in response to determining that the initial interaction information prediction model reaches the optimization target, taking the initial interaction information prediction model as a trained interaction information prediction model.

2. The method of claim 1, wherein, after the response to determining that the initial interaction information prediction model reaches the optimization target, taking the initial interaction information prediction model as a trained interaction information prediction model, the training steps further comprise: in response to determining that the initial interaction information prediction model does not reach the optimization target, adjusting network parameters of the initial interaction information prediction model, and using a sample set composed of unused samples to perform the training steps again using the adjusted initial interaction information prediction model as the initial interaction information prediction model.

3. The method of claim 1, wherein, the generating, based on each sample user-item interaction information included in at least one sample in the sample set, a respective dynamic user-item modal graph information, comprises: generate, based on the individual sample user-item interaction information included in the at least one sample in the sample set, individual sample user-item interaction feature information, wherein each piece of sample user-item interaction feature information in the individual sample user-item interaction feature information corresponds to a sample, and each piece of sample user-item interaction feature information includes a sample user modality feature information group and a sample item modality feature information group; generate, based on the individual sample user-item interaction feature information, individual predefined user-item modality graph information, wherein each piece of predefined user-item modality graph information in the individual predefined user-item modality graph information corresponds to a sample, and each piece of predefined user-item modality graph information includes a predefined user modality graph and a predefined item modality graph; generate, based on the individual predefined user-item modality graph information, individual dynamic user-item modality graph information, wherein each piece of dynamic user-item modality graph information in the individual dynamic user-item modality graph information corresponds to a sample, and each piece of dynamic user-item modality graph information includes a dynamic user modality graph and a dynamic item modality graph.

4. The method of claim 1, wherein, The method further includes the following steps: generate, based on pre-stored user-item interaction data, user-item interaction data feature information, wherein the user-item interaction data feature information includes user interaction data feature information and item interaction data feature information; for the node feature information corresponding to each sample in the at least one sample, the following steps are performed: determine a sum of individual user node feature information included in the node feature information as accumulated user node feature information; determine a sum of the accumulated user node feature information and user interaction data feature information included in the user-item interaction data feature information as to-be-fused user node feature information; determine a ratio of the to-be-fused user node feature information to a preset quantity as fused user node feature information; determine a sum of individual item node feature information included in the node feature information as accumulated item node feature information; determine a sum of the accumulated item node feature information and item interaction data feature information included in the user-item interaction data feature information as to-be-fused item node feature information; determine a ratio of the to-be-fused item node feature information to a preset quantity as fused item node feature information; generate, based on the fused user node feature information and the fused item node feature information, interaction information between the fused user node feature information and the fused item node feature information.

5. A recommendation information generation method, comprising: for each piece of user information in a user information set, the following steps are performed: obtain individual user-item interaction information corresponding to the user information; for each piece of user-item interaction information in the individual user-item interaction information, determine a preset quantity of item information in a preset item information set that satisfies a preset interaction condition as interaction item information corresponding to the user-item interaction information; integrating the respective user-item interaction information and the determined respective interaction item information to obtain an integrated user-item interaction information set; generating, based on a pre-trained interaction information prediction model, interaction information corresponding to each integrated user-item interaction information in the integrated user-item interaction information set, wherein the interaction information prediction model is trained by the method of any one of claims 1-4; sorting the generated respective interaction information to obtain an interaction information sequence; generating item recommendation information corresponding to the user information according to the interaction information sequence.

6. The method of claim 5, wherein, After the item recommendation information corresponding to the user information is generated according to the interaction information sequence, the method further comprises: sending the item recommendation information to a user terminal corresponding to the user information. 7.An interaction information prediction model training apparatus, comprising: an acquisition unit configured to acquire a sample set, wherein a sample in the sample set comprises sample user-item interaction information, the sample user-item interaction information comprises sample user information, sample item information and sample interaction information, the sample user information comprises a sample user modality information set, and the sample item information comprises a sample item modality information set; a training unit configured to perform the following training steps based on the sample set: generating respective dynamic user-item modality graph information based on respective sample user-item interaction information included in at least one sample in the sample set, wherein each dynamic user-item modality graph information in the respective dynamic user-item modality graph information corresponds to one sample, and each dynamic user-item modality graph information comprises a dynamic user modality graph and a dynamic item modality graph; for each sample in the at least one sample, generating integrated user-item modality graph information corresponding to the sample based on pre-defined user-item modality graph information corresponding to the sample in respective pre-defined user-item modality graph information generated in advance and dynamic user-item modality graph information corresponding to the sample in the respective dynamic user-item modality graph information, wherein the integrated user-item modality graph information comprises an integrated user modality graph and an integrated item modality graph; generating node feature information based on the integrated user modality graph and the integrated item modality graph included in the integrated user-item modality graph information corresponding to each sample in the at least one sample, wherein the node feature information comprises respective user node feature information corresponding to the integrated user modality graph and respective item node feature information corresponding to the integrated item modality graph; inputting respective node feature information corresponding to the at least one sample to an initial interaction information prediction model to obtain interaction information corresponding to each sample in the at least one sample; comparing the interaction information corresponding to each sample in the at least one sample with sample interaction information corresponding to the sample; determining whether the initial interaction information prediction model reaches a preset optimization target according to a comparison result; and in response to determining that the initial interaction information prediction model reaches the optimization target, taking the initial interaction information prediction model as a trained interaction information prediction model. 8.A recommendation information generation apparatus, comprising: The execution unit is configured to, for each user information in the user information set, perform the following steps: obtain respective user-item interaction information corresponding to the user information; For each user-item interaction information in the respective user-item interaction information, determine a preset number of item information in a preset item information set that satisfies a preset interaction condition as interaction item information corresponding to the user-item interaction information; perform integration processing on the respective user-item interaction information and the determined respective interaction item information to obtain an integrated user-item interaction information group; generate interaction information corresponding to each integrated user-item interaction information in the integrated user-item interaction information group based on a pre-trained interaction information prediction model, wherein the interaction information prediction model is trained by the method of any one of claims 1-4; perform sorting processing on the generated respective interaction information to obtain an interaction information sequence; and generate item recommendation information corresponding to the user information according to the interaction information sequence. 9.An electronic device, comprising: one or more processors; a memory device having stored thereon one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-4 or 5-6.

10. A computer readable medium having stored thereon a computer program, wherein, The program is executed by the processor to implement the method of any one of claims 1-4 or 5-6.

Images