Resource recommendation method, and deep learning model training method and device

Abstract

The invention provides a resource recommendation method and a deep learning model training method and device, and relates to the technical field of artificial intelligence, in particular to the technical field of deep learning and content recommendation. According to the specific implementation scheme, a historical behavior sequence of a target object is converted into an input feature sequence, and respective resource features of multiple candidate resources are determined; feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features are calculated, and the historical input features are input features located before the input features in the input feature sequence; determining an output feature sequence of the input feature sequence according to the feature correlation and the relative information correlation; calculating the correlation between the resource feature and each output feature in the output feature sequence; and determining a target resource from the plurality of candidate resources according to the correlation between the resource features and the output features. The invention further provides electronic equipment, a storage medium and a computer program product.

Description

Technical Field

[0001] This disclosure relates to the field of artificial intelligence technology, and more particularly to the fields of deep learning and content recommendation technology. More specifically, this disclosure provides a resource recommendation method, a deep learning model training method, an apparatus, an electronic device, a storage medium, and a computer program product. Background Technology

[0002] With the rapid development of artificial intelligence technology, users can conveniently browse news, videos, e-commerce products, and more using smartphones and other terminal devices. Related internet platforms can also recommend resources and information to users based on their needs. Summary of the Invention

[0003] This disclosure provides a resource recommendation method, a deep learning model training method, an apparatus, an electronic device, a storage medium, and a computer program product.

[0004] According to the first aspect, a resource recommendation method is provided, which includes: converting a historical behavior sequence of a target object into an input feature sequence, and determining the resource features of multiple candidate resources; calculating the feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features, wherein historical input features are input features that precede other input features in the input feature sequence; determining an output feature sequence of the input feature sequence based on the feature correlation and relative information correlation; calculating the correlation between resource features and each output feature in the output feature sequence; and determining the target resource from multiple candidate resources based on the correlation between resource features and output features.

[0005] According to the second aspect, a method for training a deep learning model is provided, the method comprising: converting a historical behavior sequence of sample objects in sample data into an input feature sequence, and determining resource features of the sample resources; determining a first evaluation value of the sample resources based on the correlation between each input feature in the input feature sequence and historical input features, and the correlation between the resource features and each input feature in the input feature sequence, wherein the historical input features are the input features in the input feature sequence that precede the input features; determining a first loss based on the first evaluation value; and adjusting the parameters of a first deep learning model based on the first loss.

[0006] According to a third aspect, a resource recommendation apparatus is provided, comprising: a first feature determination module, configured to convert a historical behavior sequence of a target object into an input feature sequence and determine the resource features of multiple candidate resources; a feature calculation module, configured to calculate the feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features, wherein historical input features are input features preceding input features in the input feature sequence; a feature sequence determination module, configured to determine an output feature sequence of the input feature sequence based on the feature correlation and relative information correlation; a correlation calculation module, configured to calculate the correlation between resource features and each output feature in the output feature sequence; and a resource determination module, configured to determine a target resource from multiple candidate resources based on the correlation between resource features and output features.

[0007] According to a fourth aspect, a training apparatus for a deep learning model is provided, the apparatus comprising: a second feature determination module, configured to convert a historical behavior sequence of sample objects in sample data into an input feature sequence and determine resource features of sample resources; a first evaluation value determination module, configured to determine a first evaluation value of the sample resources based on the correlation between each input feature in the input feature sequence and historical input features, and the correlation between resource features and each input feature in the input feature sequence, wherein historical input features are input features that precede other input features in the input feature sequence; a first loss determination module, configured to determine a first loss based on the first evaluation value; and a first parameter adjustment module, configured to adjust the parameters of a first deep learning model based on the first loss.

[0008] According to a fifth aspect, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a method provided according to the present disclosure.

[0009] According to a sixth aspect, a non-transitory computer-readable storage medium is provided storing computer instructions for causing a computer to perform the methods provided in this disclosure.

[0010] According to a seventh aspect, a computer program product is provided, comprising a computer program stored on at least one of a readable storage medium and an electronic device, the computer program implementing the method provided in this disclosure when executed by a processor.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0012] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein:

[0013] Figure 1 This is an exemplary system architecture diagram of a resource recommendation method, a deep learning model training method and apparatus that can be applied according to an embodiment of the present disclosure;

[0014] Figure 2 This is a flowchart of a resource recommendation method according to an embodiment of the present disclosure;

[0015] Figure 3 This is a schematic diagram of a resource recommendation method according to an embodiment of the present disclosure;

[0016] Figure 4 This is a schematic diagram of attention calculation for an input feature sequence according to an embodiment of the present disclosure;

[0017] Figure 5 This is a flowchart of a training method for a deep learning model according to an embodiment of the present disclosure;

[0018] Figure 6 This is a schematic diagram of a method for training a deep learning model according to an embodiment of the present disclosure;

[0019] Figure 7 This is a schematic diagram of multi-channel modeling according to an embodiment of the present disclosure;

[0020] Figure 8 This is a schematic diagram of a resource recommendation method according to one embodiment of this disclosure and another embodiment;

[0021] Figure 9 This is a block diagram of a resource recommendation apparatus according to an embodiment of the present disclosure;

[0022] Figure 10 This is a block diagram of a training apparatus for a deep learning model according to an embodiment of the present disclosure; and

[0023] Figure 11 This is a block diagram of an electronic device for a resource recommendation method and a deep learning model training method according to an embodiment of the present disclosure. Detailed Implementation

[0024] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0025] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0026] In the technical solution disclosed herein, the user's authorization or consent is obtained before acquiring or collecting the user's personal information.

[0027] Advertising recommendation is a common form of resource recommendation. Online advertising recommendation systems are widely used in advertising recommendation scenarios. The core function of such systems is to retrieve data from a massive advertising database, recommend ads to users based on their intent, build a bridge between advertisers and target users, and provide users with more personalized information services. Therefore, the prediction of the effectiveness of targeted advertising directly affects the efficiency of online retrieval systems and user experience. Existing ranking models in advertising recommendation systems rely on a large number of discrete features related to users and products, analyzing and understanding user behavior through methods such as aggregated statistical features. However, in such systems, each feature is meticulously designed, resulting in a strong reliance on expert experience and insufficient adaptability.

[0028] With the development of Large Language Models (LLMs), attempts have been made to improve the advertising prediction ability by fine-tuning LLMs in advertising recommendation tasks. However, due to the influence of user behavior patterns and content formats in recommendation scenarios, the logic and generalization ability of large language models are not entirely applicable to advertising recommendation scenarios.

[0029] The existing search recommendation models in online advertising retrieval systems have the following problems: excessive reliance on manual features; excessive loss of detail due to serialization modeling; relatively fragmented serialization modeling with poor modeling adaptability, and high cost of introducing external data or new features in serialization; and insufficient characterization of user historical behavior, which limits the model's ability to reason about user behavior sequences and habits.

[0030] Based on this, this disclosure provides an end-to-end generative modeling method based on user sequences, which aims to model user behavior sequences through a Transformer structure, reduce information loss, reconstruct existing serialization modeling forms based on adaptive information supply, and improve the advertising prediction capability of recommendation systems.

[0031] Figure 1 This is a schematic diagram of an exemplary system architecture for applying a resource recommendation method, a deep learning module training and apparatus according to an embodiment of this disclosure. It should be noted that... Figure 1 The examples shown are merely examples of system architectures that can be applied to the embodiments of this disclosure, in order to help those skilled in the art understand the technical content of this disclosure, but do not mean that the embodiments of this disclosure cannot be used in other devices, systems, environments or scenarios.

[0032] like Figure 1 As shown, the system architecture 100 according to this embodiment may include terminal devices 101, 102, and 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the terminal devices 101, 102, and 103 and the server 105. The network 104 may include various connection types, such as wired and / or wireless communication links, etc.

[0033] Users can use terminal devices 101, 102, and 103 to interact with server 105 via network 104 to receive or send messages, etc. Terminal devices 101, 102, and 103 can be various electronic devices, including but not limited to smartphones, tablets, laptops, etc.

[0034] Server 105 can be a server providing various services, such as providing advertising recommendations, browsing, and purchasing services to websites visited by users using terminal devices 101, 102, and 103. The backend management server can analyze received user requests, user behavior, advertising information, and other data, and based on the analysis results, determine which advertisements match the user's intent and push the advertisements to the terminal devices. Server 105 can be deployed with a pre-trained model, which can be a deep learning model.

[0035] At least one of the resource recommendation method and deep learning model training method provided in this disclosure can generally be executed by server 105. Correspondingly, the resource recommendation device and deep learning model training device provided in this disclosure can generally be located in server 105. The resource recommendation method and deep learning model training method provided in this disclosure can also be executed by a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105. Correspondingly, the resource recommendation device and deep learning model training device provided in this disclosure can also be located in a server or server cluster that is different from server 105 and capable of communicating with terminal devices 101, 102, 103 and / or server 105.

[0036] Figure 2This is a flowchart of a resource recommendation method according to an embodiment of the present disclosure.

[0037] like Figure 2 As shown, the resource recommendation method 200 includes operations S210 to S250.

[0038] In operation S210, the historical behavior sequence of the target object is converted into an input feature sequence, and the resource characteristics of multiple candidate resources are determined.

[0039] The target audience can refer to users who request web page services. Historical behavior analysis can be performed on these users, and resources can be recommended based on the analysis results. A historical behavior sequence can refer to a series of actions performed by the target audience regarding resources at different points in time, such as browsing, clicking, and purchasing. For example, user A clicked on the first advertisement at the first moment, user A viewed the second advertisement at the second moment, user B clicked on the third advertisement at the third moment, and so on.

[0040] The input feature sequence can refer to the feature sequence obtained after preprocessing historical behavior sequences such as discretization, word segmentation, and combination, followed by embedding processing and multilayer perceptron mapping. For example, after discretizing, segmenting, and combining a user's behavior information, discrete features such as user ID, ad ID, ad title, behavior category, and the time of the behavior are obtained. Subsequently, these discrete plaintext features are processed by the embedding layer to obtain continuous embedded features. Then, the multilayer perceptron processing layer converts the embedded features into fixed-dimensional feature representations (Tokens). Each Token can represent a user's behavior. Arranging the feature representations of multiple user behaviors in chronological order yields the user's feature sequence. For example, the feature sequence of a target object can be {Token1, Token2, ...}, where Token1 can represent that the target object clicked on ad A at the first moment, Token2 can represent that the target object viewed ad B at the second moment, and so on.

[0041] Candidate resources can be resources in the resource library that are to be pushed. Resource features can be feature information obtained by preprocessing candidate resources such as discretization, word segmentation, and combination, followed by embedding and multilayer perceptron mapping.

[0042] In operation S220, the feature correlation and relative information correlation between each input feature and historical input features in the input feature sequence are calculated.

[0043] In the embodiments of this disclosure, historical input features are input features that precede other input features in the input feature sequence. Calculating feature correlation can refer to performing attention calculations on the input features and historical input features. Calculating relative information correlation refers to performing attention calculations on the relative information between the input features and historical input features.

[0044] Relative information can include relative time information. For example, input features indicate that a user purchased a product from an advertisement at time B, while historical input features include when the user clicked on the advertisement at time A. The feature correlation between input features and historical input features can be determined based on the attention between them. Furthermore, the relative time information between input features and historical input features can be determined based on time B in the input features and time A in the historical input features, thus allowing for the calculation of relative temporal attention and obtaining relative temporal correlation.

[0045] In addition, relative information can also include relative location information, relative behavioral signal information, etc. Similar to relative time information, relative location correlation and relative behavioral correlation can be calculated.

[0046] In operation S230, the output feature sequence of the input feature sequence is determined based on feature correlation and relative information correlation.

[0047] For example, for each input feature, the overall correlation between the input feature and historical input features can be determined based on the feature correlation and relative information correlation between the input feature and historical input features. Based on the overall correlation between the input feature and each historical input feature, the input feature can be further characterized to obtain its output feature. The output features of each input feature in the input feature sequence constitute the output feature sequence.

[0048] In operation S240, the correlation between resource features and each output feature in the output feature sequence is calculated.

[0049] Calculating the correlation between resource features and each output feature can refer to performing attention calculations between resource features and each output feature to obtain the correlation between the features of candidate resources and the output feature sequence of the target object.

[0050] In operation S250, the target resource is determined from multiple candidate resources based on the correlation between resource characteristics and output characteristics.

[0051] After obtaining the correlation between the features of the candidate resource and the output feature sequence of the target object, the features of the candidate resource can be further characterized based on this correlation to obtain the output features of the candidate resource. Based on the output features of the candidate resource, the evaluation value of the candidate resource can be determined. For example, the evaluation value represents the probability that the target object clicks on or converts the candidate resource (e.g., an advertisement).

[0052] There can be multiple candidate resources. After obtaining the evaluation value of each candidate resource, the candidate resource with the highest probability of being clicked or converted can be determined from the multiple candidate resources and selected as the target resource. This target resource can then be recommended to the user.

[0053] According to embodiments of this disclosure, the behavior sequence of the target object is converted into an input feature sequence, and the correlation between each input feature in the input feature sequence and historical input features is calculated to obtain an output feature sequence after attention calculation. While calculating the attention between sequences, the relative information between the input features and historical input features is also introduced, thereby enabling the model to learn richer sequence information and reducing the information loss of the sequence. Subsequently, the correlation between the resource features of the candidate resources and the output features in the output feature sequence is calculated, and the target resource is determined from the candidate resources based on the correlation, which can improve the accuracy of resource recommendation.

[0054] Figure 3 This is a schematic diagram of a resource recommendation method according to an embodiment of the present disclosure.

[0055] like Figure 3 As shown, the input feature sequence can be obtained by sorting behavioral features 301 in chronological order, and may also include resource features 302 of candidate resources. During attention calculation within the input feature sequence, relative position information 303, relative time information 304, and relative behavioral information 305 between the input features and historical input features can be obtained to calculate relative information correlation. By sequentially passing the input feature sequence, relative position information 303, relative time information 304, and relative behavioral information 305 through a multi-head attention layer, a residual connection / normalization layer, a forward feedback layer, and another residual connection / normalization layer, the output feature sequence of the input feature sequence can be obtained.

[0056] According to embodiments of this disclosure, calculating the feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features includes: calculating the relative positional correlation, relative temporal correlation, and relative behavioral correlation between the input feature and historical input features as relative information correlation.

[0057] Relative positional relevance refers to attention information calculated by applying attention to the positions of input features and historical input features within the input feature sequence. Relative temporal relevance refers to attention information calculated by applying attention to the times when the actions corresponding to input features and historical input features occurred. Relative behavioral relevance refers to attention information calculated by applying attention to the categories of actions corresponding to input features and historical input features.

[0058] In the embodiments of this disclosure, when performing attention calculation on input features, the attention score between the query feature corresponding to the i-th input feature and the key feature corresponding to the j-th input feature is... It can be determined by the following formula (1):

[0059] (1)

[0060] in, It refers to the transpose of the query feature corresponding to the i-th input feature. It refers to the key feature corresponding to the j-th input feature. This refers to the relative position information between the i-th input feature and the j-th input feature. This refers to the relative behavioral information between the i-th input feature and the j-th input feature. This refers to the relative time information between the i-th input feature and the j-th input feature.

[0061] For example, the input feature is the third input feature in the input feature sequence, representing that the target object purchased item number 1 at time N. The historical input feature is the first input feature in the input feature sequence, representing that the target object browsed item number 1 at time M. The relative temporal correlation between the input feature and the historical input feature can be determined by attention calculation based on time N and time M. The relative positional correlation between the input feature and the historical input feature can be determined by attention calculation based on the position of the input feature and the historical input feature in the input feature sequence. The relative behavioral correlation between the input feature and the historical input feature can be determined by attention calculation based on the target object's browsing behavior and the target object's purchasing behavior.

[0062] According to embodiments of this disclosure, in order to improve computational efficiency and reduce temporary variables in the computation process, the relative positional correlation, relative temporal correlation, and relative behavioral correlation between input features and historical input features can be calculated in the following ways.

[0063] Calculate the position correlation matrix between the input feature and the position information encoding table of the input feature sequence, and determine the relative position correlation from the position correlation matrix using the relative position information of the input feature and the historical input feature as an index.

[0064] The location correlation matrix determined based on the location information encoding table includes the relative location information between each input feature and all input features in the input feature sequence. Using the positions of the input feature and historical input features in the input feature sequence as indices, the relative location information corresponding to the input feature and historical input features can be obtained from the location correlation matrix.

[0065] For example, an input feature sequence includes X input features. Based on a location information encoding table, X×X relative location information can be obtained, resulting in a location correlation matrix. When determining the relative location information between the input feature and historical input features, if the input feature is the 5th input feature in the input feature sequence (X≥5), and the historical input feature is the 3rd input feature in the input feature sequence, then the relative location information in the 5th row and 3rd column of the location correlation matrix can be used as the relative location information between the input feature and the historical input feature.

[0066] Calculate the temporal correlation matrix between the input feature and the temporal information encoding table of the input feature sequence, and determine the relative temporal correlation from the temporal correlation matrix using the relative temporal information of the input feature and historical input features as an index.

[0067] The time correlation matrix determined based on the time information encoding table includes the relative time information between each input feature and all input features in the input feature sequence. Using the positions of the input features and historical input features in the input feature sequence as indices, the relative time information corresponding to the input features and historical input features can be obtained from the time correlation matrix.

[0068] For example, an input feature sequence includes Y input features. Based on a time information encoding table, Y×Y relative time information can be obtained, resulting in a time correlation matrix. When determining the relative time information between the input feature and historical input features, if the input feature is the 5th input feature in the input feature sequence (Y≥5), and the historical input feature is the 3rd input feature in the input feature sequence, then the relative time information in the 5th row and 3rd column of the time correlation matrix can be used as the relative time information between the input feature and the historical input feature.

[0069] Calculate the behavior correlation matrix between the input features and the behavior information encoding table of the input feature sequence, and determine the relative behavior correlation from the behavior correlation matrix using the relative behavior information of the input features and historical input features as indexes.

[0070] The behavior correlation matrix determined based on the behavior information coding table includes the relative behavior information between each input feature and all input features in the input feature sequence. Using the positions of the input feature and historical input features in the input feature sequence as indices, the relative behavior information corresponding to the input feature and historical input features can be obtained from the behavior correlation matrix.

[0071] For example, an input feature sequence includes Z input features. Based on the behavior information encoding table, Z×Z relative behavior information can be obtained, resulting in a behavior correlation matrix. When determining the relative behavior information between input features and historical input features, if the input feature is the 5th input feature in the input feature sequence (Z≥5), and the historical input feature is the 3rd input feature in the input feature sequence, then the relative behavior information in the 5th row and 3rd column of the behavior correlation matrix can be used as the relative behavior information between the input feature and the historical input feature.

[0072] Based on this, the attention score between the query feature corresponding to the i-th input feature and the key feature corresponding to the j-th input feature is... It can be determined by the following formula (2):

[0073] (2)

[0074] in, Used to represent the feature correlation between the i-th input feature and the j-th input feature. Represents the location correlation matrix. This represents the index information corresponding to the i-th row and j-th column in the location relevance matrix. Represents the behavioral correlation matrix. This represents the index information corresponding to the i-th row and j-th column in the behavior relevance matrix. Represents the time correlation matrix. This represents the index information corresponding to the i-th row and j-th column in the time correlation matrix. Used to represent the relative positional correlation between the i-th input feature and the j-th input feature. Used to represent the relative behavioral correlation between the i-th input feature and the j-th input feature. Used to represent the relative temporal correlation between the i-th input feature and the j-th input feature.

[0075] According to embodiments of this disclosure, determining a target resource from multiple candidate resources based on the correlation between resource features and output features includes: for each candidate resource, determining the output features of the candidate resource based on the correlation between the resource features of the candidate resource and each output feature in the output feature sequence, and determining the evaluation value of the candidate resource based on the output features of the candidate resource; and determining the target resource from multiple candidate resources based on the evaluation values ​​of each of the multiple candidate resources.

[0076] By performing attention calculations on each output feature in the resource feature and output feature sequence of the candidate resource, the attention score between the resource feature and each output feature is obtained. Based on the attention score between the resource feature and each output feature, the output feature of the candidate resource can be obtained.

[0077] Based on the output characteristics of the candidate resource, an evaluation value can be obtained, such as the probability of the candidate resource being clicked. Based on the evaluation values ​​of multiple candidate resources, a target resource can be determined from the candidate resources; for example, the target resource is the candidate resource with the highest evaluation value. Subsequently, the target resource can be recommended to [the relevant authority / organization].

[0078] According to embodiments of this disclosure, converting a target object's historical behavior sequence into an input feature sequence includes: extracting discrete features from the target object's historical behavior sequence; mapping the discrete features into continuous embedded features; converting the embedded features into behavioral features 301 of a specific dimension; and arranging the behavioral features in chronological order to obtain an input feature sequence.

[0079] Discrete features can be feature information obtained by discretely splitting, segmenting, and combining the historical behavior sequences of a target object. Embedded features can be feature information obtained by making discrete features continuous through an embedding layer. Behavioral features can be feature information obtained by mapping embedded features through a multilayer perceptron. The input feature sequence can be a feature sequence composed of behavioral features arranged in chronological order.

[0080] According to embodiments of this disclosure, by converting a user's historical behavior sequence into a behavior feature sequence, features with minimal information loss are obtained. Subsequently, attention between features of the sequence is calculated. By introducing relative time, relative location, and relative behavior signals for attention calculation, the accuracy of feature calculation can be improved, thereby improving the accuracy of resource recommendation.

[0081] Next, combined Figure 4 The attention calculation of the input feature sequence in this disclosure is further explained.

[0082] Figure 4 This is a schematic diagram of attention calculation for an input feature sequence according to an embodiment of the present disclosure.

[0083] According to embodiments of this disclosure, the input features include partial features and full features, wherein the partial features are features from which redundant information has been removed; calculating the feature correlation and relative information correlation between each input feature and historical input features in the input feature sequence includes: for each input feature, calculating the feature correlation and relative information correlation between the full features of the input feature and the partial features of historical input features; and calculating the feature correlation and relative information correlation between the full features of the input feature and the full features of the input feature itself.

[0084] In embodiments of this disclosure, some features may be features obtained by removing highly homogeneous information such as the gender of the target object, the year information in the time frame, and the platform information where the behavior occurred from the full set of features. When performing attention calculation on the full set of input features, attention is calculated between the full set of input features, some features of historical input features, and the full set of input features themselves.

[0085] For example, such as Figure 4 As shown, the input feature sequence includes the full features T1, T2, and T3 of three input features, and the partial features T1', T2', and T3' of three input features. When calculating the attention of the full feature T3 of the third input feature, the full feature T3 of the third input feature is treated with the partial features T1' and T2' of the first and second input features, and the full feature T3 of the third input feature is treated with itself.

[0086] According to embodiments of this disclosure, calculating the feature correlation and relative information correlation between the full features of an input feature and the partial features of historical input features includes: for each input feature, inserting the partial features of the input feature before the full features of the input feature; and using the full features of each input feature as the current feature, calculating the feature correlation and relative information correlation between the current feature and the partial features of historical features.

[0087] When performing attention calculations on input features and input feature sequences, a portion of each input feature can be inserted before the full set of input features to obtain the inserted input feature sequence. Performing attention calculations on the full set of each input feature in the input feature sequence can improve the computational efficiency of the model while avoiding information loss.

[0088] According to embodiments of this disclosure, determining the output feature sequence of an input feature sequence based on feature correlation and relative information correlation includes: performing attention processing on all features of each input feature in the input feature sequence based on feature correlation and relative information correlation to obtain the output feature sequence.

[0089] For example, such as Figure 4As shown, the input feature sequence includes the full features T1, T2, and T3 of three input features and the partial features T1', T2', and T3' of three input features. Attention is calculated for the full features of each input feature: the full features T1 of the first input feature are treated with itself; the full features T2 of the second input feature are treated with itself and the partial features T1' of the first input feature; and the full features T3 of the third input feature are treated with itself and the partial features T1' and T2' of the first and second input features. The attention calculation results for the full features of each input feature are used as the output feature sequence.

[0090] According to embodiments of this disclosure, when extracting discrete features, the features can be divided into partial features and full features. After the discrete features are processed by the embedding layer and the multilayer perceptron processing layer, a full feature sequence and a partial feature sequence are obtained. When performing attention calculation between sequences, each full feature calculates attention with the historical partial features, and the full feature calculates attention with itself, which can improve computational efficiency.

[0091] Figure 5 This is a flowchart of a training method for a deep learning model according to an embodiment of the present disclosure.

[0092] like Figure 5 As shown, the training method 500 of this deep learning model includes operations S510 to S540. This deep learning model is a first deep learning model. The first deep learning model may include a discrete feature processing layer, an embedding layer, a first feature processing layer, and an attention processing layer.

[0093] In operation S510, the historical behavior sequence of sample objects in the sample data is converted into an input feature sequence, and the resource characteristics of the sample resources are determined.

[0094] The sample object can be a historical user who requested the web page service, and the sample resource can be a resource from which historical users generated clicks, views, purchases, and other behaviors. The historical behavior sequence can refer to a sequence of browsing, clicks, purchases, and other behaviors that the sample object performed on the resource at a historical moment.

[0095] Discrete features can be obtained by preprocessing historical behavior sequences through discrete feature processing layers such as discretization, word segmentation, and combination. Then, embedding layers are used to embed these discrete features, resulting in embedded features. Next, a first feature processing layer, such as a multilayer perceptron, converts the embedded features into fixed-dimensional feature representations (tokens). The tokens representing user behavior output by the first feature processing layer are arranged in chronological order to obtain the input feature sequence.

[0096] Resource features can be the features that represent sample resources obtained by processing the sample resources through the discrete feature processing layer, embedding layer, first feature processing layer and attention processing layer.

[0097] In operation S520, based on the correlation between each input feature in the input feature sequence and historical input features, and the correlation between resource features and each input feature in the input feature sequence, the first evaluation value of the sample resource is determined, wherein historical input features are the input features in the input feature sequence that precede the input features.

[0098] Attention processing layers can be used to compute attention within the input feature sequence and attention between resource features and input features of the input feature sequence. The attention processing layer can be a Transformer structure.

[0099] Intra-sequence attention calculation refers to the correlation between each input feature and historical features within the input feature sequence. The correlation between each input feature and historical input features can refer to performing attention calculations on each input feature and historical input features. The correlation between resource features and each input feature in the input feature sequence can refer to performing attention calculations on the resource features and each input feature after attention calculations. The first evaluation value can be the probability that a sample object clicks on or converts that sample resource.

[0100] In operation S530, the first loss is determined based on the first assessment value.

[0101] The first evaluation value is the probability that the sample resource output by the model is clicked or converted. Based on this first evaluation value and whether the sample resource is actually clicked or converted by the sample object, the loss of the first deep learning model can be determined, i.e., the first loss.

[0102] When operating the S540, the parameters of the first deep learning model are adjusted based on the first loss.

[0103] The parameters of the first deep learning model can be adjusted based on the loss of the first deep learning model. For example, the parameters of the attention processing layer, i.e., the parameters of the Transformer, can be adjusted.

[0104] According to embodiments of this disclosure, by converting the historical behavior sequence of a sample object into an input feature sequence, and utilizing an attention processing layer to perform attention calculation within the input feature sequence and attention between the sample resource and the input feature sequence, it is possible to perform serialized modeling of user behavior, improve the information utilization rate of the user feature sequence, and improve the accuracy of the model's ability to predict resource clicks, conversions, etc.

[0105] Next, combined Figure 6 The training of the deep learning model disclosed herein will be further explained.

[0106] Figure 6 This is a schematic diagram of a method for training a deep learning model according to an embodiment of the present disclosure.

[0107] like Figure 6 As shown, during the training phase, the deep learning model disclosed herein includes a first deep learning model 610 and a second deep learning model 620. The first deep learning model 610 includes an embedding layer 611, a first feature processing layer 612, and an attention processing layer 613. The second deep learning model 620 includes an embedding layer 611 and a second feature processing layer 622.

[0108] According to embodiments of this disclosure, the first deep learning model 610 determines a first evaluation value 606 of a sample resource by: using an attention processing layer 613 to perform the following operations: calculating the feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features; determining an output feature sequence of the input feature sequence based on the feature correlation and relative information correlation; calculating the correlation between resource features and each output feature in the output feature sequence; determining the output feature of the sample resource based on the correlation between resource features and output features; and determining a first evaluation value of the sample resource based on the output feature of the sample resource.

[0109] In embodiments of this disclosure, the first deep learning model 610 may be a model used to execute the resource recommendation method 200. The attention processing layer 613 obtains output features of the sample resource based on the input feature sequence, similar to obtaining output features of candidate resources based on the input feature sequence in the resource recommendation method 200; for simplicity, this will not be repeated. Based on the output features of the sample resource, a first evaluation value of the sample resource can be determined, such as the probability that the sample object clicks on or converts the sample resource.

[0110] According to embodiments of this disclosure, calculating the feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features includes: calculating the relative positional correlation, relative temporal correlation, and relative behavioral correlation between the input feature and historical input features as relative information correlation.

[0111] In the embodiments of this disclosure, the calculation of the relative positional correlation, relative temporal correlation, and relative behavioral correlation between input features and historical input features is similar to the calculation of the relative positional correlation, relative temporal correlation, and relative behavioral correlation between input features and historical input features in resource recommendation method 200, and will not be repeated for the sake of brevity.

[0112] According to an embodiment of this disclosure, in the training of the first deep learning model 610, converting the historical behavior sequence of the sample object into an input feature sequence includes: extracting dynamic discrete features 602 representing the dynamic behavior of the sample object from the sample data 601; converting the dynamic discrete features 602 into dynamic embedded features 604 using an embedding layer 611; converting the dynamic embedded features 604 into behavioral features of a specific dimension using a first feature processing layer 612; and arranging the behavioral features in chronological order to obtain an input feature sequence 605.

[0113] Dynamic discrete features 602 can refer to discrete features characterizing the dynamic behavior of a sample object, and can be feature information obtained by preprocessing the dynamic behavior of the sample object through word segmentation, combination, etc. Dynamic discrete features characterize the continuous sequence of user behavior information. The dynamic discrete features can be made continuous through the embedding layer to obtain dynamic embedded features 604. The first feature processing layer 612 can be, for example, a multilayer perceptron, which can convert the dynamic embedded features 604 into fixed-dimensional behavioral features and obtain the input feature sequence 605 arranged in chronological order.

[0114] According to embodiments of this disclosure, adjusting the parameters of the first deep learning model 606 based on the first loss may include adjusting the parameters of the first feature processing layer 612 and the attention processing layer 613 based on the first loss.

[0115] In the training of the first deep learning model 610, the attention processing layer 613 is used for serialization modeling, and can also be called the serialization modeling layer. The serialization modeling layer is trained using feature sequences from the same user, which may lead to overfitting of the discrete representation of the model due to the skewed distribution of discrete features. In view of this, embodiments of this disclosure also provide a second deep learning model 620, which is trained in conjunction with the first deep learning model 610.

[0116] According to embodiments of this disclosure, a second deep learning model 620 includes an embedding layer 611 and a second feature processing layer 622. Training the second deep learning model 620 includes: extracting static discrete features 603 representing static attributes of sample data 601 from sample data 601; converting the static discrete features 603 into static embedded features 603 using the embedding layer 611; determining a second evaluation value 608 of the sample resources based on the static embedded features 603 and resource features using the second feature processing layer 622; determining a second loss based on the second evaluation value 608; and adjusting the embedding layer 611 and the second feature processing layer 622 based on the second loss.

[0117] In the embodiments of this disclosure, the first deep learning model 610 may share the same embedding layer 611 with the second deep learning model 620. The embedding layer 611 is used to convert the dynamic discrete features 602 of the first deep learning model and the static discrete features 603 of the second deep learning model into dynamic embedding features 604 and static embedding features 607, respectively.

[0118] Static discrete features 603 can be feature information extracted from the static attributes of sample data 601. The static attributes of sample data 601 can characterize the attribute information of the sample data itself, such as resource category, resource source information (e.g., advertiser information), user profile, user IP address, statistical characteristics of user historical behavior, etc. The static discrete features 603 can be made continuous through the embedding layer 611 to obtain static embedded features 607. The second feature processing layer 622, for example, can be a multilayer perceptron, which can calculate a second evaluation value 608 for the sample resource based on the static embedded features 607 and the resource features of the sample resource. The second evaluation value can characterize the probability that the sample object clicks on or converts the sample resource. Based on the second evaluation value 608 of the sample resource and whether the sample object actually clicks on or converts the sample resource, the loss of the second deep learning model, i.e., the second loss, can be determined. The parameters of the embedding layer 611 and the second feature processing layer 622 of the second deep learning model 620 can be adjusted using the second loss to obtain the trained second deep learning model 620.

[0119] In the embodiments of this disclosure, for the same batch of sample data 601, dynamic discrete features 602 can be extracted based on the sample data 601, and the first deep learning model 610 can be trained to obtain a first evaluation value 606. A first loss is determined based on the first evaluation value 606, and the parameters of the first feature processing layer 612 and the attention processing layer 613 are adjusted based on the first loss. Subsequently, static discrete features 603 are extracted from the sample data 601, and the second deep learning model 620 is trained based on the embedding layer to obtain a second evaluation value 608. A second loss is determined based on the second evaluation value 608, and the embedding layer 611 and the second feature processing layer 622 are adjusted based on the second loss. Subsequently, for the next batch of samples, the first deep learning model 610 is trained again based on the adjusted embedding layer 611, and so on, iterating continuously until the first deep learning model 610 is trained, resulting in a trained first deep learning model, which is then used to execute the resource recommendation method 200 described above.

[0120] According to embodiments of this disclosure, dynamic discrete features representing user behavior sequences are constructed for the first deep learning model. Sequential modeling is performed based on these dynamic discrete features, and the first deep learning model is then trained. The training of the first deep learning model does not update the parameters of the embedding layer to prevent overfitting caused by learning features from only the same user. Static discrete features representing the static attributes of a single sample are constructed for the second deep learning model. The second deep learning model is trained based on these static discrete features, and the training of the second deep learning model avoids the overfitting problem inherent in sequential modeling. Therefore, the first and second deep learning models are trained together, and the loss function of the second deep learning model is used to adjust the shared embedding layer. This avoids the overfitting problem caused by the first deep learning model learning features from only the same user, improving the accuracy of the model in predicting metrics such as resource clicks and conversions.

[0121] According to embodiments of this disclosure, the input feature sequence includes input feature sequences of multiple sample objects; calculating the feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features includes: for each sample object's input feature sequence, calculating the feature correlation and relative information correlation between each input feature in the sample object's input feature sequence and historical input features; and masking the correlation calculation between the input feature sequences of multiple sample objects.

[0122] The input feature sequence of multiple sample objects can be a feature sequence obtained from the dynamic behavior of multiple sample objects in the sample data. The input feature sequence of each sample object is independent and unrelated. When performing attention calculation between sequences for each sample object, the input feature sequences of other sample objects can be masked to prevent the model from learning irrelevant feature information, thereby improving computational efficiency and accuracy.

[0123] According to embodiments of this disclosure, the input feature sequence includes input feature sequences of multiple scenarios, and the first deep learning model includes multiple attention processing layers corresponding to the multiple scenarios respectively; determining the first evaluation value of the sample resources includes: for each scenario's input feature sequence, using the attention processing layer corresponding to that scenario to calculate the output features of the sample resources based on the input feature sequence and the resource features of the sample resources; determining the fused output features of the sample resources based on the multiple output features of the sample resources corresponding to the multiple scenarios respectively; and determining the first evaluation value of the sample resources based on the fused output features.

[0124] Multiple scenarios can include commercial scenarios and user scenarios. Commercial scenarios can include e-commerce scenarios, advertiser scenarios, etc. User scenarios can include search scenarios, information scenarios, etc. Commercial scenario signals are usually sparse, but user scenario behaviors are rich. In view of this, the embodiments of this disclosure use multi-channel modeling to process the commercial side user behavior sequence and the user side behavior sequence through their respective sequence modeling layers for attention processing. Before estimating the click-through rate or conversion rate of the advertisement, feature fusion is performed through a multilayer perceptron, and the click-through rate or conversion rate of the advertisement is calculated based on the fused features.

[0125] Figure 7 This is a schematic diagram of multichannel modeling according to an embodiment of the present disclosure.

[0126] In the embodiments of this disclosure, the first deep learning model can construct multiple serialization modeling layers corresponding to multiple scenarios respectively, and each serialization modeling layer has an independent Transformer structure as an independent channel.

[0127] like Figure 7 As shown, the scenarios can include three scenarios: e-commerce scenario 710, search scenario 720, and information scenario 730. Each scenario has an independent channel of serialization modeling layer, which can be the attention processing layer mentioned above.

[0128] The input feature sequence can be divided according to the scenario to obtain input feature sequences under multiple scenarios. For example, the first input feature sequence 701 is the feature sequence under the e-commerce scenario, the second input feature sequence 702 is the feature sequence under the search scenario, and so on.

[0129] The input feature sequence and candidate resource features 703 for each scenario can be input into the sequence modeling layer of the corresponding channel for attention calculation. For example, the first input feature sequence 701 and candidate resource features 703 can be input into the sequence modeling layer of channel 710 in the e-commerce scenario. The second input feature sequence 702 and candidate resource features 703 can be input into the sequence modeling layer of channel 720 in the search scenario, and so on.

[0130] The candidate resource features calculated by attention in each channel can be fused to form a representation of the final candidate resource, which is then used to calculate the click-through rate or conversion rate of the candidate resource.

[0131] For example, in each channel, attention is first calculated within the input feature sequence to obtain the output feature sequence. Then, the attention between the candidate resource features and the output feature sequence is calculated to obtain the output features of the candidate resources. Subsequently, the features of candidate resources from multiple channels are concatenated to obtain the final fused features of the candidate resources. The click-through rate of the candidate resources can be calculated based on the fused features.

[0132] According to embodiments of this disclosure, multi-channel technology is used to fuse output features from different scenarios, constructing more comprehensive features with more information, thereby improving the efficiency and recommendation effect of the overall advertising recommendation system.

[0133] Next, combined Figure 8 The resource recommendation method disclosed herein will be further explained.

[0134] Figure 8 This is a schematic diagram of a resource recommendation method according to another embodiment of the present disclosure.

[0135] like Figure 8 As shown, during the online service phase, a resource recommendation model is set up on the server 810 to execute the resource recommendation method 200 described above. When the server receives a page browsing request from the user terminal, the server obtains the user's recent behavior data and calculates the click-through rate of candidate resources based on the user's recent behavior data using the resource recommendation model. For example, the click-through rate is the probability that the user clicks on the candidate resource. Subsequently, the server can perform resource matching and sorting based on the click-through rate of the candidate resources, and filter out the target resources to recommend to the user.

[0136] In this embodiment of the disclosure, resources displayed to the target object will form a display log, which records the resource information viewed by the target object. At the same time, when the target object performs actions such as clicking, browsing, or purchasing on the target resource, a behavior log can be generated based on the target object's actions.

[0137] During the offline service phase, discrete features can be extracted from the display logs and behavior logs of the target object to update the training data. The updated training data can then be used to train the model offline. Once training is complete, the latest model can be pushed online periodically (e.g., hourly).

[0138] During the offline service phase, display logs and behavior logs of the target object can be collected to form a historical sequence of user behavior, which is then stored. When recent user behavior data is needed during the online service phase, it can be retrieved from the stored historical behavior data.

[0139] Figure 9 This is a block diagram of a resource recommendation apparatus according to an embodiment of the present disclosure.

[0140] like Figure 9 As shown, the resource recommendation device 900 includes a first feature determination module 910, a feature calculation module 920, a feature sequence determination module 930, a correlation calculation module 940, and a resource determination module 950.

[0141] The first feature determination module 910 is used to convert the historical behavior sequence of the target object into an input feature sequence and determine the resource features of each of the multiple candidate resources.

[0142] The feature calculation module 920 is used to calculate the feature correlation and relative information correlation between each input feature in the input feature sequence and the historical input features, wherein the historical input features are the input features that precede the input features in the input feature sequence.

[0143] The feature sequence determination module 930 is used to determine the output feature sequence of the input feature sequence based on feature correlation and relative information correlation.

[0144] The correlation calculation module 940 is used to calculate the correlation between resource features and each output feature in the output feature sequence.

[0145] The resource determination module 950 is used to determine the target resource from multiple candidate resources based on the correlation between resource characteristics and output characteristics.

[0146] According to an embodiment of this disclosure, the feature calculation module 920 includes a first calculation submodule for calculating the relative positional correlation, relative temporal correlation, and relative behavioral correlation between the input features and historical input features, as relative information correlation.

[0147] According to embodiments of this disclosure, the first calculation submodule includes a first calculation unit, a second calculation unit, and a third calculation unit. The first calculation unit is used to calculate the positional correlation matrix between the input feature and the positional information encoding table of the input feature sequence, and to determine the relative positional correlation from the positional correlation matrix using the relative positional information of the input feature and historical input features as an index. The second calculation unit is used to calculate the temporal correlation matrix between the input feature and the temporal information encoding table of the input feature sequence, and to determine the relative temporal correlation from the temporal correlation matrix using the relative temporal information of the input feature and historical input features as an index. The third calculation unit is used to calculate the behavioral correlation matrix between the input feature and the behavioral information encoding table of the input feature sequence, and to determine the relative behavioral correlation from the behavioral correlation matrix using the relative behavioral information of the input feature and historical input features as an index.

[0148] According to embodiments of this disclosure, the input features include partial features and full features, where partial features are features from which redundant information has been removed; the feature calculation module 920 includes a second calculation submodule and a third calculation submodule. The second calculation submodule is used to calculate the feature correlation and relative information correlation between the full features of the input features and the partial features of historical input features; the third calculation submodule is used to calculate the feature correlation and relative information correlation between the full features of the input features and the full features of the input features themselves.

[0149] According to embodiments of this disclosure, the second calculation submodule includes an insertion unit and a fourth calculation unit. The insertion unit is used to insert a partial feature of the input feature before the full feature of the input feature. The fourth calculation unit is used to calculate the feature correlation and relative information correlation between the current feature and the partial features of historical features, using the full feature of each input feature as the current feature.

[0150] According to an embodiment of this disclosure, the feature sequence determination module 930 includes a sequence determination submodule, which performs attention processing on all features of each input feature in the input feature sequence based on feature correlation and relative information correlation to obtain an output feature sequence.

[0151] According to embodiments of this disclosure, the resource determination module 950 includes an evaluation value determination submodule and a resource determination submodule. The evaluation value determination submodule is used to determine the output features of the candidate resources based on the correlation between the resource features of the candidate resources and each output feature in the output feature sequence, and to determine the evaluation value of the candidate resources based on the output features of the candidate resources; the resource determination submodule is used to determine the target resource from the multiple candidate resources based on the evaluation values ​​of the multiple candidate resources respectively.

[0152] According to embodiments of this disclosure, the first feature determination module 910 includes an extraction submodule, a mapping submodule, a transformation submodule, and a sorting submodule. The extraction submodule is used to extract discrete features from the historical behavior sequence of the target object; the mapping submodule is used to map the discrete features into continuous embedded features; the transformation submodule is used to convert the embedded features into behavioral features of a specific dimension; and the sorting submodule is used to arrange the behavioral features in chronological order to obtain an input feature sequence.

[0153] Figure 10 This is a block diagram of a training apparatus for a deep learning model according to an embodiment of the present disclosure.

[0154] like Figure 10 As shown, the training device 1000 for the deep learning model includes a second feature determination module 1010, a first evaluation value determination module 1020, a first loss determination module 1030, and a first parameter adjustment module 1040.

[0155] The second feature determination module 1010 is used to convert the historical behavior sequence of sample objects in the sample data into an input feature sequence and determine the resource characteristics of the sample resources.

[0156] The first evaluation value determination module 1020 is used to determine the first evaluation value of the sample resource based on the correlation between each input feature in the input feature sequence and the historical input features, and the correlation between the resource feature and each input feature in the input feature sequence, wherein the historical input features are the input features in the input feature sequence that precede the input features.

[0157] The first loss determination module 1030 is used to determine the first loss based on the first evaluation value.

[0158] The first parameter adjustment module 1040 is used to adjust the parameters of the first deep learning model according to the first loss.

[0159] According to embodiments of this disclosure, a first deep learning model includes an attention processing layer; the first evaluation value determination module 1020 is further configured to perform the following operations: calculate the feature correlation and relative information correlation between each input feature in the input feature sequence and historical input features; determine the output feature sequence of the input feature sequence based on the feature correlation and relative information correlation; calculate the correlation between resource features and each output feature in the output feature sequence; determine the output feature of the sample resource based on the correlation between the resource feature and the output feature; and determine the first evaluation value of the sample resource based on the output feature of the sample resource.

[0160] According to an embodiment of this disclosure, the first evaluation value determination module 1020 is further configured to perform the following operations: calculate the relative positional correlation, relative temporal correlation, and relative behavioral correlation between the input features and historical input features as relative information correlation.

[0161] According to embodiments of this disclosure, the first deep learning model further includes an embedding layer and a first feature processing layer; the second feature determination module 1010 includes a dynamic feature extraction submodule, a dynamic embedding feature transformation submodule, and a dimension transformation submodule. The dynamic feature extraction submodule is used to extract dynamic discrete features representing the dynamic behavior of sample objects from sample data; the dynamic embedding feature transformation submodule is used to convert the dynamic discrete features into dynamic embedded features using the embedding layer; the dimension transformation submodule is used to convert the dynamic embedded features into behavioral features of a specific dimension using the first feature processing layer, and arrange the behavioral features in chronological order to obtain an input feature sequence.

[0162] According to an embodiment of this disclosure, the first parameter adjustment module 1040 includes a first adjustment submodule, used to adjust the parameters of the first feature processing layer and the attention processing layer according to the first loss.

[0163] According to embodiments of this disclosure, the deep learning model further includes a second deep learning model, which includes an embedding layer and a second feature processing layer. The training apparatus 1000 for the deep learning model also includes a static feature extraction module, a static embedding feature conversion module, a second evaluation value determination module, a second loss determination module, and a second parameter adjustment module. The static feature extraction module extracts static discrete features characterizing the static attributes of the sample data from the sample data. The static embedding feature conversion module converts the static discrete features into static embedding features using the embedding layer. The second evaluation value determination module determines a second evaluation value of the sample resource based on the static embedding features and resource features using the second feature processing layer. The second loss determination module determines a second loss based on the second evaluation value. The second parameter adjustment module adjusts the embedding layer and the second feature processing layer based on the second loss.

[0164] According to an embodiment of this disclosure, the input feature sequence includes input feature sequences of multiple sample objects; the first evaluation value determination module 1020 is further configured to perform the following operations: for each sample object's input feature sequence, calculate the feature correlation and relative information correlation between each input feature in the sample object's input feature sequence and historical input features; and mask the correlation calculation between the input feature sequences of multiple sample objects.

[0165] According to embodiments of this disclosure, the input feature sequence includes input feature sequences of multiple scenarios, and the first deep learning model includes multiple attention processing layers corresponding to the multiple scenarios respectively; the first evaluation value determination module 1020 is further configured to perform the following operations: for each scenario's input feature sequence, using the attention processing layer corresponding to that scenario, calculate the output features of the sample resources based on the input feature sequence and the resource features of the sample resources; determine the fused output features of the sample resources based on the multiple output features of the sample resources corresponding to the multiple scenarios respectively; and determine the first evaluation value of the sample resources based on the fused output features.

[0166] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0167] Figure 11A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0168] like Figure 11 As shown, device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in read-only memory (ROM) 1102 or a computer program loaded into random access memory (RAM) 1103 from storage unit 1108. The RAM 1103 may also store various programs and data required for the operation of device 1100. The computing unit 1101, ROM 1102, and RAM 1103 are interconnected via bus 1104. Input / output (I / O) interface 1105 is also connected to bus 1104.

[0169] Multiple components in device 1100 are connected to I / O interface 1105, including: input unit 1106, such as keyboard, mouse, etc.; output unit 1107, such as various types of monitors, speakers, etc.; storage unit 1108, such as disk, optical disk, etc.; and communication unit 1109, such as network card, modem, wireless transceiver, etc. Communication unit 1109 allows device 1100 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0170] The computing unit 1101 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 1101 performs the various methods and processes described above, such as resource recommendation methods and deep learning model training methods. For example, in some embodiments, the resource recommendation methods and deep learning model training methods can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 1108. In some embodiments, part or all of the computer program can be loaded and / or installed on device 1100 via ROM 1102 and / or communication unit 1109. When the computer program is loaded into RAM 1103 and executed by the computing unit 1101, one or more steps of the resource recommendation methods and deep learning model training methods described above can be performed. Alternatively, in other embodiments, the computing unit 1101 may be configured in any other suitable manner (e.g., by means of firmware) to perform resource recommendation methods or deep learning model training methods.

[0171] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0172] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0173] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0174] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0175] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0176] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0177] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0178] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A resource recommendation method, comprising: converting a historical behavior sequence of a target object into an input feature sequence, and determining resource features of a plurality of candidate resources respectively; calculating feature correlation and relative information correlation between each input feature in the input feature sequence and a historical input feature, wherein the historical input feature is an input feature located before the input feature in the input feature sequence; determining an output feature sequence of the input feature sequence according to the feature correlation and the relative information correlation; calculating correlation between the resource features and each output feature in the output feature sequence; and determining a target resource from the plurality of candidate resources according to the correlation between the resource features and the output features.

2. The method of claim 1, wherein, The calculating the feature correlation and the relative information correlation between each input feature in the input feature sequence and a historical input feature comprises: calculating relative position correlation, relative time correlation and relative behavior correlation between the input feature and the historical input feature as the relative information correlation.

3. The method of claim 2, wherein, The calculating the relative position correlation, the relative time correlation and the relative behavior correlation between the input feature and the historical input feature as the relative information correlation comprises: calculating a position correlation matrix between the input feature and a position information encoding table of the input feature sequence, and determining the relative position correlation from the position correlation matrix with the relative position information between the input feature and the historical input feature as an index; calculating a time correlation matrix between the input feature and a time information encoding table of the input feature sequence, and determining the relative time correlation from the time correlation matrix with the relative time information between the input feature and the historical input feature as an index; calculating a behavior correlation matrix between the input feature and a behavior information encoding table of the input feature sequence, and determining the relative behavior correlation from the behavior correlation matrix with the relative behavior information between the input feature and the historical input feature as an index.

4. The method of claim 1, wherein, The input feature comprises a partial feature and a full feature, the partial feature being a feature with redundant information removed from the full feature; the calculating the feature correlation and the relative information correlation between each input feature in the input feature sequence and a historical input feature comprises, for each input feature, calculating the feature correlation and the relative information correlation between the full feature of the input feature and the partial feature of the historical input feature; calculating the feature correlation and the relative information correlation between the full feature of the input feature and the full feature of the input feature itself.

5. The method of claim 4, wherein, The calculating the feature correlation and the relative information correlation between the full feature of the input feature and the partial feature of the historical input feature comprises, for each input feature, inserting the partial feature of the input feature before the full feature of the input feature; and taking the full feature of each input feature as a current feature, calculating the feature correlation and the relative information correlation between the current feature and the partial feature of the historical feature.

6. The method of claim 4, wherein, determining, according to the feature correlation and the relative information correlation, an output feature sequence of the input feature sequence includes: performing attention processing on a full amount feature of each input feature in the input feature sequence according to the feature correlation and the relative information correlation to obtain the output feature sequence.

7. The method of claim 1, wherein, determining, according to the correlation between the resource feature and the output feature, a target resource from the plurality of candidate resources includes: for each candidate resource, determining an output feature of the candidate resource according to the correlation between the resource feature of the candidate resource and each output feature in the output feature sequence, and determining an evaluation value of the candidate resource according to the output feature of the candidate resource; determining a target resource from the plurality of candidate resources according to the evaluation values of the plurality of candidate resources.

8. The method of claim 1, wherein, the conversion of the historical behavior sequence of the target object into the input feature sequence includes: extracting discrete features from the historical behavior sequence of the target object; mapping the discrete features into continuous embedding features; converting the embedding features into behavior features of a specific dimension; and arranging the behavior features in a time sequence to obtain the input feature sequence.

9. A training method of a deep learning model, comprising: converting a historical behavior sequence of a sample object in sample data into an input feature sequence, and determining a resource feature of a sample resource; determining a first evaluation value of the sample resource according to a correlation between each input feature in the input feature sequence and a historical input feature, and a correlation between the resource feature and each input feature in the input feature sequence, wherein the historical input feature is an input feature located before the input feature in the input feature sequence; determining a first loss according to the first evaluation value; and adjusting parameters of a first deep learning model according to the first loss.

10. The method of claim 9, wherein, the first deep learning model includes an attention processing layer; and the determination of the first evaluation value of the sample resource includes: using the attention processing layer to perform the following operations: calculating feature correlation and relative information correlation between each input feature in the input feature sequence and a historical input feature; determining an output feature sequence of the input feature sequence according to the feature correlation and the relative information correlation; calculating a correlation between the resource feature and each output feature in the output feature sequence; determining an output feature of the sample resource according to the correlation between the resource feature and the output feature; determining the first evaluation value of the sample resource according to the output feature of the sample resource.

11. The method of claim 10, wherein, the calculation of the feature correlation and the relative information correlation between each input feature in the input feature sequence and a historical input feature includes: calculating relative position correlation, relative time correlation and relative behavior correlation between the input feature and the historical input feature as the relative information correlation.

12. The method of claim 10, wherein, the first deep learning model further includes an embedding layer and a first feature processing layer; the conversion of the historical behavior sequence of the sample object into the input feature sequence includes: extracting dynamic discrete features representing dynamic behaviors of the sample object from sample data; mapping the dynamic discrete features into continuous embedding features; converting the embedding features into behavior features of a specific dimension; and arranging the behavior features in a time sequence to obtain the input feature sequence. The embedding layer is used to convert the dynamic discrete features into dynamic embedding features. The first feature processing layer is used to convert the dynamic embedding features into behavior features of a specific dimension, and arrange the behavior features in a time sequence to obtain the input feature sequence.

13. The method of claim 12, wherein, The adjusting the parameters of the first deep learning model according to the first loss includes: The parameters of the first feature processing layer and the attention processing layer are adjusted according to the first loss.

14. The method of claim 12, wherein, The deep learning model further includes a second deep learning model, and the second deep learning model includes an embedding layer and a second feature processing layer; the method includes: The static discrete features representing the static attributes of the sample data are extracted from the sample data; The embedding layer is used to convert the static discrete features into static embedding features; The second feature processing layer is used to determine a second evaluation value of the sample resource based on the static embedding features and the resource features; A second loss is determined according to the second evaluation value; and The embedding layer and the second feature processing layer are adjusted according to the second loss.

15. The method of claim 10, wherein, The input feature sequence includes input feature sequences of a plurality of sample objects; and the calculating the feature correlation and the relative information correlation between each input feature and historical input features in the input feature sequence includes: The feature correlation and the relative information correlation between each input feature and historical input features in the input feature sequence of each sample object are calculated; The correlation calculation between the input feature sequences of the plurality of sample objects is shielded.

16. The method of claim 10, wherein, The input feature sequence includes input feature sequences of a plurality of scenes, and the first deep learning model includes a plurality of attention processing layers corresponding to the plurality of scenes respectively; The determining the first evaluation value of the sample resource includes: For the input feature sequence of each scene, the output feature of the sample resource is calculated based on the input feature sequence and the resource features of the sample resource by using the attention processing layer corresponding to the scene; A fusion output feature of the sample resource is determined according to a plurality of output features of the sample resource corresponding to the plurality of scenes respectively; The first evaluation value of the sample resource is determined according to the fusion output feature.

17. A resource recommendation apparatus, comprising: a first feature determination module configured to convert a historical behavior sequence of a target object into an input feature sequence, and determine resource features of a plurality of candidate resources respectively; a feature calculation module configured to calculate a feature correlation and a relative information correlation between each input feature and historical input features in the input feature sequence, wherein the historical input features are input features located before the input feature in the input feature sequence; a feature sequence determination module configured to determine an output feature sequence of the input feature sequence according to the feature correlation and the relative information correlation; a correlation calculation module configured to calculate a correlation between the resource features and each output feature in the output feature sequence; and a feature sequence determination module configured to determine an output feature sequence of the input feature sequence according to the feature correlation and the relative information correlation; a correlation calculation module configured to calculate a correlation between the resource features and each output feature in the output feature sequence; and The resource determining module is configured to determine a target resource from the plurality of candidate resources according to the correlation between the resource feature and the output feature.

18. An apparatus for training a deep learning model, comprising: a second feature determining module configured to convert a historical behavior sequence of a sample object in sample data into an input feature sequence, and determine a resource feature of a sample resource; a first evaluation value determining module configured to determine a first evaluation value of the sample resource according to a correlation between each input feature in the input feature sequence and a historical input feature, and a correlation between the resource feature and each input feature in the input feature sequence, wherein the historical input feature is an input feature located before the input feature in the input feature sequence; a first loss determining module configured to determine a first loss according to the first evaluation value; and a first parameter adjusting module configured to adjust a parameter of a first deep learning model according to the first loss.

19. An electronic device, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 16.

20. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to perform the method of any one of claims 1 to 16.

21. A computer program product comprising a computer program stored on at least one of a readable storage medium and an electronic device, the computer program, when executed by a processor, implements the method of any one of claims 1 to 16.

Images