Information matching method, device, computer equipment and storage medium

Abstract

The present application can be applied to various scenarios such as cloud technology, artificial intelligence, intelligent transportation, assisted driving, etc., and relates to an information matching method, device, computer device, computer-readable storage medium, and computer program product. The method includes: obtaining the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information; calculating the information similarity features between the information features and each type of interaction behavior sequence, and calculating the behavior similarity features between various types of interaction behavior sequences; performing weighted summation on the object features, information similarity features, and behavior similarity features to obtain the target object interaction features of the target object; performing similarity matching between the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result of the target information and the target object. Using the above method for information matching has the advantage of high accuracy.

Description

Technical Field

[0001] This application relates to the field of computer application technologies, and in particular, to an information matching method, apparatus, computer device, computer-readable storage medium, and computer program product. Background Art

[0002] With the development of computer technologies, more and more users obtain information through the network. Stimulated by the strong information needs of Internet users, information service platforms avoid user loss by continuously providing users with a vast amount of information resources. However, faced with a vast amount of information resources, it is difficult for users to quickly and effectively extract their required information from a large amount of information, resulting in the problem of information overload. To alleviate the problem of information overload, information matching methods have emerged as the times require.

[0003] Traditional information matching methods establish user preferences based on the user's historical interaction behaviors such as clicks, browsing, and purchases, and then screen information resources that match the user's interests based on the user preferences. Using traditional information matching methods, the information obtained by matching is highly similar to the associated information of the user's historical interaction behaviors, and the information type is single, unable to meet the diverse needs of users' interests. Therefore, traditional information matching methods have the disadvantage of low accuracy of matching results. Summary of the Invention

[0004] Based on this, it is necessary to provide an information matching method, apparatus, computer device, computer-readable storage medium, and computer program product that can improve the accuracy of matching results for the above technical problems.

[0005] In a first aspect, this application provides an information matching method. The method includes:

[0006] Obtain the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information;

[0007] Calculate the information similarity features between the information features and each type of the interaction behavior sequences, and calculate the behavior similarity features between various types of the interaction behavior sequences;

[0008] Perform weighted summation on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object;

[0009] Perform similarity matching between the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result of the target information and the target object.

[0010] Second aspect, the present application further provides an information matching device, characterized in that the device includes:

[0011] An acquisition module, configured to acquire the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information;

[0012] A similarity feature calculation module, configured to calculate the information similarity features between the information features and each type of the interaction behavior sequences, and calculate the behavior similarity features between various types of the interaction behavior sequences;

[0013] An interaction feature determination module, configured to perform weighted summation on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object;

[0014] A matching module, configured to perform similarity matching between the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result between the target information and the target object.

[0015] Third aspect, the present application further provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0016] Acquire the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information;

[0017] Calculate the information similarity features between the information features and each type of the interaction behavior sequences, and calculate the behavior similarity features between various types of the interaction behavior sequences;

[0018] Perform weighted summation on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object;

[0019] Perform similarity matching between the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result between the target information and the target object.

[0020] Fourth aspect, the present application further provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0021] Acquire the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information;

[0022] Calculate the information similarity features between the information features and each type of the interaction behavior sequences, and calculate the behavior similarity features between each type of the interaction behavior sequences;

[0023] Perform a weighted sum on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object;

[0024] Perform a similarity match between the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result between the target information and the target object.

[0025] In a fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the following steps are implemented:

[0026] Obtain the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information;

[0027] Calculate the information similarity features between the information features and each type of the interaction behavior sequences, and calculate the behavior similarity features between each type of the interaction behavior sequences;

[0028] Perform a weighted sum on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object;

[0029] Perform a similarity match between the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result between the target information and the target object.

[0030] The above-mentioned information matching method, device, computer equipment, computer-readable storage medium and computer program product obtain information characteristics of target information, object characteristics of target object, and at least two types of interactive behavior sequences generated by the target object for historical information, calculate the information similarity characteristics between the information characteristics and each type of interactive behavior sequence, and calculate the behavioral similarity characteristics between various types of interactive behavior sequences, and then perform weighted summation on the object characteristics, the information similarity characteristics, and the behavioral similarity characteristics to obtain the target object interaction characteristics of the target object, which is equivalent to comprehensively considering the object characteristics of the target object, the information similarity characteristics of the target information and each type of interactive behavior sequence, and the behavioral similarity characteristics between various types of interactive behavior sequences, and determining the target object interaction characteristics of the target object from multiple different similarity dimensions, which can mine the potential interactive characteristics of the target object, is conducive to improving the accuracy of the target object interaction characteristics, and thereby improves the accuracy of the matching results of the target information and the target object obtained by similarly matching the target object interaction characteristics and the target information interaction characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 An application environment diagram of an information matching method in an embodiment;

[0032] Figure 2 is a schematic diagram of a flow chart of an information matching method in an embodiment;

[0033] Figure 3 A schematic diagram of a process of obtaining information features in one embodiment;

[0034] Figure 4 A schematic diagram of a process of obtaining object features in one embodiment;

[0035] Figure 5 A schematic diagram of a process for calculating information features and information similarity features of a certain type of interactive behavior sequence based on a self-attention mechanism in one embodiment;

[0036] Figure 6 Schematic diagram of a process for respectively calculating information similarity features of information features and explicit positive interaction behavior sequences, explicit negative interaction behavior sequences, and implicit interaction behavior sequences, and behavior similarity features of implicit interaction behavior sequences and explicit interaction behavior sequences based on an attention mechanism in one embodiment;

[0037] Figure 7 is a flow chart of an information matching method in another embodiment;

[0038] Figure 8 A schematic diagram of a process of information matching in an information push application scenario in one embodiment;

[0039] Figure 9 It is a structural block diagram of an information matching device in an embodiment;

[0040] Figure 10 It is an internal structure diagram of a computer device in an embodiment. Specific implementation manners

[0041] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0042] In one embodiment, the information matching method provided by the present application can be applied to an application environment as Figure 1 shown. Among them, the terminal 102 communicates with the server 104 through a network. The data storage system can store the data that the server 104 needs to process. The data storage system can be integrated on the server 104, or can be placed in the cloud or on other servers. Specifically, in the process of the server 104 performing information matching: obtaining the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information; calculating the information similarity features between the information features and each type of interaction behavior sequence, and calculating the behavior similarity features between various types of interaction behavior sequences; performing weighted summation on the object features, information similarity features, and behavior similarity features to obtain the target object interaction features of the target object; performing similarity matching on the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result of the target information and the target object.

[0043] In one embodiment, when the data processing capability of the terminal 102 meets the data processing requirements, the information method provided by the embodiments of the present application may only involve the terminal 102 in its application environment. Specifically, the terminal 102 obtains the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information, calculates the information similarity features between the information features and each type of interaction behavior sequence, and calculates the behavior similarity features between various types of interaction behavior sequences. Then, the terminal 102 performs weighted summation on the object features, information similarity features, and behavior similarity features to obtain the target object interaction features of the target object. Finally, the terminal 102 performs similarity matching on the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result of the target information and the target object.

[0044] Among them, the terminal 102 includes, but is not limited to, mobile phones, computers, intelligent voice interaction devices, smart home appliances, vehicle-mounted terminals, aircraft, etc. Embodiments of the present invention can be applied to various scenarios, including, but not limited to, cloud technology, artificial intelligence, intelligent transportation, assisted driving, etc. The server 104 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The terminal 102 and the server 104 can be directly or indirectly connected through wired or wireless communication methods, and this application does not limit this here.

[0045] In one embodiment, as Figure 2 shown, an information matching method is provided. In this embodiment, this method is exemplified by being applied to the server 104. It can be understood that this method can also be applied to the terminal 102, and can also be applied to a system including the terminal 102 and the server 104, and is implemented through the interaction between the terminal 102 and the server 104. In this embodiment, the method includes the following steps:

[0046] Step S201, obtain the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information.

[0047] Among them, the target object refers to the object for which information recommendation needs to be made based on the information matching result. The target information refers to the information to be determined whether it needs to be recommended to the target object. The types of the target information include, but are not limited to, advertisements, articles, news, short videos, application programs, etc. The information features of the target information refer to the features used to characterize the characteristics of the target information, specifically including the information identifier, category features, brand features, and item features of the target information, etc. Among them, the item features include, but are not limited to, semantic features, image features, etc. included in the target information. In some embodiments, the information features further include context features, and the context features include, but are not limited to, the context information, time, and terminal device features of the access request associated with the target information, etc. The object features of the target object refer to the features used to characterize the characteristics of the target object, specifically including, but not limited to, the object identifier, basic attribute features, behavioral interest features, etc. Among them, the basic attribute features include features such as name, gender, age, and city where the user is located, and the behavioral interest features include browsing behavioral interest features, click behavioral interest features, etc. It should be noted that the information (including but not limited to identification information, feature information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0048] Further, the interaction behavior refers to the behavior generated by the target object during the interaction process with historical information. The interaction behavior may specifically include an explicit interaction behavior that has performed an operation on the historical information, and an implicit interaction behavior that has not performed an operation on the historical information. The explicit interaction behavior further includes an explicit positive interaction behavior that has performed a positive feedback operation on the historical information, such as behaviors of clicking, favoriting, liking, giving a good review, downloading, and purchasing, etc., and an explicit negative interaction behavior that has performed a negative feedback operation on the historical information, such as behaviors of rejecting, giving a bad review, uninstalling, etc. The interaction behavior sequence refers to a behavior sequence composed of the interaction behaviors generated by the target object for multiple historical information. For example, the explicit positive interaction behavior sequence can be expressed as {c 1 ,…,c n1}, where n1 represents the sequence length, and each item in the sequence, such as c 1 , can represent an independent interaction behavior, carrying information such as a timestamp, an object identifier, an information identifier, and a behavior type. Among them, the timestamp is used to characterize the occurrence time of the interaction behavior, the object identifier is used to characterize the object that initiates the interaction behavior, the information identifier is used to characterize the information associated with the interaction behavior, and the behavior type is used to characterize the type of the interaction behavior. At least two types of interaction behavior sequences refer to at least two of multiple types of behavior sequences such as the explicit positive interaction behavior sequence, the explicit negative interaction behavior sequence, and the implicit interaction behavior sequence.

[0049] Specifically, the server can sort the interaction behavior features corresponding to multiple interaction behaviors of the same category according to the time stamp or behavior type based on the category of the interaction behaviors generated by the target object for the historical information, so as to obtain the interaction behavior sequence of this category. Among them, the categories of interaction behaviors include explicit positive interaction behaviors, explicit negative interaction behaviors, implicit interaction behaviors, etc., and the behavior types of interaction behaviors include click, favorite, like, good review, download, purchase, rejection, bad review, uninstall, etc. Further, the server can also map each interaction behavior feature into a low-dimensional embedding vector through a vectorization processing layer, that is, an Embedding layer, so as to improve the data processing efficiency. The specific ways for the server to obtain the information features of the target information, the object features of the target object, and at least two categories of interaction behavior sequences generated by the target object for the historical information can be to actively obtain or passively receive.

[0050] In addition, a dataset of information to be matched can be preset in advance, and multiple candidate information to be matched are included in this dataset of information to be matched. When the server receives an access request, it obtains the object features of the target object corresponding to the access request and at least two categories of interaction behavior sequences generated by the target object for the historical information, and obtains the candidate information to be matched corresponding to the target object from this dataset of information to be matched, determines it as the target information for the target object, and further obtains the information features of this target information.

[0051] Step S203, calculate the information similarity features between the information features and each category of interaction behavior sequences, and calculate the behavior similarity features between different categories of interaction behavior sequences.

[0052] Among them, the information similarity feature is used to characterize the similarity between the target information and the interaction behavior sequence in the information dimension. The behavior similarity feature is used to characterize the similarity between different categories of interaction behavior sequences in the behavior dimension. Since the interaction behavior sequence contains features in the information dimension such as information identifiers, and features in the behavior dimension such as behavior types, the server can calculate the information similarity features between the information features and each category of interaction behavior sequences, and calculate the behavior similarity features between different categories of interaction behavior sequences based on a convolutional neural network (CNN) or an attention network.

[0053] Taking the case where there are at least two types of interaction behavior sequences, including an explicit positive interaction behavior sequence and an explicit negative interaction behavior sequence, as an example. During the historical interaction process, the target object may click or like historical information due to misoperation, or be interested in historical information during a specific historical period and perform collection, purchase, positive review, or download. For example, the target object purchases pregnancy-related items during pregnancy. Similarly, the target object may reject historical information due to misoperation, or uninstall or delete historical information due to hardware limitations (such as insufficient device memory). Therefore, the explicit positive interaction behavior sequence cannot fully represent the information that the target object is interested in, and the explicit negative interaction behavior sequence cannot fully represent the information that the target object is not interested in. To represent the true interests of the target object as comprehensively as possible, the server can calculate the behavior similarity between the explicit positive interaction behavior sequence and the explicit negative interaction behavior sequence, eliminate the noise in the explicit positive interaction behavior sequence and the explicit negative interaction behavior sequence, and even mine the substantial negative interaction behavior in the explicit positive interaction behavior sequence and the substantial positive interaction behavior in the explicit negative interaction behavior sequence to improve the matching effect.

[0054] Step S205: Perform weighted summation on the object feature, the information similarity feature, and the behavior similarity feature to obtain the target object interaction feature of the target object.

[0055] Among them, the target object interaction feature is used to represent the interaction behavior feature associated with the target object.

[0056] In one embodiment, corresponding weights are assigned to the object feature, the information similarity feature, and the behavior similarity feature respectively. Then, based on the respective weights of the object feature, the information similarity feature, and the behavior similarity feature, the server performs weighted summation on the object feature, the information similarity feature, and the behavior similarity feature to obtain the target object interaction feature of the target object.

[0057] In another embodiment, the server first performs splicing processing on the information similarity feature and the behavior similarity feature to obtain a comprehensive similarity feature, and then performs weighted summation on the comprehensive similarity feature and the object feature to obtain the target object interaction feature of the target object.

[0058] Further, after obtaining the weighted summation result of the object feature, the information similarity feature, and the behavior similarity feature, dimensionality reduction processing can be further performed on the weighted summation result to obtain the target object interaction feature of the target object. For example, an embedding representation of the weighted summation result can be obtained through a normalization layer (Normalization) to improve the data processing efficiency of subsequent similarity matching.

[0059] Step S207, performing similarity matching on the target information interaction feature obtained by recursively processing the information feature and the target object interaction feature to obtain a matching result between the target information and the target object.

[0060] Among them, the target information interaction feature is used to characterize the interactive behavior feature associated with the target information. Specifically, the server can recursively process the information feature to obtain the target information interaction feature, and then perform similarity matching on the target information interaction feature and the target object interaction feature to obtain the similarity score of the target information interaction feature and the target object interaction feature, and then determine the matching result of the target information and the target object according to the similarity score. It should be noted that the specific algorithm for similarity matching is not unique. For example, similarity matching can be performed by calculating the cosine similarity of the target information interaction feature and the target object interaction feature, or by calculating the Euclidean distance of the target information interaction feature and the target object interaction feature. It can be understood that the higher the similarity score of the target information interaction feature and the target object interaction feature, the higher the matching degree between the target information and the target user. Based on this, the server can determine the target information with a similarity score higher than a preset threshold as the information to be recommended for the target object.

[0061] The above-mentioned information matching method obtains the information characteristics of the target information, the object characteristics of the target object, and at least two types of interactive behavior sequences generated by the target object for historical information, calculates the information similarity characteristics between the information characteristics and each type of interactive behavior sequence, and calculates the behavioral similarity characteristics between various types of interactive behavior sequences, and then performs weighted summation on the object characteristics, the information similarity characteristics, and the behavioral similarity characteristics to obtain the target object interaction characteristics of the target object, which is equivalent to comprehensively considering the object characteristics of the target object, the information characteristics of the target information and the information similarity characteristics of each type of interactive behavior sequence, and the behavioral similarity characteristics between various types of interactive behavior sequences, and determining the target object interaction characteristics of the target object from multiple different similarity dimensions, which can mine the potential interactive characteristics of the target object, is conducive to improving the accuracy of the target object interaction characteristics, and thereby improving the accuracy of the matching results of the target information and the target object obtained by similarly matching the target object interaction characteristics and the target information interaction characteristics.

[0062] In one embodiment, obtaining information features of target information includes: obtaining information parameters of the target information and auxiliary features of the target information; mapping the information parameters to obtain information mapping features of the target information; and concatenating the information mapping features and the auxiliary features of the target information to obtain information features of the target information.

[0063] Among them, the target information auxiliary feature is used to represent a set of object interaction features that match the target information. The information parameters of the target information refer to the parameters corresponding to the information features and used to represent the characteristics of the target information, specifically including but not limited to the information identification parameter, category parameter, brand parameter, item parameter, context parameter, etc. of the target information. These information parameters can be multi-valued parameters with non-unique values, such as the brand parameter in the case of co-production, or single-valued parameters with unique values, such as the information identification parameter.

[0064] Specifically, since there may be some information parameters that are high-dimensional parameters, in order to improve the data processing efficiency, as Figure 3 shown, the server obtains the information parameters of the target information and performs mapping processing on each information parameter to obtain the corresponding low-dimensional features of each information parameter. For example, through the Embedding layer, the corresponding low-dimensional embedding vectors of each information parameter are obtained, and then the low-dimensional vectors are concatenated to form the information mapping feature of the target information. Finally, the information mapping feature and the target information auxiliary feature are concatenated to obtain the information feature of the target information. Further, in the process of the server obtaining the embedding vectors corresponding to each information parameter, for multi-valued parameters, the embedding vector corresponding to the multi-valued parameter can be obtained by weighted summation of the embedding sub-vectors corresponding to each parameter value.

[0065] Taking the information identification parameter as an example, the hash value (hash id) can be obtained by converting the information identification parameter through a hash function, and then using the hash value as the key and the corresponding feature embedding value as the value to store in the query table (embedding table). During the mapping process, the corresponding feature embedding value is obtained from the query table based on the information identification parameter of the target information, which is the embedding vector corresponding to the information identification parameter.

[0066] Further, after obtaining the target object interaction feature of the target object, the target information auxiliary feature of the target information can be updated based on the target object interaction feature for use in the subsequent information matching process for the target information to improve the accuracy of the updated target information auxiliary feature. In the process of the first information matching for the target information, the initialized target information auxiliary feature can be obtained by random initialization. This random initialization can be Gaussian distribution initialization or mean distribution initialization.

[0067] In the above embodiments, in the process of determining information features, information auxiliary features for splicing a set of object interaction features characterizing matching with target information are equivalent to integrating object features with a high degree of matching with target information into information features, which is beneficial to improving the accuracy of information interaction features determined based on information features.

[0068] In one embodiment, the process of obtaining object features of a target object includes: obtaining object parameters of the target object and target object auxiliary features; performing a mapping process on the object parameters to obtain object mapping features of the target object; splicing the object mapping features and the target object auxiliary features to obtain object features of the target object.

[0069] Among them, the target object auxiliary features are used to characterize a set of information interaction features matching the target object. The object parameters of the target object refer to the parameters corresponding to the object features and used to characterize the characteristics of the target object, specifically including but not limited to object identification parameters, basic attribute parameters, behavior interest parameters, etc. For example, the basic attribute parameters include name parameters, gender parameters, age parameters, city parameters where located, etc. Similar to information parameters, object parameters can be multi-valued parameters with non-unique values, such as behavior interest parameters, or single-valued parameters with unique values, such as object identification parameters.

[0070] Specifically, as Figure 4 shown, the server obtains object parameters of the target object, performs a mapping process on each object parameter to obtain low-dimensional features corresponding to each object parameter respectively. For example, through an Embedding layer, low-dimensional embedding vectors corresponding to each object parameter are obtained respectively, and then they are spliced to form the object mapping features of the target object. Finally, the object mapping features and the target object auxiliary features are spliced to obtain the object features of the target object. Further, in the process of the server obtaining the embedding vectors corresponding to each object parameter, for multi-valued parameters, the embedding vector corresponding to the multi-valued parameter can be obtained by weighted summation of the embedding sub-vectors corresponding to each parameter value.

[0071] Further, after obtaining the target information interaction feature of the target information, the target object auxiliary feature of the target object can be updated based on the target information interaction feature for subsequent information matching processes for the target object, so as to improve the accuracy of the updated target object auxiliary feature. In the process of performing the first information matching for the target object, the initialized target object auxiliary feature can be obtained by means of random initialization. In addition, the server can obtain the information mapping feature and the object mapping feature based on a pre-trained mapping model. Specifically, all embedding values are initialized before model training, and the values in the embedding table are updated according to the reverse gradient during the model training process to ensure the accuracy of the information mapping feature.

[0072] In the above embodiments, in the process of determining the object feature, the object auxiliary feature that concatenates the set of information interaction features characterizing the match with the target object is equivalent to fusing information features with a high degree of match with the target object in the object feature, which is beneficial to improving the accuracy of the object interaction feature determined based on the object feature.

[0073] In one embodiment, the process of obtaining the target object interaction feature and the target information interaction feature includes: based on a neural network model, performing a weighted sum of the object feature, the information similarity feature, and the behavior similarity feature to obtain the target object interaction feature of the target object, and performing a recursive process on the information feature to obtain the target information interaction feature of the target information.

[0074] Among them, the neural network model refers to a model obtained based on neural networks and machine learning algorithms. It reflects many basic characteristics of the human brain function and is a highly complex non-linear dynamic learning system that can implement information processing with multiple factors and multiple conditions. Specifically, the server can use training samples containing sample object interaction features and sample information interaction features for model training to obtain a neural network model for determining object interaction features and information interaction features. Then, based on this neural network model, a weighted sum of the object feature, the information similarity feature, and the behavior similarity feature is performed to obtain the target object interaction feature of the target object, and a recursive process is performed on the information feature to obtain the target information interaction feature of the target information.

[0075] In the above embodiments, obtaining the target object interaction feature of the target object and the target information interaction feature of the target information based on the neural network model benefits from many advantages of the neural network model such as large-scale parallelism, distributed storage and processing, self-organization, self-adaptation, and self-learning ability, which is beneficial to improving the data processing ability of the information matching method.

[0076] In one embodiment, the process of training a neural network model includes: using a training sample set composed of sample object interaction features and sample object auxiliary features of a sample object, and sample information interaction features and sample information auxiliary features of sample information to perform model training based on a comprehensive loss function determined by a first auxiliary loss function, a second auxiliary loss function, and a cross-entropy loss function, so as to obtain a neural network model for determining object interaction features and information interaction features.

[0077] Among them, the sample object auxiliary feature is used to represent a set of information interaction features matching the sample object; the sample information auxiliary feature is used to represent a set of object interaction features matching the sample information. The first auxiliary loss function is used to calculate the distance between the sample object auxiliary feature and the sample information interaction feature, and the second auxiliary loss function is used to calculate the distance between the sample information auxiliary feature and the sample object interaction feature. The first auxiliary loss function and the second auxiliary loss function can be any one or a combination of mean squared error (MSE) loss function, root mean squared error (RMSE) loss function, and mean absolute error (MAE) loss function, and the specific types of the first auxiliary loss function and the second auxiliary loss function can be the same or different. The cross-entropy loss function is used to calculate the gap between the actual matching degree and the expected matching degree of the sample object interaction feature and the sample information interaction feature.

[0078] Specifically, the server determines a comprehensive loss function by performing weighted summation on the first auxiliary loss function, the second auxiliary loss function, and the cross-entropy loss function, and then based on this comprehensive loss function, uses a training sample set composed of sample object interaction features and sample object auxiliary features of a sample object, and sample information interaction features and sample information auxiliary features of sample information to perform model training, so as to obtain a neural network model for determining object interaction features and information interaction features. Further, in the process of determining the comprehensive loss function, the server can dynamically adjust the weights of each loss function based on split testing (A / B Test) to improve the accuracy of the neural network model.

[0079] In the above embodiment, in the process of training the neural network model, three types of loss functions are comprehensively considered, which is beneficial to improving the model accuracy and further enhancing the accuracy of the information matching result.

[0080] In one embodiment, step S203 includes: calculating information similarity features between the information feature and each type of interaction behavior sequence based on a self-attention mechanism; calculating behavior similarity features between various interaction behavior sequences based on a soft-attention mechanism.

[0081] Among them, the self-attention mechanism refers to constructing an attention model by using the relationships within the input samples themselves. Specifically in this application, when the server calculates the information similarity features based on the self-attention mechanism, each type of interaction behavior sequence is used to represent the information features, so as to obtain the information similarity features between the information features and each type of interaction behavior sequence.

[0082] The specific manner of obtaining the information similarity features based on the self-attention mechanism is not unique. For example, non-parametric statistical methods such as average pooling and max pooling can be used to calculate the weights for each interaction behavior feature in the information features and the interaction behavior sequence, and then the information similarity features between the information features and the interaction behavior sequence are obtained through weighted summation; it can also be a parametric neural network to learn the information similarity features between each interaction behavior sequence and the information features. In one embodiment, based on the self-attention mechanism, calculating the information similarity features between the information features and each type of interaction behavior sequence includes: concatenating the information features and each interaction behavior feature in each type of interaction behavior sequence to obtain an interaction feature matrix corresponding to each type of interaction behavior sequence; respectively performing query weight matrix transformation, key weight matrix transformation, and value weight matrix transformation on each interaction feature matrix to obtain the matrix transformation result corresponding to each interaction feature matrix; and respectively calculating the information similarity features between the information features and each type of interaction behavior sequence according to the matrix transformation result corresponding to each interaction feature matrix.

[0083] Among them, the query weight matrix, the key weight matrix, and the value weight matrix are parameter matrices learned during the model training process. Performing query weight matrix transformation, key weight matrix transformation, and value weight matrix transformation on the interaction feature matrix respectively means: multiplying the interaction feature matrix by the query weight matrix, the key weight matrix, and the value weight matrix respectively to obtain the corresponding matrix transformation result. The matrix transformation result can specifically include the query matrix, the key matrix, and the value matrix of the interaction feature matrix.

[0084] Specifically, the server splices the information features and the interaction behavior features in each type of interaction behavior sequence to obtain an interaction feature matrix respectively matched to each type of interaction behavior sequence, then performs query weight (Q) matrix transformation, key weight (K) matrix transformation, and value weight (V) matrix transformation on each interaction feature matrix respectively to obtain the matrix transformation results respectively corresponding to each interaction feature matrix. Then, according to the matrix transformation results respectively corresponding to each interaction feature matrix, the relevance between any two matrices among the matrix transformation results is calculated, and the weight corresponding to the relevance calculation result is output through the Softmax function. The weight is multiplied by the third matrix in the matrix transformation result to obtain the initial information similarity features between the information features and each interaction behavior sequence. Finally, the server obtains the information similarity features between the information features and each type of interaction behavior sequence through pooling processing. The pooling processing can specifically be average pooling (Avg-pooling) processing or maximum (max-pooling) pooling processing.

[0085] Taking the explicit positive interaction behavior sequence as an example, such as Figure 5 shown, the server splices the information feature Z v of the target information and the interaction behavior features c i in the explicit positive interaction behavior sequence to obtain an interaction feature matrix, then performs Q matrix transformation, K matrix transformation, and V matrix transformation on the interaction feature matrix, calculates the relevance between the transformed query matrix and key matrix, outputs the weight corresponding to the relevance calculation result through the Softmax function, multiplies the weight by the value matrix to obtain the initial information similarity features between the information features and the explicit positive interaction behavior sequence, and finally outputs the information similarity feature F c corresponding to the initial information similarity features based on Avg-pooling. By splicing the information features and the interaction behavior features in each type of interaction behavior sequence to obtain an interaction feature matrix respectively matched to each type of interaction behavior sequence, and calculating the information similarity features between the information features and each type of interaction behavior sequence based on various matrix transformation results of the interaction feature matrix, the correlation between the interaction behavior features in the interaction behavior sequence can be mined, which is beneficial to improving the accuracy of the information similarity features, and further improving the accuracy of the information matching results.

[0086] Furthermore, the soft-attention mechanism refers to the fusion of attention scores of multiple input features to obtain output features. Specifically in the present application, the server calculates the behavioral similarity features between various types of interactive behavior sequences based on the soft attention mechanism. By calculating the attention scores of each interactive behavior feature in a certain category of interactive behavior sequence and another category of interactive behavior sequence, the server performs feature fusion on each interactive behavior feature based on its corresponding attention score to obtain the behavioral similarity features of the two categories of interactive behavior sequences.

[0087] In one embodiment, at least two types of interactive behavior sequences include an explicit positive interactive behavior sequence, an explicit negative interactive behavior sequence, and an implicit interactive behavior sequence. In the case of this embodiment, based on the soft attention mechanism, the behavioral similarity features between the various types of interactive behavior sequences are calculated, including: based on the soft attention mechanism, the first behavioral similarity feature between the implicit interactive behavior sequence and the explicit positive interactive behavior sequence is calculated, and the second behavioral similarity feature between the implicit interactive behavior sequence and the explicit negative interactive behavior sequence is calculated.

[0088] Among them, the specific definitions of explicit positive interaction behavior sequences, explicit negative interaction behavior sequences, and implicit interaction behavior sequences can be found above and will not be repeated here. Specifically, since the target object does not operate on historical information in the interaction behavior represented by the implicit interaction behavior sequence, it is impossible to clearly define the contribution of the interaction behavior represented by the implicit interaction behavior sequence to the interest of the target object. If the implicit interaction behavior sequence is simply attributed to positive feedback or negative feedback, it is impossible to represent the true interest of the target object, which is likely to cause deviations in the matching results. Therefore, if Figure 6 As shown, after calculating the information features Z of each interaction behavior sequence and target information v The information similarity feature F between C 、F d and F u Afterwards, the server calculates the first behavior similarity feature F between the implicit interaction behavior sequence and the explicit positive interaction behavior sequence based on the soft attention mechanism. uc , in order to mine the substantial positive interaction behaviors in the implicit interaction behavior sequence, and calculate the second behavior similarity feature F between the implicit interaction behavior sequence and the explicit negative interaction behavior sequence ud , in order to mine the actual negative interaction behaviors in the implicit interaction behavior sequence, thereby improving the accuracy of the target object interaction features calculated based on the behavior similarity features, and improving the accuracy of the information matching method.

[0089] In the above embodiments, based on the characteristics of different similarity features, different attention mechanisms are used to calculate the information similarity features and the behavior similarity features, which is beneficial to improving the scientific nature of the information matching method.

[0090] In one embodiment, at least one of the at least two types of interactive behavior sequences is an explicit interactive behavior sequence. In the case of this embodiment, based on the soft attention mechanism, the behavioral similarity features between the various types of interactive behavior sequences are calculated, including: calculating the behavioral similarity weights between each interactive behavior feature in at least another type of interactive behavior sequence and the information similarity features associated with the explicit interactive behavior sequence; and performing weighted summation on each interactive behavior feature according to the behavioral similarity weight corresponding to each interactive behavior feature, to obtain the behavioral similarity features between another type of interactive behavior sequence corresponding to each interactive behavior feature and the explicit interactive behavior sequence.

[0091] Among them, the explicit interactive behavior sequence can be an explicit positive interactive behavior sequence or an explicit negative interactive behavior sequence. The behavioral similarity weight is used to characterize the similarity between the interactive behavior feature and the information similarity feature. Since the explicit interactive behavior sequence can reflect the interest of the target object to a certain extent, the server takes the explicit interactive behavior sequence as the similarity learning object, calculates the behavioral similarity weight between each interactive behavior feature in at least another type of interactive behavior sequence and the information similarity feature associated with the explicit interactive behavior sequence, and then performs weighted summation on each interactive behavior feature according to the behavioral similarity weight corresponding to each interactive behavior feature, and obtains the behavioral similarity feature between another type of interactive behavior sequence corresponding to each interactive behavior feature and the explicit interactive behavior sequence, so as to explore the potential interest of the target object represented by another type of interactive behavior sequence.

[0092] In one embodiment, the behavior similarity weight between each interactive behavior feature in at least another type of interactive behavior sequence and the information similarity feature associated with the explicit interactive behavior sequence is calculated, including: the information similarity feature associated with the explicit interactive behavior sequence is spliced ​​with each interactive behavior feature in at least another type of interactive behavior sequence to obtain a spliced ​​feature corresponding to each interactive behavior feature; and the behavior similarity weight between each interactive behavior feature and the information similarity feature associated with the explicit interactive behavior sequence is determined based on each spliced ​​feature and the sum of the spliced ​​features.

[0093] Specifically, the server splices the information similarity features associated with the explicit interactive behavior sequence with each interactive behavior feature in at least another type of interactive behavior sequence, and obtains the spliced ​​features corresponding to each interactive behavior feature. For example, the server can output the spliced ​​feature fusion result through a multilayer perceptron (MLP) neural network to obtain the spliced ​​features corresponding to each interactive behavior feature. Then, the server determines the behavior similarity weight between each interactive behavior feature and the information similarity feature associated with the explicit interactive behavior sequence based on each spliced ​​feature and the sum of each spliced ​​feature. For example, the server can determine the ratio of the spliced ​​feature to the sum of each spliced ​​feature as the behavior similarity weight associated with the interactive behavior feature corresponding to the spliced ​​feature.

[0094] In the above embodiment, the information similarity features associated with the explicit interactive behavior sequence are spliced ​​and fused with the interactive behavior features in at least another type of interactive behavior sequence, and then the behavior similarity weight is obtained based on the fusion result. The algorithm is simple and is conducive to improving data processing efficiency.

[0095] In one embodiment, the information similarity feature associated with the explicit interactive behavior sequence is spliced ​​with each interactive behavior feature in at least another type of interactive behavior sequence to obtain a spliced ​​feature corresponding to each interactive behavior feature, including: performing feature operations on the information similarity feature associated with the explicit interactive behavior sequence and each interactive behavior feature in at least another type of interactive behavior sequence to obtain a feature operation result corresponding to each interactive behavior feature; splicing the interactive behavior feature, the information similarity feature, and the feature operation result corresponding to the interactive behavior feature to obtain a spliced ​​feature corresponding to each interactive behavior feature.

[0096] The specific content of the feature operation can be a combination of one or more operations such as addition, subtraction, dot multiplication, Hadamard product, etc. Specifically, the server performs feature operations on the information similarity feature associated with the explicit interactive behavior sequence and each interactive behavior feature in at least another type of interactive behavior sequence, obtains the feature operation results corresponding to each interactive behavior feature, and then splices the interactive behavior feature, the information similarity feature, and the feature operation results corresponding to the interactive behavior feature to obtain the spliced ​​features corresponding to each interactive behavior feature.

[0097] In the above embodiment, the interactive behavior features, the information similarity features, and the feature operation results corresponding to the interactive behavior features are spliced ​​to obtain the spliced ​​features corresponding to each interactive behavior feature. The detailed features of the interactive behavior features and the information similarity features can be mined, which is beneficial to improving the feature fusion effect and further improving the accuracy of the behavior similarity weight.

[0098] In one embodiment, as Figure 7 shown, the information matching method includes:

[0099] Step S701, obtaining object parameters of a target object and auxiliary features of the target object;

[0100] Step S702, performing a mapping process on the object parameters to obtain object mapping features of the target object;

[0101] Step S703, concatenating the object mapping features and the auxiliary features of the target object to obtain object features of the target object;

[0102] Step S704, obtaining information parameters of target information and auxiliary features of the target information;

[0103] Step S705, performing a mapping process on the information parameters to obtain information mapping features of the target information;

[0104] Step S706, concatenating the information mapping features and the auxiliary features of the target information to obtain information features of the target information;

[0105] Step S707, obtaining at least two types of interaction behavior information generated by the target object for historical information;

[0106] Step S708, performing a mapping process on the interaction behavior information to obtain at least two types of interaction behavior sequences generated by the target object for historical information;

[0107] Step S709, performing a concatenation process on the information features and the interaction behavior features in each type of interaction behavior sequence to obtain an interaction feature matrix respectively matched by each type of interaction behavior sequence;

[0108] Step S710, respectively performing a query weight matrix transformation, a key weight matrix transformation, and a value weight matrix transformation on each interaction feature matrix to obtain a matrix transformation result respectively corresponding to each interaction feature matrix;

[0109] Step S711, respectively calculating information similarity features between the information features and each type of interaction behavior sequence according to the matrix transformation result respectively corresponding to each interaction feature matrix;

[0110] Step S712, performing a feature operation on the information similarity feature associated with the explicit interaction behavior sequence and the interaction behavior features in at least another type of interaction behavior sequence respectively to obtain a feature operation result respectively corresponding to each interaction behavior feature;

[0111] Step S713: Concatenate the interaction behavior features, information similarity features, and the feature operation results corresponding to the interaction behavior features to obtain the concatenated features corresponding to each interaction behavior feature.

[0112] Step S714: Determine the behavior similarity weights between each interaction behavior feature and the information similarity features associated with the explicit interaction behavior sequence based on each concatenated feature and the sum of all concatenated features.

[0113] Step S715: Perform weighted summation on each interaction behavior feature according to the behavior similarity weights corresponding to each interaction behavior feature to obtain the behavior similarity features between another type of interaction behavior sequence corresponding to each interaction behavior feature and the explicit interaction behavior sequence.

[0114] Step S716: Based on the comprehensive loss function determined by the first auxiliary loss function, the second auxiliary loss function, and the cross-entropy loss function, use the training sample set composed of the sample object interaction features and sample object auxiliary features of the sample object, as well as the sample information interaction features and sample information auxiliary features of the sample information to train the model, and obtain a neural network model for determining object interaction features and information interaction features.

[0115] Step S717: Based on the neural network model, perform weighted summation on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object, and perform recursive processing on the information features to obtain the target information interaction features of the target information.

[0116] Step S718: Perform similarity matching between the target information interaction features and the target object interaction features to obtain the matching result between the target information and the target object.

[0117] The present application also provides an application scenario for application push, which applies the above information matching method. In this application scenario for application push, when the target object opens the application market through the control terminal to search for interesting applications, and the server determines the application programs to be pushed, first, the object characteristics of the target object, at least two types of interaction behavior sequences generated by the target object for the historically pushed application programs, and the information characteristics of the target application program to be matched are obtained. Then, the information similarity characteristics between the information characteristics and each type of interaction behavior sequence, and the behavior similarity characteristics between various types of interaction behavior sequences are calculated, and the object characteristics, information similarity characteristics, and behavior similarity characteristics are weighted and summed to obtain the target object interaction characteristics of the target object. Finally, the server further performs similarity matching between the target information interaction characteristics of the target application program obtained by recursively processing the information characteristics and the target object interaction characteristics to obtain the matching result between the target application program and the target object. This matching result is used to represent the matching degree between the target application program and the target object. The server can, based on the matching degrees between multiple target application programs and the target object, push the target application programs with high matching degrees to the target object.

[0118] The present application also provides an application scenario for news push. In this application scenario, the server executes the above information matching method to determine the matching result between the target news and the target object, so as to subsequently determine the news to be pushed to the target object according to the matching result.

[0119] Specifically, as Figure 8 shown, the server first constructs the news characteristics of the target news, the object characteristics of the target object, and the interaction behavior sequences. Among them, the interaction behavior sequences of the target object include explicit positive interaction behavior sequences, explicit negative interaction behavior sequences, and implicit interaction behavior sequences.

[0120] In the process of constructing information features and object features, the server obtains the information parameters of the target information and the object parameters of the target object, and respectively performs mapping processing on each information parameter and object parameter through the Embedding layer to obtain the low-dimensional embedding vectors corresponding to each parameter. Then, after concatenating the low-dimensional vectors corresponding to each object parameter with the target object auxiliary vector, the object feature of the target object is obtained. After concatenating the low-dimensional vectors corresponding to each information parameter with the target information auxiliary vector, the information feature of the target information is obtained. Among them, the object parameters specifically include but are not limited to the identification parameter of the target object, basic attribute parameters (such as name, gender, age, and city where located), behavioral interest parameters (such as browsing behavioral interest parameters and click behavioral interest parameters), etc.; the information parameters include but are not limited to the identification parameter of the target information, category parameters (such as news category, entertainment category), brand parameters (such as the platform for publishing information), item parameters (such as semantic parameters and image parameters), and context parameters (such as the context information of the access request associated with the target information, time, and terminal device parameters), etc.

[0121] Object feature z u and information feature z v can be expressed as:

[0122] z u =[e u1 ||e u2 ||...||e ui ||....||e uk ||a u (1)

[0123] z v =[e v1 ||e v2 ||...||e vi ||...||e vl ||a v (2)

[0124] In the formula, || represents feature concatenation, e ui is the embedding vector corresponding to the i-th object parameter, a u is the target object auxiliary vector, e vi is the embedding vector corresponding to the i-th information parameter, a v is the target information auxiliary vector. Further, for multi-valued parameters such as behavioral interest parameters, the server can obtain the embedding vector corresponding to the multi-valued parameter by performing weighted summation on the embedding sub-vectors corresponding to each parameter value.

[0125] In the process of constructing the interactive behavior sequence, the server maps each interactive behavior feature into a low-dimensional embedding vector based on the Embedding layer, and then sorts according to the timestamps carried by each embedding vector to obtain the interactive behavior sequence. Specifically, the explicit positive interactive behavior sequence can be expressed as {c 1 ,..., c n1}, the explicit negative interactive behavior sequence can be expressed as {d 1 ,..., d n2}, and the implicit interactive behavior sequence can be expressed as {u 1 ,..., u n3}, where n1, n2, and n3 represent the sequence lengths, and each item in the sequence, such as c 1 , can represent an independent interactive behavior and carry information such as timestamps, object identifiers, information identifiers, and behavior types.

[0126] After completing the construction of the information features, object features, and interactive behavior sequences, the server then calculates the feature similarity, specifically including the calculation of the information similarity features between the information features and each interactive behavior sequence, and the calculation of the first behavior similarity features between the implicit interactive behavior sequence and the explicit positive interactive behavior sequence, and the calculation of the second behavior similarity features between the implicit interactive behavior sequence and the explicit negative interactive behavior sequence.

[0127] Taking the explicit positive interactive behavior sequence as an example below, the calculation process of the information similarity features is described. Specifically, the server concatenates the information feature Z v and each interactive behavior feature c i in the explicit positive interactive behavior sequence to obtain the interactive feature matrix B C = Z v , c 1 ,..., c n1 , and calculates the query matrix Q, key matrix K, and value matrix V of B C respectively:

[0128] Q = W Q B C (3)

[0129] K = W K B C (4)

[0130] V = W V B C (5)

[0131] In the formula, W Q is the query weight matrix, W K is the key weight matrix, and w Vis the value weight matrix, and the above parameter matrices can all be learned through model training.

[0132] Then, the server calculates the initial information similarity feature Attention(Q, K, V) of the information feature and the explicit positive interaction behavior sequence based on the self-attention mechanism:

[0133]

[0134] where n h is the dimension of Q, K, and V.

[0135] Finally, the server outputs the information similarity feature F corresponding to the initial information similarity feature based on Avg-pooling c :

[0136] F c = Avg_pooling(Attention(Q, K, V)) (7)

[0137] The information similarity feature F of the information feature and the explicit negative interaction behavior sequence d , and the information similarity feature F of the information feature and the implicit interaction behavior sequence u are calculated in a similar process to that of F c , which will not be elaborated here.

[0138] Furthermore, since each interaction behavior feature in the implicit interaction behavior sequence does not come from the direct feedback of the user, therefore, compared with the explicit interaction behavior sequence, the implicit interaction behavior sequence {u 1 ,..., u n3} has more noise. Based on this, noise filtering can be achieved by learning the behavior similarity between the implicit interaction behavior sequence and each explicit interaction behavior sequence. Specifically, each interaction behavior feature u i in the implicit interaction behavior sequence is concatenated with F c to obtain the concatenated feature f(F i , u c ) corresponding to each interaction behavior feature u i :

[0139] f(F c , u i ) = MLP(concat(F c , u i , F c - u i , F c ⊙ u i )) (8)

[0140] Then, the server determines the ratio of the splicing feature to the sum of all splicing features as the behavior similarity weight α associated with the interaction behavior feature corresponding to the splicing feature i :

[0141]

[0142] Finally, the server performs a weighted sum of each interaction behavior feature according to the behavior similarity weight corresponding to each interaction behavior feature to obtain the behavior similarity feature f between the implicit interaction behavior sequence and the explicit positive interaction behavior sequence uc :

[0143]

[0144] Based on the same method, the behavior similarity feature f between the implicit interaction behavior sequence and the explicit negative interaction behavior sequence can be calculated ud :

[0145] f(F d , u i ) = MLP(concat(F d , u i , F d - u i , F d ⊙ u i )) (11)

[0146]

[0147]

[0148] Among them, f(F d , u i ) is the splicing feature of the interaction behavior feature u i and the explicit negative interaction behavior sequence, and β i is the behavior similarity weight of u i and the explicit negative interaction behavior sequence

[0149] After obtaining the information similarity feature and the behavior similarity feature, the server performs a weighted sum of the object feature, the information similarity feature, and the behavior similarity feature to obtain the target object interaction vector of the target object. Specifically, the object feature, the information similarity feature, and the behavior similarity feature are connected to a one-layer neural network and a one-layer RELU activation function, and the final output score h u is:

[0150] h u = RELU(W 1 * (Feed + z u ) + b 1 ) (14)

[0151] Where Feed = {F c , F d , F u , f uc , f ud}, W 1 is the weight matrix for Feed obtained through training, and b 1 is the error coefficient learned during the training process.

[0152] Then, through a multi-layer fully connected network and the final L2 normalization layer, an output embedding representation is obtained, and the output target object interaction vector p u is:

[0153] p u = L2Norm(h u ) (15)

[0154] On the other hand, the server inputs the information feature z v into a multi-layer neural network and the RELU activation function, and through recursive and dimensionality reduction processing, the target information interaction vector p v is obtained:

[0155]

[0156]

[0157]

[0158] Wherein, is the intermediate result of the recursive processing, is the final result of the recursive processing, W i is the weight matrix obtained through training, and b i is the error coefficient learned during the training process.

[0159] Finally, the server performs cosine similarity matching on the target object interaction vector p u and the target information interaction vector p v to obtain the similarity score Score between the two, and determines the target information with a similarity score higher than the preset threshold as the information to be recommended for the target object.

[0160] Score = Cosine(p u , p v ) (19)

[0161] Further, since the target information auxiliary vector is used to represent a set of object interaction vectors matching the target information, and the target object auxiliary vector is used to represent a set of information interaction vectors matching the target object, the server obtains the target object interaction vector p u and the target information interaction vector p v After that, based on the target object interaction vector p u Update the target information auxiliary vector a v , and based on the target information interaction vector p v Update the target object auxiliary vector a u , for subsequent information matching.

[0162] Specifically, the server constructs the first auxiliary loss function loss u by learning the distance between the target object auxiliary vector a v and the target information interaction vector p u :

[0163]

[0164] In the formula, T is the number of samples, y represents the label in the training data. If the sample object has performed an action on the sample information, the label y = 1; otherwise, y = 0. Specifically, based on the target information interaction vector p v , update the target object auxiliary vector a u using gradient clipping.

[0165] Based on the same method, the target information auxiliary vector a v can be updated, and the corresponding second auxiliary loss function is:

[0166]

[0167] In addition, the server can obtain the target object interaction vector and the target information interaction vector based on the neural network model. The main loss function loss P of this neural network model is:

[0168]

[0169] In the formula, y is the label corresponding to the interaction behavior, taking values in [0, 1]. Among them, 0 represents an explicit negative interaction behavior, 1 represents an explicit positive interaction behavior, and the label value range of the implicit interaction behavior is (0, 1). σ is the sigmoid function, <p u , p v > represents the dot product operation on the vectors.

[0170] Fusing the above first auxiliary loss function, second auxiliary loss function and main loss function, the comprehensive loss function loss for model training is obtained as:

[0171] loss = loss P + λ 1 loss u + λ 2 loss v (23)

[0172] where λ 1 represents the weight of the first auxiliary loss function, and λ 2 represents the weight of the second auxiliary loss function, and λ 1 , λ 2 has a value range between [0.1, 0.8], and can be dynamically adjusted based on split testing (A / B Test) to improve the accuracy of the neural network model.

[0173] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the indications of the arrows, these steps do not necessarily need to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limitation, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages, and these steps or stages do not necessarily need to be executed at the same moment, but can be executed at different moments, and the execution order of these steps or stages does not necessarily need to be sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0174] Based on the same inventive concept, the embodiments of the present application also provide an information matching device for implementing the above-mentioned information matching method. The implementation solutions provided by this device to solve problems are similar to the implementation solutions described in the above method. Therefore, the specific limitations in one or more embodiments of the information matching device provided below can refer to the limitations on the information matching method in the above text, and will not be repeated here.

[0175] In one embodiment, as Figure 9 shown, an information matching device 900 is provided, including: an acquisition module 901, a similarity feature calculation module 902, an interaction feature determination module 903, and a matching module 904, where:

[0176] The acquisition module 901 is configured to acquire the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information;

[0177] A similarity feature calculation module 902 is configured to calculate the information similarity features between the information features and each type of interaction behavior sequence, and calculate the behavior similarity features between various interaction behavior sequences;

[0178] An interaction feature determination module 903 is configured to perform weighted summation on the object features, information similarity features, and behavior similarity features to obtain the target object interaction features of the target object;

[0179] A matching module 904 is configured to perform similarity matching between the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result between the target information and the target object.

[0180] In one embodiment, the acquisition module 901 includes: an information feature acquisition unit configured to acquire the information parameters and target information auxiliary features of the target information; perform mapping processing on the information parameters to obtain the information mapping features of the target information; splice the information mapping features and the target information auxiliary features to obtain the information features of the target information.

[0181] In one embodiment, the acquisition module 901 further includes: an object feature acquisition unit configured to acquire the object parameters and target object auxiliary features of the target object; perform mapping processing on the object parameters to obtain the object mapping features of the target object; splice the object mapping features and the target object auxiliary features to obtain the object features of the target object.

[0182] In one embodiment, the interaction feature determination module 903 is specifically configured to: based on a neural network model, perform weighted summation on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object, and perform recursive processing on the information features to obtain the target information interaction features of the target information.

[0183] In one embodiment, the information matching device 900 further includes: a training module, which is configured to perform model training on a training sample set composed of sample object interaction features and sample object auxiliary features of sample objects, and sample information interaction features and sample information auxiliary features of sample information, based on a comprehensive loss function determined by a first auxiliary loss function, a second auxiliary loss function, and a cross-entropy loss function, to obtain a neural network model for determining object interaction features and information interaction features. Among them, the sample object auxiliary features are used to represent a set of information interaction features matching the sample object; the sample information auxiliary features are used to represent a set of object interaction features matching the sample information. The first auxiliary loss function is used to calculate the distance between the sample object auxiliary features and the sample information interaction features, and the second auxiliary loss function is used to calculate the distance between the sample information auxiliary features and the sample object interaction features. The cross-entropy loss function is used to calculate the gap between the actual matching degree and the expected matching degree of the sample object interaction features and the sample information interaction features.

[0184] In one embodiment, the similarity feature calculation module 902 includes: an information similarity feature calculation unit, which is configured to calculate information similarity features between the information feature and each type of interaction behavior sequence based on a self-attention mechanism; a behavior similarity calculation unit, which is configured to calculate behavior similarity features between various interaction behavior sequences based on a soft-attention mechanism.

[0185] In one embodiment, the information similarity calculation unit includes: an interaction feature matrix determination component, which is configured to splice the information feature and the interaction behavior features in each type of interaction behavior sequence to obtain an interaction feature matrix respectively matched by each type of interaction behavior sequence; a matrix transformation component, which is configured to perform query weight matrix transformation, key weight matrix transformation, and value weight matrix transformation on each interaction feature matrix respectively to obtain a matrix transformation result corresponding to each interaction feature matrix; an information similarity feature determination component, which is configured to calculate information similarity features between the information feature and each type of interaction behavior sequence respectively according to the matrix transformation result corresponding to each interaction feature matrix.

[0186] In one embodiment, at least two types of interaction behavior sequences include an explicit positive interaction behavior sequence, an explicit negative interaction behavior sequence, and an implicit interaction behavior sequence. In this embodiment, the behavior similarity calculation unit is specifically configured to: calculate a first behavior similarity feature between the implicit interaction behavior sequence and the explicit positive interaction behavior sequence based on a soft-attention mechanism, and calculate a second behavior similarity feature between the implicit interaction behavior sequence and the explicit negative interaction behavior sequence.

[0187] In one embodiment, at least one of at least two types of interaction behavior sequences is an explicit interaction behavior sequence. In this embodiment, the behavior similarity calculation unit includes: a behavior similarity weight determination component, configured to calculate the behavior similarity weight between each interaction behavior feature in at least another type of interaction behavior sequence and the information similarity feature associated with the explicit interaction behavior sequence; a behavior similarity feature determination component, configured to perform a weighted sum on each interaction behavior feature according to the behavior similarity weight corresponding to each interaction behavior feature, to obtain the behavior similarity feature between another type of interaction behavior sequence corresponding to each interaction behavior feature and the explicit interaction behavior sequence.

[0188] In one embodiment, the behavior similarity weight determination component includes: a splicing feature determination sub-component, configured to perform splicing processing on the information similarity feature associated with the explicit interaction behavior sequence and each interaction behavior feature in at least another type of interaction behavior sequence respectively, to obtain the splicing feature corresponding to each interaction behavior feature; a behavior similarity weight determination sub-component, configured to determine the behavior similarity weight between each interaction behavior feature and the information similarity feature associated with the explicit interaction behavior sequence according to each splicing feature and the sum of all splicing features.

[0189] In one embodiment, the splicing feature determination sub-component is specifically configured to: perform feature operation on the information similarity feature associated with the explicit interaction behavior sequence and each interaction behavior feature in at least another type of interaction behavior sequence respectively, to obtain the feature operation result corresponding to each interaction behavior feature; perform splicing processing on the interaction behavior feature, the information similarity feature, and the feature operation result corresponding to the interaction behavior feature, to obtain the splicing feature corresponding to each interaction behavior feature.

[0190] Each module in the above information matching device can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor in the computer device in the form of hardware or be independent of it, or can be stored in the memory in the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0191] In one embodiment, a computer device is provided. The computer device can be a server, and its internal structure diagram can be as Figure 10As shown in the figure. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O), and a communication interface. Among them, the processor, the memory, and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store object parameters, information parameters, and interaction behavior record data. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through a network connection. The computer program, when executed by the processor, implements an information matching method. Those skilled in the art can understand, Figure 10 The structure shown in the figure is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. Specifically, the computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0192] In one embodiment, a computer device is further provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in the above method embodiments are implemented.

[0193] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0194] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by the processor, the steps in the above method embodiments are implemented.

[0195] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0196] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0197] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. An information matching method, characterized in that, the method includes: obtaining the information feature of the target information, the object feature of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information; the information feature is obtained by splicing the target information auxiliary feature; the object feature is obtained by splicing the target object auxiliary feature of the target object; the target information auxiliary feature is used to represent a set of object interaction features matching the target information; the target object auxiliary feature is used to represent a set of information interaction features matching the target object; calculating the information similarity feature between the information feature and each type of the interaction behavior sequences, and calculating the behavior similarity feature between various types of the interaction behavior sequences; performing weighted summation on the object feature, the information similarity feature, and the behavior similarity feature to obtain the target object interaction feature of the target object; performing similarity matching on the target information interaction feature obtained by recursively processing the information feature and the target object interaction feature to obtain the matching result of the target information and the target object.

2. The method according to claim 1, wherein obtaining the information feature of the target information, includes: obtaining the information parameter of the target information and the target information auxiliary feature; performing mapping processing on the information parameter to obtain the information mapping feature of the target information; splicing the information mapping feature and the target information auxiliary feature to obtain the information feature of the target information.

3. The process of obtaining the object feature of the target object according to the method of claim 1, includes: obtaining the object parameter of the target object and the target object auxiliary feature; performing mapping processing on the object parameter to obtain the object mapping feature of the target object; splicing the object mapping feature and the target object auxiliary feature to obtain the object feature of the target object.

4. The method according to claim 1, characterized in that, the process of obtaining the target object interaction feature and the target information interaction feature includes: based on a neural network model, performing weighted summation on the object feature, the information similarity feature, and the behavior similarity feature to obtain the target object interaction feature of the target object, and recursively processing the information feature to obtain the target information interaction feature of the target information.

5. The method according to claim 4, characterized in that, the process of training to obtain the neural network model includes: based on a comprehensive loss function determined by a first auxiliary loss function, a second auxiliary loss function, and a cross-entropy loss function, using a training sample set composed of the sample object interaction feature and sample object auxiliary feature of the sample object, and the sample information interaction feature and sample information auxiliary feature of the sample information for model training to obtain a neural network model for determining the object interaction feature and the information interaction feature; the sample object auxiliary feature is used to represent a set of information interaction features matching the sample object; the sample information auxiliary feature is used to represent a set of object interaction features matching the sample information; The first auxiliary loss function is used to calculate the distance between the auxiliary features of the sample object and the interactive features of the sample information; The second auxiliary loss function is used to calculate the distance between the auxiliary features of the sample information and the interactive features of the sample object; The cross-entropy loss function is used to calculate the gap between the actual matching degree and the expected matching degree of the interactive features of the sample object and the interactive features of the sample information.

6. The method according to any one of claims 1 to 5, wherein calculating the information similarity features between the information features and each type of the interactive behavior sequences, and calculating the behavior similarity features between each type of the interactive behavior sequences, comprises: Based on the self-attention mechanism, calculating the information similarity features between the information features and each type of the interactive behavior sequences; Based on the soft attention mechanism, calculating the behavior similarity features between each type of the interactive behavior sequences.

7. The method according to claim 6, wherein, the calculating the information similarity features between the information features and each type of the interactive behavior sequences based on the self-attention mechanism comprises: Performing splicing processing on the information features and the interactive behavior features in each type of the interactive behavior sequences to obtain an interactive feature matrix respectively matched by each type of the interactive behavior sequences; Performing query weight matrix transformation, key weight matrix transformation and value weight matrix transformation on each of the interactive feature matrices respectively to obtain a matrix transformation result respectively corresponding to each of the interactive feature matrices; According to the matrix transformation results respectively corresponding to each of the interactive feature matrices, calculating the information similarity features between the information features and each type of the interactive behavior sequences respectively.

8. The method according to claim 6, wherein, the at least two types of interactive behavior sequences include an explicit positive interactive behavior sequence, an explicit negative interactive behavior sequence and an implicit interactive behavior sequence; the calculating the behavior similarity features between each type of the interactive behavior sequences based on the soft attention mechanism comprises: Based on the soft attention mechanism, calculating a first behavior similarity feature between the implicit interactive behavior sequence and the explicit positive interactive behavior sequence, and calculating a second behavior similarity feature between the implicit interactive behavior sequence and the explicit negative interactive behavior sequence.

9. The method according to claim 6, wherein, at least one type of the at least two types of interactive behavior sequences is an explicit interactive behavior sequence; the calculating the behavior similarity features between each type of the interactive behavior sequences based on the soft attention mechanism comprises: Calculating the behavior similarity weights between the interactive behavior features in at least another type of interactive behavior sequences and the information similarity features associated with the explicit interactive behavior sequence; According to the behavior similarity weights respectively corresponding to each of the interactive behavior features, performing weighted summation on each of the interactive behavior features to obtain the behavior similarity features between the other type of interactive behavior sequences corresponding to each of the interactive behavior features and the explicit interactive behavior sequence.

10. The method according to claim 9, wherein, Calculating the behavior similarity weights between the interaction behavior features in at least another type of interaction behavior sequence and the information similarity features associated with the explicit interaction behavior sequence includes: Performing splicing processing on the information similarity features associated with the explicit interaction behavior sequence and each interaction behavior feature in at least another type of interaction behavior sequence respectively to obtain the splicing features corresponding to each of the interaction behavior features; Determining the behavior similarity weights between each interaction behavior feature and the information similarity features associated with the explicit interaction behavior sequence according to each splicing feature and the sum of the splicing features.

11. The method according to claim 10, wherein, The performing splicing processing on the information similarity features associated with the explicit interaction behavior sequence and each interaction behavior feature in at least another type of interaction behavior sequence respectively to obtain the splicing features corresponding to each of the interaction behavior features includes: Performing feature operations on the information similarity features associated with the explicit interaction behavior sequence and each interaction behavior feature in at least another type of interaction behavior sequence respectively to obtain the feature operation results corresponding to each of the interaction behavior features; Performing splicing processing on the interaction behavior feature, the information similarity feature, and the feature operation result corresponding to the interaction behavior feature to obtain the splicing features corresponding to each of the interaction behavior features.

12. An information matching device, wherein, The device includes: An acquisition module, configured to acquire the information features of the target information, the object features of the target object, and at least two types of interaction behavior sequences generated by the target object for historical information; the information features are obtained by splicing the target information auxiliary features; the object features are obtained by splicing the target object auxiliary features of the target object; the target information auxiliary features are used to represent the set of object interaction features matching the target information; the target object auxiliary features are used to represent the set of information interaction features matching the target object; A similarity feature calculation module, configured to calculate the information similarity features between the information features and each type of the interaction behavior sequences, and calculate the behavior similarity features between different types of the interaction behavior sequences; An interaction feature determination module, configured to perform weighted summation on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object; A matching module, configured to perform similarity matching on the target information interaction features obtained by recursively processing the information features and the target object interaction features to obtain the matching result between the target information and the target object.

13. The device according to claim 12, wherein the acquisition module includes an information feature acquisition unit, configured to: Acquire the information parameters and target information auxiliary features of the target information; Perform mapping processing on the information parameters to obtain the information mapping features of the target information; Splice the information mapping features and the target information auxiliary features to obtain the information features of the target information.

14. The device according to claim 12, wherein the obtaining module comprises an object feature obtaining unit, configured to: obtain the object parameters of the target object and the target object auxiliary features; perform mapping processing on the object parameters to obtain the object mapping features of the target object; concatenate the object mapping features and the target object auxiliary features to obtain the object features of the target object.

15. The device according to claim 12, wherein, the interaction feature determining module is specifically configured to: based on a neural network model, perform weighted summation on the object features, the information similarity features, and the behavior similarity features to obtain the target object interaction features of the target object, and perform recursive processing on the information features to obtain the target information interaction features of the target information.

16. The device according to claim 15, wherein, the device further comprises a training module, configured to: based on a comprehensive loss function determined by a first auxiliary loss function, a second auxiliary loss function, and a cross-entropy loss function, use a training sample set composed of the sample object interaction features and sample object auxiliary features of sample objects, and the sample information interaction features and sample information auxiliary features of sample information to perform model training, so as to obtain a neural network model for determining object interaction features and information interaction features; the sample object auxiliary features are used to represent a set of information interaction features matching the sample object; the sample information auxiliary features are used to represent a set of object interaction features matching the sample information; the first auxiliary loss function is used to calculate the distance between the sample object auxiliary features and the sample information interaction features; the second auxiliary loss function is used to calculate the distance between the sample information auxiliary features and the sample object interaction features; the cross-entropy loss function is used to calculate the gap between the actual matching degree and the expected matching degree of the sample object interaction features and the sample information interaction features.

17. The device according to any one of claims 12 to 16, wherein the similarity feature calculation module comprises: an information similarity feature calculation unit, configured to calculate, based on a self-attention mechanism, the information similarity features between the information features and each class of the interaction behavior sequences; a behavior similarity calculation unit, configured to calculate, based on a soft attention mechanism, the behavior similarity features between various classes of the interaction behavior sequences.

18. The device according to claim 17, wherein, the information similarity feature calculation unit comprises: an interaction feature matrix determination component, configured to perform concatenation processing on the information features and the interaction behavior features in each class of the interaction behavior sequences to obtain an interaction feature matrix respectively matching each class of the interaction behavior sequences; a matrix transformation component, configured to perform query weight matrix transformation, key weight matrix transformation, and value weight matrix transformation on each of the interaction feature matrices respectively to obtain a matrix transformation result respectively corresponding to each of the interaction feature matrices; An information similarity feature determination component, configured to calculate, respectively, information similarity features between the information features and each type of the interaction behavior sequences according to matrix transformation results corresponding to each of the interaction feature matrices.

19. The apparatus according to claim 17, wherein, the at least two types of interaction behavior sequences include an explicit positive interaction behavior sequence, an explicit negative interaction behavior sequence, and an implicit interaction behavior sequence; and the behavior similarity calculation unit is specifically configured to: calculate, based on a soft attention mechanism, a first behavior similarity feature between the implicit interaction behavior sequence and the explicit positive interaction behavior sequence, and calculate a second behavior similarity feature between the implicit interaction behavior sequence and the explicit negative interaction behavior sequence.

20. The apparatus according to claim 17, wherein, at least one type of the at least two types of interaction behavior sequences is an explicit interaction behavior sequence; and the behavior similarity calculation unit is specifically configured to: calculate behavior similarity weights between each interaction behavior feature in at least another type of interaction behavior sequence and the information similarity feature associated with the explicit interaction behavior sequence; perform weighted summation on each of the interaction behavior features according to the behavior similarity weights corresponding to each of the interaction behavior features, to obtain behavior similarity features between each of the interaction behavior features and the another type of interaction behavior sequence and the explicit interaction behavior sequence.

21. The apparatus according to claim 20, wherein, the behavior similarity weight determination component includes: a splicing feature determination sub-component, configured to perform splicing processing on the information similarity feature associated with the explicit interaction behavior sequence and each interaction behavior feature in at least another type of interaction behavior sequence respectively, to obtain a splicing feature corresponding to each of the interaction behavior features; a behavior similarity weight determination sub-component, configured to determine behavior similarity weights between each of the interaction behavior features and the information similarity feature associated with the explicit interaction behavior sequence according to each of the splicing features and the sum of the splicing features.

22. The apparatus according to claim 21, wherein, the splicing feature determination sub-component is specifically configured to: perform feature operation on the information similarity feature associated with the explicit interaction behavior sequence and each interaction behavior feature in at least another type of interaction behavior sequence respectively, to obtain a feature operation result corresponding to each of the interaction behavior features; perform splicing processing on the interaction behavior feature, the information similarity feature, and the feature operation result corresponding to the interaction behavior feature, to obtain a splicing feature corresponding to each of the interaction behavior features.

23. A computer device, including a memory and a processor, where the memory stores a computer program, wherein, when the processor executes the computer program, the steps of the method according to any one of claims 1 to 11 are implemented.

24. A computer-readable storage medium, on which a computer program is stored, wherein, when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 11 are implemented.

25. A computer program product, comprising a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 11.

Images