A data processing method, a computer device and a readable storage medium

Abstract

Embodiments of the present application provide a data processing method, a computer device and a readable storage medium. The method can be applied to various scenes such as cloud technology, artificial intelligence, intelligent transportation and multimedia. The method comprises: obtaining target business data associated with a target object, determining matching features between the target object and the target business data; determining target business trigger features and target business conversion features; performing splicing processing on the matching features and the target business trigger features to obtain initial splicing trigger features, determining a trigger probability estimate of the target object for the target business data based on the initial splicing trigger features; performing splicing processing on the matching features and the target business conversion features to obtain initial splicing conversion features, and determining a conversion probability estimate of the target object for the target business data based on the initial splicing conversion features. The present application can improve the accuracy of predicting the trigger probability estimate and the conversion probability estimate of the target business data.

Description

Technical Field

[0001] This application relates to the field of Internet technology, and in particular to a data processing method, computer device, and readable storage medium. Background Technology

[0002] With the development of multimedia technology, multimedia data (such as text and images) has become the main medium for obtaining information and enjoying entertainment in daily life. Business data recommendation systems can distribute text and images to application clients while simultaneously delivering business data. It's understood that business data recommendation systems typically determine the business data to be delivered to application clients based on click-through rates and conversion rates.

[0003] Existing business data recommendation systems use deep learning models to train click-through rate (CTR) prediction models and conversion rate (CTR) prediction models, respectively. These models then extract click and conversion features from business data (e.g., business data G). The conversion rate prediction model is trained using sample business data that shows a click-through rate to a conversion rate, meaning all sample business data exhibits click behavior. However, the business data G used for prediction in the conversion rate prediction model does not exhibit click behavior. This inconsistency between the training and prediction data distributions leads to inaccurate conversion features extracted from business data G, thus reducing the model's overall prediction accuracy. Summary of the Invention

[0004] This application provides a data processing method, a computer device, and a readable storage medium, which can improve the accuracy of the predicted trigger probability and conversion probability of target business data.

[0005] One embodiment of this application provides a data processing method, including:

[0006] Acquire target business data associated with the target object, and determine the matching characteristics between the target object and the target business data;

[0007] Determine the target business trigger characteristics and target business conversion characteristics of the target object for the target business data; the target business conversion characteristics are obtained by performing feature sharing and transfer processing on the target business trigger characteristics;

[0008] The matching features and the target business trigger features are concatenated to obtain the initial concatenated trigger features of the target object for the target business data. Based on the initial concatenated trigger features, the estimated trigger probability of the target object for the target business data is determined.

[0009] The matching features and target business conversion features are concatenated to obtain the initial concatenated conversion features of the target object for the target business data. Based on the initial concatenated conversion features, the conversion probability estimate of the target object for the target business data is determined.

[0010] One embodiment of this application provides a data processing apparatus, including:

[0011] The first determining module is used to acquire target business data associated with the target object and determine the matching features between the target object and the target business data.

[0012] The second determining module is used to determine the target business triggering characteristics and target business conversion characteristics of the target object for the target business data; the target business conversion characteristics are obtained by performing feature sharing and transmission processing on the target business triggering characteristics;

[0013] The first prediction module is used to concatenate the matching features and the target business trigger features to obtain the initial concatenated trigger features of the target object for the target business data. Based on the initial concatenated trigger features, the predicted trigger probability of the target object for the target business data is determined.

[0014] The second prediction module is used to concatenate the matching features and the target business conversion features to obtain the initial concatenated conversion features of the target object for the target business data. Based on the initial concatenated conversion features, the conversion probability prediction value of the target object for the target business data is determined.

[0015] The first determining module includes:

[0016] The data acquisition unit is used to acquire a set of business data associated with the target object and to obtain the target business data from the set of business data.

[0017] The feature acquisition unit is used to acquire pre-trained object features associated with the target object and pre-trained business features associated with the business data sequence; the business data sequence is associated with the target object;

[0018] The feature embedding unit is used to input the object attributes of the target object, the business attributes of the target business data, the pre-trained object features, and the pre-trained business features into the matching sub-network of the probability prediction network model; the matching sub-network includes a matching embedding layer.

[0019] The feature embedding unit is used to embed features of object attributes and business attributes through the matching embedding layer to obtain object features corresponding to object attributes and business features corresponding to business attributes.

[0020] The vector determination unit is used in the matching subnetwork to determine the object feature vector associated with the target object based on object features and pre-trained object features, and to determine the business feature vector associated with the target business data based on business features and pre-trained business features.

[0021] The dot product operation unit is used to perform a dot product operation on the object feature vector and the business feature vector to obtain the matching features between the target object and the target business data.

[0022] The business data sequence includes a business trigger sequence and a business conversion sequence; the business trigger sequence includes business data of the target object that has triggering behavior, and the business conversion sequence includes business data of the target object that has conversion behavior.

[0023] The feature acquisition unit is specifically used to input the object attributes of the target object, the business trigger sequence associated with the target object, and the business conversion sequence associated with the target object into the feature recognition network model; the feature recognition network model includes a pre-trained embedding layer, an attention processing layer, and a pre-trained connection layer;

[0024] The feature acquisition unit is specifically used to embed features into object attributes, business data in business trigger sequences, and business data in business conversion sequences through a pre-trained embedding layer, respectively, to obtain object features corresponding to object attributes, trigger business features corresponding to business data in business trigger sequences, and conversion business features corresponding to business data in business conversion sequences.

[0025] The feature acquisition unit is specifically used to input object features, triggering business features and conversion business features into the attention processing layer. The attention processing layer performs attention processing on the object features, triggering business features and conversion business features to obtain the attention object features corresponding to the object features, the attention triggering features corresponding to the triggering business features and the attention conversion features corresponding to the conversion business features.

[0026] The feature acquisition unit is specifically used to input object features, attention object features, attention trigger features, and attention conversion features into the pre-trained connection layer. The pre-trained connection layer performs feature fusion on the object features and attention object features to obtain pre-trained object features associated with the target object. In the pre-trained connection layer, based on the attention trigger features and attention conversion features, pre-trained business features corresponding to the business data contained in the business trigger sequence and business conversion sequence are generated.

[0027] The matching sub-network includes an object fully connected layer, a business fully connected layer, and a feature averaging layer; the business data sequence includes one or more business data, each business data corresponding to a pre-trained business feature.

[0028] The vector determination unit is specifically used in the matching sub-network to input object features into the fully connected object layer, and then to process the object features through the fully connected object layer to obtain fully connected object features.

[0029] The vector determination unit is specifically used to perform feature fusion on the fully connected object features and the pre-trained object features to obtain the object feature vector associated with the target object.

[0030] The vector determination unit is specifically used to input business features into the business fully connected layer, and then perform fully connected processing on the business features through the business fully connected layer to obtain fully connected business features;

[0031] The vector determination unit is specifically used to input one or more pre-trained business features into the feature averaging layer, and to average the one or more pre-trained business features to obtain the average pre-trained business features.

[0032] The vector determination unit is specifically used to perform feature fusion on fully connected business features and average pre-trained business features to obtain a business feature vector associated with the target business data.

[0033] The probability prediction network model also includes a ranking sub-network; the ranking sub-network includes an input network layer, a parameter sharing network layer, and a multi-layer perceptual network layer.

[0034] The second determining module includes:

[0035] The feature determination unit is used to determine the shared attribute features of the target object for the target business data by inputting the network layer;

[0036] The feature generation unit is used to input shared attribute features into the parameter sharing network layer, and generate the initial business trigger features and initial business transformation features of the target object for the target business data through the parameter sharing network layer.

[0037] The feature transfer unit is used to input the initial service trigger feature and the initial service conversion feature into the multilayer perception network layer. The multilayer perception network layer performs fully connected processing on the initial service trigger feature to obtain the target service trigger feature of the target object for the target service data. In the multilayer perception network layer, the target service trigger feature and the initial service conversion feature are processed by feature sharing to obtain the target service conversion feature of the target object for the target service data.

[0038] The input network layer includes a feature embedding layer and a feature concatenation layer; the business data sequence includes one or more business data, each business data corresponding to a pre-trained business feature.

[0039] The feature determination unit is specifically used to input the object attributes of the target object, the business attributes of the target business data, and the context attributes associated with the target object into the feature embedding layer. The feature embedding layer performs feature embedding on the object attributes, business attributes, and context attributes to obtain the object features corresponding to the object attributes, the business features corresponding to the business attributes, and the context features corresponding to the context attributes.

[0040] The feature determination unit is specifically used to average one or more pre-trained business features to obtain the average pre-trained business features.

[0041] The feature determination unit is specifically used to input object features, pre-trained object features, business features and average pre-trained business features into the feature concatenation layer. In the feature concatenation layer, the object features and pre-trained object features are concatenated to obtain concatenated object features, and the business features and average pre-trained business features are concatenated to obtain concatenated business features.

[0042] The feature determination unit is specifically used to perform feature splicing on splicing object features, splicing business features and context features to obtain the shared attribute features of the target object for the target business data.

[0043] The parameter-sharing network layer includes a shared fully connected layer, a weight learning layer, a feature classification layer, a first trigger fully connected layer, a first shallow transformation fully connected layer, and a first deep transformation fully connected layer; the initial business transformation features include initial shallow transformation features and initial deep transformation features.

[0044] The feature generation unit is specifically used to input the shared attribute features into the shared fully connected layer, and to perform fully connected processing on the shared attribute features through the shared fully connected layer to obtain fully connected shared features.

[0045] The feature generation unit is specifically used to input the fully connected shared features into the weight learning layer, and the weight learning layer performs feature weighting on the fully connected shared features to obtain the weighted shared features.

[0046] The feature generation unit is specifically used to input the weighted shared features into the feature classification layer, and to classify the weighted shared features through the feature classification layer to obtain the trigger distribution vector, shallow transformation distribution vector and deep transformation distribution vector of the target object for the target business data.

[0047] The feature generation unit is specifically used to input the trigger distribution vector into the first trigger fully connected layer, and to perform full connection processing on the trigger distribution vector through the first trigger fully connected layer to generate the initial business trigger features of the target object for the target business data.

[0048] The feature generation unit is specifically used to input the shallow transformation distribution vector into the first shallow transformation fully connected layer, and to perform fully connected processing on the shallow transformation distribution vector through the first shallow transformation fully connected layer to generate the initial shallow transformation features of the target object for the target business data.

[0049] The feature generation unit is specifically used to input the deep transformation distribution vector into the first deep transformation fully connected layer, and to perform fully connected processing on the deep transformation distribution vector through the first deep transformation fully connected layer to generate the initial deep transformation features of the target object for the target business data.

[0050] The weight learning layer includes weight learning components corresponding to object features, business features, context features, pre-trained object features, and average pre-trained business features, respectively; the feature classification layer includes a trigger feature classification layer, a shallow transformation feature classification layer, and a deep transformation feature classification layer.

[0051] The feature generation unit is specifically used to input the fully connected shared features into each weight learning component, and to perform feature weighting on the fully connected shared features through each weight learning component to obtain the weighted shared features corresponding to each weight learning component.

[0052] The feature generation unit is specifically used to perform feature fusion on the weight-shared features corresponding to each weight learning component to obtain fused shared features.

[0053] The feature generation unit is specifically used to input the fused shared features into the trigger feature classification layer, and to classify the fused shared features through the trigger feature classification layer to obtain the trigger distribution vector of the target object for the target business data.

[0054] The feature generation unit is specifically used to input the fused shared features into the shallow transformation feature classification layer, and to classify the fused shared features through the shallow transformation feature classification layer to obtain the shallow transformation distribution vector of the target object for the target business data;

[0055] The feature generation unit is specifically used to input the fused shared features into the deep transformation feature classification layer. The deep transformation feature classification layer performs feature classification on the fused shared features to obtain the deep transformation distribution vector of the target object for the target business data.

[0056] The multilayer perception network layer includes a second trigger fully connected layer, a second shallow transformation fully connected layer, a shallow connection layer, a second deep transformation fully connected layer, and a deep connection layer; the initial service transformation features include initial shallow transformation features and initial deep transformation features; the target service transformation features include target shallow transformation features and target deep transformation features.

[0057] The feature transfer unit is specifically used to input the initial business trigger feature to the second trigger fully connected layer, and to perform fully connected processing on the initial business trigger feature through the second trigger fully connected layer to obtain the target business trigger feature of the target object for the target business data.

[0058] The feature transfer unit is specifically used to input the target business trigger feature and the initial shallow transformation feature into the shallow connection layer. The shallow connection layer performs feature concatenation on the target business trigger feature and the initial shallow transformation feature to obtain the shared shallow concatenation feature.

[0059] The feature transfer unit is specifically used to input the shared shallow splicing features into the second shallow transformation fully connected layer, and to perform fully connected processing on the shared shallow splicing features through the second shallow transformation fully connected layer to obtain the target shallow transformation features of the target object for the target business data.

[0060] The feature transfer unit is specifically used to input the target shallow transformation features and the initial deep transformation features into the deep connection layer. The deep connection layer performs feature concatenation on the target shallow transformation features and the initial deep transformation features to obtain shared deep concatenation features.

[0061] The feature transfer unit is specifically used to input the shared deep concatenation features into the second deep transformation fully connected layer. The shared deep concatenation features are then fully connected through the second deep transformation fully connected layer to obtain the target deep transformation features of the target object for the target business data.

[0062] The sorting subnetwork also includes a feature connection layer and a dot product fully connected layer;

[0063] The first prediction module includes:

[0064] The first splicing unit is used to input the matching features and the target business triggering features into the feature connection layer. The feature connection layer splices the matching features and the target business triggering features to obtain the initial splicing triggering features of the target object for the target business data.

[0065] The first processing unit is used to input the initial splicing trigger features into the dot-product fully connected layer, and to perform fully connected processing on the initial splicing trigger features through the dot-product fully connected layer to obtain the target splicing trigger features;

[0066] The first estimation unit is used to determine the estimated probability of the target object triggering the target business data based on the target splicing trigger characteristics.

[0067] The ranking subnetwork also includes a feature connection layer and a dot product fully connected layer; the conversion probability prediction includes shallow conversion probability prediction and deep conversion probability prediction; the target business conversion features include target shallow conversion features and target deep conversion features.

[0068] The second prediction module includes:

[0069] The second splicing unit is used to input the matching features and the target shallow transformation features into the feature connection layer. The feature connection layer splices the matching features and the target shallow transformation features to obtain the initial spliced ​​shallow transformation features of the target object for the target business data.

[0070] The second processing unit is used to input the initial splicing shallow transformation features into a dot-product fully connected layer, and to perform fully connected processing on the initial splicing shallow transformation features through the dot-product fully connected layer to obtain the target splicing shallow transformation features.

[0071] The second estimation unit is used to determine the shallow conversion probability estimate of the target object for the target business data based on the shallow conversion characteristics of the target splicing.

[0072] The third splicing unit is used to input the matching features and the target deep transformation features into the feature connection layer. The feature connection layer splices the matching features and the target deep transformation features to obtain the initial spliced ​​deep transformation features of the target object for the target business data.

[0073] The third processing unit is used to input the initial deep transformation features of the splicing into the dot-product fully connected layer, and to perform fully connected processing on the initial deep transformation features of the splicing through the dot-product fully connected layer to obtain the target deep transformation features of the splicing.

[0074] The third estimation unit is used to determine the estimated probability of deep conversion of the target object for the target business data based on the deep conversion characteristics of the target splicing.

[0075] One embodiment of this application provides a data processing method, including:

[0076] Obtain sample business data associated with the sample object, and determine the sample matching features between the sample object and the sample business data through an initial probability prediction network model;

[0077] In the initial probability prediction network model, the sample service triggering features and sample service conversion features of the sample objects for the sample service data are determined; the sample service conversion features are obtained by performing feature sharing and transfer processing on the sample service triggering features.

[0078] The sample matching features and sample business triggering features are concatenated to obtain the sample concatenated triggering features of the sample object in response to the sample business data. Based on the sample concatenated triggering features, the sample triggering probability prediction of the sample object in response to the sample business data is determined.

[0079] The sample matching features and sample business conversion features are concatenated to obtain the sample concatenated conversion features of the sample object for the sample business data. Based on the sample concatenated conversion features, the sample conversion probability prediction value of the sample object for the sample business data is determined.

[0080] Based on the sample label information, sample trigger probability prediction, and sample conversion probability prediction of the sample object for the sample business data, the parameters of the initial probability prediction network model are adjusted, and the parameter-adjusted initial probability prediction network model is used as the probability prediction network model; the probability prediction network model is used to predict the trigger probability prediction and conversion probability prediction of the target object for the target business data.

[0081] One embodiment of this application provides a data processing apparatus, including:

[0082] The first sample determination module is used to acquire sample business data associated with the sample object and determine the sample matching features between the sample object and the sample business data through the initial probability prediction network model.

[0083] The second sample determination module is used in the initial probability prediction network model to determine the sample service triggering features and sample service transformation features of the sample objects for the sample service data; the sample service transformation features are obtained by performing feature sharing and transmission processing on the sample service triggering features.

[0084] The first sample prediction module is used to concatenate the sample matching features and the sample business triggering features to obtain the sample concatenation triggering features of the sample object in relation to the sample business data. Based on the sample concatenation triggering features, the sample triggering probability prediction value of the sample object in relation to the sample business data is determined.

[0085] The second sample prediction module is used to concatenate the sample matching features and sample business conversion features to obtain the sample concatenated conversion features of the sample object for the sample business data. Based on the sample concatenated conversion features, the sample conversion probability prediction value of the sample object for the sample business data is determined.

[0086] The parameter adjustment module is used to adjust the parameters of the initial probability prediction network model based on the sample label information, sample trigger probability prediction value, and sample conversion probability prediction value of the sample object for the sample business data. The parameter-adjusted initial probability prediction network model is used as the probability prediction network model. The probability prediction network model is used to predict the trigger probability prediction value and conversion probability prediction value of the target object for the target business data.

[0087] One embodiment of this application provides a computer device, including: a processor and a memory;

[0088] The processor is connected to a memory, which stores a computer program. When the computer program is executed by the processor, it causes the computer device to perform the method provided in the embodiments of this application.

[0089] One aspect of this application provides a computer-readable storage medium storing a computer program adapted to be loaded and executed by a processor, so that a computer device having the processor performs the method provided in this application.

[0090] One embodiment of this application provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method provided in this application embodiment.

[0091] This application embodiment can determine the matching features between the target object and the target business data after acquiring the target business data associated with the target object, and then determine the target business triggering features and target business conversion features of the target object for the target business data. The target business conversion features are obtained by performing feature sharing and transfer processing on the target business triggering features. It can be understood that by concatenating the matching features and the target business triggering features, an initial concatenated triggering feature of the target object for the target business data can be obtained. This initial concatenated triggering feature can be used to generate a predicted trigger probability value for the target object for the business data. Similarly, by concatenating the matching features and the target business conversion features, an initial concatenated conversion feature of the target object for the target business data can be obtained. This initial concatenated conversion feature can be used to generate a predicted conversion probability value for the target object for the business data. Based on this, by performing feature sharing and transfer processing on the target business triggering features, the target business triggering features corresponding to the business triggering task can be transferred to the target business conversion features corresponding to the business conversion task. This is equivalent to sharing the underlying parameters corresponding to the business triggering task with the business conversion task, thereby generating target business conversion features with higher accuracy. Furthermore, embodiments of this application can incorporate the matching features between the target object and the target business data into the click probability prediction and conversion probability prediction, making full use of the correlation information between the target object and the target business data, while generating the trigger probability prediction and conversion probability prediction, and improving the accuracy of the trigger probability prediction and conversion probability prediction of the target business data. Attached Figure Description

[0092] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0093] Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application;

[0094] Figure 2 This is a schematic diagram of a data interaction scenario provided in an embodiment of this application;

[0095] Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0096] Figure 4 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0097] Figure 5 This is a schematic diagram of the structure of a feature recognition network model provided in an embodiment of this application;

[0098] Figure 6 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0099] Figure 7 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0100] Figure 8a This is a schematic diagram of the structure of a probability prediction network model provided in an embodiment of this application;

[0101] Figure 8b This is a schematic diagram of the structure of a probability prediction network model provided in an embodiment of this application;

[0102] Figure 9 This is a schematic diagram of a business data recommendation process provided in an embodiment of this application;

[0103] Figure 10 This is a flowchart illustrating a data processing method provided in an embodiment of this application;

[0104] Figure 11 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0105] Figure 12 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application;

[0106] Figure 13This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0107] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0108] It should be understood that Artificial Intelligence (AI) is the theory, methods, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0109] Artificial intelligence (AI) is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0110] The solutions provided in this application primarily relate to Machine Learning (ML) technology in artificial intelligence. Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and many other disciplines. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0111] For details, please see Figure 1 , Figure 1 This is a schematic diagram of a network architecture provided in an embodiment of this application. Figure 1As shown, this network architecture may include a service server 2000 and a terminal device cluster. The terminal device cluster may specifically include one or more terminal devices; the number of terminal devices in the cluster is not limited here. Figure 1 As shown, the multiple terminal devices may specifically include terminal device 3000a, terminal device 3000b, terminal device 3000c, ..., terminal device 3000n; terminal device 3000a, terminal device 3000b, terminal device 3000c, ..., terminal device 3000n can be directly or indirectly connected to the business server 2000 via wired or wireless communication, so that each terminal device can interact with the business server 2000 through the network connection.

[0112] Each terminal device in the terminal device cluster can include: smart TVs, smartphones, tablets, laptops, desktop computers, smart home devices, wearable devices, in-vehicle terminals, and other smart terminals with data processing capabilities. It should be understood that, for example... Figure 1 Each terminal device in the terminal device cluster shown can be integrated with an application client. When the application client runs on each terminal device, it can interact with the aforementioned... Figure 1 The business servers 2000 shown interact with each other. The application client can be a standalone client or an embedded sub-client integrated into another client; this application does not impose any limitations on this.

[0113] Specifically, the application clients can include browsers, in-vehicle clients, smart home clients, entertainment clients, multimedia clients (e.g., video clients), social clients, and information clients (e.g., news clients) with data processing capabilities. The in-vehicle terminal can be a smart terminal in a smart transportation scenario, and the application clients on this in-vehicle terminal can be the aforementioned in-vehicle clients.

[0114] Among them, the business server 2000 can be the server corresponding to the application client. The business server 2000 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms.

[0115] For ease of understanding, the embodiments of this application may be described in detail below. Figure 1 From the multiple terminal devices shown, one terminal device is selected as the target terminal device. For example, in the embodiments of this application, a terminal device can be selected as the target terminal device. Figure 1The terminal device 3000b shown serves as the target terminal device, which may integrate an application client with data processing capabilities. In this case, the target terminal device can interact with the business server 2000 through this application client.

[0116] For ease of understanding, in this application embodiment, the user corresponding to the target terminal device can be collectively referred to as the target object; in this application embodiment, the business data recommended by the business server 2000 to the target object can be collectively referred to as recommended business data; and in this application embodiment, the business data sent by the advertiser (i.e., a legal person, other economic organization, or individual who designs, produces, or publishes advertisements themselves or entrusts others to do so for the purpose of promoting goods or providing services) to the business server 2000 can be collectively referred to as business data to be recommended. In this application embodiment, the business data can be advertising data (i.e., advertisements).

[0117] Understandably, the target object can send a business data retrieval request to the business server 2000 through the application client in the target terminal device. Upon receiving the request, the business server 2000 can retrieve a set of business data associated with the target object, extract recommended business data from this set, and then send this recommended business data to the target object's corresponding application client. Therefore, after receiving the recommended business data from the business server 2000, the application client can display the recommended business data, allowing the target object to determine whether to perform a trigger operation based on the recommended business data.

[0118] In this context, ad impressions represent the products (i.e., recommendation data) that a target audience observes in the application client. The triggering actions performed by the target audience in response to this recommendation data can be categorized into three types: The first type is ad clicks, which represent the target audience's click behavior on the recommendation data; the second type is shallow ad conversions, which represent the shallow conversion behavior of the target audience after clicking on the recommendation data, such as downloading, activating, or registering; and the third type is deep ad conversions, which represent the deep conversion behavior of the target audience after clicking on the recommendation data, such as paying for services or maintaining user retention the following day.

[0119] Among them, shallow conversion behavior is based on click behavior, that is, click behavior comes first and then shallow conversion behavior; deep conversion behavior is based on shallow conversion behavior, that is, shallow conversion behavior comes first and then deep conversion behavior.

[0120] It should be understood that the embodiments of this application can measure the behaviors corresponding to the above three types of triggering operations using different metrics. CTR (Click Through Rate) represents the ratio from ad impression to ad click, i.e., ad clicks divided by impressions; CVR (Conversion Rate) represents the ratio from ad clicks to ad conversions, i.e., ad conversions divided by clicks. CVR can be divided into CVR1 (shallow conversion rate) and CVR2 (deep conversion rate). CVR1 represents the ratio from ad clicks to shallow ad conversions, i.e., shallow conversions divided by clicks; CVR2 represents the ratio from ad clicks to deep ad conversions, i.e., deep conversions divided by clicks.

[0121] It should be understood that when advertisers send data for recommended services, they need to set the corresponding business objectives. These business objectives can be divided into shallow-target ads and deep-target ads. Shallow-target ads refer to ads where the bidding objective includes a shallow-target bid, i.e., a single bid for shallow metrics. This shallow-target bid can represent a bid for shallow conversion behaviors (i.e., shallow metrics, such as downloads, activations, and registrations). Shallow-target ads require estimated click-through rates and shallow conversion rates. Deep-target ads refer to ads where the bidding objective includes both shallow-target bids and deep-target bids, i.e., a bid that involves bidding again for deep metrics based on a single bid. This deep-target bid can represent a bid for deep conversion behaviors (i.e., deep metrics, such as payments and next-day retention). Deep-target ads require estimated click-through rates, shallow conversion rates, and deep conversion rates. Furthermore, in this application's embodiments, shallow-target ads can be referred to as single-target ads, and deep-target ads as multi-target ads.

[0122] In this application, the Bayesian network is a probabilistic graphical model, also known as a directed acyclic graph model. A Bayesian network is a graphical network model used to describe uncertain causal relationships between variables. It consists of nodes, directed connections, and a node probability table, where the directed connections represent causal dependencies between nodes.

[0123] It is understood that the aforementioned network framework can be applied to the field of advertising recommendation. Specific business scenarios in this field may include: advertising distribution scenarios, advertising search scenarios, and advertising viewing scenarios, etc., which will not be listed here. It should be understood that this application does not limit the format in which business data is displayed.

[0124] For example, in an advertising distribution scenario, a computer device (e.g., the aforementioned business server 2000) can analyze the user profile of the target audience to determine recommended business data (e.g., business data G1) that the target audience may be interested in. Then, when the target audience opens the application client, the distributed content and business data G1 are pushed to the application client. Here, the distributed content can be video data, text and image data, or audio data; there are no specific limitations.

[0125] For example, in an advertising search scenario, a computer device (e.g., the aforementioned business server 2000) can analyze the search text entered by the target audience in the application client, determine recommended business data (e.g., business data G2) that the target audience may be interested in, and then push the search content corresponding to the search text and business data G2 to the application client. Here, the search content can be video data, image and text data, or audio data; there are no specific limitations.

[0126] For example, in an advertising viewing scenario, a computer device (e.g., the aforementioned business server 2000) can analyze the contextual content viewed by the target object in the application client, determine the recommended business data (e.g., business data G3) that the target object may be interested in, and then push the business data G3 to the application client.

[0127] It is understood that in the specific implementation of this application, data related to object attributes, context attributes, etc. are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0128] For better understanding, please refer to [link / reference]. Figure 2 , Figure 2 This is a schematic diagram illustrating a data interaction scenario provided in an embodiment of this application. For example... Figure 2 The server 20a shown can be the one described above. Figure 1 The corresponding business server 2000 in the embodiment, such as Figure 2 The terminal device 20b shown can be the one described above. Figure 1 The target terminal device in the corresponding embodiment. The terminal device 20b has an application client installed, which can be used to display recommended business data associated with the object 20c corresponding to the terminal device 20b.

[0129] Understandably, object 20c can send a business data retrieval request to server 20a through an application client in terminal device 20b. This allows it to view multimedia data (e.g., video data, text and image data) tailored to its interests, along with recommended business data matching its preferences, within the application client. For example, when object 20c performs a search operation on the application client, the application client can send a business data retrieval request to server 20a, so that server 20a returns recommended business data associated with the search results along with the search results. Similarly, when object 20c performs an open operation on the application client, the application client can send a business data retrieval request to server 20a, so that server 20a returns recommended business data associated with the distributed content along with the distributed content. Likewise, when object 20c performs a news feed viewing operation on the application client, the application client can send a business data retrieval request to server 20a, so that server 20a returns recommended business data associated with the news feed content along with the news feed content.

[0130] For ease of understanding, this application embodiment uses the example of a business data acquisition request being triggered by a search operation. The specific processes of opening operations, information flow viewing operations, etc., can be found in the description of the search operation, and will not be repeated here.

[0131] like Figure 2 As shown, server 20a can receive business data retrieval requests sent by application clients, and then retrieve a set of business data associated with object 20c (e.g., business data set 21b) from business database 21a. Business database 21a can include multiple databases, specifically databases 22a, 22b, ..., 22n. Databases 22a, 22b, ..., 22n can be used to store different types of business data. For example, database 22a can be used to store business data related to animals, database 22b can be used to store business data related to plants, ..., database 22n can be used to store business data related to space.

[0132] When the search text for object 20c is "cat," server 20a can retrieve a set of business data 21b associated with object 20c from database 22a. This means server 20a can construct an object vector for object 20c and a business vector for the business data in database 22a, and then use methods such as cosine similarity or Euclidean distance to determine the vector similarity between the object vector and the business vector. This allows server 20a to obtain business data whose vector similarity meets a threshold, and the business data retrieved from database 22a constitutes the business data set 21b. The business data set 21b can include multiple sets of business data, specifically business data 23a, business data 23b, ..., business data 23m.

[0133] Optionally, it should be understood that server 20a can also simultaneously obtain business data from databases 22a, 22b, ..., 22n to form business data set 21b. Here, we will take the example of business data 23a, business data 23b, ..., business data 23m being obtained from database 22a for illustration.

[0134] like Figure 2 As shown, after obtaining the business data set 21b, server 20a can sequentially retrieve target business data from the business data set 21b. Here, we will use business data 23a as an example to illustrate the process. Server 20a can determine the matching features between object 20c and business data 23a, and then determine the target business trigger features and target business conversion features of object 20c for business data 23a. Specifically, the target business trigger features correspond to the features of object 20c for click behavior, and the target business conversion features correspond to the features of object 20c for conversion behavior. Furthermore, the target business conversion features are obtained by performing feature sharing and transfer processing on the target business trigger features.

[0135] Furthermore, server 20a can concatenate the matching features and target business trigger features to obtain the initial concatenated trigger features of object 20c for business data 23a, and then determine the estimated trigger probability of object 20c for business data 23a based on the initial concatenated trigger features. Simultaneously, server 20a can concatenate the matching features and target business conversion features to obtain the initial concatenated conversion features of object 20c for business data 23a, and then determine the estimated conversion probability of object 20c for business data 23a based on the initial concatenated conversion features. Here, the estimated trigger probability corresponds to the estimated value of object 20c for click behavior, and the estimated conversion probability corresponds to the estimated value of object 20c for conversion behavior.

[0136] like Figure 2As shown, after determining the estimated trigger probability and the estimated conversion probability for each business data in the business data set 21b, server 20a sorts each business data in the business data set 21b based on the estimated trigger probability and the estimated conversion probability for each business data, obtains recommended business data from each sorted business data, uses the recommended business data as the business data that matches object 20c, and returns the recommended business data to the application client.

[0137] Therefore, the embodiments of this application can, after determining the matching features between the target object and the target business data, integrate the matching features into the target business triggering features corresponding to the business triggering task to obtain a trigger probability prediction with high accuracy, and integrate the matching features into the target business conversion features corresponding to the business conversion task to obtain a conversion probability prediction with high accuracy. Furthermore, by performing feature sharing and transfer processing on the target business triggering features, the accuracy of determining the target business conversion features can be improved, and thus, based on the target business triggering features and target business conversion features, the accuracy of predicting the trigger probability prediction and conversion probability prediction of the target business data can be improved.

[0138] Further, please see Figure 3 , Figure 3 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The method can be executed by a server, by an application client, or by both a server and an application client. The server can be one of the aforementioned... Figure 2 The server 20a in the corresponding embodiment can be the application client described above. Figure 2 The corresponding application client in the embodiments. For ease of understanding, this application embodiment uses the method executed by the server as an example for illustration. The data processing method may include the following steps S101-S104:

[0139] Step S101: Obtain target business data associated with the target object and determine the matching features between the target object and the target business data;

[0140] Specifically, the server can acquire a set of business data associated with the target object, and then extract the target business data from this set. Further, the server can acquire pre-trained object features associated with the target object, and pre-trained business features associated with the business data sequence. The business data sequence is associated with the target object. Further, the server can input the object attributes of the target object, the business attributes of the target business data, the pre-trained object features, and the pre-trained business features into the matching sub-network of the probability prediction network model. The matching sub-network includes a matching embedding layer. Further, the server can embed features of the object attributes and business attributes through the matching embedding layer to obtain object features (i.e., user-side features) corresponding to the object attributes and business features (i.e., advertising-side features, item features) corresponding to the business attributes. Further, in the matching sub-network, the server can determine the object feature vector associated with the target object based on the object features and the pre-trained object features, and determine the business feature vector associated with the target business data based on the business features and the pre-trained business features. Further, the server can perform a dot product operation on the object feature vector and the business feature vector to obtain the matching features between the target object and the target business data.

[0141] Understandably, a server can use a recall model to retrieve a set of business data associated with a target object. This recall model can identify as much relevant business data as possible from massive amounts of business data. The recall model can comprehensively consider the object attributes of the target object, the business attributes of the business data, the contextual attributes associated with the target object, and the search terms associated with the target object. The search terms can represent the search text entered by the target object when performing a search operation in the application client.

[0142] Recall models can be broadly categorized into two types. The first type uses keyword-based hard matching, such as gender (male, female) in business data targeting, which divides business data into two categories: one for males and the other for females. Another example is search terms in search text (e.g., "cat"), which retrieves business data associated with cats. The second type uses vector-based soft matching, constructing object vectors for the target object and business vectors for the business data, and then using methods such as cosine similarity or Euclidean distance to determine the vector similarity between the object vector and the business vector.

[0143] It is understandable that object attributes can be used to represent target objects, and business attributes can be used to represent business data. Among them, object attributes can include, but are not limited to, the object identifier, basic attributes, and behavioral attributes (i.e., behavioral interests) of the target object; business attributes can include, but are not limited to, the business identifier (i.e., advertisement identifier), the business data object identifier (i.e., advertiser identifier), the business data category (i.e., advertisement category, for example, application client advertisement), the business data text (i.e., semantic features), and the business data image (i.e., image features).

[0144] Among these, the object identifier of the target object is unique, with one target object corresponding to one object identifier; the business identifier of the business data is unique, with one piece of business data corresponding to one business identifier. It should be understood that the target object does not necessarily have all object attributes, and the business data does not necessarily have all business attributes.

[0145] The matching sub-network comprises an object fully connected layer, a business fully connected layer, and a feature averaging layer. The business data sequence includes one or more business data points, each corresponding to a pre-trained business feature. The specific process by which the server determines the object feature vector and business feature vector in the matching sub-network can be described as follows: The server can input object features into the object fully connected layer in the matching sub-network, perform fully connected processing on the object features, and obtain fully connected object features. Further, the server can perform feature fusion on the fully connected object features and the pre-trained object features to obtain an object feature vector associated with the target object. Further, the server can input business features into the business fully connected layer, perform fully connected processing on the business features, and obtain fully connected business features. Further, the server can input one or more pre-trained business features into the feature averaging layer, perform average processing on the one or more pre-trained business features, and obtain averaged pre-trained business features. Further, the server can perform feature fusion on the fully connected business features and the averaged pre-trained business features to obtain a business feature vector associated with the target business data.

[0146] The probability prediction network model is obtained by iteratively training the initial probability prediction network model. The specific process of the server iteratively training the initial probability prediction network model to obtain the probability prediction network model can be found below. Figure 10 The description of steps S301-S305 in the corresponding embodiment.

[0147] It should be understood that the probability prediction network model also includes a ranking subnetwork, which includes an input network layer, a parameter sharing network layer, and a multilayer sensing network layer. The input network layer, parameter sharing network layer, and multilayer sensing network layer can be used to perform the following step S102.

[0148] Step S102: Determine the target business triggering characteristics and target business conversion characteristics of the target object for the target business data;

[0149] Specifically, the server can determine the shared attribute characteristics of the target object for the target business data by inputting the network layer. Further, the server can input these shared attribute characteristics into the parameter sharing network layer, which generates the initial business trigger characteristics and initial business transformation characteristics of the target object for the target business data. Further, the server can input these initial business trigger characteristics and initial business transformation characteristics into the multilayer perceptron network layer. The multilayer perceptron network layer performs fully connected processing on the initial business trigger characteristics to obtain the target business trigger characteristics of the target object for the target business data. In the multilayer perceptron network layer, feature sharing and transfer processing is performed on the target business trigger characteristics and initial business transformation characteristics to obtain the target business transformation characteristics of the target object for the target business data.

[0150] It should be understood that the sorting subnetwork also includes feature connection layers and dot product fully connected layers, which can be used to perform the following step S103.

[0151] Step S103: The matching features and the target business trigger features are concatenated to obtain the initial concatenated trigger features of the target object for the target business data. Based on the initial concatenated trigger features, the estimated trigger probability of the target object for the target business data is determined.

[0152] Specifically, the server can input matching features and target business trigger features into a feature connection layer. The feature connection layer concatenates the matching features and target business trigger features to obtain the initial concatenated trigger features of the target object for the target business data. Further, the server can input the initial concatenated trigger features into a dot-product fully connected layer. The dot-product fully connected layer performs fully connected processing on the initial concatenated trigger features to obtain the target concatenated trigger features. Furthermore, the server can determine the estimated trigger probability of the target object for the target business data based on the target concatenated trigger features.

[0153] The target splicing trigger feature is composed of a first trigger feature value and a second trigger feature value. The first trigger feature value represents the probability that the target business data has a click behavior, and the second trigger feature value represents the probability that the target business data does not have a click behavior. Therefore, the server can use the first trigger feature value as an estimated value of the trigger probability of the target object for the target business data.

[0154] Step S104: The matching features and target business conversion features are concatenated to obtain the initial concatenated conversion features of the target object for the target business data. Based on the initial concatenated conversion features, the conversion probability estimate of the target object for the target business data is determined.

[0155] Specifically, the server can input matching features and target business conversion features into a feature connection layer. The feature connection layer concatenates these features to obtain the initial concatenated conversion features of the target object for the target business data. Further, the server can input these initial concatenated conversion features into a dot-product fully connected layer. This layer performs a fully connected processing on the initial concatenated conversion features to obtain the target concatenated conversion features. Finally, the server can determine the estimated conversion probability of the target object for the target business data based on these target concatenated conversion features.

[0156] The target concatenation conversion feature is composed of a first conversion feature value and a second conversion feature value. The first conversion feature value represents the probability that the target business data will undergo conversion behavior, and the second conversion feature value represents the probability that the target business data will not undergo conversion behavior. Therefore, the server can use the first conversion feature value as an estimated value of the conversion probability of the target object for the target business data.

[0157] Therefore, this application embodiment can determine the matching features between the target object and the target business data after obtaining the target business data associated with the target object, and then determine the target business triggering features and target business conversion features of the target object for the target business data. The target business conversion features are obtained by performing feature sharing and transfer processing on the target business triggering features. It can be understood that by concatenating the matching features and the target business triggering features, the initial concatenated triggering features of the target object for the target business data can be obtained. These initial concatenated triggering features can be used to generate a predicted trigger probability value for the target object for the business data. Similarly, by concatenating the matching features and the target business conversion features, the initial concatenated conversion features of the target object for the target business data can be obtained. These initial concatenated conversion features can be used to generate a predicted conversion probability value for the target object for the business data. Based on this, by performing feature sharing and transfer processing on the target business triggering features, the target business triggering features corresponding to the business triggering task can be transferred to the target business conversion features corresponding to the business conversion task. This is equivalent to sharing the underlying parameters corresponding to the business triggering task with the business conversion task, thereby generating target business conversion features with higher accuracy. Furthermore, embodiments of this application can incorporate the matching features between the target object and the target business data into the click probability prediction and conversion probability prediction, making full use of the correlation information between the target object and the target business data, while generating the trigger probability prediction and conversion probability prediction, and improving the accuracy of the trigger probability prediction and conversion probability prediction of the target business data.

[0158] Further, please see Figure 4 , Figure 4 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The data processing method may include the following steps S1011-S1019, and steps S1011-S1019 are... Figure 3 A specific embodiment of step S101 in the corresponding embodiment.

[0159] Step S1011: Obtain the business data set associated with the target object, and retrieve the target business data from the business data set;

[0160] The specific process by which the server obtains the set of business data associated with the target object can be found in the above. Figure 3 The description of step S101 in the corresponding embodiments will not be repeated here.

[0161] Step S1012: Input the object attributes of the target object, the business trigger sequence associated with the target object, and the business conversion sequence associated with the target object into the feature recognition network model;

[0162] It should be understood that business data sequences include business trigger sequences and business conversion sequences. Specifically, business trigger sequences include business data that triggers actions in the target object, while business conversion sequences include business data that enables conversion actions in the target object. These conversion actions can include shallow conversion actions and deep conversion actions. In other words, business trigger sequences and business conversion sequences can be collectively referred to as business data sequences.

[0163] Optionally, the business conversion sequence may include a shallow business conversion sequence and a deep business conversion sequence. The shallow business conversion sequence may include business data of the target object with shallow conversion behavior, and the deep business conversion sequence may include business data of the target object with deep conversion behavior.

[0164] It should be understood that the feature recognition network model includes a pre-trained embedding layer, an attention processing layer, and a pre-trained connection layer, which can be used to perform the following steps S1013-S1015.

[0165] Step S1013: The pre-trained embedding layer is used to embed features into the object attributes, business data in the business trigger sequence, and business data in the business conversion sequence, respectively, to obtain the object features corresponding to the object attributes, the trigger business features corresponding to the business data in the business trigger sequence, and the conversion business features corresponding to the business data in the business conversion sequence.

[0166] The object attributes may include, but are not limited to, object identifiers, basic attributes, and behavioral attributes. It should be understood that the server can perform a hash operation on the object identifier in the pre-trained embedding layer to obtain the hash object identifier corresponding to the object identifier. Further, the server can obtain an object identifier lookup table associated with the object identifier, and search for the object identifier feature corresponding to the object identifier in the object identifier lookup table based on the hash object identifier. Further, the server can obtain a basic attribute lookup table associated with the basic attributes, and search for the basic attribute feature corresponding to the basic attribute in the basic attribute lookup table. Further, the server can obtain a behavior attribute lookup table associated with the behavior attributes, and search for the behavior attribute feature corresponding to the behavior attribute in the behavior attribute lookup table. Further, the server can perform feature fusion on the object identifier feature, basic attribute feature, and behavior attribute feature to obtain the object feature corresponding to the object attribute.

[0167] It is understandable that the object identifier lookup table, basic attribute lookup table, and behavioral attribute lookup table can each include multiple target features (i.e., object identifier features, basic attribute features, and behavioral attribute features). The multiple target features in the object identifier lookup table are obtained by training on the initial features in the initial object identifier lookup table; the multiple target features in the basic attribute lookup table are obtained by training on the initial features in the initial basic attribute lookup table; and the multiple target features in the behavioral attribute lookup table are obtained by training on the initial features in the initial behavioral attribute lookup table. The initial object identifier lookup table, initial basic attribute lookup table, and initial behavioral attribute lookup table are the lookup tables in the initial probability prediction network model.

[0168] The basic attributes can include, but are not limited to, the gender and age of the target object. Therefore, the basic attribute lookup table can include an age lookup table corresponding to age and a gender lookup table corresponding to gender. In this way, the server can look up the age feature corresponding to age in the age lookup table and the gender feature corresponding to gender in the gender lookup table, and then perform feature fusion on the age feature and gender feature to obtain the basic attribute feature corresponding to the basic attribute.

[0169] The number of behavioral attributes can be one or more, where "multiple" can mean at least two. Understandably, when there is only one behavioral attribute (e.g., basketball), the server can look up the behavioral attribute feature corresponding to that attribute in the behavioral attribute lookup table (i.e., look up the behavioral attribute feature corresponding to basketball). Optionally, when there are multiple behavioral attributes (e.g., volleyball, vegetables), the server can look up the behavioral attribute features corresponding to each attribute in the behavioral attribute lookup table, and then perform feature fusion on these multiple behavioral attribute features (i.e., fuse the behavioral attribute features corresponding to volleyball and the behavioral attribute features corresponding to vegetables) to obtain the behavioral attribute feature corresponding to each attribute.

[0170] It is understood that the dimensions of object identification features, basic attribute features, and behavioral attribute features are the same. The server can perform feature fusion on object identification features, basic attribute features, and behavioral attribute features in the following ways: feature concatenation (e.g., concatenating three 128-dimensional features to obtain a 384-dimensional object feature), feature addition (e.g., adding the corresponding dimensions of three 128-dimensional features to obtain a 128-dimensional object feature), or weighted averaging (e.g., obtaining a 128-dimensional object feature based on the three 128-dimensional features and their respective weights). It should be understood that the embodiments of this application do not limit the dimensions of object identification features, basic attribute features, and behavioral attribute features, nor do they limit the specific method of feature fusion.

[0171] Step S1014: Input the object features, triggering business features and conversion business features into the attention processing layer. The attention processing layer performs attention processing on the object features, triggering business features and conversion business features to obtain the attention object features corresponding to the object features, the attention triggering features corresponding to the triggering business features and the attention conversion features corresponding to the conversion business features.

[0172] In this context, business data in the business trigger sequence can be referred to as trigger business data, and business data in the business conversion sequence can be referred to as conversion business data. Trigger business data corresponds to trigger business features and attention trigger features, while conversion business data corresponds to conversion business features and attention conversion features.

[0173] The specific process by which the server performs attention processing on object features through the attention processing layer can be found in the following formula (1):

[0174]

[0175] Among them, V user V can represent object characteristics. item It can represent triggering business characteristics and conversion business characteristics, g(V) user V item ) can represent the model parameters that give weights to different business data for each user, then V′ user This can represent the characteristics of the attention object, and N can represent the number of triggering business data and conversion business data. Where g(V) user V item The weights for attention processing of object features can be represented by ). Based on the weights of the triggering business data and the conversion business data on the object features, attention processing can be performed on the object features to obtain the attention object features. It should be understood that the formula for the server to perform attention processing on the triggering business features and the conversion business features through the attention processing layer can be found in the above formula (1), and will not be repeated here.

[0176] Step S1015: Input the object features, attention object features, attention trigger features and attention conversion features into the pre-trained connection layer. The object features and attention object features are fused through the pre-trained connection layer to obtain the pre-trained object features associated with the target object. In the pre-trained connection layer, the pre-trained business features corresponding to the business data contained in the business trigger sequence and business conversion sequence are generated according to the attention trigger features and attention conversion features.

[0177] The transformation business data includes triggering business data. The server can use business data other than triggering business data from the transformation business data as auxiliary business data. The triggering business data and auxiliary business data constitute the business data included in the business trigger sequence and business transformation sequence. It can be understood that the server can fuse the attention triggering features and attention transformation features corresponding to the triggering business data to generate pre-trained business features corresponding to the triggering business data; the server can also directly use the attention transformation features corresponding to the auxiliary business data as pre-trained business features corresponding to the auxiliary business data.

[0178] The feature recognition network model is obtained by iteratively training an initial feature recognition network model. The specific process of the server iteratively training the initial feature recognition network model to obtain the feature recognition network model can be described as follows: The server can acquire sample business data associated with the sample object, and determine the pre-trained sample object features and pre-trained sample business features associated with the sample business data through the initial feature recognition network model. Further, the server can generate sample quality for the sample object relative to the sample business data based on the pre-trained sample object features and pre-trained sample business features. Further, the server can adjust the parameters of the initial feature recognition network model based on the sample label information and sample quality of the sample object relative to the sample business data, and use the parameter-adjusted initial feature recognition network model as the feature recognition network model.

[0179] Here, the sample tag information indicates whether the target object has clicked or converted on the sample business data. If the target object has clicked or converted on the sample business data, the sample tag information can be 1; optionally, if the target object has not clicked or converted on the sample business data, the sample tag information can be 0.

[0180] For easier understanding, please refer to Figure 5 , Figure 5 This is a schematic diagram of the structure of a feature recognition network model provided in an embodiment of this application. For example... Figure 5 As shown, the server can input the object attributes of the target object, the trigger business data in the business trigger sequence (i.e., the ad click sequence), and the conversion business data in the business conversion sequence (i.e., the ad conversion sequence) into the pre-trained embedding layer in the feature recognition network model. The pre-trained embedding layer performs feature embedding on the object attributes, trigger business data, and conversion business data to obtain the object feature 51a corresponding to the object attributes, the trigger business feature (not shown in the figure) corresponding to the trigger business data, and the conversion business feature (not shown in the figure) corresponding to the conversion business data.

[0181] like Figure 5As shown, the server can input object feature 51a, triggering business feature, and conversion business feature into the attention processing layer. The attention processing layer performs attention processing on object feature 51a, triggering business feature, and conversion business feature, that is, it performs attention processing on object feature 51a, triggering business feature, and conversion business feature based on the attention mechanism, obtaining the attention object feature corresponding to object feature 51a, the attention triggering feature corresponding to the triggering business feature, and the attention conversion feature corresponding to the conversion business feature. Among these, the attention object feature, attention triggering feature, and attention conversion feature can be... Figure 5 Feature 51b shown includes multiple sets of attention object features, attention trigger features, and attention transformation features. These features are obtained by performing multiple attention processing steps.

[0182] like Figure 5 As shown, the server can input object feature 51a and attention object feature into the pre-trained connection layer. The pre-trained connection layer performs feature fusion on object feature 51a and attention object feature to obtain pre-trained object feature associated with the target object. Similarly, the server can input attention trigger feature and attention transformation feature into the pre-trained connection layer. In the pre-trained connection layer, based on the attention trigger feature and attention transformation feature, pre-trained business features corresponding to the business data contained in the business trigger sequence and business transformation sequence are generated.

[0183] Step S1016: Input the object attributes of the target object, the business attributes of the target business data, the pre-trained object features, and the pre-trained business features into the matching sub-network of the probability prediction network model;

[0184] The matching subnetwork includes a matching embedding layer.

[0185] Step S1017: By matching the embedding layer, feature embedding is performed on the object attributes and business attributes to obtain the object features corresponding to the object attributes and the business features corresponding to the business attributes.

[0186] The business attributes may include, but are not limited to, business identifiers, business data object identifiers, business data categories, business data text, and business data images. It should be understood that the server can perform a hash operation on the business identifier in the matching embedding layer to obtain the hash business identifier corresponding to the business identifier. Further, the server can obtain a business identifier lookup table associated with the business identifier, and search for the corresponding business identifier feature in the business identifier lookup table based on the hash business identifier. Further, the server can obtain a business data object identifier lookup table associated with the business data object identifier, and search for the corresponding business data object identifier feature in the business data object identifier lookup table. Further, the server can obtain a business data category lookup table associated with the business data category, and search for the corresponding business data category feature in the business data category lookup table. Further, the server can obtain a business data text lookup table associated with the business data text, and search for the corresponding business data text feature in the business data text lookup table. Further, the server can obtain a business data image lookup table associated with the business data image, and search for the corresponding business data image feature in the business data image lookup table. Furthermore, the server can perform feature fusion on business identification features, business data object identification features, business data category features, business data text features, and business data image features to obtain business features corresponding to business attributes.

[0187] It is understood that the business identifier lookup table, business data object identifier lookup table, business data category lookup table, business data text lookup table, and business data image lookup table can each include multiple target features (i.e., business identifier features, business data object identifier features, business data category features, business data text features, and business data image features), and these multiple target features are obtained after training the initial features. It should be understood that the specific process by which the server trains the initial features to obtain the target features can be found in the description of step S1013 above, and will not be repeated here.

[0188] It should be understood that the specific process of the server searching the business data category lookup table can be found in the description of searching the behavior attribute lookup table above, and will not be repeated here. Similarly, it should be understood that the specific process of the server searching the business data object identifier lookup table can be found in the description of searching the basic attribute lookup table above, and will not be repeated here.

[0189] Specifically, the server can perform text analysis on business data text to obtain the corresponding business data text key, and then look up the corresponding business data text feature in a business data text lookup table based on the business data text key. Similarly, the server can perform image analysis on business data images to obtain the corresponding business data image key, and then look up the corresponding business data image feature in a business data image lookup table based on the business data image key.

[0190] It should be understood that the specific process of the server performing feature fusion on business identification features, business data object identification features, business data category features, business data text features, and business data image features can be found in the above description of feature fusion on object identification features, basic attribute features, and behavioral attribute features, and will not be repeated here.

[0191] The specific process of the server embedding features of object attributes through the matching embedding layer can be found in the description of embedding features of object attributes through the pre-trained embedding layer above, and will not be repeated here.

[0192] Step S1018: In the matching sub-network, based on object features and pre-trained object features, determine the object feature vector associated with the target object, and based on business features and pre-trained business features, determine the business feature vector associated with the target business data.

[0193] The specific process by which the server determines the object feature vector and the business feature vector can be found in the above. Figure 3 The description of step S101 in the corresponding embodiments will not be repeated here.

[0194] Step S1019: Perform a dot product operation on the object feature vector and the business feature vector to obtain the matching features between the target object and the target business data.

[0195] Among them, the matching feature can represent the matching probability between the target object and the target business data.

[0196] Therefore, the embodiments of this application can generate matching features between a target object and target business data based on the object attributes of the target object, the business trigger sequence associated with the target object, and the business conversion sequence associated with the target object. It is understood that the matching features can be used to incorporate target business trigger features and target business conversion features to integrate the correlation information between the target object and target business data into the trigger probability estimate and conversion probability estimate, thereby improving the accuracy of the predicted trigger probability estimate and conversion probability estimate of the target business data.

[0197] Further, please see Figure 6, Figure 6 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The data processing method may include the following steps S1021-S1023, and steps S1021-S1023 are... Figure 3 A specific embodiment of step S102 in the corresponding embodiment.

[0198] Step S1021: Determine the shared attribute characteristics of the target object for the target business data by inputting the network layer;

[0199] Specifically, the server can input the object attributes of the target object, the business attributes of the target business data, and the context attributes associated with the target object into the feature embedding layer. The feature embedding layer embeds features from the object attributes, business attributes, and context attributes to obtain object features corresponding to the object attributes, business features corresponding to the business attributes, and context features corresponding to the context attributes. The input network layer includes a feature embedding layer and a feature concatenation layer. Further, the server can average one or more pre-trained business features to obtain averaged pre-trained business features. Further, the server can input the object features, pre-trained object features, business features, and averaged pre-trained business features into the feature concatenation layer. In this layer, the object features and pre-trained object features are concatenated to obtain concatenated object features, and the business features and averaged pre-trained business features are concatenated to obtain concatenated business features. Further, the server can concatenate the concatenated object features, concatenated business features, and context features to obtain shared attribute features of the target object for the target business data.

[0200] Optionally, the server can also directly concatenate object features, pre-trained object features, business features, average pre-trained business features, and context features to obtain the shared attribute features of the target object for the target business data.

[0201] The business data sequence includes one or more business data points, each corresponding to a pre-trained business feature. It is understood that the input network layer may also include a feature averaging layer. The server can input one or more pre-trained business features into the feature averaging layer, which then averages these features to obtain the averaged pre-trained business feature.

[0202] The context attributes may include, but are not limited to, device content and device type. It should be understood that the server can obtain a device content lookup table associated with the device content in the feature embedding layer, and search for the device content feature corresponding to the device content in the device content lookup table. Furthermore, the server can obtain a device type lookup table associated with the device type, and search for the device type feature corresponding to the device type in the device type lookup table. Furthermore, the server can perform feature fusion on the device content feature and the device type feature to obtain the context features corresponding to the context attributes.

[0203] It is understood that the device content lookup table and the device type lookup table can each include multiple target features (i.e., device content features and device type features), which are obtained after training the initial features. It should be understood that the specific process by which the server trains the initial features to obtain the target features can be found in the description of step S1013 above, and will not be repeated here.

[0204] It should be understood that the specific process of the server looking up the device content lookup table can be found in the description of looking up the business data text lookup table above, and will not be repeated here. Similarly, it should be understood that the specific process of the server looking up the device content type lookup table can be found in the description of looking up the basic attribute lookup table above, and will not be repeated here.

[0205] The specific process of the server embedding features of object attributes through the feature embedding layer can be found in the description of embedding features of object attributes through the pre-trained embedding layer above, and will not be repeated here. The specific process of the server embedding features of business attributes through the feature embedding layer can also be found in the description of embedding features of business attributes through the pre-trained embedding layer above, and will not be repeated here.

[0206] The dimensions of object features, business features, context features, pre-trained object features, and average pre-trained business features can be N-dimensional, where N can be a positive integer. Therefore, the dimension of the shared attribute features is 5N (i.e., 5*N). It should be understood that this application does not limit the value of N.

[0207] It should be understood that before training the initial probability prediction network model, the server can initialize the initial features in the lookup table (i.e., the embedding table). Then, during the training of the initial probability prediction network model, the server can update the initial features in the lookup table based on the backpropagation gradient to obtain a lookup table containing the target features. Furthermore, for high-dimensional features (e.g., business identifier features), the server can map the high-dimensional features to low-dimensional features before storing the low-dimensional features in the lookup table.

[0208] Step S1022: Input the shared attribute features into the parameter sharing network layer, and generate the initial business trigger features and initial business conversion features of the target object for the target business data through the parameter sharing network layer;

[0209] Specifically, the server can input shared attribute features into a shared fully connected layer, which performs fully connected processing on the shared attribute features to obtain fully connected shared features. The parameter sharing network layer includes a shared fully connected layer, a weight learning layer, a feature classification layer, a first trigger fully connected layer, and an initial transformation fully connected layer. Further, the server can input the fully connected shared features into a weight learning layer, which performs feature weighting on the fully connected shared features to obtain weighted shared features. Further, the server can input the weighted shared features into a feature classification layer, which performs feature classification on the weighted shared features to obtain the trigger distribution vector and transformation distribution vector of the target object for the target business data. Further, the server can input the trigger distribution vector into the first trigger fully connected layer, which performs fully connected processing on the trigger distribution vector to generate the initial business trigger features of the target object for the target business data. Further, the server can input the transformation distribution vector into the initial transformation fully connected layer, which performs fully connected processing on the transformation distribution vector to generate the initial business transformation features of the target object for the target business data.

[0210] The weight learning layer comprises weight learning components (gates) corresponding to object features, business features, context features, pre-trained object features, and average pre-trained business features, respectively. The feature classification layer includes a trigger feature classification layer and a transformation feature classification layer. Therefore, the server can input the fully connected shared features into each weight learning component, and each weight learning component performs feature weighting on the fully connected shared features to obtain the weighted shared features corresponding to each weight learning component. Further, the server can perform feature fusion on the weighted shared features corresponding to each weight learning component to obtain fused shared features. Further, the server can input the fused shared features into the trigger feature classification layer, and perform feature classification on the fused shared features to obtain the trigger distribution vector of the target object for the target business data. Further, the server can input the fused shared features into the transformation feature classification layer, and perform feature classification on the fused shared features to obtain the transformation distribution vector of the target object for the target business data.

[0211] Step S1023: Input the initial service trigger feature and the initial service conversion feature into the multilayer perception network layer. The initial service trigger feature is processed by the multilayer perception network layer through full connection to obtain the target service trigger feature of the target object for the target service data. In the multilayer perception network layer, the target service trigger feature and the initial service conversion feature are processed by feature sharing to obtain the target service conversion feature of the target object for the target service data.

[0212] Specifically, the server can input the initial business trigger features into the second trigger fully connected layer. The second trigger fully connected layer performs fully connected processing on the initial business trigger features to obtain the target business trigger features of the target object for the target business data. The multilayer perceptron layer includes the second trigger fully connected layer, the target transformation fully connected layer, and the transformation connection layer. Further, the server can input the target business trigger features and the initial business transformation features into the transformation connection layer. The transformation connection layer performs feature concatenation on the target business trigger features and the initial business transformation features to obtain shared concatenated features. Further, the server can input the shared concatenated features into the target transformation fully connected layer. The target transformation fully connected layer performs fully connected processing on the shared concatenated features to obtain the target business transformation features of the target object for the target business data.

[0213] Therefore, the embodiments of this application can determine the shared attribute characteristics of the target object for the target business data through a sorting subnetwork, and then generate the target business trigger characteristics and target business conversion characteristics of the target object for the target business data based on the shared attribute characteristics. It is understood that the target business conversion characteristics are obtained by sharing and transmitting the target business trigger characteristics. When generating trigger probability estimates through the target business trigger characteristics and conversion probability estimates through the target business conversion characteristics, the accuracy of predicting the trigger probability estimates and conversion probability estimates of the target business data can be improved.

[0214] Further, please see Figure 7 , Figure 7 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The method can be executed by a server, by an application client, or by both a server and an application client. The server can be one of the aforementioned... Figure 2 The server 20a in the corresponding embodiment can be the application client described above. Figure 2 The corresponding application client in the embodiments. For ease of understanding, this application embodiment uses the method executed by the server as an example for illustration. The data processing method may include the following steps S201-S207:

[0215] Step S201: Obtain target business data associated with the target object, and determine the matching features between the target object and the target business data;

[0216] The specific process by which the server determines the matching characteristics between the target object and the target business data can be found in the above. Figure 3 The description of step S101 in the corresponding embodiments will not be repeated here.

[0217] Step S202: Determine the shared attribute characteristics of the target object for the target business data by inputting the network layer;

[0218] The specific process by which the server determines shared attribute characteristics can be found in the above. Figure 6 The description of step S1021 in the corresponding embodiments will not be repeated here.

[0219] Step S203: Input the shared attribute features into the parameter sharing network layer, and generate the initial business trigger features and initial business conversion features of the target object for the target business data through the parameter sharing network layer;

[0220] Specifically, the server can input shared attribute features into a shared fully connected layer, which performs fully connected processing on the shared attribute features to obtain fully connected shared features. The parameter sharing network layer includes a shared fully connected layer, a weight learning layer, a feature classification layer, a first trigger fully connected layer, a first shallow transformation fully connected layer, and a first deep transformation fully connected layer; the initial business transformation features include initial shallow transformation features and initial deep transformation features. Further, the server can input the fully connected shared features into a weight learning layer, which performs feature weighting on the fully connected shared features to obtain weighted shared features. Further, the server can input the weighted shared features into a feature classification layer, which performs feature classification on the weighted shared features to obtain the trigger distribution vector, shallow transformation distribution vector, and deep transformation distribution vector for the target object's target business data. Further, the server can input the trigger distribution vector into the first trigger fully connected layer, which performs fully connected processing on the trigger distribution vector to generate the initial business trigger features for the target object's target business data. Furthermore, the server can input the shallow transformation distribution vector into a first shallow transformation fully connected layer, and perform fully connected processing on the shallow transformation distribution vector through the first shallow transformation fully connected layer to generate the initial shallow transformation features of the target object for the target business data. Furthermore, the server can input the deep transformation distribution vector into a first deep transformation fully connected layer, and perform fully connected processing on the deep transformation distribution vector through the first deep transformation fully connected layer to generate the initial deep transformation features of the target object for the target business data.

[0221] The system comprises three weight learning layers: a weight learning layer (including object features, business features, context features, pre-trained object features, and average pre-trained business features), and a feature classification layer (including a trigger feature classification layer, a shallow transformation feature classification layer, and a deep transformation feature classification layer). The server can input the fully connected shared features into each weight learning component, and each weight learning component can perform feature weighting on the fully connected shared features to obtain the weighted shared features corresponding to each weight learning component. Furthermore, the server can perform feature fusion on the weighted shared features corresponding to each weight learning component to obtain fused shared features. Further, the server can input the fused shared features into the trigger feature classification layer, and the trigger feature classification layer can classify the fused shared features to obtain the trigger distribution vector of the target object for the target business data. Finally, the server can input the fused shared features into the shallow transformation feature classification layer, and the shallow transformation feature classification layer can classify the fused shared features to obtain the shallow transformation distribution vector of the target object for the target business data. Furthermore, the server can input the fused shared features into the deep transformation feature classification layer, and classify the fused shared features through the deep transformation feature classification layer to obtain the deep transformation distribution vector of the target object for the target business data.

[0222] Step S204: Input the initial service triggering feature and the initial service conversion feature into the multilayer perception network layer. The initial service triggering feature is processed by the multilayer perception network layer through full connection to obtain the target service triggering feature of the target object for the target service data. In the multilayer perception network layer, the target service triggering feature and the initial service conversion feature are processed by feature sharing to obtain the target service conversion feature of the target object for the target service data.

[0223] Specifically, the server can input the initial service trigger features into the second trigger fully connected layer. The second trigger fully connected layer performs fully connected processing on the initial service trigger features to obtain the target service trigger features for the target object regarding the target service data. The multilayer perceptron layer includes a second trigger fully connected layer, a second shallow transformation fully connected layer, a shallow connection layer, a second deep transformation fully connected layer, and a deep connection layer. The initial service transformation features include initial shallow transformation features and initial deep transformation features; the target service transformation features include target shallow transformation features and target deep transformation features. Further, the server can input the target service trigger features and the initial shallow transformation features into the shallow connection layer. The shallow connection layer performs feature concatenation on the target service trigger features and the initial shallow transformation features to obtain shared shallow concatenated features. Further, the server can input the shared shallow concatenated features into the second shallow transformation fully connected layer. The second shallow transformation fully connected layer performs fully connected processing on the shared shallow concatenated features to obtain the target shallow transformation features for the target object regarding the target service data. Furthermore, the server can input the target shallow transformation features and the initial deep transformation features into a deep connection layer. The deep connection layer then concatenates these features to obtain shared deep concatenated features. Further, the server can input these shared deep concatenated features into a second fully connected deep transformation layer. This second fully connected layer performs a full connection process on the shared deep concatenated features to obtain the target object's deep transformation features for the target business data.

[0224] Step 205: The matching features and the target business trigger features are concatenated to obtain the initial concatenated trigger features of the target object for the target business data. Based on the initial concatenated trigger features, the estimated trigger probability of the target object for the target business data is determined.

[0225] The specific process by which the server determines the estimated trigger probability can be found in the above. Figure 3 The description of step S103 in the corresponding embodiments will not be repeated here.

[0226] The sorting subnetwork further includes a feature connection layer and a dot product fully connected layer, which can be used to perform steps S206 and S207 below.

[0227] Step 206: The matching features and the target shallow conversion features are concatenated to obtain the initial concatenated shallow conversion features of the target object for the target business data. Based on the initial concatenated shallow conversion features, the shallow conversion probability estimate of the target object for the target business data is determined.

[0228] Specifically, the server can input the matching features and the target shallow transformation features into a feature connection layer. The feature connection layer concatenates the matching features and the target shallow transformation features to obtain the initial concatenated shallow transformation features of the target object for the target business data. Further, the server can input the initial concatenated shallow transformation features into a dot-product fully connected layer. The dot-product fully connected layer performs fully connected processing on the initial concatenated shallow transformation features to obtain the target concatenated shallow transformation features. Furthermore, the server can determine the estimated shallow transformation probability of the target object for the target business data based on the target concatenated shallow transformation features.

[0229] The target shallow conversion feature is composed of a first shallow conversion feature value and a second shallow conversion feature value. The first shallow conversion feature value represents the probability that the target business data has shallow conversion behavior, and the second shallow conversion feature value represents the probability that the target business data does not have shallow conversion behavior. Therefore, the server can use the first shallow conversion feature value as an estimated value of the shallow conversion probability of the target object for the target business data.

[0230] Step 207: The matching features and target deep conversion features are concatenated to obtain the initial concatenated deep conversion features of the target object for the target business data. Based on the initial concatenated deep conversion features, the predicted probability of deep conversion of the target object for the target business data is determined.

[0231] Specifically, the server can input the matching features and the target deep transformation features into a feature connection layer. The feature connection layer concatenates the matching features and the target deep transformation features to obtain the initial concatenated deep transformation features of the target object for the target business data. Further, the server can input the initial concatenated deep transformation features into a dot-product fully connected layer. The dot-product fully connected layer performs fully connected processing on the initial concatenated deep transformation features to obtain the target concatenated deep transformation features. Furthermore, the server can determine the predicted deep transformation probability of the target object for the target business data based on the target concatenated deep transformation features.

[0232] The target deep conversion feature is composed of a first deep conversion feature value and a second deep conversion feature value. The first deep conversion feature value represents the probability that the target business data has deep conversion behavior, and the second deep conversion feature value represents the probability that the target business data does not have deep conversion behavior. Therefore, the server can use the first deep conversion feature value as an estimate of the probability of deep conversion of the target object for the target business data.

[0233] It should be understood that conversion probability estimates include shallow conversion probability estimates and deep conversion probability estimates; target business conversion characteristics include target shallow conversion characteristics and target deep conversion characteristics. In other words, shallow conversion probability estimates and deep conversion probability estimates can be collectively referred to as conversion probability estimates, and target shallow conversion characteristics and target deep conversion characteristics can be collectively referred to as target business conversion characteristics.

[0234] Understandably, the specific process by which the server determines the predicted trigger probability, shallow conversion probability, and deep conversion probability for the target object based on business data can be found in [reference needed]. Figure 8a and Figure 8b , Figure 8a and Figure 8b This is a schematic diagram of the structure of a probability prediction network model provided in an embodiment of this application. Figure 8a The model structure corresponding to the matching subnetwork, Figure 8b The model structure corresponding to the sorting subnetwork.

[0235] like Figure 8a As shown, the matching subnetwork can be a dual-tower structure. The server can embed features of object attributes and business attributes through the matching embedding layer to obtain object features corresponding to object attributes and business features corresponding to business attributes. Then, the object features are input to the triggering fully connected layer, which performs fully connected processing on the object features to obtain fully connected object features. The business features are input to the business fully connected layer, which performs fully connected processing on the business features to obtain fully connected business features.

[0236] like Figure 8a As shown, the server can input pre-trained business features into a feature averaging layer, which averages the pre-trained business features to obtain averaged pre-trained business features. Further, the server can perform feature fusion on pre-trained object features and fully connected object features to obtain an object feature vector, and then perform feature fusion on the averaged pre-trained business features and fully connected business features to obtain a business feature vector. Finally, the server can perform a dot product operation on the business feature vector and the object feature vector to obtain the matching features between the target object and the target business data.

[0237] like Figure 8bAs shown, the server can input object attributes, business attributes, and context attributes into the input network layer 80b. The feature embedding layer in input network layer 80b embeds features of the object attributes, business attributes, and context attributes to obtain object features corresponding to the object attributes, business features corresponding to the business attributes, and context features corresponding to the context attributes. Further, the server can input pre-trained business features into the feature averaging layer in input network layer 80b. The feature averaging layer outputs pre-trained business features, which are then concatenated with the pre-trained business features to obtain concatenated business features. Similarly, object features and pre-trained object features are concatenated to obtain concatenated object features. Further, the server can use the feature concatenation layer in input network layer 80b (not shown in the diagram; for ease of understanding, the diagram shows the feature embedding layer and feature concatenation layer as the same layer) to concatenate the concatenated object features, concatenated business features, and context features to obtain shared attribute features of the target object for the target business data.

[0238] like Figure 8b As shown, the server can input the shared attribute features output from the input network layer 80b to the parameter sharing network layer 80c, where the fully connected layer outputs fully connected shared features. Further, the server can input the fully connected shared features to weight learning components 1, 2, 3, 4, and 5 in the sharing network layer 80c, respectively. These weight learning components output their corresponding weight shared features, and then perform feature fusion on the weight shared features corresponding to weight learning components 1, 2, 3, 4, and 5 to obtain fused shared features. These fused shared features are then input to the trigger feature classification layer, shallow transformation feature classification layer, and deep transformation feature classification layer in the parameter sharing network layer 80c. Specifically, weight learning component 1 can correspond to object features, weight learning component 2 to business features, weight learning component 3 to context features, weight learning component 4 to pre-trained object features, and weight learning component 5 to pre-trained business features. Furthermore, the server can input the trigger distribution vector corresponding to the trigger feature classification layer into the first trigger fully connected layer in the parameter sharing network layer 80c, input the shallow transformation distribution vector corresponding to the shallow transformation feature classification layer into the first shallow transformation fully connected layer in the parameter sharing network layer 80c, and input the deep transformation vector corresponding to the deep transformation feature classification layer into the first deep transformation fully connected layer in the parameter sharing network layer 80c. The first trigger fully connected layer outputs the initial service trigger feature, the first shallow transformation fully connected layer outputs the initial shallow transformation feature, and the first deep transformation fully connected layer outputs the initial deep transformation feature. Figure 8bThe first shallow transformation fully connected layer and the first deep transformation fully connected layer shown can be collectively referred to as the initial transformation fully connected layer.

[0239] like Figure 8b As shown, the server can input the initial service trigger feature into the multilayer perceptron layer 80d, and the second fully connected trigger layer in the multilayer perceptron layer 80d outputs the target service trigger feature. Further, the server can perform feature concatenation between the target service trigger feature and the initial shallow transformation feature through the shallow connection layer in the multilayer perceptron layer 80d to obtain a shared shallow concatenation feature, which is then output by the second fully connected shallow transformation layer in the multilayer perceptron layer 80d. Further still, the server can perform feature concatenation between the target shallow transformation feature and the initial deep transformation feature through the deep connection layer in the multilayer perceptron layer 80d to obtain a shared deep concatenation feature, which is then output by the second fully connected deep transformation layer in the multilayer perceptron layer 80d.

[0240] like Figure 8b As shown, the server can obtain matching features between the target object and the target business data. A feature connection layer concatenates the matching features and the target business trigger features to obtain initial concatenated trigger features. Then, a dot-product fully connected layer outputs the predicted trigger probability of the target object for the target business data. Similarly, a feature connection layer concatenates the matching features and the target shallow transformation features to obtain initial concatenated shallow transformation features. This dot-product fully connected layer then outputs the predicted shallow transformation probability of the target object for the target business data. Finally, a feature connection layer concatenates the matching features and the target deep transformation features to obtain initial concatenated deep transformation features. This dot-product fully connected layer then outputs the predicted deep transformation probability of the target object for the target business data. The dot-product fully connected layer can use the sigmoid function as the activation function.

[0241] It should be understood that the server can process the estimated trigger probability, estimated conversion probability, and estimated resources corresponding to the target business data to obtain the business data quality of the target object for the target business data. Here, estimated resources can represent the advertiser's shallow target bid and / or deep target bid for the target business data, and business data quality can represent the ECPM (Effective Cost Per Mille) metric, which represents the advertising revenue that can be obtained per thousand ad impressions. Furthermore, the server can perform a first sorting process on the target business data based on the business data quality to obtain the initial sorting information corresponding to the target business data. Further, the server can perform a second sorting process on the target business data using the sorting strategy and the initial sorting information to obtain the target sorting information corresponding to the target business data.

[0242] Among them, the server can perform calculations on the estimated trigger probability, estimated conversion probability, and estimated resources using multiplication. That is, the server can use the product of the estimated trigger probability, estimated conversion probability, and estimated resources as the business data quality.

[0243] It is understandable that when the server receives a request for business data acquisition from the target object, it can determine the matching probability between the target object and the target business data (i.e., the matching score of the target object for the target business data logit0), the trigger probability estimate of the target object for the target business data (i.e., the click-through rate estimate ectr(ExpectClick Through Rate)), the shallow conversion probability estimate (i.e., the shallow conversion rate estimate ecvr1(ExpectConversion Rate 1)) and the deep conversion probability estimate (i.e., the deep conversion rate estimate ecvr2(ExpectConversion Rate 2)). Among them, the click-through rate estimate (hereinafter referred to as click rate) can be represented by the following formula (2), the shallow conversion rate estimate (hereinafter referred to as shallow conversion rate) can be represented by the following formula (3), and the deep conversion rate estimate (hereinafter referred to as deep conversion rate) can be represented by the following formula (4):

[0244] ectr=P(ectr|x,H,logit0) (2)

[0245] ecvr1=P(ecvr1|ectr,x,H,logit0) (3)

[0246] ecvr2=P(ecvr2|ectr,ecvr1,x,H,logit0) (4)

[0247] Here, x can represent all input features (i.e., object attributes, business attributes, context attributes, triggering business features, and conversion business features), and H can represent the parameters of the probabilistic prediction network model. Therefore, P(ectr|x,H,logit0) can represent the click-through rate prediction value output by the probabilistic prediction network model, P(ecvr1|ectr,x,H,logit0) can represent the shallow conversion rate prediction value output by the probabilistic prediction network model, and P(ecvr2|ectr,ecvr1,x,H,logit0) can represent the deep conversion rate prediction value output by the probabilistic prediction network model.

[0248] Understandably, when the business data belongs to shallow target advertising, the server can determine the quality of the business data by referring to the following formula (5):

[0249] ecpm=ectr*ecvr1*bid1 (5)

[0250] Here, ectr can represent the estimated click probability (i.e., the estimated click-through rate) of a single-target ad, ecvr1 can represent the estimated shallow conversion probability (i.e., the estimated shallow conversion rate) of a single-target ad, and bid1 can represent the shallow target bid of a single-target ad.

[0251] Optionally, when the business data belongs to deep target advertising, if the shallow conversion rate estimate is greater than the deep conversion rate estimate, the server can determine the quality of the business data by referring to the following formula (6). If the shallow conversion rate estimate is not greater than (i.e. less than or equal to) the deep conversion rate estimate, the server can determine the quality of the business data by referring to the following formula (7).

[0252] ecpm=ectr*ecvr1*bid1 (6)

[0253] ecpm=ectr*ecvr2*bid2 (7)

[0254] Here, ectr can represent the estimated click probability (i.e., the estimated click-through rate) of a multi-objective ad, ecvr1 can represent the estimated shallow conversion probability (i.e., the estimated shallow conversion rate) of a multi-objective ad, ecvr2 can represent the estimated deep conversion probability (i.e., the estimated deep conversion rate) of a multi-objective ad, bid1 can represent the shallow target bid of a multi-objective ad, and bid2 can represent the deep target bid of a multi-objective ad.

[0255] For easier understanding, please refer to Figure 9 , Figure 9 This is a schematic diagram of a business data recommendation process provided in an embodiment of this application. For example... Figure 9 As shown, the application client can initiate a request (i.e., a business data retrieval request) to the server through step S91. Then, the server can retrieve and recall advertisements from the advertisement database through step S92. This advertisement database can be the one described above. Figure 2 In the corresponding embodiment, the business database 21a can be business data (i.e., advertisements) in the business data set obtained through the recall model.

[0256] like Figure 9As shown, the server can execute step S93, in which the click-through rate (CTR) of the advertisement is estimated using a probabilistic prediction network model, and then execute step S94, in which the business type of the business data is determined. This business type can include single-target advertisements and multi-target advertisements. It can be understood that when the business data is a single-target advertisement, the server can execute step S95, in which the shallow conversion rate of the advertisement is estimated using a probabilistic prediction network model, and then execute step S96, calculating the quality of the single-target advertisement (i.e., ECPM) based on the CTR and shallow conversion rate. Optionally, it can be understood that when the business data is a multi-target advertisement, the server can execute step S97, in which the shallow and deep conversion rates of the advertisement are estimated using a probabilistic prediction network model, and then execute step S98, calculating the quality of the multi-target advertisement based on the CTR, shallow conversion rate, and deep conversion rate.

[0257] Therefore, after calculating the quality of a single-target advertisement, the server can sort the single-target advertisements based on their quality to obtain initial sorting information; similarly, after calculating the quality of a multi-target advertisement, the server can sort the multi-target advertisements based on their quality to obtain initial sorting information.

[0258] Furthermore, such as Figure 9 As shown, in step S99, the server can reorder the advertisements according to ranking strategies such as ad diversity, frequency control, category control, and specific result weighting to obtain target ranking information for single-target or multi-target advertisements. Then, step S100 is executed, whereby the server retrieves the final top-ranked advertisement from the target ranking information and exposes this advertisement (i.e., recommendation business data) to the user (i.e., the target audience) on media (i.e., multimedia data displayed in the application client). Ad diversity, frequency control, and category control can be collectively referred to as ranking strategies. Optionally, the server can also retrieve the final top-ranked advertisement from the initial ranking information without considering ranking strategies and expose this advertisement (i.e., recommendation business data) to the target audience on media.

[0259] Therefore, the embodiments of this application can include three tasks: a click-through rate (CTR) prediction task, a shallow conversion rate (SCR) prediction task, and a deep conversion rate (DCR) prediction task. Each task can share the same shared fully connected layer. This sharing mechanism allows the shallow and deep conversion rate prediction tasks to share the underlying parameters of the CTR prediction task. Specifically, the target business triggering features of the CTR prediction task are shared with the conversion rate prediction task (i.e., the shallow and deep conversion rate prediction models). More specifically, the target business triggering features of the CTR prediction task are shared with the shallow conversion rate prediction task, and the target shallow conversion features of the shallow conversion rate prediction task are shared with the deep conversion rate prediction task. This improves the accuracy of the predicted trigger probability, shallow conversion probability, and deep conversion probability. Therefore, when using the aforementioned highly accurate trigger probability estimate and conversion probability estimate to sort the target business data, a more accurate sorting information of the target business data can be obtained. In turn, when using this sorting information to recommend business data to the target object, the accuracy of business data recommendation can be improved, similar advertisements can be avoided, and the user experience of the advertising recommendation system (i.e., the business data recommendation system) can be enhanced.

[0260] Further, please see Figure 10 , Figure 10 This is a flowchart illustrating a data processing method provided in an embodiment of this application. The method can be executed by a server, by an application client, or by both a server and an application client. The server can be one of the aforementioned... Figure 2 The server 20a in the corresponding embodiment can be the application client described above. Figure 2 The corresponding application client in the embodiments. For ease of understanding, this application embodiment uses the method executed by the server as an example for explanation. The data processing method may include the following steps S301-S305:

[0261] Step S301: Obtain sample business data associated with the sample object, and determine the sample matching features between the sample object and the sample business data through the initial probability prediction network model;

[0262] Among them, the sample objects have click behavior, shallow conversion behavior or deep conversion behavior for the sample business data, and the number of sample business data is one or more.

[0263] It should be understood that the specific process by which the server determines the sample matching features through the initial probability prediction network model can be found in the above description of determining matching features through the probability prediction network model, and will not be repeated here.

[0264] Step S302: In the initial probability prediction network model, determine the sample business triggering features and sample business conversion features of the sample object for the sample business data;

[0265] Among them, the sample business conversion features are obtained by sharing and transmitting the features of the sample business trigger features. The sample business conversion features include shallow conversion features and deep conversion features.

[0266] The specific process by which the server determines the sample service triggering characteristics and sample service conversion characteristics through the initial probability prediction network model can be found in the description above of determining the target service triggering characteristics and target service conversion characteristics through the probability prediction network model, and will not be repeated here.

[0267] Step S303: The sample matching features and sample business triggering features are concatenated to obtain the sample concatenated triggering features of the sample object in response to the sample business data. Based on the sample concatenated triggering features, the sample triggering probability prediction of the sample object in response to the sample business data is determined.

[0268] It should be understood that the specific process by which the server determines the sample trigger probability estimate through the initial probability prediction network model can be found in the above description of determining the trigger probability estimate through the probability prediction network model, and will not be repeated here.

[0269] Step S304: The sample matching features and sample business conversion features are concatenated to obtain the sample concatenated conversion features of the sample object for the sample business data. Based on the sample concatenated conversion features, the sample conversion probability prediction value of the sample object for the sample business data is determined.

[0270] Among them, the sample conversion probability prediction value includes the sample shallow conversion probability prediction value and the sample deep conversion probability prediction value, and the sample splicing conversion feature includes the sample shallow splicing conversion feature and the sample deep splicing conversion feature.

[0271] It should be understood that the specific process by which the server determines the sample conversion probability estimate through the initial probability prediction network model can be found in the description above of determining the conversion probability estimate through the probability prediction network model, and will not be repeated here.

[0272] Step S305: Based on the sample label information, sample trigger probability prediction value, and sample conversion probability prediction value of the sample object for the sample business data, adjust the parameters of the initial probability prediction network model, and use the parameter-adjusted initial probability prediction network model as the probability prediction network model.

[0273] Specifically, the server can determine the trigger loss value of the initial probability prediction network model based on the business trigger tag information of the sample object for the sample business data and the sample trigger probability prediction value. The sample tag information includes business trigger tag information and business conversion tag information. Further, the server can determine the conversion loss value of the initial probability prediction network model based on the business conversion tag information of the sample object for the sample business data and the sample conversion probability prediction value. Further, the server can determine the model loss value of the initial probability prediction network model based on the trigger loss value and the conversion loss value. Further, the server can adjust the parameters of the initial probability prediction network model based on the model loss value. When the parameter-adjusted initial probability prediction network model meets the model convergence condition, the parameter-adjusted initial probability prediction network model is used as the probability prediction network model. The probability prediction network model is used to predict the trigger probability prediction value and conversion probability prediction value of the target object for the target business data.

[0274] The business conversion tag information includes shallow conversion tag information and deep conversion tag information. It can be understood that the server can determine the shallow conversion loss value of the initial probability prediction network model based on the shallow conversion tag information of the sample object for the sample business data and the sample shallow conversion probability prediction value. Furthermore, the server can determine the deep conversion loss value of the initial probability prediction network model based on the deep conversion tag information of the sample object for the sample business data and the sample deep conversion probability prediction value. Furthermore, the server can determine the conversion loss value of the initial probability prediction network model based on the shallow conversion loss value and the deep conversion loss value.

[0275] It should be understood that the specific process of the server adjusting the parameters of the initial probability prediction network model based on the model loss value (i.e., iterative training) can be described as follows: When the model loss value of the initial probability prediction network model does not meet the model convergence condition, the server can adjust the model parameters of the initial probability prediction network model based on the model loss value that does not meet the model convergence condition. Furthermore, the server can determine the initial probability prediction network model after adjusting the model parameters as the transitional probability prediction network model, and iteratively train the transitional probability prediction network model until the model loss value of the iteratively trained transitional probability prediction network model meets the model convergence condition. The transitional probability prediction network model that meets the model convergence condition is then used as the probability prediction network model.

[0276] It should be understood that the initial probability prediction network model and the probability prediction network model can be collectively referred to as the generalized probability prediction network model. The initial probability prediction network model and the probability prediction network model are names of the generalized probability prediction network model at different times. During the training phase, the generalized probability prediction network model can be called the initial probability prediction network model, and during the prediction phase, the generalized probability prediction network model can be called the probability prediction network model.

[0277] Understandably, when the server iteratively trains the initial probability prediction network model, it can output multiple target probability values ​​P(ectr,ecvr1,ecvr2|x,H) (i.e., sample conversion probability prediction, shallow sample conversion probability prediction, and deep sample conversion probability prediction) through the initial probability prediction network model. The server can decompose the joint distribution into smaller individual distributions and obtain the probability values ​​output by the initial probability prediction network model based on the assumptions of the Bayesian probability model. The probability values ​​output by the initial probability prediction network model can be found in the following formula (8):

[0278]

[0279] Here, x can represent all input features (i.e., sample object attributes, sample business attributes, and sample context attributes), and H can represent the parameters of the initial probability prediction network model. Therefore, P(ectr|x,H,logit0) can represent the sample conversion probability prediction value output by the initial probability prediction network model, P(ecvr1|ectr,x,H,logit0) can represent the sample shallow conversion probability prediction value output by the target probability prediction network model, and P(ecvr2|ectr,ecvr1,x,H,logit0) can represent the sample deep conversion probability prediction value output by the target probability prediction network model.

[0280] It is understandable that taking the negative log-likelihood loss function (i.e., negative log-likelihood) of formula (8) yields the loss function L(x,H) shown in formula (9) below:

[0281]

[0282] It is understandable that by adding weight parameters (or simply weights) to different objectives in formula (9), we can obtain the loss function L(x,H) containing weight parameters as shown in formula (10) below:

[0283]

[0284] Here, W1, W2, and W3 can represent the weights of the three loss functions respectively. W1 represents the weight corresponding to the loss function P(ectr|x,H,logit0), W2 can represent the weight corresponding to P(ecvr1|ectr,x,H,logit0), and W3 can represent the weight corresponding to P(ecvr2|ectr,ecvr1,x,H,logit0).

[0285] Understandably, the weight parameters can represent the click behavior, shallow conversion behavior, or deep conversion behavior of the sample object in response to the sample business data. For example, when the sample object has click behavior but no shallow or deep conversion behavior, W1 can be equal to 1, and W2 and W3 can be equal to 0; or, when the sample object has click behavior and shallow conversion behavior but no deep conversion behavior, W1 and W2 can be equal to 1, and W3 can be equal to 0; or, when the sample object has click behavior, shallow conversion behavior, and deep conversion behavior, W1, W2, and W3 can be equal to 1.

[0286] It should be understood that the initial probability prediction network model can be derived from... Figure 8b It consists of three parts, as shown. Figure 8b The diagram shows a portion that can be used to fit ectr, another portion to fit ecvr1, and yet another portion to fit ecvr2. Therefore, the initial probability prediction network model has three sub-tasks, which can be used to output ectr, ecvr1, and ecvr2 respectively. Since this embodiment can predict multiple targets—ectr, ecvr1, and ecvr2—based on a Bayesian network, the goal of the server training the initial probability prediction network model is to optimize P(ectr, ecvr1, ecvr2 | x, H, logit0).

[0287] Therefore, this application proposes a multi-task model based on a pre-trained Bayesian network, which integrates the matching score of the target object and the target business data, the click-through rate, the shallow conversion rate, and the deep conversion rate into one model. Furthermore, it incorporates the intermediate training results of the click-through rate into the training of the shallow conversion rate target, the intermediate training results of the shallow conversion rate into the training of the deep conversion rate target, the correlation information between the target object and the target business data into the training of the click-through rate, the shallow conversion rate, and the deep conversion rate, and the matching score into the training of the click-through rate, the shallow conversion rate, and the deep conversion rate. Based on this, by simultaneously using three objectives—click-through rate (CTR), shallow conversion rate, and deep conversion rate—the initial probability prediction network model can fully utilize the underlying parameters, alleviating the problem of sparse training data. This is equivalent to sharing the click prediction training samples of the CTR prediction model with the conversion rate prediction model, and sharing the shallow conversion samples of the shallow conversion rate prediction model with the deep conversion rate prediction model (in fact, the CTR prediction model, conversion rate prediction model, shallow conversion rate prediction model, and deep conversion rate prediction model all belong to the initial probability prediction network model). This can improve the accuracy of CTR, shallow conversion rate, and deep conversion rate predictions, thereby improving the accuracy of business data recommendations.

[0288] Further, please see Figure 11 , Figure 11 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. The data processing device 1 may include: a first determining module 11, a second determining module 12, a first estimating module 13, and a second estimating module 14;

[0289] The first determining module 11 is used to acquire target business data associated with the target object and determine the matching features between the target object and the target business data.

[0290] The first determining module 11 includes: a data acquisition unit 111, a feature acquisition unit 112, a feature embedding unit 113, a vector determining unit 114, and a dot product operation unit 115.

[0291] The data acquisition unit 111 is used to acquire a set of business data associated with the target object and to acquire the target business data from the set of business data.

[0292] The feature acquisition unit 112 is used to acquire pre-trained object features associated with the target object and pre-trained business features associated with the business data sequence; the business data sequence is associated with the target object;

[0293] The business data sequence includes a business trigger sequence and a business conversion sequence; the business trigger sequence includes business data of the target object that has triggering behavior, and the business conversion sequence includes business data of the target object that has conversion behavior.

[0294] The feature acquisition unit 112 is specifically used to input the object attributes of the target object, the business trigger sequence associated with the target object, and the business conversion sequence associated with the target object into the feature recognition network model; the feature recognition network model includes a pre-trained embedding layer, an attention processing layer, and a pre-trained connection layer;

[0295] The feature acquisition unit 112 is specifically used to embed features into object attributes, business data in business trigger sequences, and business data in business conversion sequences through a pre-trained embedding layer, so as to obtain object features corresponding to object attributes, trigger business features corresponding to business data in business trigger sequences, and conversion business features corresponding to business data in business conversion sequences.

[0296] The feature acquisition unit 112 is specifically used to input object features, triggering business features and conversion business features into the attention processing layer, and to perform attention processing on the object features, triggering business features and conversion business features through the attention processing layer to obtain the attention object features corresponding to the object features, the attention triggering features corresponding to the triggering business features and the attention conversion features corresponding to the conversion business features.

[0297] The feature acquisition unit 112 is specifically used to input object features, attention object features, attention trigger features and attention transformation features into the pre-trained connection layer, and to perform feature fusion on the object features and attention object features through the pre-trained connection layer to obtain pre-trained object features associated with the target object. In the pre-trained connection layer, based on the attention trigger features and attention transformation features, pre-trained business features corresponding to the business data contained in the business trigger sequence and business transformation sequence are generated.

[0298] The feature embedding unit 113 is used to input the object attributes of the target object, the business attributes of the target business data, the pre-trained object features, and the pre-trained business features into the matching sub-network in the probability prediction network model; the matching sub-network includes a matching embedding layer.

[0299] The feature embedding unit 113 is used to embed features of object attributes and business attributes through the matching embedding layer to obtain object features corresponding to object attributes and business features corresponding to business attributes.

[0300] The vector determination unit 114 is used in the matching subnetwork to determine the object feature vector associated with the target object based on object features and pre-trained object features, and to determine the business feature vector associated with the target business data based on business features and pre-trained business features.

[0301] The matching sub-network includes an object fully connected layer, a business fully connected layer, and a feature averaging layer; the business data sequence includes one or more business data, each business data corresponding to a pre-trained business feature.

[0302] The vector determination unit 114 is specifically used in the matching sub-network to input the object features into the fully connected object layer, and to perform fully connected processing on the object features through the fully connected object layer to obtain fully connected object features;

[0303] The vector determination unit 114 is specifically used to perform feature fusion on the fully connected object features and the pre-trained object features to obtain an object feature vector associated with the target object.

[0304] The vector determination unit 114 is specifically used to input business features into the business fully connected layer, and to perform fully connected processing on the business features through the business fully connected layer to obtain fully connected business features;

[0305] The vector determination unit 114 is specifically used to input one or more pre-trained business features into the feature averaging layer, and to average the one or more pre-trained business features through the feature averaging layer to obtain the average pre-trained business features.

[0306] The vector determination unit 114 is specifically used to perform feature fusion on the fully connected business features and the average pre-trained business features to obtain a business feature vector associated with the target business data.

[0307] The dot product operation unit 115 is used to perform a dot product operation on the object feature vector and the business feature vector to obtain the matching features between the target object and the target business data.

[0308] The specific implementation methods of the data acquisition unit 111, feature acquisition unit 112, feature embedding unit 113, vector determination unit 114, and dot product operation unit 115 can be found in the above description. Figure 3 In the corresponding embodiment, steps S101 and Figure 4 The descriptions of steps S1011-S1019 in the corresponding embodiments will not be repeated here.

[0309] The second determining module 12 is used to determine the target business triggering features and target business conversion features of the target object for the target business data; the target business conversion features are obtained by performing feature sharing and transmission processing on the target business triggering features;

[0310] The probability prediction network model also includes a ranking sub-network; the ranking sub-network includes an input network layer, a parameter sharing network layer, and a multi-layer perceptual network layer.

[0311] The second determining module 12 includes: a feature determining unit 121, a feature generating unit 122, and a feature transmitting unit 123;

[0312] Feature determination unit 121 is used to determine the shared attribute features of the target object for the target business data by inputting the network layer;

[0313] The input network layer includes a feature embedding layer and a feature concatenation layer; the business data sequence includes one or more business data, each business data corresponding to a pre-trained business feature.

[0314] The feature determination unit 121 is specifically used to input the object attributes of the target object, the business attributes of the target business data, and the context attributes associated with the target object into the feature embedding layer. The feature embedding layer performs feature embedding on the object attributes, business attributes, and context attributes to obtain the object features corresponding to the object attributes, the business features corresponding to the business attributes, and the context features corresponding to the context attributes.

[0315] The feature determination unit 121 is specifically used to perform average processing on one or more pre-trained business features to obtain average pre-trained business features.

[0316] The feature determination unit 121 is specifically used to input object features, pre-trained object features, business features and average pre-trained business features into the feature concatenation layer. In the feature concatenation layer, the object features and pre-trained object features are concatenated to obtain concatenated object features, and the business features and average pre-trained business features are concatenated to obtain concatenated business features.

[0317] The feature determination unit 121 is specifically used to perform feature splicing on the splicing object features, splicing business features and context features to obtain the shared attribute features of the target object for the target business data.

[0318] The feature generation unit 122 is used to input shared attribute features into the parameter sharing network layer, and generate initial business trigger features and initial business transformation features of the target object for the target business data through the parameter sharing network layer.

[0319] The parameter-sharing network layer includes a shared fully connected layer, a weight learning layer, a feature classification layer, a first trigger fully connected layer, a first shallow transformation fully connected layer, and a first deep transformation fully connected layer; the initial business transformation features include initial shallow transformation features and initial deep transformation features.

[0320] The feature generation unit 122 is specifically used to input the shared attribute features into the shared fully connected layer, and to perform fully connected processing on the shared attribute features through the shared fully connected layer to obtain fully connected shared features.

[0321] The feature generation unit 122 is specifically used to input the fully connected shared features into the weight learning layer, and to perform feature weighting on the fully connected shared features through the weight learning layer to obtain the weighted shared features.

[0322] The feature generation unit 122 is specifically used to input the weighted shared features into the feature classification layer, and to classify the weighted shared features through the feature classification layer to obtain the trigger distribution vector, shallow transformation distribution vector and deep transformation distribution vector of the target object for the target business data.

[0323] The feature generation unit 122 is specifically used to input the trigger distribution vector into the first trigger fully connected layer, and to perform full connection processing on the trigger distribution vector through the first trigger fully connected layer to generate the initial business trigger features of the target object for the target business data.

[0324] The feature generation unit 122 is specifically used to input the shallow transformation distribution vector into the first shallow transformation fully connected layer, and to perform fully connected processing on the shallow transformation distribution vector through the first shallow transformation fully connected layer to generate the initial shallow transformation features of the target object for the target business data.

[0325] The feature generation unit 122 is specifically used to input the deep transformation distribution vector into the first deep transformation fully connected layer, and to perform full connection processing on the deep transformation distribution vector through the first deep transformation fully connected layer to generate the initial deep transformation features of the target object for the target business data.

[0326] The weight learning layer includes weight learning components corresponding to object features, business features, context features, pre-trained object features, and average pre-trained business features, respectively; the feature classification layer includes a trigger feature classification layer, a shallow transformation feature classification layer, and a deep transformation feature classification layer.

[0327] The feature generation unit 122 is specifically used to input the fully connected shared features into each weight learning component, and to perform feature weighting on the fully connected shared features through each weight learning component to obtain the weight shared features corresponding to each weight learning component.

[0328] The feature generation unit 122 is specifically used to perform feature fusion on the weight-shared features corresponding to each weight learning component to obtain fused shared features.

[0329] The feature generation unit 122 is specifically used to input the fused shared features into the trigger feature classification layer, and to classify the fused shared features through the trigger feature classification layer to obtain the trigger distribution vector of the target object for the target business data.

[0330] The feature generation unit 122 is specifically used to input the fused shared features into the shallow transformation feature classification layer, and to classify the fused shared features through the shallow transformation feature classification layer to obtain the shallow transformation distribution vector of the target object for the target business data;

[0331] The feature generation unit 122 is specifically used to input the fused shared features into the deep transformation feature classification layer, and to classify the fused shared features through the deep transformation feature classification layer to obtain the deep transformation distribution vector of the target object for the target business data.

[0332] The feature transfer unit 123 is used to input the initial service trigger feature and the initial service conversion feature into the multilayer perception network layer. The initial service trigger feature is processed by the multilayer perception network layer through full connection to obtain the target service trigger feature of the target object for the target service data. In the multilayer perception network layer, the target service trigger feature and the initial service conversion feature are processed by feature sharing to obtain the target service conversion feature of the target object for the target service data.

[0333] The multilayer perception network layer includes a second trigger fully connected layer, a second shallow transformation fully connected layer, a shallow connection layer, a second deep transformation fully connected layer, and a deep connection layer; the initial service transformation features include initial shallow transformation features and initial deep transformation features; the target service transformation features include target shallow transformation features and target deep transformation features.

[0334] The feature transfer unit 123 is specifically used to input the initial business trigger feature to the second trigger fully connected layer, and to perform full connection processing on the initial business trigger feature through the second trigger fully connected layer to obtain the target business trigger feature of the target object for the target business data.

[0335] The feature transfer unit 123 is specifically used to input the target business trigger feature and the initial shallow transformation feature into the shallow connection layer, and to perform feature splicing on the target business trigger feature and the initial shallow transformation feature through the shallow connection layer to obtain the shared shallow splicing feature;

[0336] The feature transfer unit 123 is specifically used to input the shared shallow splicing features into the second shallow transformation fully connected layer, and to perform full connection processing on the shared shallow splicing features through the second shallow transformation fully connected layer to obtain the target shallow transformation features of the target object for the target business data.

[0337] The feature transfer unit 123 is specifically used to input the target shallow transformation feature and the initial deep transformation feature into the deep connection layer, and to perform feature splicing on the target shallow transformation feature and the initial deep transformation feature through the deep connection layer to obtain the shared deep splicing feature;

[0338] The feature transfer unit 123 is specifically used to input the shared deep splicing features into the second deep transformation fully connected layer, and to perform full connection processing on the shared deep splicing features through the second deep transformation fully connected layer to obtain the target deep transformation features of the target object for the target business data.

[0339] The specific implementation methods of the feature determination unit 121, the feature generation unit 122, and the feature transfer unit 123 can be found in the above description. Figure 3 In the corresponding embodiment, steps S102 and Figure 6 The descriptions of steps S1021-S1023 in the corresponding embodiments will not be repeated here.

[0340] The first prediction module 13 is used to concatenate the matching features and the target business trigger features to obtain the initial concatenated trigger features of the target object for the target business data, and to determine the trigger probability prediction value of the target object for the target business data based on the initial concatenated trigger features.

[0341] The sorting subnetwork also includes a feature connection layer and a dot product fully connected layer;

[0342] The first prediction module 13 includes: a first splicing unit 131, a first processing unit 132, and a first prediction unit 133;

[0343] The first splicing unit 131 is used to input the matching features and the target business triggering features into the feature connection layer, and to splice the matching features and the target business triggering features through the feature connection layer to obtain the initial splicing triggering features of the target object for the target business data.

[0344] The first processing unit 132 is used to input the initial splicing trigger features into the dot-product fully connected layer, and perform fully connected processing on the initial splicing trigger features through the dot-product fully connected layer to obtain the target splicing trigger features;

[0345] The first estimation unit 133 is used to determine the estimated probability of the target object triggering the target business data based on the target splicing trigger characteristics.

[0346] The specific implementation methods of the first splicing unit 131, the first processing unit 132, and the first estimation unit 133 can be found in the above description. Figure 3 The description of step S103 in the corresponding embodiments will not be repeated here.

[0347] The second prediction module 14 is used to concatenate the matching features and the target business conversion features to obtain the initial concatenated conversion features of the target object for the target business data, and to determine the conversion probability prediction value of the target object for the target business data based on the initial concatenated conversion features.

[0348] The ranking subnetwork also includes a feature connection layer and a dot product fully connected layer; the conversion probability prediction includes shallow conversion probability prediction and deep conversion probability prediction; the target business conversion features include target shallow conversion features and target deep conversion features.

[0349] The second prediction module 14 includes: a second splicing unit 141, a second processing unit 142, a second prediction unit 143, a third splicing unit 144, a third processing unit 145, and a third prediction unit 146.

[0350] The second splicing unit 141 is used to input the matching features and the target shallow transformation features into the feature connection layer, and to splice the matching features and the target shallow transformation features through the feature connection layer to obtain the initial spliced ​​shallow transformation features of the target object for the target business data.

[0351] The second processing unit 142 is used to input the initial splicing shallow transformation features into the dot product fully connected layer, and to perform fully connected processing on the initial splicing shallow transformation features through the dot product fully connected layer to obtain the target splicing shallow transformation features.

[0352] The second estimation unit 143 is used to determine the shallow conversion probability estimate of the target object for the target business data based on the shallow conversion characteristics of the target splicing.

[0353] The third splicing unit 144 is used to input the matching features and the target deep transformation features into the feature connection layer, and to splice the matching features and the target deep transformation features through the feature connection layer to obtain the initial spliced ​​deep transformation features of the target object for the target business data.

[0354] The third processing unit 145 is used to input the initial splicing deep transformation features into the dot product fully connected layer, and to perform fully connected processing on the initial splicing deep transformation features through the dot product fully connected layer to obtain the target splicing deep transformation features;

[0355] The third estimation unit 146 is used to determine the estimated probability of deep conversion of the target object for the target business data based on the deep conversion characteristics of the target splicing.

[0356] The specific implementation methods of the second splicing unit 141, the second processing unit 142, the second estimation unit 143, the third splicing unit 144, the third processing unit 145, and the third estimation unit 146 can be found in the above description. Figure 3 The description of step S104 in the corresponding embodiment will not be repeated here.

[0357] The specific implementation methods of the first determining module 11, the second determining module 12, the first estimating module 13, and the second estimating module 14 can be found in the above description. Figure 3In the corresponding embodiment, steps S101-S104, Figure 4 In the corresponding embodiment, steps S1011-S1019, Figure 6 In the corresponding embodiment, steps S1021-S1023 and Figure 7 The descriptions of steps S201-S207 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0358] Further, please see Figure 12 , Figure 12 This is a schematic diagram of the structure of a data processing device provided in an embodiment of this application. The data processing device 2 may include: a first sample determination module 21, a second sample determination module 22, a first sample estimation module 23, a second sample estimation module 24, and a parameter adjustment module 25;

[0359] The first sample determination module 21 is used to acquire sample business data associated with the sample object and determine the sample matching features between the sample object and the sample business data through the initial probability prediction network model.

[0360] The second sample determination module 22 is used to determine the sample service triggering features and sample service transformation features of the sample object for the sample service data in the initial probability prediction network model; the sample service transformation features are obtained by performing feature sharing and transmission processing on the sample service triggering features.

[0361] The first sample prediction module 23 is used to concatenate the sample matching features and the sample business triggering features to obtain the sample concatenation triggering features of the sample object in relation to the sample business data, and to determine the sample triggering probability prediction value of the sample object in relation to the sample business data based on the sample concatenation triggering features.

[0362] The second sample prediction module 24 is used to concatenate the sample matching features and the sample business conversion features to obtain the sample concatenated conversion features of the sample object for the sample business data, and to determine the sample conversion probability prediction value of the sample object for the sample business data based on the sample concatenated conversion features.

[0363] The parameter adjustment module 25 is used to adjust the parameters of the initial probability prediction network model based on the sample label information, sample trigger probability prediction value, and sample conversion probability prediction value of the sample object for the sample business data. The parameter-adjusted initial probability prediction network model is used as the probability prediction network model. The probability prediction network model is used to predict the trigger probability prediction value and conversion probability prediction value of the target object for the target business data.

[0364] The specific implementation methods of the first sample determination module 21, the second sample determination module 22, the first sample prediction module 23, the second sample prediction module 24, and the parameter adjustment module 25 can be found in the above description. Figure 10 The descriptions of steps S301-S305 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0365] Further, please see Figure 13 , Figure 13 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 13 As shown, the computer device 1000 may include a processor 1001, a network interface 1004, and a memory 1005. Furthermore, the computer device 1000 may also include a user interface 1003 and at least one communication bus 1002. The communication bus 1002 is used to enable communication between these components. In some embodiments, the user interface 1003 may include a display screen and a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. Optionally, the network interface 1004 may include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be high-speed RAM or non-volatile memory, such as at least one disk storage device. Optionally, the memory 1005 may also be at least one storage device located remotely from the processor 1001. Figure 13 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a device control application.

[0366] In such Figure 13 In the computer device 1000 shown, the network interface 1004 provides network communication functions; the user interface 1003 is mainly used to provide an input interface for users; and the processor 1001 can be used to call the device control application stored in the memory 1005.

[0367] It should be understood that the computer device 1000 described in the embodiments of this application can execute the foregoing text. Figure 3 , Figure 4 , Figure 6 , Figure 7 and Figure 10 The description of the data processing method in the corresponding embodiments can also be performed as described above. Figure 11 In the corresponding embodiments, data processing device 1 and Figure 12The description of the data processing device 2 in the corresponding embodiments will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated.

[0368] Furthermore, it should be noted that this application embodiment also provides a computer-readable storage medium, which stores a computer program executed by the aforementioned data processing apparatus 1 and data processing apparatus 2. The computer program includes program instructions, and when the processor executes the program instructions, it can execute the aforementioned... Figure 3 , Figure 4 , Figure 6 , Figure 7 and Figure 10 The description of the data processing method in the corresponding embodiments is already provided and will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer-readable storage medium embodiments related to this application, please refer to the description of the method embodiments of this application.

[0369] Furthermore, it should be noted that this application also provides a computer program product or computer program, which may include computer instructions, which may be stored in a computer-readable storage medium. The processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor may execute the computer instructions, causing the computer device to perform the aforementioned actions. Figure 3 , Figure 4 , Figure 6 , Figure 7 and Figure 10 The description of the data processing method in the corresponding embodiments is already provided and will not be repeated here. Furthermore, the beneficial effects of using the same method will also not be repeated. For technical details not disclosed in the computer program products or computer program embodiments related to this application, please refer to the description of the method embodiments of this application.

[0370] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0371] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. A data processing method, characterized in that, include: Obtain target business data associated with the target object, and determine the matching features between the target object and the target business data; The shared attribute features of the target object for the target business data are determined by the input network layer in the ranking subnetwork of the probability prediction network model; the ranking subnetwork also includes a parameter sharing network layer and a multi-layer perception network layer. The shared attribute features are input into the parameter sharing network layer, and the parameter sharing network layer generates the initial business trigger features and initial business conversion features of the target object for the target business data. The initial service triggering feature and the initial service conversion feature are input into the multilayer perception network layer. The initial service triggering feature is processed by the multilayer perception network layer through full connection to obtain the target service triggering feature of the target object for the target service data. In the multilayer perception network layer, the target service triggering feature and the initial service conversion feature are processed by feature sharing and transmission to obtain the target service conversion feature of the target object for the target service data. The matching features and the target service triggering features are concatenated to obtain the initial concatenated triggering features of the target object for the target service data. Based on the initial concatenated triggering features, the estimated triggering probability of the target object for the target service data is determined. The matching features and the target business conversion features are concatenated to obtain the initial concatenated conversion features of the target object for the target business data. Based on the initial concatenated conversion features, the conversion probability estimate of the target object for the target business data is determined.

2. The method according to claim 1, characterized in that, The step of acquiring target business data associated with the target object and determining the matching features between the target object and the target business data includes: Obtain a set of business data associated with the target object, and retrieve the target business data from the set of business data; Obtain pre-trained object features associated with the target object, and pre-trained business features associated with the business data sequence; the business data sequence is associated with the target object; The object attributes of the target object, the business attributes of the target business data, the pre-trained object features, and the pre-trained business features are input into the matching sub-network of the probability prediction network model; the matching sub-network includes a matching embedding layer; The object attributes and business attributes are embedded using the matching embedding layer to obtain object features corresponding to the object attributes and business features corresponding to the business attributes. In the matching sub-network, an object feature vector associated with the target object is determined based on the object features and the pre-trained object features, and a business feature vector associated with the target business data is determined based on the business features and the pre-trained business features. Perform a dot product operation on the object feature vector and the business feature vector to obtain the matching features between the target object and the target business data.

3. The method according to claim 2, characterized in that, The business data sequence includes a business trigger sequence and a business conversion sequence; the business trigger sequence includes business data of the target object that has triggering behavior, and the business conversion sequence includes business data of the target object that has conversion behavior; The acquisition of pre-trained object features associated with the target object and pre-trained business features associated with the business data sequence includes: The object attributes of the target object, the business trigger sequence associated with the target object, and the business conversion sequence associated with the target object are input into the feature recognition network model; the feature recognition network model includes a pre-trained embedding layer, an attention processing layer, and a pre-trained connection layer; The pre-trained embedding layer performs feature embedding on the object attributes, the business data in the business trigger sequence, and the business data in the business conversion sequence, respectively, to obtain the object features corresponding to the object attributes, the trigger business features corresponding to the business data in the business trigger sequence, and the conversion business features corresponding to the business data in the business conversion sequence. The object feature, the triggering business feature, and the conversion business feature are input into the attention processing layer. The attention processing layer performs attention processing on the object feature, the triggering business feature, and the conversion business feature to obtain the attention object feature corresponding to the object feature, the attention triggering feature corresponding to the triggering business feature, and the attention conversion feature corresponding to the conversion business feature. The object features, attention object features, attention trigger features, and attention conversion features are input into the pre-trained connection layer. The object features and attention object features are fused through the pre-trained connection layer to obtain pre-trained object features associated with the target object. In the pre-trained connection layer, pre-trained business features corresponding to the business data contained in the business trigger sequence and the business conversion sequence are generated based on the attention trigger features and the attention conversion features.

4. The method according to claim 2, characterized in that, The matching sub-network includes an object fully connected layer, a business fully connected layer, and a feature averaging layer; the business data sequence includes one or more business data, each business data corresponding to a pre-trained business feature. In the matching sub-network, determining the object feature vector associated with the target object based on the object features and the pre-trained object features, and determining the business feature vector associated with the target business data based on the business features and the pre-trained business features, includes: In the matching sub-network, the object features are input to the fully connected object layer, and the object features are processed by the fully connected object layer to obtain fully connected object features; The fully connected object features and the pre-trained object features are fused to obtain an object feature vector associated with the target object; The business features are input into the fully connected business layer, and the fully connected business layer performs full-connection processing on the business features to obtain fully connected business features; One or more pre-trained business features are input into the feature averaging layer, and the feature averaging layer is used to average the one or more pre-trained business features to obtain the average pre-trained business features. The fully connected service features and the average pre-trained service features are fused to obtain a service feature vector associated with the target service data.

5. The method according to claim 1, characterized in that, The input network layer includes a feature embedding layer and a feature concatenation layer; the business data sequence includes one or more business data, each business data corresponding to a pre-trained business feature; The step of determining the shared attribute features of the target object for the target business data through the input network layer in the ranking subnetwork of the probability prediction network model includes: The object attributes of the target object, the business attributes of the target business data, and the context attributes associated with the target object are input into the feature embedding layer. The feature embedding layer performs feature embedding on the object attributes, the business attributes, and the context attributes to obtain the object features corresponding to the object attributes, the business features corresponding to the business attributes, and the context features corresponding to the context attributes. Average one or more pre-trained business features to obtain the average pre-trained business features; The object features, pre-trained object features, business features, and average pre-trained business features are input into the feature concatenation layer. In the feature concatenation layer, the object features and the pre-trained object features are concatenated to obtain concatenated object features, and the business features and the average pre-trained business features are concatenated to obtain concatenated business features. The splicing object features, the splicing business features, and the context features are spliced ​​together to obtain the shared attribute features of the target object for the target business data.

6. The method according to claim 5, characterized in that, The parameter-sharing network layer includes a shared fully connected layer, a weight learning layer, a feature classification layer, a first trigger fully connected layer, a first shallow transformation fully connected layer, and a first deep transformation fully connected layer; the initial business transformation features include initial shallow transformation features and initial deep transformation features; The step of inputting the shared attribute features into the parameter sharing network layer, and generating the initial business triggering features and initial business conversion features of the target object for the target business data through the parameter sharing network layer, includes: The shared attribute features are input into the shared fully connected layer, and the shared attribute features are processed by the shared fully connected layer to obtain fully connected shared features; The fully connected shared features are input into the weight learning layer, and the weight learning layer performs feature weighting on the fully connected shared features to obtain weighted shared features. The weight-sharing features are input into the feature classification layer, and the feature classification layer performs feature classification on the weight-sharing features to obtain the trigger distribution vector, shallow conversion distribution vector and deep conversion distribution vector of the target object for the target business data; The trigger distribution vector is input to the first trigger fully connected layer, and the trigger distribution vector is fully connected through the first trigger fully connected layer to generate the initial business trigger features of the target object for the target business data; The shallow transformation distribution vector is input into the first shallow transformation fully connected layer, and the shallow transformation distribution vector is fully connected through the first shallow transformation fully connected layer to generate the initial shallow transformation features of the target object for the target business data; The deep transformation distribution vector is input into the first deep transformation fully connected layer, and the deep transformation distribution vector is fully connected through the first deep transformation fully connected layer to generate the initial deep transformation features of the target object for the target business data.

7. The method according to claim 6, characterized in that, The weight learning layer includes weight learning components corresponding to the object features, business features, context features, pre-trained object features, and average pre-trained business features, respectively; the feature classification layer includes a trigger feature classification layer, a shallow transformation feature classification layer, and a deep transformation feature classification layer; The step of inputting the fully connected shared features into the weight learning layer, and then performing feature weighting processing on the fully connected shared features through the weight learning layer to obtain weighted shared features includes: The fully connected shared features are input into each weight learning component, and each weight learning component performs feature weighting on the fully connected shared features to obtain the weight shared features corresponding to each weight learning component. The step of inputting the weighted shared features into the feature classification layer, and classifying the weighted shared features through the feature classification layer to obtain the trigger distribution vector, shallow conversion distribution vector, and deep conversion distribution vector of the target object for the target business data, includes: The weight-shared features corresponding to each weight learning component are fused to obtain fused shared features. The fused shared features are input into the trigger feature classification layer, and the trigger feature classification layer performs feature classification on the fused shared features to obtain the trigger distribution vector of the target object for the target business data; The fused shared features are input into the shallow conversion feature classification layer, and the shallow conversion feature classification layer performs feature classification on the fused shared features to obtain the shallow conversion distribution vector of the target object for the target business data; The fused shared features are input into the deep conversion feature classification layer, and the deep conversion feature classification layer performs feature classification on the fused shared features to obtain the deep conversion distribution vector of the target object for the target business data.

8. The method according to claim 1, characterized in that, The multilayer perception network layer includes a second triggered fully connected layer, a second shallow conversion fully connected layer, a shallow connection layer, a second deep conversion fully connected layer, and a deep connection layer; the initial service conversion features include initial shallow conversion features and initial deep conversion features; the target service conversion features include target shallow conversion features and target deep conversion features; The process involves inputting the initial service triggering feature and the initial service conversion feature into the multilayer perceptron layer, performing fully connected processing on the initial service triggering feature through the multilayer perceptron layer to obtain the target service triggering feature of the target object for the target service data, and then performing feature sharing and transfer processing on the target service triggering feature and the initial service conversion feature within the multilayer perceptron layer to obtain the target service conversion feature of the target object for the target service data. This includes: The initial service trigger feature is input to the second trigger fully connected layer, and the initial service trigger feature is processed by the second trigger fully connected layer to obtain the target service trigger feature of the target object for the target service data; The target service trigger feature and the initial shallow conversion feature are input into the shallow connection layer. The shallow connection layer performs feature concatenation on the target service trigger feature and the initial shallow conversion feature to obtain a shared shallow concatenation feature. The shared shallow splicing feature is input into the second shallow transformation fully connected layer, and the shared shallow splicing feature is fully connected through the second shallow transformation fully connected layer to obtain the target shallow transformation feature of the target object for the target business data; The target shallow transformation feature and the initial deep transformation feature are input into the deep connection layer. The deep connection layer performs feature concatenation on the target shallow transformation feature and the initial deep transformation feature to obtain a shared deep concatenation feature. The shared deep splicing features are input into the second deep transformation fully connected layer, and the shared deep splicing features are fully connected through the second deep transformation fully connected layer to obtain the target deep transformation features of the target object for the target business data.

9. The method according to claim 1, characterized in that, The sorting subnetwork also includes a feature connection layer and a dot product fully connected layer; The step of concatenating the matching features and the target service triggering features to obtain the initial concatenated triggering features of the target object for the target service data, and determining the estimated triggering probability of the target object for the target service data based on the initial concatenated triggering features, includes: The matching feature and the target service trigger feature are input into the feature connection layer. The feature connection layer concatenates the matching feature and the target service trigger feature to obtain the initial concatenation trigger feature of the target object for the target service data. The initial splicing trigger features are input into the dot-product fully connected layer, and the initial splicing trigger features are fully connected through the dot-product fully connected layer to obtain the target splicing trigger features; Based on the target splicing trigger characteristics, determine the estimated trigger probability of the target object for the target business data.

10. The method according to claim 1, characterized in that, The ranking subnetwork further includes a feature connection layer and a dot product fully connected layer; the conversion probability prediction value includes a shallow conversion probability prediction value and a deep conversion probability prediction value; the target business conversion features include target shallow conversion features and target deep conversion features; The step of concatenating the matching features and the target business conversion features to obtain the initial concatenated conversion features of the target object for the target business data, and determining the conversion probability estimate of the target object for the target business data based on the initial concatenated conversion features, includes: The matching feature and the target shallow transformation feature are input into the feature connection layer. The feature connection layer concatenates the matching feature and the target shallow transformation feature to obtain the initial concatenated shallow transformation feature of the target object for the target business data. The initial shallow layer transformation feature is input into the dot product fully connected layer, and the initial shallow layer transformation feature is fully connected through the dot product fully connected layer to obtain the target shallow layer transformation feature. Based on the target splicing shallow conversion characteristics, determine the shallow conversion probability estimate of the target object for the target business data; The matching features and the target deep transformation features are input into the feature connection layer. The feature connection layer concatenates the matching features and the target deep transformation features to obtain the initial concatenated deep transformation features of the target object for the target business data. The initial deep transformation feature of the splicing is input into the dot-product fully connected layer, and the initial deep transformation feature of the splicing is fully connected through the dot-product fully connected layer to obtain the target deep transformation feature of the splicing. Based on the target splicing deep conversion characteristics, determine the estimated value of the deep conversion probability of the target object for the target business data.

11. A data processing method, characterized in that, include: Obtain sample business data associated with the sample object, and determine the sample matching features between the sample object and the sample business data through an initial probability prediction network model; The shared attribute features of the sample object for the sample business data are determined by the input network layer in the ranking subnetwork of the initial probability prediction network model; the ranking subnetwork also includes a parameter sharing network layer and a multi-layer perception network layer. The shared attribute features are input into the parameter sharing network layer, and the parameter sharing network layer generates the initial business trigger features and initial business conversion features of the sample object for the sample business data. The initial service triggering feature and the initial service conversion feature are input into the multilayer perception network layer. The initial service triggering feature is processed by the multilayer perception network layer through full connection to obtain the sample service triggering feature of the sample object for the sample service data. In the multilayer perception network layer, the sample service triggering feature and the initial service conversion feature are processed by feature sharing to obtain the sample service conversion feature of the sample object for the sample service data. The sample matching features and the sample service triggering features are concatenated to obtain the sample concatenated triggering features of the sample object in response to the sample service data. Based on the sample concatenated triggering features, the estimated sample triggering probability of the sample object in response to the sample service data is determined. The sample matching features and the sample business conversion features are concatenated to obtain the sample concatenated conversion features of the sample object for the sample business data. Based on the sample concatenated conversion features, the sample conversion probability prediction value of the sample object for the sample business data is determined. Based on the sample tag information of the sample object for the sample business data, the sample trigger probability estimate, and the sample conversion probability estimate, the parameters of the initial probability prediction network model are adjusted, and the parameter-adjusted initial probability prediction network model is used as the probability prediction network model; the probability prediction network model is used to predict the trigger probability estimate and conversion probability estimate of the target object for the target business data.

12. A computer device, characterized in that, include: Processor and memory; The processor is connected to the memory, wherein the memory is used to store a computer program, and the processor is used to invoke the computer program to cause the computer device to perform the method according to any one of claims 1-11.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded and executed by a processor to cause a computer device having the processor to perform the method according to any one of claims 1-11.

14. A computer program product, characterized in that, The computer program product includes computer instructions stored in a computer-readable storage medium and adapted to be read and executed by a processor to cause a computer device having the processor to perform the method of any one of claims 1-11.

Images