Information pushing method and device, computer equipment and storage medium

Abstract

The invention relates to an information pushing method and device, computer equipment, a storage medium and a computer program product. The method comprises the steps of obtaining candidate product features and object description features; the candidate product features and the object description features are combined to obtain shared features, at least two autocorrelation degrees of the shared features are calculated, the shared features are transformed based on the at least two autocorrelation degrees, and autocorrelation features corresponding to the at least two autocorrelation degrees are obtained; taking the candidate product features and the object description features as input features, and fusing the input features with the self-correlation features corresponding to the at least two self-correlation degrees in sequence to obtain target fusion features; and calculating an interaction degree of the to-be-pushed object to the product information based on the target fusion feature, and pushing the product information of the candidate product to a terminal of the to-be-pushed object when the interaction degree meets a preset pushing condition. By adopting the method, the information pushing accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to an information push method, apparatus, computer equipment, storage medium, and computer program product. Background Art

[0002] With the development of artificial intelligence (AI), information push technology has emerged. This technology can search and filter information based on a subject's interests and push it to the subject, helping them efficiently discover information of interest. Currently, when pushing information, the decision to push product information is typically based on the subject's level of interest in the product being pushed. However, when there is less product information to push, the accuracy of determining the subject's level of interest decreases, making it easy to push product information that the subject is not interested in, resulting in a waste of push resources and reduced accuracy in information push. Summary of the Invention

[0003] Based on this, it is necessary to provide an information push method, device, computer equipment, computer-readable storage medium and computer program product that can save push resources and improve accuracy in response to the above technical problems.

[0004] In a first aspect, the present application provides an information push method. The method comprises:

[0005] Obtain product information of the candidate product and object description information of the object to be pushed, and extract features of the candidate product information and object description information to obtain candidate product features and object description features;

[0006] Combine the candidate product features and the object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees;

[0007] The candidate product features and object description features are used as input features, and the autocorrelation features corresponding to at least two autocorrelation levels are sequentially fused. The result of the previous feature fusion is used as the input feature of the next feature fusion until the target fusion feature is obtained.

[0008] The degree of interaction of the target object with the product information is calculated based on the target fusion features, and when the degree of interaction meets the preset push conditions, the product information of the candidate product is pushed to the terminal of the target object.

[0009] In a second aspect, the present application also provides an information push device. The device includes:

[0010] A feature extraction module is used to obtain product information of candidate products and object description information of the object to be pushed, and extract features of the candidate product information and object description information to obtain candidate product features and object description features;

[0011] An autocorrelation extraction module is configured to combine candidate product features and object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees;

[0012] A feature fusion module is used to take the candidate product features and object description features as input features, and sequentially fuse them with the autocorrelation features corresponding to at least two autocorrelation levels, with the result of the previous feature fusion being used as the input feature of the next feature fusion until the target fusion feature is obtained;

[0013] The information push module is used to calculate the degree of interaction of the object to be pushed with the product information based on the target fusion features, and when the degree of interaction meets the preset push conditions, push the product information of the candidate product to the terminal of the object to be pushed.

[0014] In a third aspect, the present application further provides a computer device. The computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are performed:

[0015] Obtain product information of the candidate product and object description information of the object to be pushed, and extract features of the candidate product information and object description information to obtain candidate product features and object description features;

[0016] Combine the candidate product features and the object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees;

[0017] The candidate product features and object description features are used as input features, and the autocorrelation features corresponding to at least two autocorrelation levels are sequentially fused. The result of the previous feature fusion is used as the input feature of the next feature fusion until the target fusion feature is obtained.

[0018] The degree of interaction of the target object with the product information is calculated based on the target fusion features, and when the degree of interaction meets the preset push conditions, the product information of the candidate product is pushed to the terminal of the target object.

[0019] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps:

[0020] Obtain product information of the candidate product and object description information of the object to be pushed, and extract features of the candidate product information and object description information to obtain candidate product features and object description features;

[0021] Combine the candidate product features and the object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees;

[0022] The candidate product features and object description features are used as input features, and the autocorrelation features corresponding to at least two autocorrelation levels are sequentially fused. The result of the previous feature fusion is used as the input feature of the next feature fusion until the target fusion feature is obtained.

[0023] The degree of interaction of the target object with the product information is calculated based on the target fusion features, and when the degree of interaction meets the preset push conditions, the product information of the candidate product is pushed to the terminal of the target object.

[0024] In a fifth aspect, the present application further provides a computer program product. The computer program product includes a computer program that, when executed by a processor, implements the following steps:

[0025] Obtain product information of the candidate product and object description information of the object to be pushed, and extract features of the candidate product information and object description information to obtain candidate product features and object description features;

[0026] Combine the candidate product features and the object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees;

[0027] The candidate product features and object description features are used as input features, and the autocorrelation features corresponding to at least two autocorrelation levels are sequentially fused. The result of the previous feature fusion is used as the input feature of the next feature fusion until the target fusion feature is obtained.

[0028] The degree of interaction of the target object with the product information is calculated based on the target fusion features, and when the degree of interaction meets the preset push conditions, the product information of the candidate product is pushed to the terminal of the target object.

[0029] The above-described information push method, apparatus, computer device, storage medium, and computer program product combine candidate product features and object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees. The candidate product features and object description features are then used as input features and sequentially fused with the autocorrelation features corresponding to the at least two autocorrelation degrees. The result of the previous feature fusion serves as the input feature for the next feature fusion, until a target fused feature is obtained. Specifically, by calculating multiple different autocorrelation degrees of the shared features, transforming the shared features separately, and then sequentially performing feature fusion, the relevant information between the candidate product and the object to be pushed can be effectively utilized, fully exploring the potential connections between the object and the item. This allows the obtained target fused feature to contain rich relevant information between the candidate product and the object to be pushed, thereby improving the accuracy of the obtained target fused feature. Finally, the target fused feature is used to calculate the degree of interaction of the object to be pushed with the product information, avoiding the problem of being unable to accurately determine the object's interest level when there is less product information, thereby improving the accuracy of the obtained interaction level. And when the degree of interaction meets the preset push conditions, the product information of the candidate product is pushed to the terminal of the object to be pushed, thereby avoiding pushing product information that the object is not interested in, saving push resources, and improving the accuracy of product information push. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is a diagram of an application environment of an information push method in one embodiment;

[0031] Figure 2 Schematic diagram of a flow chart of an information push method in one embodiment;

[0032] Figure 3 Schematic diagram of a process for obtaining autocorrelation features in one embodiment;

[0033] Figure 4 Schematic diagram of a process for obtaining target fusion features in one embodiment;

[0034] Figure 5 Schematic diagram of the principle of obtaining target fusion features in a specific embodiment;

[0035] Figure 6 A schematic diagram of a process for screening candidate products in one embodiment;

[0036] Figure 7 A schematic diagram of a process for screening objects to be pushed in one embodiment;

[0037] Figure 8A schematic diagram of a process for pushing product information in a push model according to an embodiment;

[0038] Figure 9 A schematic diagram of the network architecture of a product information push model in a specific embodiment;

[0039] Figure 10 Schematic diagram of a flow chart of an information push method in a specific embodiment;

[0040] Figure 11 A schematic diagram of a framework of a product information push model in a specific embodiment;

[0041] Figure 12 This is a schematic diagram of a page triggered by an information push event in a specific embodiment;

[0042] Figure 13 This is a schematic diagram of a page displaying product discount information in a specific embodiment;

[0043] Figure 14 This is a schematic diagram of a page for receiving product discount information in a specific embodiment;

[0044] Figure 15 is a structural block diagram of an information push device in one embodiment;

[0045] Figure 16 is a diagram of the internal structure of a computer device in one embodiment;

[0046] Figure 17 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0048] Natural language processing (NLP) is a key area of ​​research in computer science and artificial intelligence. It studies theories and methods that enable effective communication between humans and computers using natural language. Natural language processing involves natural language, the language we use daily, and is closely related to linguistics, while also involving computer science and mathematics. Pre-trained models, a key technology for model training in artificial intelligence, are derived from large language models (LLMs) in the NLP field. After fine-tuning, LLMs can be widely applied to downstream tasks. Natural language processing technologies generally include text processing, semantic understanding, machine translation, robotic question-answering, and knowledge graphs.

[0049] The solutions provided in the embodiments of this application involve technologies such as artificial intelligence semantic understanding, which are specifically described through the following embodiments:

[0050] The information push method provided in the embodiment of the present application can be applied to Figure 1 In the application environment shown, the terminal 102 to be pushed communicates with the server 104 via a network. The data storage system can store data that the server 104 needs to process. The data storage system can be integrated on the server 104, or placed on the cloud or other servers. The server 104 receives the information push request sent by the terminal 102. The server 104 obtains the product information of the candidate product and the object description information of the object to be pushed according to the information push request, and extracts the features of the candidate product information and the object description information to obtain the candidate product features and the object description features; the server 104 combines the candidate product features and the object description features to obtain shared features, calculates at least two autocorrelation degrees of the shared features, and transforms the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees respectively; the server 104 uses the candidate product features and the object description features as input features, and sequentially fuses them with the autocorrelation features corresponding to the at least two autocorrelation degrees respectively, and the result obtained by the previous feature fusion is used as the input feature of the next feature fusion, until the target fusion feature is obtained; the server 104 calculates the degree of interaction of the object to be pushed with the product information based on the target fusion feature, and when the interaction degree meets the preset push conditions, pushes the product information of the candidate product to the terminal 102 of the object to be pushed. The terminals may include, but are not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices may include smart speakers, smart TVs, smart air conditioners, smart car-mounted devices, etc. Portable wearable devices may include smart watches, smart bracelets, head-mounted devices, etc. The server may be an independent physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services. The terminals and servers may be connected directly or indirectly via wired or wireless communication, which is not limited in this application.

[0051] In one embodiment, Figure 2 As shown, a method for pushing information is provided, which is applied to Figure 1 The server in the example is used for illustration. It is understood that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and is implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0052] S202 , obtaining product information of the candidate product and object description information of the object to be pushed, and extracting features of the candidate product information and the object description information to obtain candidate product features and object description features.

[0053] Candidate products are products whose information needs to be pushed to the recipients. These products can be physical or virtual. Physical products are products with physical form and clear physical properties. They can be visible and tangible, such as food, clothing, and electrical appliances. Virtual products are intangible products that exist in the form of information and can be distributed and traded online. Examples include music, e-books, games, and software. Product information is information used to describe a product. This information may include, but is not limited to, basic product attributes and content. Basic product attribute information characterizes the product's basic properties, such as category, price, and origin. Product content information characterizes the product's content, such as title, description, instructions, and labels. Product information can be of different types or the same type. For example, product information can include at least one of text, images, audio, and video. Product information can also be in different languages, such as Chinese, Russian, or English. The candidate product features refer to feature vectors extracted from the candidate product information, and the candidate product features may include but are not limited to basic attribute features and content features.

[0054] The object to be pushed refers to the object to which product information is being pushed. The object can be a physical object or a virtual object. A physical object is an object that exists in reality, for example, it can be a smart computer, a smart phone, or a person. A virtual object can be a virtualized object, for example, it can be a virtual character, a virtual robot, a virtual idol, etc. Object description information refers to information used to describe the object. The object description product information may include but is not limited to the basic attribute information of the object, the behavior information of the object, etc. The basic attribute information of the object is used to describe the basic attributes of the object. The behavior information of the object refers to the specific information that describes the behavior of the object, which may include but is not limited to behavior information such as browsing, clicking, collecting, and purchasing. Object description features refer to feature vectors extracted from object description information. The object description features include but are not limited to the basic attribute features of each object and the behavior features of each object.

[0055] Specifically, the server can obtain an information push request sent by the terminal of the object to be pushed. In response to the information push request, the server can obtain the object description information of the object to be pushed and the product information of the candidate product from the database. The server can also respond to the information push request, parse the information push request sent by the terminal of the object to be pushed to obtain the object description information of the object to be pushed, and then obtain the product information of the candidate product from the database. At this time, the server extracts the features of the candidate product information and the object description information to obtain the candidate product features and the object description features. The candidate product information and object description information can be vectorized, that is, discrete, textual, and categorical information can be converted into numerical, continuous, and vector feature data. When vectorizing, a hot encoding algorithm, a bag-of-words model algorithm, or a neural network algorithm can be used for vectorization.

[0056] In one embodiment, the server may encode the candidate product information and object description information through an embedded neural network, or may encode the candidate product information and object description information through a fully connected neural network to obtain candidate product features and object description features.

[0057] In one embodiment, the server can extract features of image and video type object description information or product information through a neural network algorithm to obtain corresponding feature vectors. The server can convert audio type information into text information and then encode the text information to obtain feature vectors.

[0058] S204: Combine the candidate product features and the object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees.

[0059] Shared features are fixed features used when extracting autocorrelation features. They are used to extract features that contain relevant information between the target object and the product. The degree of autocorrelation represents the degree of correlation between features in the shared features. Autocorrelation features are extracted features that contain relevant information between the target object and the product.

[0060] Specifically, the server may combine the candidate product features and the object description features. This combination may be achieved by directly concatenating the candidate product features and the object description features into a feature matrix to obtain shared features, or by concatenating the candidate product features and the object description features end-to-end to obtain the shared features. The server then extracts the relevant information between the features in the shared features using the shared features. Multiple different autocorrelation levels of the shared features may be extracted using different parameter information used for related information extraction. Each different parameter information for related information extraction may be obtained from a data service provider or pre-set. The server may also extract multiple different autocorrelation levels of the shared features using a trained neural network for different related information extraction, where multiple refers to at least two. In one embodiment, the server may simultaneously calculate each autocorrelation level corresponding to the shared features, or sequentially calculate each autocorrelation level corresponding to the shared features in a pre-set order. Each autocorrelation level is then used to weight the shared features, i.e., the product of each autocorrelation level and the shared feature is calculated to obtain the autocorrelation feature extracted for each autocorrelation level.

[0061] S206: The candidate product features and the object description features are used as input features, and the autocorrelation features corresponding to at least two autocorrelation levels are sequentially fused, and the result of the previous feature fusion is used as the input feature of the next feature fusion, until the target fusion feature is obtained.

[0062] Fusion refers to the optimal combination of different features, specifically the input features and autocorrelation features. The result of the previous feature fusion is the fusion of the input features and the previous autocorrelation features. The next feature fusion is the fusion of the previous feature fusion result as the input feature and the next autocorrelation feature. The previous autocorrelation feature and the next autocorrelation feature are two different autocorrelation features. The target fusion feature is the feature obtained by fusing the candidate product features and object description features with all the autocorrelation features.

[0063] Specifically, the server can use candidate product features and object description features as input features, then sequentially fuse the input features with at least two autocorrelation features corresponding to their respective autocorrelation levels. Specifically, the server fuses the input features with the previous autocorrelation feature to obtain a feature fusion result, and then fuses the feature fusion result with the next autocorrelation feature. This process continues until all autocorrelation features have been fused, resulting in a final fusion result, which serves as the target fusion feature. During the fusion process, the input features can be fused with the autocorrelation features sequentially in the order of the calculated autocorrelation features, or the autocorrelation feature currently to be fused can be selected from all autocorrelation features and fused with the input features. The fusion process can then be performed according to pre-set feature fusion parameters, with each fusion process using different feature fusion parameters. These feature fusion parameters can be pre-set parameters for feature fusion. The server can also directly concatenate the input features with the autocorrelation features to obtain a fused result. The server can also perform vector operations, such as vector product operations or vector sum operations, on the input features and the autocorrelation features to obtain a fused result. Then when all the autocorrelation features are fused, the result of the last fusion is used as the target fusion feature.

[0064] In one embodiment, the server can extract contextual features from the candidate product features and object description features. The contextual features are used to characterize the relationship between the candidate product features and the object description features. The server then combines the candidate product features, the object description features, and the contextual features to obtain shared features, and uses the candidate product features, the object description features, and the contextual features as input features for feature fusion.

[0065] S208 , calculating the degree of interaction of the target object with respect to the product information based on the target fusion feature, and when the degree of interaction meets the preset push condition, pushing the product information of the candidate product to the terminal of the target object.

[0066] The degree of interaction refers to the likelihood that the recipient of the push will interact with the product information. The higher the degree of interaction, the more likely the recipient will interact with the product information. Preset push conditions refer to pre-set conditions for pushing information to the recipient's terminal. These pre-set push conditions can be set by the recipient or by the server administrator. The pre-set push condition can be that the degree of interaction exceeds a pre-set threshold for pushing information. The degree of interaction can be, for example, click likelihood, browse likelihood, forwarding likelihood, or comment likelihood.

[0067] Specifically, the server converts the target fusion features into an interaction degree to obtain the degree of interaction of the target object with the product information. The server then compares the interaction degree with the preset push conditions. If the interaction degree does not meet the preset push conditions, it indicates that the product information of the candidate product is not of interest to the target object. Even if the product information of the candidate product is pushed to the target object, the target object will not interact with the product information. To avoid wasting push resources, the product information of the candidate product is not pushed to the target object's terminal. If the interaction degree meets the preset push conditions, it indicates that the product information of the candidate product is of interest to the target object. At this time, the server can push the product information of the candidate product to the target object's terminal.

[0068] In the above-mentioned information push method, a shared feature is obtained by combining candidate product features and object description features, and at least two autocorrelation degrees of the shared feature are calculated. The shared feature is then transformed based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees. The candidate product features and object description features are then used as input features and sequentially fused with the autocorrelation features corresponding to the at least two autocorrelation degrees. The result of the previous feature fusion serves as the input feature for the next feature fusion, until a target fused feature is obtained. Specifically, by calculating multiple different autocorrelation degrees of the shared feature, transforming the shared features separately, and then sequentially performing feature fusion, the relevant information between the candidate product and the object to be pushed can be effectively utilized, fully exploring the potential connections between the object and the item. This allows the obtained target fused feature to contain rich relevant information between the candidate product and the object to be pushed, thereby improving the accuracy of the obtained target fused feature. Finally, the target fused feature is used to calculate the degree of interaction of the object to be pushed with the product information, avoiding the problem of being unable to accurately determine the object's level of interest when there is less product information, thereby improving the accuracy of the obtained interaction degree. And when the degree of interaction meets the preset push conditions, the product information of the candidate product is pushed to the terminal of the object to be pushed, thereby avoiding pushing product information that the object is not interested in, saving push resources, and improving the accuracy of product information push.

[0069] In one embodiment, S204, combining the candidate product features and the object description features to obtain shared features includes the following steps:

[0070] Calculate the feature importance of the candidate product features and the feature importance of the object description features, and perform feature selection on the candidate product features according to their feature importance to obtain the candidate product features after feature selection, and perform feature selection on the object description features according to their feature importance to obtain the object description features after feature selection; concatenate the candidate product features after feature selection and the object description features after feature selection to obtain shared features.

[0071] Feature importance is a metric used in feature selection to characterize the relative importance of a given feature relative to all other features. Feature selection involves selecting N features from a set of M to optimize a specific metric. This metric is the process of selecting the most effective features from the original set. This metric can be anything from feature importance to feature information.

[0072] Specifically, the server calculates the feature importance of candidate product features and object description features using a feature selection algorithm. It then performs feature selection on the candidate product features based on their feature importance to obtain selected candidate product features. It also performs feature selection on the object description features based on their feature importance to obtain selected object description features. Feature selection algorithms include the Chi-Square Test, Mutual Information, and Correlation Coefficient. Specifically, the server performs feature selection on the basic attribute features and content features in the candidate product features to obtain filtered candidate product features. It also performs feature selection on the basic attribute features and behavioral features in the object description features to obtain filtered candidate product features. Finally, the server concatenates the selected candidate product features and the selected object description features into a matrix to obtain a feature matrix, which is used as the shared feature matrix.

[0073] In the above embodiment, by calculating the feature importance, selecting features according to the feature importance and then splicing them together, shared features are obtained. That is, important features can be selected for splicing, which improves the accuracy of the obtained shared features. In addition, through feature selection, the feature dimension can be reduced, thereby reducing the complexity of subsequent calculations.

[0074] In one embodiment, S204, calculating at least two autocorrelation degrees of the shared feature, and transforming the shared feature based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees, includes the steps of:

[0075] At least two autocorrelation calculation parameters are obtained, and the correlation degree between shared features and shared features is calculated based on the at least two autocorrelation calculation parameters to obtain at least two autocorrelation degrees of the shared features; based on the at least two autocorrelation degrees, the shared features are respectively transformed to obtain autocorrelation features corresponding to the at least two autocorrelation degrees.

[0076] The autocorrelation calculation parameters are used to extract correlation information between features in the shared features. The autocorrelation calculation parameters may include at least two parameters. The autocorrelation calculation parameters may be obtained from the data service provider or may be parameters of a pre-trained neural network for autocorrelation feature extraction.

[0077] Specifically, the server can directly obtain each autocorrelation calculation parameter from the database, or obtain each autocorrelation calculation parameter from the service provider providing the data service, or obtain each deployed autocorrelation calculation parameter. Each autocorrelation calculation parameter is then used to calculate the degree of correlation between shared features and shared features, and the degree of autocorrelation calculated by each autocorrelation calculation parameter is obtained, that is, at least two degrees of autocorrelation are obtained. The server then calculates the product of each autocorrelation degree and the shared feature to obtain each autocorrelation feature, and then uses each autocorrelation degree to weight the shared feature, that is, calculates the product of each autocorrelation degree and the shared feature to obtain the autocorrelation feature corresponding to each autocorrelation degree.

[0078] In the above embodiment, different autocorrelation degrees of shared features are calculated using different autocorrelation calculation parameters, and then different autocorrelation features are extracted using different autocorrelation degrees. This can effectively extract the potential connection between the candidate product features and the object description features in the shared features, that is, use multiple different autocorrelation features to characterize the relevant information between the candidate product and the object, thereby improving the accuracy of relevant information extraction.

[0079] In one embodiment, the current autocorrelation calculation parameter of the at least two autocorrelation calculation parameters includes a current query parameter, a current key parameter, and a current value parameter;

[0080] like Figure 3 The step of calculating the correlation between shared features based on at least two autocorrelation calculation parameters to obtain at least two autocorrelation levels of the shared features includes:

[0081] S302: transform the shared features based on the current query parameters to obtain the current query features, and transform the shared features based on the current key parameters to obtain the current key features.

[0082] The current autocorrelation calculation parameters refer to the currently used autocorrelation calculation parameters. The current query parameters refer to the query parameters in the current autocorrelation calculation parameters, which are used to calculate the query features of the shared features. The current query features refer to the query features of the currently calculated shared features. The current key parameters refer to the key parameters in the current autocorrelation calculation parameters, which are used to calculate the key features of the shared features. The current key features refer to the key features of the currently calculated shared features.

[0083] Specifically, the server can use the current query parameters to weight the shared features, that is, calculate the product of the current query parameters and the shared features to obtain the current query features, and at the same time use the current key parameters to weight the shared features, that is, calculate the product of the current key parameters and the shared features to obtain the current key features. The server can also perform different linear transformations on the shared features to obtain the current query features and the current key features. The server can also directly use the shared features as the current query features, and use the shared features as the current key features. The server can also input the shared features into a convolutional neural network that has been trained for feature transformation for adaptive transformation to obtain the current query features and the current key features.

[0084] S304, calculating the correlation between the current query feature and the current key feature to obtain a current correlation matrix, normalizing the current correlation matrix to obtain the current autocorrelation degree of the shared feature;

[0085] S306 , traversing each autocorrelation calculation parameter to obtain at least two autocorrelation degrees of the shared features.

[0086] The current correlation matrix is ​​a matrix used to represent the correlation between the current query feature and the current key feature. The current autocorrelation level refers to the currently calculated autocorrelation level.

[0087] Specifically, the server can also use a similarity algorithm to calculate the correlation between the current query feature and the current key feature. For example, the distance similarity between the current query feature and the current key feature can be calculated, or the cosine similarity between the current query feature and the current key feature can be calculated to obtain the current correlation matrix. The current correlation matrix is ​​then normalized by a normalization algorithm to obtain the current autocorrelation degree of the shared feature. The normalization algorithm can be a minimum-maximum normalization algorithm, a standardized normalization algorithm, and the like. Finally, the server traverses each autocorrelation calculation parameter, that is, uses each autocorrelation calculation parameter to calculate the query feature and key feature of the shared feature, and calculates the correlation matrix between the query feature and the key feature. Finally, the correlation matrix is ​​normalized to obtain the autocorrelation degree calculated by each autocorrelation calculation parameter, that is, the current autocorrelation degree of the shared feature is obtained.

[0088] S204, transforming the shared features based on at least two autocorrelation levels to obtain autocorrelation features corresponding to the at least two autocorrelation levels, including:

[0089] S308, transforming the shared feature based on the current value parameter to obtain a current value feature, and transforming the current value feature based on the current autocorrelation degree of the shared feature to obtain a current autocorrelation feature corresponding to the current autocorrelation degree;

[0090] S310 , traversing each autocorrelation calculation parameter to obtain at least two autocorrelation features corresponding to the autocorrelation degrees.

[0091] The current value parameter refers to the value parameter in the current autocorrelation calculation parameter, which is used to calculate the value feature of the shared feature. The current value feature refers to the value feature currently calculated.

[0092] Specifically, the server can use the current value parameter to weight the shared feature, that is, it can calculate the product of the current value parameter and the shared feature to obtain the current value feature. The server can also directly use the shared feature as the current value feature. The server can also input the shared feature into a neural network that performs value feature transformation for adaptive transformation to obtain the value feature. The current autocorrelation degree of the shared feature is then used to weight the current value feature, that is, the product of the current autocorrelation degree and the current value feature can be calculated to obtain the current autocorrelation feature corresponding to the current autocorrelation degree. The server traverses each autocorrelation calculation parameter, that is, uses the value parameter in each autocorrelation calculation parameter to weight the shared feature to obtain the value feature, and then uses the autocorrelation degree of the shared feature to weight the value feature to obtain the autocorrelation feature corresponding to each autocorrelation degree.

[0093] In one embodiment, the server may also use a pre-trained neural network for autocorrelation feature extraction to extract autocorrelation features. Each autocorrelation feature is extracted using a different neural network for autocorrelation feature extraction, that is, the shared features are input into different neural networks for autocorrelation feature extraction to obtain output autocorrelation features. The neural network for autocorrelation feature extraction may be a multilayer perceptron (MLP) neural network or a self-attention neural network.

[0094] In the above embodiment, by using shared features to calculate query features and key features, and then calculating the correlation between the query features and the key features to obtain a correlation matrix, the correlation matrix is ​​normalized to obtain the autocorrelation degree of the shared features, which can fully mine the relevant information between the features in the shared features and improve the accuracy of the obtained autocorrelation degree. Then, the autocorrelation degree is used to transform the value features to obtain the autocorrelation features corresponding to the autocorrelation degree, thereby improving the accuracy of the obtained autocorrelation features.

[0095] In one embodiment, Figure 4 As shown, in S206, the candidate product features and the object description features are used as input features, and the autocorrelation features corresponding to at least two autocorrelation levels are sequentially fused, and the result of the previous feature fusion is used as the input feature of the next feature fusion until the target fusion feature is obtained, including:

[0096] S402, taking the candidate product features and the object description features as input features, and determining a current autocorrelation feature from at least two autocorrelation features corresponding to respective autocorrelation levels;

[0097] S404, fusing the input feature with the current autocorrelation feature to obtain the current fused feature;

[0098] S406, taking the current fused feature as the input feature, and returning to the step of determining the current autocorrelation feature from the autocorrelation features corresponding to at least two autocorrelation levels, until all the autocorrelation features corresponding to at least two autocorrelation levels are fused, thereby obtaining the target fused feature.

[0099] The current autocorrelation feature refers to the autocorrelation feature that currently needs to be fused. This current autocorrelation feature can be randomly selected from all unfused autocorrelation features, or can be selected from unfused autocorrelation features in a pre-set order. The current fused feature refers to the result of the current fusion. The dimension of the current fused feature can be the same as or different from the dimension of the input feature.

[0100] Specifically, the server can combine the candidate product features and the object description features into a feature matrix to obtain the input features. At the same time, the server can randomly select an autocorrelation feature that has not yet been fused from at least two autocorrelation features corresponding to the respective correlation levels as the current autocorrelation feature. The server can also select an autocorrelation feature from at least two autocorrelation features corresponding to the respective correlation levels according to a pre-set fusion order to obtain the current autocorrelation feature. The server then fuses the input features with the current autocorrelation features. The fusion can be by concatenating the input features with the current autocorrelation features, calculating the sum of the input features and the current autocorrelation features, calculating the product of the input features and the current autocorrelation features, or inputting the input features and the current autocorrelation features into a neural network for feature fusion for fusion, thereby obtaining the current fused feature. At this time, the server determines that there are still unfused autocorrelation features, then the server uses the previous feature fusion result, that is, the current fused feature as the input feature for subsequent feature fusion, and returns to the step of determining the current autocorrelation feature from the autocorrelation features corresponding to at least two autocorrelation degrees, that is, the server again selects the current autocorrelation feature from the unfused autocorrelation features, and fuses the current autocorrelation feature with the input feature to obtain the current fused feature, and continuously iterates until all autocorrelation features are fused, and the result obtained by the last fusion is used as the target fused feature.

[0101] In the above embodiment, the current fused feature is obtained by fusing the input feature with the current autocorrelation feature. The current fused feature is then used as the input feature, and the step of determining the current autocorrelation feature from the autocorrelation features corresponding to at least two autocorrelation levels is repeated until all the autocorrelation features corresponding to at least two autocorrelation levels are completely fused, thereby obtaining the target fused feature. This allows the target fused feature to fuse the different autocorrelation features, enabling the target fused feature to fully represent the relevant information between the object and the candidate product, thereby improving the accuracy of the obtained target fused feature.

[0102] In one embodiment, S404, fusing the input feature with the current autocorrelation feature to obtain the current fused feature includes the following steps:

[0103] The input feature is spliced ​​with the current autocorrelation feature to obtain the current spliced ​​feature; the full connection operation parameter is obtained, and the current spliced ​​feature is subjected to feature interaction through the full connection operation parameter to obtain the current fusion feature.

[0104] Among them, the current splicing feature refers to the splicing of the input feature and the current autocorrelation feature, and the splicing can be up and down splicing to obtain the current splicing feature matrix. The splicing can also be head-to-tail splicing, for example, the input feature can be used as the head and the current autocorrelation feature can be used as the tail to perform head-to-tail splicing to obtain the current splicing feature vector, or the current autocorrelation feature can be used as the head and the input feature can be used as the tail to perform head-to-tail splicing to obtain the current splicing feature vector. The fully connected operation parameters are used to map the columns and rows in the features to achieve information fusion in the spatial domain and the channel domain. The fully connected operation parameters can be parameters in a pre-selected trained fully connected neural network, for example, they can be parameters in an MLP network. The fully connected operation parameters can also be obtained from the service provider providing the data service.

[0105] Specifically, the server can concatenate the input feature and the current autocorrelation feature end to end to obtain the current concatenated feature. The server can also calculate the sum of the input feature and the current autocorrelation feature, or calculate the product of the input feature and the current autocorrelation feature to obtain the current concatenated feature. The server can then obtain the full-connection operation parameters from the database or from the data service provider, and then use the full-connection operation parameters to map and transform the columns and rows in the current concatenated feature to obtain the current fused feature.

[0106] In the above embodiment, by obtaining the full connection operation parameters, the current splicing features are subjected to feature interaction through the full connection operation parameters to obtain the current fusion features. That is, the full connection operation parameters can fully interact with each dimension in the current splicing features, thereby improving the accuracy of the current fusion features obtained.

[0107] In a specific embodiment, Figure 5 As shown, a schematic diagram of the principle of obtaining the target fusion feature is provided. Specifically, the server combines the object description feature and the candidate product feature to obtain a shared feature, and then uses the shared feature to extract n different autocorrelation features, including autocorrelation feature 1, autocorrelation feature 2, and so on. At the same time, the object description feature and the candidate product feature are used as input features to perform a first feature fusion with the autocorrelation feature 1 to obtain the first feature fusion result. The first feature fusion result is then used as input features to perform a second feature fusion with the autocorrelation feature 2 to obtain the second feature fusion result. The feature fusion result is then continuously fused with the autocorrelation feature to obtain the n-1th feature fusion result. The n-1th feature fusion result is used as input features to perform feature fusion with the autocorrelation feature n to obtain the output feature fusion result, i.e., the target fusion feature. This can fully utilize the relevant information between the object and the candidate product, thereby improving the accuracy of the obtained target fusion feature.

[0108] In one embodiment, S208, calculating the degree of interaction of the target object with the product information based on the target fusion feature, includes the following steps:

[0109] Perform a full connection operation on the target fusion feature to obtain the full connection operation result, and normalize the full connection operation result to obtain the degree of interaction of the object to be pushed to the product information.

[0110] The fully connected operation is used to map the target fusion feature space to the output result space, which may be a nonlinear transformation of the target fusion feature. The result of the fully connected operation may be a scalar.

[0111] Specifically, the server obtains computation parameters and uses them to perform a fully connected computation on the target fusion features, obtaining a fully connected computation result. These computation parameters can be parameters from a pre-trained fully connected neural network, including weights and biases. These computation parameters can also be obtained from the data service provider. Finally, the server uses a normalization algorithm to normalize the fully connected computation result and determine the level of interaction of the target audience with the product information.

[0112] In a specific embodiment, the server can calculate the degree of interaction of the target audience with the product information using a fully connected neural network layer and a sigmoid function. Specifically, the server can input the target fusion feature into the fully connected neural network layer, performing matrix multiplication and bias addition operations on the target fusion feature and the connection weights between each neuron in the trained fully connected neural network layer to obtain a fully connected operation result. The sigmoid (S-type activation function) function is then used to map the fully connected operation result to the range of 0 to 1 to obtain the degree of interaction.

[0113] In the above embodiment, by performing a full connection operation on the target fusion feature to obtain a full connection operation result, and normalizing and mapping the full connection operation result, the interaction degree of the object to be pushed to the product information is obtained, thereby improving the accuracy of the obtained interaction degree.

[0114] In one embodiment, the object description information includes object basic description information and object interaction description information, and the object description features include object basic description features and object interaction description features;

[0115] S202, extracting features of candidate product information and object description information to obtain candidate product features and object description features, including the following steps:

[0116] Obtain the object original information of the object to be pushed and the product original information of the candidate product, and preprocess the object original information and the product original information respectively to obtain the object basic description information, the object interaction description information and the product information of the candidate product; encode the object basic description information to obtain the object basic description feature, encode the object interaction description information to obtain the object interaction description feature, and encode the product information of the candidate product to obtain the candidate product feature.

[0117] Among them, object raw information is unprocessed object description information, which can be information collected directly from the data source. Product raw information refers to unprocessed product information, which can be information collected directly from the data source. Object basic description information refers to information describing the basic attributes of the object. Object interaction description information refers to information describing the interaction between the object and the product. Object basic description features are features extracted from the object basic description information. Object interaction description features are features extracted from the object interaction description information. Features can be feature vectors or feature matrices. Preprocessing refers to a series of processing processes such as cleaning, conversion, integration, and normalization of raw data before data analysis or modeling. Data preprocessing aims to reduce errors and deviations in the data analysis or modeling process and improve the quality and reliability of data.

[0118] Specifically, the server can collect the original object information of the object to be pushed and the original product information of the candidate product from various data sources. The server can also obtain the original object information of the object to be pushed and the original product information of the candidate product from the database. The server then pre-processes the original object information and the original product information respectively to obtain the basic object description information, the object interaction description information and the product information of the candidate product. Among them, the basic object attribute information in the original object information can be pre-processed, that is, the useless, repeated, erroneous and missing data in the basic object attribute information are removed to ensure the quality and integrity of the data and obtain the basic object description information. At the same time, the object interaction behavior information in the original object information can be pre-processed, that is, the useless, repeated, erroneous and missing data are removed to obtain the object interaction description information. At the same time, the useless, repeated, erroneous and missing data in the original product information can be removed to obtain the product information of the candidate product. In a specific embodiment, the server can directly use RDD (Resilient Distributed Datasets) operators in Spark (a general-purpose big data distributed computing engine), such as filter, map, flatMap, and other operation functions, to perform preprocessing to obtain basic object description information, object interaction description information, and product information of candidate products. Spark can achieve efficient data cleaning and preprocessing through distributed computing capabilities and a rich data processing library. The server then encodes the basic object description information, object interaction description information, and product information of candidate products separately, for example, through an embedding layer or fully connected layer for feature encoding, or through hot encoding, bag-of-words model, etc., to obtain basic object description features, object interaction description features, and candidate product features.

[0119] In the above embodiment, the quality of the data can be improved by preprocessing the original information of the object and the original information of the product respectively, and then the preprocessed data are encoded respectively to obtain the basic description features of the object, the interaction description features of the object and the candidate product features, thereby improving the accuracy of the encoded features.

[0120] In one embodiment, the candidate product includes at least two, such as Figure 6 As shown, the information push method further includes:

[0121] S602 , obtaining product information of at least two candidate products, and extracting features of the product information of the at least two candidate products to obtain candidate product features of the at least two candidate products.

[0122] Specifically, there may be multiple candidate products. These multiple candidate products may be pre-set or obtained by screening existing products based on the historical data of the object. The server may then obtain product information for each candidate product from a database, from a service provider providing product services, or from a data provider providing data. The server may then extract features from the product information of each candidate product. For example, the server may extract features from the product information using a trained feature encoding neural network to obtain candidate product features for each candidate product.

[0123] In one embodiment, the server can filter existing products based on the object's historical behavior and preferences to obtain a candidate product set for the object to be pushed, which includes multiple candidate products. The server can then obtain corresponding product information based on each candidate product in the candidate product set, and perform feature extraction to obtain candidate product features for each candidate product.

[0124] S604: Obtain target fusion features of at least two candidate products based on the candidate product features and object description features of the at least two candidate products.

[0125] Specifically, for each candidate product, the server combines the candidate product features and the object description features to obtain shared features, calculates at least two autocorrelation degrees of the shared features, and transforms the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees. The candidate product features and the object description features are then used as input features and sequentially fused with the autocorrelation features corresponding to the at least two autocorrelation degrees. The result of the previous feature fusion is used as the input feature for the next feature fusion until the target fusion feature between the candidate product and the object is obtained. The server then traverses the candidate product features and object description features of each candidate product to obtain the target fusion feature between each candidate product and the object. That is, the server can use the method steps in the above embodiment of obtaining the target fusion feature to obtain the target fusion feature between the candidate product and the object.

[0126] S606: Calculate the interaction degree of the object to be pushed to the at least two candidate products based on the target fusion features of the at least two candidate products;

[0127] S608: Filter target candidate products corresponding to the target interaction level from the at least two candidate products according to the interaction level of the target object with the at least two candidate products, and push the product information of the target candidate products to the terminal of the target object.

[0128] The target interaction level refers to the interaction level of the target product candidate with the target product candidate. The target product candidate is the product candidate that needs to push information to the target product candidate. The target interaction level can be the maximum interaction level among all the candidate product interaction levels.

[0129] Specifically, the server uses the target fusion features of each candidate product to calculate the interaction level of the target product recipient with each candidate product. It then ranks each candidate product from highest to lowest according to their corresponding interaction level to obtain a candidate product sequence. The top-ranked candidate product is then selected to obtain the target candidate product corresponding to the target interaction level. Finally, the server pushes the product information of the target candidate product to the target recipient's device.

[0130] In one embodiment, after the server obtains the candidate product sequence, it can select the same number of candidate products from large to small according to the pre-set number of product information to be pushed to the object to be pushed, and then push the product information of each selected candidate product to the terminal of the object to be pushed. The product information of each candidate product can be pushed at the same time, or the product information of each candidate product can be pushed in sequence.

[0131] In the above embodiment, product information for at least two candidate products is obtained, and then a target fusion feature is calculated between each candidate product and the target to be pushed. This target fusion feature is then used to calculate the degree of interaction of the target to be pushed with each candidate product. Each candidate product is then screened based on the degree of interaction to obtain target candidate products corresponding to the target degree of interaction. Finally, the product information of the target candidate products is pushed to the target to be pushed terminal. In other words, the target candidate product with the highest degree of interaction is screened from multiple candidate products and pushed to the target to be pushed terminal. This avoids pushing product information that the target is not interested in, saving push resources and further improving the accuracy of information push.

[0132] In one embodiment, the objects to be pushed include at least two, such as Figure 7 As shown, the information push method further includes:

[0133] S702, obtaining object description information of at least two objects to be pushed, and extracting features of the object description information of the at least two objects to be pushed to obtain object description features of the at least two objects to be pushed;

[0134] S704 : Acquire target fusion features of at least two objects to be pushed based on the candidate product features and the object description features of at least two objects to be pushed.

[0135] The number of objects to be pushed may also include multiple objects, where multiple refers to at least two. The multiple objects to be pushed may be pre-set or determined based on the objects to be pushed in multiple information push requests obtained within a certain period of time.

[0136] Specifically, the server can obtain object description information for each object to be pushed from a database, object description information uploaded by the object terminal, or object description information from a data service provider. The server then extracts object description features for each object to be pushed, encoding the object description information for each object to be pushed. For example, encoding can be performed using a neural network encoder to obtain object description features. The server then combines the object description features with candidate product features for each object to be pushed to obtain shared features. The server then calculates multiple autocorrelation levels for the shared features and weights the shared features according to each autocorrelation level to obtain individual autocorrelation features. The object description features and candidate product features are then used as input features and sequentially fused with the individual autocorrelation features to obtain a target fused feature between the object to be pushed and the candidate product. The server traverses each object to be pushed to obtain a target fused feature between each object to be pushed and the candidate product. That is, the server can sequentially obtain the target fused feature between each object to be pushed and the candidate product according to the method and steps for obtaining the target fused feature in the above-described embodiment.

[0137] S706: Calculate the interaction degree of the at least two objects to be pushed with the candidate product based on the target fusion features of the at least two objects to be pushed.

[0138] S708: Filter a target push object from the at least two push objects according to the interaction degree of the at least two push objects with the candidate product, and push the product information of the candidate product to the terminal of the target push object.

[0139] The target push object refers to the object to be pushed to which the product information of the candidate product can be pushed. The target push object can be the object to be pushed that has the greatest degree of interaction with the candidate product.

[0140] Specifically, the server sorts the candidate products based on the interaction levels of at least two of the candidates with the candidate products, generating a sequence of candidates. The server then selects the first-ranked candidate from the sequence to obtain a target candidate, and pushes the candidate product information to the target candidate's terminal. The server may also select a preset number of candidates from the sequence, in descending order, and then simultaneously send the candidate product information to the terminals of the preset number of candidates.

[0141] In the above embodiment, by obtaining the object contempt information of at least two objects to be pushed, and then calculating the target fusion features between each object to be pushed and the candidate product, the target fusion features are used to calculate the degree of interaction of each object to be pushed with the candidate product, and finally the objects to be pushed are screened according to the degree of interaction to obtain the target push object, so that the candidate product can be pushed to the object to be pushed with the highest degree of interaction, avoiding pushing the product information of the candidate product to the object to be pushed with a low degree of interaction, saving push resources, and improving the accuracy of information push.

[0142] In one embodiment, Figure 8 As shown, the information push method further includes:

[0143] S802, inputting product information and object description information into the product information push model;

[0144] S804 , extracting features of the candidate product information and object description information through a feature extraction network in the product information push model to obtain candidate product features and object description features.

[0145] The product information push model refers to a neural network model used to predict product information push notifications. This model is trained using training data using a neural network algorithm, such as a convolutional neural network, a recurrent neural network, or a feedforward neural network. The training data can be derived from historical objects and corresponding historical product information push data. The feature extraction network is a neural network used for feature extraction and can be either an embedded neural network or a fully connected neural network.

[0146] Specifically, the server pre-acquires training data to train the product information push model, initializing the model parameters. Upon completion, the product information push model is generated. This model, which can also be initialized through pre-training, is then deployed and used. When product information push predictions are needed, the server invokes the deployed product information push model and extracts candidate product features and object description features from the model using the network parameters in the feature extraction network.

[0147] S806, combining the candidate product features and the object description features through the product information push model to obtain shared features, and calculating at least two autocorrelation degrees of the shared features through the autocorrelation feature extraction network in the product information push model, and transforming the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees.

[0148] S808: The feature fusion network in the product information push model uses the candidate product features and object description features as input features, and fuses the autocorrelation features corresponding to at least two autocorrelation levels in sequence. The result of the previous feature fusion is used as the input feature of the next feature fusion until the target fusion feature is obtained.

[0149] The autocorrelation feature extraction network refers to a neural network used to extract autocorrelation features. The autocorrelation feature network can be a self-attention neural network or a fully connected neural network. The feature fusion network is a neural network used to fuse different features. The feature fusion network can be composed of multiple different fully connected neural networks, for example, multiple MLP networks.

[0150] Specifically, the server's product information push model combines candidate product features and object description features to obtain shared features. These shared features are then input into the autocorrelation feature extraction network to calculate the autocorrelation levels of the shared features. The shared features are then transformed based on the autocorrelation levels to obtain the output autocorrelation features. Each autocorrelation feature is then input into the feature fusion network. The feature fusion network uses the candidate product features and object description features as input features and fuses them with each autocorrelation feature in sequence. This involves fusing one autocorrelation feature before fusing it with the next, and this continues until all autocorrelation features are fused. This results in the fusion result, which is the target fused feature.

[0151] S810, calculating the degree of interaction of the target object with the product information based on the target fusion feature through the interaction degree calculation network in the product information push model; when the interaction degree meets the preset push conditions, pushing the product information of the candidate product to the terminal of the target object.

[0152] The interaction degree calculation network refers to a neural network that calculates the degree of interaction of the target recipient with the product information. This interaction degree calculation network can also be used for classification. By inputting the target fusion features into the classification neural network, a classification result is obtained as to whether the product information of the candidate product should be pushed to the target recipient's terminal. The classification result includes a push result and a non-push result.

[0153] Specifically, the product information push model in the server inputs the target fusion features into the interaction degree calculation network for calculation, and outputs the degree of interaction of the target recipient with the product information. Then, if the degree of interaction exceeds a preset interaction degree threshold, it indicates that the product information can be pushed to the target recipient. At this point, the server pushes the candidate product information to the target recipient's terminal. If the degree of interaction is less than or equal to the preset interaction degree threshold, it indicates that the target recipient is not interested in the product information, and in this case, the product information does not need to be pushed to the target recipient.

[0154] In one embodiment, Figure 9 As shown in the figure, it is a schematic diagram of the network architecture of the product information push model. Specifically: the server inputs the product information and object description information into the product information push model. The product information push model extracts the candidate product features and object description features through the feature extraction network, extracts each autocorrelation feature through the autocorrelation feature extraction network, and then fuses the features through the feature fusion network to obtain the target fusion feature. Finally, the interaction degree calculation network calculates and outputs the interaction degree of the object to be pushed to the candidate product, thereby improving the efficiency of obtaining the interaction degree.

[0155] In the above embodiment, the product information push model uses the feature extraction network, autocorrelation feature extraction network, feature fusion network, and interaction degree calculation network to calculate the degree of interaction of the target audience with the product information. This means that the product information push model can quickly calculate the degree of interaction, improving the efficiency of calculating the degree of interaction and ensuring the accuracy of the obtained degree of interaction. When the degree of interaction meets the preset push conditions, the product information of the candidate product is pushed to the target audience's terminal, improving the accuracy and efficiency of information push.

[0156] In one embodiment, the training of the product information push model includes the following steps:

[0157] Obtain training product information, training object description information, and information push training labels; input the training product information and training object description information into the initial product information push model for forward calculation to obtain the degree of training interaction of the output training object description information with the training product information; calculate the loss information between the training interaction degree and the information push training label, and reversely update the initial product information push model based on the loss information and perform loop iterations until the training completion conditions are met, thereby obtaining the product information push model.

[0158] Among them, training product information refers to product information during training, which can be historical product information. Training object description information refers to object description information during training, which can be historical object description information. The information push training label is a label used to indicate whether the training product information is pushed to the object of the training object description information, including a push label and a not-pushed label. The push label indicates that the training product information is pushed to the object of the training object description information. The not-pushed label indicates that the training product information is not pushed to the object of the training object description information. The training interaction degree refers to the degree of interaction between the object of the training object description information and the training product information, calculated during training. The loss information is used to indicate the error between the training interaction degree and the information push training label.

[0159] Specifically, the server can obtain training product information, training object description information, and information push training labels from a database, or from a data service provider, or from a terminal. The server then establishes an initial product information push model, specifically, using a neural network algorithm to establish the network architecture of the initial product information push model and initialize model parameters to obtain the initial product information push model. The server then inputs the training product information and training object description information into the initial product information push model for forward computation. Specifically, the server calculates the degree of training interaction between the training object description information and the training product information using an initial feature extraction network, an initial autocorrelation feature extraction network, an initial feature fusion network, and an initial interaction degree calculation network. The server then uses a classification loss function to calculate the loss between the training interaction degree and the information push training label. This classification loss function can be a binary classification loss function, such as a cross-entropy loss function, a 0-1 loss function, a squared loss function, a linear regression loss function, or the like. The server then uses this loss information to reversely update the initialization parameters of the initial product information push model using a gradient descent algorithm, thereby obtaining an updated product information push model. At this point, the server iterates, updating the product information push model as the initial product information push model and returning to the steps of obtaining training product information, training object description information, and information push training labels. This process continues until the training completion condition is met, resulting in the product information push model. This training completion condition can be when the loss information reaches a preset threshold, when the training number of iterations reaches the maximum, or when the model parameters no longer change. The server then deploys and uses the trained product information push model.

[0160] In the above embodiment, by obtaining training product information, training object description information and information push training labels, and then training the initial product information push model until the training completion conditions are met, the product information push model is obtained, that is, the initial product information push model is updated through the loss information between the training interaction degree and the information push training label and a loop iteration is performed to ensure the accuracy of the obtained product information push model.

[0161] In one embodiment, the candidate product includes a network traffic product; S208, calculating the degree of interaction of the target object with the product information based on the target fusion feature, and when the degree of interaction meets the preset push condition, pushing the product information of the candidate product to the target object, including the steps of:

[0162] The degree of interaction of the object to be pushed with the network traffic product is calculated based on the target fusion characteristics, and when the degree of interaction meets the preset push conditions, the product information of the network traffic product is pushed to the terminal of the object to be pushed.

[0163] Network traffic products are products based on network traffic, which refers to the amount of data transmitted over a network. For example, a 30GB network traffic product (a G is a unit of measurement for traffic, meaning gigabytes) would be considered a network traffic product. Product information for this network traffic product would include the available 30GB of network traffic, the price, the validity period, and information about the operator.

[0164] Specifically, the server obtains the product information of the network traffic product, which may include the size, validity period, and operation information of the network traffic product, etc. At the same time, the object description information of the object to be pushed is obtained. The server can then use the steps of the information push method in any of the above embodiments to calculate the degree of interaction of the object to be pushed with the network traffic product. For example, the product information and object description information of the network traffic product are input into the product information push model for forward calculation to obtain the output of the degree of interaction of the object to be pushed with the network traffic product. The degree of interaction is then compared to see whether it meets the preset push conditions. For example, when the degree of interaction is less than or equal to the preset interaction degree threshold, the product information of the network traffic product does not need to be pushed to the terminal of the object to be pushed, that is, the server does not process it. When the server determines that the degree of interaction is greater than the preset interaction degree threshold, the product information of the network traffic product is pushed to the terminal of the object to be pushed.

[0165] In the above embodiment, by calculating the degree of interaction of the object to be pushed with the network traffic product, and then determining whether to push the network traffic product to the object to be pushed according to the degree of interaction, it is avoided that the product information of the network traffic product is pushed to the object to be pushed when the degree of interaction is low, thereby saving push resources and improving the accuracy of pushing product information of the network traffic product.

[0166] In a specific embodiment, Figure 10 FIG. 1 is a flow chart of an information push method, which is executed by a computer device, which may be a server or a terminal, and specifically includes the following steps:

[0167] S1002: Input product information and object description information into a feature extraction network of a product information push model to extract features of candidate product information and object description information, thereby obtaining candidate product features and object description features.

[0168] S1004, combining the candidate product features and the object description features through the product information push model to obtain shared features, and transforming the shared features based on the current query parameters through the autocorrelation feature extraction network to obtain the current query features, transforming the shared features based on the current key parameters to obtain the current key features, calculating the correlation between the current query features and the current key features to obtain the current correlation matrix, normalizing the current correlation matrix to obtain the current autocorrelation degree of the shared features, traversing each autocorrelation calculation parameter to obtain at least two autocorrelation degrees of the shared features.

[0169] S1006, transforming the shared feature based on the current value parameter through the autocorrelation feature extraction network to obtain the current value feature, and transforming the current value feature based on the current autocorrelation degree of the shared feature to obtain the current autocorrelation feature corresponding to the current autocorrelation degree, traversing each autocorrelation calculation parameter to obtain at least two autocorrelation features corresponding to each autocorrelation degree.

[0170] S1008: Using the feature fusion network in the product information push model, the candidate product features and the object description features are used as input features, and a current autocorrelation feature is determined from the autocorrelation features corresponding to at least two autocorrelation levels. The input feature and the current autocorrelation feature are concatenated to obtain a current concatenated feature. Fully connected operation parameters are obtained, and the current concatenated feature is subjected to feature interaction using the full-connection operation parameters to obtain a current fused feature. The current fused feature is used as the input feature, and the step of determining the current autocorrelation feature from the autocorrelation features corresponding to at least two autocorrelation levels is returned to and executed until all the autocorrelation features corresponding to at least two autocorrelation levels are fused to obtain a target fused feature.

[0171] S1010, perform full connection operation on the target fusion feature through the interaction degree calculation network in the product information push model to obtain the full connection operation result, and normalize the full connection operation result to obtain the degree of interaction of the object to be pushed with the product information. When the interaction degree meets the preset push conditions, the product information of the candidate product is pushed to the terminal of the object to be pushed.

[0172] In a specific embodiment, Figure 11The figure shows a schematic diagram of the framework of the product information push model. Specifically, the server obtains object features, product features, and context features through feature encoding in the feature encoding layer of the product information push model. Object features include basic attribute features and behavioral features, while product features include basic attribute features and content features. Feature selection is then performed on the object features, product features, and context features to obtain selected features. The selected features are then concatenated into a feature matrix to obtain a shared feature matrix. The shared feature matrix is ​​then input into the feature transmission network, and autocorrelation features are extracted through different transmission layers. For example, shared features can be input into transmission layer 1 to obtain output autocorrelation feature 1, shared features can be input into transmission layer 2 to obtain output autocorrelation feature 2, and shared features can be input into transmission layer n to obtain output autocorrelation feature n. There are n transmission layers in total, and autocorrelation feature extraction can be achieved through an MLP network or an attention network. At the same time, the object features, product features, and context features are used as input features and concatenated with autocorrelation feature 1 output by transport layer 1 to obtain a concatenated result. This concatenated result is then fed into the first feature fusion network, the first MLP network, for feature fusion, resulting in an output fusion result. The fusion result is then used as input features and concatenated with the autocorrelation feature output by the next transport layer, namely, autocorrelation feature 2 output by transport layer 2, to obtain a concatenated result. This concatenated result is then fed into the next feature fusion network, the second MLP network, for feature fusion, resulting in an output fusion result. The concatenated result is then continuously used as input features and concatenated with the autocorrelation feature output by the next transport layer to obtain a concatenated result. This concatenated result is then fed into the next feature fusion network, the nth MLP network, for feature fusion, resulting in an output fusion result. Finally, the target fusion feature is used to calculate the degree of interaction of the recommended object with the candidate product, which can be represented by a click-through rate. The target fusion features are predicted through the output layer, which can be implemented using a fully connected layer and a sigmoid function. This output layer then outputs the degree of interaction between the target object and the candidate product. Finally, the server determines whether to push the candidate product information to the target object based on the degree of interaction. This product information push model implements feature transmission and interaction, effectively utilizing the feature information of objects and products to fully explore the potential connections between objects and products. This avoids the problem of being unable to fully explore the potential connections between objects and products when there are fewer product features, improving the accuracy of model push. At the same time, it avoids pushing product information that the target object is not interested in, saving server network resources, such as traffic resources and computing resources.

[0173] In a specific embodiment, the information push method is applied to a network service application, and network service products can be pushed to the object through the network service application, such as pushing network traffic products, pushing video application membership product information, pushing commodity discount information, etc. Specifically: when the object is using the network service application, an information push event can be triggered, such as Figure 12 As shown, it is a page diagram for triggering an information push event after the phone charge recharge service is completed. The page diagram shows a prompt message that the phone charge recharge service is completed and a trigger button "Information Push Service" for the information push event. That is, an event for information push can be triggered after the network service is completed. At this time, the object terminal detects the operation event of the object clicking the trigger button, and in response to the operation event, sends a real-time information push request to the server of the network service application. The information push request can carry an object identifier. The server of the network service application can obtain the object description information according to the object identifier, and obtain the product information of each candidate product according to the information push request. The server of the network service application then inputs the product information of each candidate product and the object description information into the deployed product information push model for forward calculation, and obtains the output object's click rate for each candidate product. When the click rate exceeds the preset click rate threshold, the server of the network service application pushes the product information of the corresponding candidate product to the object terminal corresponding to the object identifier. The object terminal receives the pushed product information of the candidate product and displays it. The pushed product information of the candidate product can be the product discount information of the commodity, such as Figure 13 As shown in FIG, a schematic diagram of a page showing product discount information pushed to the target terminal. The target can then click the "Get" button in the product discount information. At this time, the network service application will jump to the product discount information receiving page, as shown in FIG. Figure 14 The figure below shows a schematic diagram of the product discount information redemption page. This page displays a notification indicating that the product discount has been redeemed, along with information about the 30% discount and a two-hour validity period. The user can then use the 30% discount to purchase the corresponding product within the two-hour validity period. This means that by pushing accurate online service product information to users, the convenience and accuracy of online services can be improved, enhancing the user experience.

[0174] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0175] Based on the same inventive concept, the embodiments of the present application also provide an information push device for implementing the aforementioned information push method. The implementation solution provided by this device is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations of one or more information push device embodiments provided below can be referred to the limitations of the information push method above and will not be repeated here.

[0176] In one embodiment, Figure 15 As shown, an information pushing device 1500 is provided, comprising: a feature extraction module 1502, an autocorrelation extraction module 1504, a feature fusion module 1506 and an information pushing module 1508, wherein:

[0177] Feature extraction module 1502, used to obtain product information of candidate products and object description information of the object to be pushed, and extract features of the candidate product information and object description information to obtain candidate product features and object description features;

[0178] An autocorrelation extraction module 1504 is configured to combine the candidate product features and the object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees.

[0179] A feature fusion module 1506 is configured to use the candidate product features and the object description features as input features, and sequentially fuse the autocorrelation features corresponding to at least two autocorrelation levels, with the result of the previous feature fusion being used as the input feature of the next feature fusion, until the target fused feature is obtained;

[0180] The information push module 1508 is used to calculate the degree of interaction of the target object with respect to the product information based on the target fusion feature, and push the product information of the candidate product to the terminal of the target object when the degree of interaction meets the preset push conditions.

[0181] In one embodiment, the autocorrelation extraction module 1504 is also used to calculate the feature importance of the candidate product features and the feature importance of the object description features, and perform feature selection on the candidate product features according to the feature importance of the candidate product features to obtain the candidate product features after feature selection, and perform feature selection on the object description features according to the feature importance of the object description features to obtain the object description features after feature selection; the candidate product features after feature selection and the object description features after feature selection are spliced ​​to obtain shared features.

[0182] In one embodiment, the autocorrelation extraction module 1504 is also used to obtain at least two autocorrelation calculation parameters, calculate the correlation between shared features and shared features based on the at least two autocorrelation calculation parameters, and obtain at least two autocorrelation degrees of the shared features; based on the at least two autocorrelation degrees, respectively transform the shared features to obtain autocorrelation features corresponding to the at least two autocorrelation degrees.

[0183] In one embodiment, the current autocorrelation calculation parameter of the at least two autocorrelation calculation parameters includes a current query parameter, a current key parameter, and a current value parameter;

[0184] The autocorrelation extraction module 1504 is also used to transform the shared feature based on the current query parameter to obtain the current query feature, and transform the shared feature based on the current key parameter to obtain the current key feature; calculate the correlation between the current query feature and the current key feature to obtain the current correlation matrix, normalize the current correlation matrix to obtain the current autocorrelation degree of the shared feature; traverse each autocorrelation calculation parameter to obtain at least two autocorrelation degrees of the shared feature; transform the shared feature based on the current value parameter to obtain the current value feature, and transform the current value feature based on the current autocorrelation degree of the shared feature to obtain the current autocorrelation feature corresponding to the current autocorrelation degree; traverse each autocorrelation calculation parameter to obtain the autocorrelation features corresponding to at least two autocorrelation degrees.

[0185] In one embodiment, the feature fusion module 1506 is further used to take the candidate product features and the object description features as input features, and determine the current autocorrelation feature from the autocorrelation features corresponding to at least two autocorrelation degrees respectively; fuse the input features with the current autocorrelation feature to obtain the current fused feature; take the current fused feature as the input feature, and return to the step of determining the current autocorrelation feature from the autocorrelation features corresponding to at least two autocorrelation degrees respectively, until all the autocorrelation features corresponding to at least two autocorrelation degrees are fused, and obtain the target fused feature.

[0186] In one embodiment, the feature fusion module 1506 is further configured to concatenate the input feature with the current autocorrelation feature to obtain a current concatenated feature; obtain full-connection operation parameters, and perform feature interaction on the current concatenated feature through the full-connection operation parameters to obtain a current fused feature.

[0187] In one embodiment, the information push module 1508 is further configured to perform a full connection operation on the target fusion feature to obtain a full connection operation result, and perform normalized mapping on the full connection operation result to obtain the degree of interaction of the object to be pushed with the product information.

[0188] In one embodiment, the object description information includes object basic description information and object interaction description information, and the object description features include object basic description features and object interaction description features;

[0189] The feature extraction module 1502 is also used to obtain the object original information of the object to be pushed and the product original information of the candidate product, and pre-process the object original information and the product original information respectively to obtain the object basic description information, the object interaction description information and the product information of the candidate product; encode the object basic description information to obtain the object basic description feature, encode the object interaction description information to obtain the object interaction description feature, and encode the product information of the candidate product to obtain the candidate product feature.

[0190] In one embodiment, the candidate products include at least two, and the information push device 1500 further includes:

[0191] A candidate product screening module is used to obtain product information of at least two candidate products and extract features of the product information of at least two candidate products to obtain candidate product features of at least two candidate products; obtain target fusion features of at least two candidate products based on the candidate product features and object description features of at least two candidate products; calculate the degree of interaction of the object to be pushed with the at least two candidate products respectively based on the target fusion features of at least two candidate products; according to the degree of interaction of the object to be pushed with the at least two candidate products respectively, screen target candidate products corresponding to the target interaction degree from at least two candidate products, and push the product information of the target candidate product to the terminal of the object to be pushed.

[0192] In one embodiment, the objects to be pushed include at least two, and the information pushing device 1500 further includes:

[0193] The object screening module is used to obtain object description information of at least two objects to be pushed, and extract features of the object description information of at least two objects to be pushed to obtain object description features of at least two objects to be pushed; based on the candidate product features and the object description features of at least two objects to be pushed, obtain target fusion features of at least two objects to be pushed; based on the target fusion features of at least two objects to be pushed, calculate the degree of interaction of at least two objects to be pushed with the candidate product respectively; according to the degree of interaction of at least two objects to be pushed with the candidate product respectively, screen the target push object from at least two objects to be pushed, and push the product information of the candidate product to the terminal of the target push object.

[0194] In one embodiment, the information push device 1500 further includes:

[0195] The model push module is used to input product information and object description information into the product information push model; extract features of candidate product information and object description information through the feature extraction network in the product information push model to obtain candidate product features and object description features; combine the candidate product features and object description features through the product information push model to obtain shared features, and calculate at least two autocorrelation degrees of the shared features through the autocorrelation feature extraction network in the product information push model, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees; use the candidate product features and object description features as input features through the feature fusion network in the product information push model, and fuse them with the autocorrelation features corresponding to the at least two autocorrelation degrees in sequence, with the result of the previous feature fusion serving as the input feature of the next feature fusion, until the target fusion feature is obtained; use the interaction degree calculation network in the product information push model to calculate the interaction degree of the object to be pushed with the product information based on the target fusion feature; when the interaction degree meets the preset push conditions, push the product information of the candidate product to the terminal of the object to be pushed.

[0196] In one embodiment, the information push device 1500 further includes:

[0197] The model training module is used to obtain training product information, training object description information and information push training labels; input the training product information and training object description information into the initial product information push model for forward calculation to obtain the training interaction degree of the output training object description information with the training product information; calculate the loss information between the training interaction degree and the information push training label, and reversely update the initial product information push model based on the loss information and perform loop iteration until the training completion conditions are met, thereby obtaining the product information push model.

[0198] In one embodiment, the candidate products include network traffic products; the information push module 1508 is also used to calculate the degree of interaction of the object to be pushed with the network traffic product based on the target fusion characteristics, and when the degree of interaction meets the preset push conditions, the product information of the network traffic product is pushed to the terminal of the object to be pushed.

[0199] Each module in the above-mentioned information push device can be implemented in whole or in part through software, hardware, or a combination thereof. Each module can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0200] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Figure 16 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, memory and input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. 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 computer program in the non-volatile storage medium. The database of the computer device is used to store data such as product information of candidate products, object description information of objects to be pushed, and product information push models. 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 an external terminal via a network connection. When the computer program is executed by the processor, an information push method is implemented.

[0201] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 17As shown. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are connected to the system bus via the input / output interface. 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 and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner. The wireless manner can be achieved through WIFI, a mobile cellular network, NFC (near field communication), or other technologies. When the computer program is executed by the processor, it implements an information push method. The display unit of the computer device is used to form a visually visible image, and can be a display screen, a projection device or a virtual reality imaging device. The display screen can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device casing, or an external keyboard, touchpad or mouse, etc.

[0202] Those skilled in the art will understand that Figure 16 or Figure 17 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0203] In one embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0204] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0205] In one embodiment, a computer program article is provided, comprising a computer program, which implements the steps of the above method embodiments when executed by a processor.

[0206] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of relevant data must comply with the relevant laws, regulations and standards of relevant countries and regions.

[0207] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the 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 above-mentioned embodiments. In particular, any reference to memory, database, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. 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), magnetic 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 take various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM). The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may be, but are not limited to, general-purpose processors, central processing units (CPUs), graphics processing units (GPUs), digital signal processors (DSPs), programmable logic devices (PLDs), data processing logic devices based on quantum computing, and the like.

[0208] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, 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, they should be considered to be within the scope of this specification.

[0209] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. An information push method, characterized in that: The method comprises: Acquire product information of a candidate product and object description information of an object to be pushed, and extract features of the candidate product information and the object description information to obtain candidate product features and object description features; Combining the candidate product feature and the object description feature to obtain a shared feature, calculating at least two autocorrelation degrees of the shared feature, and transforming the shared feature based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees respectively; The candidate product features and the object description features are used as input features, and the autocorrelation features corresponding to the at least two autocorrelation levels are sequentially fused, and the result of the previous feature fusion is used as the input feature of the next feature fusion, until the target fusion feature is obtained; The degree of interaction of the object to be pushed with the product information is calculated based on the target fusion feature, and when the degree of interaction meets the preset push condition, the product information of the candidate product is pushed to the terminal of the object to be pushed.

2. The method according to claim 1, characterized in that The combining of the candidate product features and the object description features to obtain shared features includes: Calculating the feature importance of the candidate product features and the feature importance of the object description features, and performing feature selection on the candidate product features according to the feature importance of the candidate product features to obtain selected candidate product features, and performing feature selection on the object description features according to the feature importance of the object description features to obtain selected object description features; The candidate product features after the feature selection and the object description features after the feature selection are spliced ​​to obtain shared features.

3. The method according to claim 1, characterized in that The calculating at least two autocorrelation degrees of the shared feature, and transforming the shared feature based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees, respectively, includes: Obtaining at least two autocorrelation calculation parameters, and calculating the correlation degree between the shared features based on the at least two autocorrelation calculation parameters to obtain at least two autocorrelation degrees of the shared features; The shared features are transformed based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees.

4. The method according to claim 3, characterized in that The current autocorrelation calculation parameter of the at least two autocorrelation calculation parameters includes a current query parameter, a current key parameter and a current value parameter; The calculating the correlation degree between the shared features based on the at least two autocorrelation calculation parameters to obtain at least two autocorrelation degrees of the shared features includes: Transforming the shared feature based on the current query parameter to obtain a current query feature, and transforming the shared feature based on the current key parameter to obtain a current key feature; Calculating the correlation between the current query feature and the current key feature to obtain a current correlation matrix, and normalizing the current correlation matrix to obtain a current autocorrelation degree of the shared feature; Traversing each autocorrelation calculation parameter to obtain at least two autocorrelation degrees of the shared feature; The transforming the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees respectively includes: Transforming the shared feature based on the current value parameter to obtain a current value feature, and transforming the current value feature based on a current autocorrelation degree of the shared feature to obtain a current autocorrelation feature corresponding to the current autocorrelation degree; Each autocorrelation calculation parameter is traversed to obtain autocorrelation features corresponding to the at least two autocorrelation degrees respectively.

5. The method according to claim 1, wherein The step of using the candidate product features and the object description features as input features and fusing the autocorrelation features corresponding to the at least two autocorrelation levels in sequence, using the result of the previous feature fusion as the input feature of the next feature fusion, until a target fused feature is obtained, includes: Taking the candidate product features and the object description features as input features, and determining a current autocorrelation feature from the autocorrelation features corresponding to the at least two autocorrelation levels respectively; Fusing the input feature with the current autocorrelation feature to obtain a current fused feature; The current fused feature is used as an input feature, and the step of determining the current autocorrelation feature from the autocorrelation features corresponding to the at least two autocorrelation degrees is returned to execution until all the autocorrelation features corresponding to the at least two autocorrelation degrees are fused, thereby obtaining the target fused feature.

6. The method according to claim 5, characterized in that The fusing the input feature with the current autocorrelation feature to obtain a current fused feature includes: Splicing the input feature with the current autocorrelation feature to obtain a current spliced ​​feature; Obtain full connection operation parameters, perform feature interaction on the current splicing feature through the full connection operation parameters, and obtain the current fusion feature.

7. The method according to claim 1, characterized in that The calculating the degree of interaction of the to-be-pushed object with the product information based on the target fusion feature includes: A full connection operation is performed on the target fusion feature to obtain a full connection operation result, and the full connection operation result is normalized and mapped to obtain the degree of interaction of the object to be pushed with the product information.

8. The method according to claim 1, characterized in that The object description information includes object basic description information and object interaction description information, and the object description features include object basic description features and object interaction description features; The extracting features of the candidate product information and the object description information to obtain candidate product features and object description features includes: Obtaining original object information of the object to be pushed and original product information of the candidate product, and preprocessing the original object information and the original product information respectively to obtain basic description information of the object, interaction description information of the object, and product information of the candidate product; The object basic description information is encoded to obtain the object basic description feature, the object interaction description information is encoded to obtain the object interaction description feature, and the product information of the candidate product is encoded to obtain the candidate product feature.

9. The method according to claim 1, characterized in that The candidate products include at least two, and the method further includes: Acquiring product information of at least two candidate products, and extracting features of the product information of the at least two candidate products to obtain candidate product features of the at least two candidate products; acquiring target fusion features of the at least two candidate products based on the candidate product features of the at least two candidate products and the object description features; Calculating the interaction degree of the to-be-pushed object to the at least two candidate products respectively based on the target fusion features of the at least two candidate products; According to the interaction degree of the object to be pushed to the at least two candidate products, a target candidate product corresponding to the target interaction degree is screened from the at least two candidate products, and product information of the target candidate product is pushed to the terminal of the object to be pushed.

10. The method according to claim 1, characterized in that The objects to be pushed include at least two, and the method further includes: Obtaining object description information of at least two objects to be pushed, and extracting features of the object description information of the at least two objects to be pushed to obtain object description features of the at least two objects to be pushed; Based on the candidate product features and the object description features of the at least two objects to be pushed, obtaining target fusion features of the at least two objects to be pushed; Calculating the interaction degree of the at least two objects to be pushed to the candidate product based on the target fusion features of the at least two objects to be pushed; According to the interaction degree of the at least two objects to be pushed to the candidate product respectively, a target push object is screened from the at least two objects to be pushed, and the product information of the candidate product is pushed to the terminal of the target push object.

11. The method according to claim 1, wherein The method further comprises: Inputting the product information and the object description information into a product information push model; Extracting features of the candidate product information and the object description information through a feature extraction network in the product information push model to obtain candidate product features and object description features; Combining the candidate product features and the object description features through the product information push model to obtain shared features, calculating at least two autocorrelation degrees of the shared features through an autocorrelation feature extraction network in the product information push model, and transforming the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees respectively; The feature fusion network in the product information push model uses the candidate product features and the object description features as input features, and sequentially fuses the autocorrelation features corresponding to the at least two autocorrelation levels, with the result of the previous feature fusion serving as the input feature of the next feature fusion, until a target fused feature is obtained; Calculating the degree of interaction of the object to be pushed with the product information based on the target fusion feature through the interaction degree calculation network in the product information push model; When the interaction level meets the preset push condition, the product information of the candidate product is pushed to the terminal of the to-be-pushed object.

12. The method according to claim 1, characterized in that The training of the product information push model includes the following steps: Obtain training product information, training object description information, and information push training tags; Inputting the training product information and the training object description information into the initial product information push model for forward calculation, and obtaining the output of the training interaction degree of the training object description information with the training product information; Calculate the loss information between the training interaction degree and the information push training label, reversely update the initial product information push model based on the loss information, and perform loop iteration until the training completion condition is met, thereby obtaining the product information push model.

13. The method according to claim 1, wherein The candidate product includes a network traffic product; the calculating the degree of interaction of the to-be-pushed object with the product information based on the target fusion feature, and when the degree of interaction meets a preset push condition, pushing the product information of the candidate product to the to-be-pushed object includes: The degree of interaction of the object to be pushed with the network traffic product is calculated based on the target fusion feature, and when the degree of interaction meets the preset push conditions, the product information of the network traffic product is pushed to the terminal of the object to be pushed.

14. An information push device, characterized in that: The device comprises: A feature extraction module is used to obtain product information of candidate products and object description information of the object to be pushed, and extract features of the candidate product information and the object description information to obtain candidate product features and object description features; an autocorrelation extraction module, configured to combine the candidate product features and the object description features to obtain shared features, calculate at least two autocorrelation degrees of the shared features, and transform the shared features based on the at least two autocorrelation degrees to obtain autocorrelation features corresponding to the at least two autocorrelation degrees; a feature fusion module, configured to use the candidate product features and the object description features as input features, and sequentially fuse the autocorrelation features corresponding to the at least two autocorrelation levels, with the result of the previous feature fusion serving as the input feature of the next feature fusion, until a target fusion feature is obtained; The information push module is used to calculate the degree of interaction of the object to be pushed with the product information based on the target fusion feature, and when the degree of interaction meets the preset push conditions, push the product information of the candidate product to the terminal of the object to be pushed.

15. A computer device comprising a memory and a processor, wherein 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 13 are implemented.

16. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 13 are implemented.

17. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 13 are implemented.

Images