Conversion rate prediction method and device, computer equipment and storage medium

Abstract

The invention relates to a conversion rate prediction method and device, computer equipment, a storage medium and a computer program product. The method comprises the steps that a directed graph representing the connection relation between task nodes of a target task is obtained, wherein the connection relation is used for representing the continuity of interactive operation for the task nodes; performing feature extraction on the shared information for each task node in the target task to obtain an initial feature of each task node; based on a directed graph, performing feature optimization processing on the respective initial feature of each task node to obtain the respective target feature of each task node; and for each task node, predicting the operation conversion rate of the interactive operation represented by the task node based on the target feature. By adopting the method, the accuracy of conversion rate prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to a conversion rate prediction method and device, computer equipment and storage medium. BACKGROUND

[0002] With the rapid development of artificial intelligence technology, artificial intelligence technology is researched and applied in many fields. In various application programs, products and activities can be promoted to push corresponding interactive information to users. Based on this, a trained machine model can be obtained by using machine model training, and the interactive information with high conversion value can be predicted and pushed according to the trained model.

[0003] At present, after extracting the displayable interactive information, the historical conversion rate and the predicted conversion rate of the interactive information in the historical time can be modeled to obtain a trained conversion rate prediction model. However, different optimization objectives and optimization tasks exist in different application scenarios, and the historical samples under different optimization objectives and optimization tasks also exist differences, thereby resulting in a large prediction error of the trained model when the optimization objectives are different, thereby reducing the accuracy of the conversion rate prediction. SUMMARY

[0004] Therefore, it is necessary to provide a conversion rate prediction method and device, computer equipment and storage medium capable of improving the accuracy of conversion rate prediction.

[0005] In a first aspect, the present application provides a conversion rate prediction method. The method comprises:

[0006] obtaining a directed graph representing the connection relationship between each task node of a target task, the connection relationship being used to represent the continuity of the interactive operation for the task node;

[0007] performing feature extraction on the shared information for each task node in the target task to obtain the initial feature of each task node;

[0008] based on the directed graph, performing feature optimization processing on the initial feature of each task node respectively to obtain the target feature of each task node;

[0009] for each task node, based on the target feature, predicting the operation conversion rate of the interactive operation represented by the task node.

[0010] In a second aspect, the present application further provides a conversion rate prediction device. The device comprises:

[0011] The directed graph acquisition module is configured to acquire a directed graph representing connection relationships between task nodes of a target task, the connection relationships representing continuity of interactive operations for the task nodes;

[0012] The feature extraction module is configured to perform feature extraction on shared information for the task nodes in the target task to obtain initial features of each task node;

[0013] The feature optimization module is configured to perform feature optimization processing on the initial features of each task node based on the directed graph to obtain target features of each task node;

[0014] The conversion rate prediction module is configured to predict, for each task node, an operation conversion rate of an interactive operation represented by the task node based on the target features.

[0015] In a third aspect, the present application also provides a computer device. The computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0016] The directed graph acquisition module is configured to acquire a directed graph representing connection relationships between task nodes of a target task, the connection relationships representing continuity of interactive operations for the task nodes;

[0017] The feature extraction module is configured to perform feature extraction on shared information for the task nodes in the target task to obtain initial features of each task node;

[0018] The feature optimization module is configured to perform feature optimization processing on the initial features of each task node based on the directed graph to obtain target features of each task node;

[0019] The conversion rate prediction module is configured to predict, for each task node, an operation conversion rate of an interactive operation represented by the task node based on the target features.

[0020] In a fourth aspect, the present application also provides a computer readable storage medium. The computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the following steps:

[0021] The directed graph acquisition module is configured to acquire a directed graph representing connection relationships between task nodes of a target task, the connection relationships representing continuity of interactive operations for the task nodes;

[0022] The feature extraction module is configured to perform feature extraction on shared information for the task nodes in the target task to obtain initial features of each task node;

[0023] The feature optimization module is configured to perform feature optimization processing on the initial features of each task node based on the directed graph to obtain target features of each task node;

[0024] For each task node, based on the target feature, the operation conversion rate of the interaction operation represented by the task node is predicted.

[0025] In a fifth aspect, the present application provides a computer program product. The computer program product comprises a computer program which, when executed by a processor, implements the following steps:

[0026] obtaining a directed graph representing the connection relationship between the task nodes of the target task, the connection relationship being used to represent the continuity of the interaction operation for the task nodes;

[0027] performing feature extraction on the shared information for the task nodes in the target task to obtain initial features of each task node;

[0028] performing feature optimization processing on the initial features of each task node based on the directed graph to obtain target features of each task node;

[0029] For each task node, based on the target feature, the operation conversion rate of the interaction operation represented by the task node is predicted.

[0030] The above conversion rate prediction method, device, computer equipment, storage medium and computer program product, obtain a directed graph representing the connection relationship between the task nodes of the target task, the connection relationship being used to represent the continuity of the interaction operation for the task nodes, that is, there is a continuous interaction operation between the task nodes with the connection relationship, then, perform feature extraction on the shared information for the task nodes in the target task to obtain initial features of each task node, at this time, the initial features at least include node attribute information of the task nodes. Perform feature optimization processing on the initial features of each task node based on the directed graph to obtain target features of each task node, the association between each task node can be described through the directed graph, so that the target features of each task node can consider the node features of the adjacent task nodes with the connection relationship with the task nodes on the basis of the node features of the task nodes themselves, so that the optimized target features can consider the association between each interaction operation. Therefore, for each task node, based on the target feature, the operation conversion rate of the interaction operation represented by the task node is predicted, which ensures that the obtained operation conversion rate considers the operation features of the interaction operations with continuity between the interaction operations on the basis of considering the features of the interaction operations themselves, thereby improving the reliability of the obtained operation conversion rate, and further improving the accuracy of the conversion rate prediction. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 An application environment diagram of the conversion rate prediction method in one embodiment;

[0032] Figure 2 A network structure diagram of a model network structure of a conversion rate prediction model in an embodiment;

[0033] Figure 3 A flowchart of a method of conversion rate prediction in an embodiment;

[0034] Figure 4 An embodiment diagram of a task node representing an interactive operation in an embodiment;

[0035] Figure 5 An embodiment diagram of sequentially connecting task nodes to obtain a directed graph in an embodiment;

[0036] Figure 6 An embodiment diagram of outputting initial features corresponding to respective task nodes in an embodiment;

[0037] Figure 7 An embodiment diagram of outputting target features corresponding to respective task nodes in an embodiment;

[0038] Figure 8 An embodiment diagram of outputting operation conversion rates of interactive operations represented by task nodes in an embodiment;

[0039] Figure 9 An embodiment diagram of outputting multi-task conversion rate prediction results in an embodiment;

[0040] Figure 10 A flowchart of obtaining target features of respective task nodes in an embodiment;

[0041] Figure 11 An embodiment diagram of an adjacency matrix matched with a directed graph in an embodiment;

[0042] Figure 12 An embodiment diagram of generating an adjacency matrix in an embodiment;

[0043] Figure 13 A flowchart of a method of conversion rate prediction in another embodiment;

[0044] Figure 14 An embodiment diagram of a relationship degree matrix in an embodiment;

[0045] Figure 15 An embodiment diagram of generating a relationship degree matrix in an embodiment;

[0046] Figure 16 A complete flowchart of a method of conversion rate prediction in an embodiment;

[0047] Figure 17A structural block diagram of a conversion rate prediction device in one embodiment;

[0048] Figure 18 An internal structural diagram of a computer device in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purposes, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0050] With the rapid development of artificial intelligence technology, artificial intelligence technology is researched and applied in many fields. In various application programs, products and activities can be promoted to push corresponding interactive information to users. Based on this, a trained machine model can be obtained by using machine model training, and interactive information with high conversion value can be predicted and pushed according to the trained model. At present, after extracting the displayable interactive information, the historical conversion rate and the predicted conversion rate of the interactive information in the historical time can be modeled to obtain a trained conversion rate prediction model. However, different optimization objectives and optimization tasks exist in different application scenarios, and the historical samples under different optimization objectives and optimization tasks also exist differences, thereby resulting in a large prediction error when the trained model is used in different optimization objectives, thereby reducing the accuracy of the conversion rate prediction.

[0051] Therefore, in order to solve the above problems, the present application provides a conversion rate prediction method capable of improving the efficiency of conversion rate prediction. The conversion rate prediction method provided by the embodiments of the present application can be applied in the application environment as shown in Figure 1 The terminal 102 communicates with the server 104 through the network. The data storage system can store the data required by the server 104 to process. The data storage system can be integrated on the server 104, or placed on the cloud or other servers.

[0052] Specifically, taking the server 104 as an example, the server 104 obtains a directed graph representing the connection relationship between each task node of a target task, the connection relationship is used to represent the continuity of the interactive operation for the task node, performs feature extraction on the shared information for each task node in the target task to obtain an initial feature of each task node, respectively performs feature optimization processing on the initial feature of each task node based on the directed graph to obtain a target feature of each task node, and predicts the operation conversion rate of the interactive operation represented by each task node based on the target feature of each task node. Through the foregoing method, the operation conversion rate obtained is ensured to consider the operation features of the interactive operations that have continuity with the interactive operation itself, thereby improving the reliability of the obtained operation conversion rate and further improving the accuracy of the conversion rate prediction.

[0053] The terminal 102 can be, but is not limited to, various desktop computers, notebook computers, smart phones, tablet computers, Internet of Things devices, and portable wearable devices. The Internet of Things device can be a smart speaker, a smart television, a smart air conditioner, a smart vehicle-mounted device, etc. The portable wearable device can be a smart watch, a smart bracelet, a head-mounted device, etc. The server 104 can be a standalone physical server, a server cluster or a distributed system composed of multiple physical servers, a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDNs, and basic cloud computing services such as big data and artificial intelligence platforms, etc. The terminal and the server can be connected directly or indirectly through wired or wireless communication, which is not limited in the present application. The conversion rate prediction method provided in the embodiments of the present application can be applied to various scenarios, including but not limited to cloud technology, artificial intelligence, etc.

[0054] The conversion rate prediction provided in the embodiments of the present application needs to be performed through a conversion rate prediction model. For ease of understanding, as shown in FIG. 1, the conversion rate prediction model is composed of a feature extraction module 101, a feature optimization module 102, and a conversion rate prediction module 103. Figure 2The model network structure of the conversion rate prediction model shown includes at least a shared embedding layer, a concatenation layer (Concat) & feature dimension reduction layer (Flatten), a deformable convolutional network (DCN), a fully connected layer (Dense) & activation layer (ReLU), a graph convolutional network (GCN), a concatenation layer, and a Sigmoid activation layer. That is, the object basic attribute information, the media basic attribute information, and the media push space information are input into the shared embedding layer, and then sequentially pass through the aforementioned feature processing layers for feature learning to obtain shared information of each task tower. The aforementioned shared information is feature extraction of the object basic attribute information, the media basic attribute information, and the media push space information to obtain corresponding object features, media features, and space features, respectively. The shared information obtained based on the object features, the media features, and the space features, that is, the shared information input into different task towers, is the same. Through the same shared information input into different task towers, the feature expression of different prediction tasks under different task towers is learned, that is, the corresponding initial features are obtained. Then, the initial features output by different task towers are updated and optimized based on the graph neural network to obtain optimized initial features (i.e., target features). Finally, the target features output by each task tower are output through the fully connected layer and the activation layer to obtain the prediction results corresponding to the task tower. Based on this, the prediction results corresponding to each task tower are spliced through the concatenation layer, and the final multi-task conversion rate prediction result is obtained through the Sigmoid activation layer. It can be understood that, Figure 2 The examples in the specification are only for understanding the scheme, that is, the present application does not limit the model type and topology structure used in the model network structure of the conversion rate prediction model, which can be replaced by other effective model structures.

[0055] The conversion rate prediction method provided by the embodiments of the present application also relates to artificial intelligence (AI) technology. The AI is introduced as follows. The AI is to use a digital computer or a machine controlled by a digital computer to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use the knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology of computer science, which tries to understand the essence of intelligence and produce a new intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machine has the functions of perception, reasoning, and decision-making.

[0056] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software level technology. Artificial intelligence basic technology generally includes, such as sensors, special artificial intelligence chips, cloud computing, distributed storage, big data processing technology, pre-training model technology, operation / interaction system, mechatronics, etc. Among them, the pre-training model is also called large model, basic model, which can be widely applied to downstream tasks in each direction of artificial intelligence after fine tuning. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology and machine learning / deep learning, etc.

[0057] The scheme provided by the embodiments of the present application relates to computer vision technology (Computer Vision, CV) and machine learning (Machine Learning, ML) technology under artificial intelligence technology. Among them, computer vision technology is a science that studies how to make machines "see", and further refers to using cameras and computers to replace human eyes to identify and measure targets and other machine vision, and further to do image processing, so that the computer processing becomes an image more suitable for human eye observation or transmission to instrument detection. As a scientific discipline, computer vision researches related theories and technologies, and tries to establish artificial intelligence systems that can obtain information from images or multidimensional data. Large model technology brings important changes to the development of computer vision technology, and pre-training models such as swin-transformer, ViT, V-MOE, and MAE in the visual field can be quickly and widely applied to downstream specific tasks after fine tuning. Computer vision technology usually includes image processing, image recognition, image semantic understanding, image retrieval, OCR, video processing, video semantic understanding, video content / behavior recognition, three-dimensional object reconstruction, 3D technology, virtual reality, augmented reality, simultaneous localization and mapping, and other technologies, as well as common face recognition, fingerprint recognition and other biometric identification technologies.

[0058] Machine learning is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It is a discipline that studies how computers simulate or implement human learning behavior to acquire new knowledge or skills, and reorganize existing knowledge structure to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental approach to making computers intelligent, and its applications are widespread in various fields of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and example-based learning. Pre-training model is the latest development of deep learning, which integrates the above technologies.

[0059] The embodiments are specifically described as follows: in one embodiment, as Figure 3As shown, a conversion rate prediction method is provided, and the method is applied to Figure 1 The server 104 in the above embodiment is taken as an example for illustration, and it can be understood that the method can also be applied to the terminal 102, and can also be applied to a system including the terminal 102 and the server 104, and is implemented through the interaction of the terminal 102 and the server 104. In the embodiment, the method includes the following steps:

[0060] In step 302, a directed graph representing the connection relationship between the task nodes of the target task is obtained, and the connection relationship is used to represent the continuity of the interactive operations for the task nodes.

[0061] The directed graph is composed of task nodes and directed edges between the connected task nodes. The foregoing task nodes correspond to the task towers shown in Figure 2 The task towers correspond to the task nodes, and one task tower is specifically for one interactive operation. Taking the scenario of applying the method to the commodity pushing for the object as an example, different task towers in the commodity pushing scenario can be towers of different pushing targets, and are used for conversion rate prediction of different interactive operations, such as conversion rate prediction of ordering, conversion rate prediction of detail page browsing, and conversion rate prediction of ordering, that is, different conversion rate prediction tasks are estimated by different task towers. Based on this, the interactive operations can include “click”, “detail page browsing”, and “ordering”, so there can be a task tower A1 corresponding to an interactive operation B1 “click”, a task tower A2 corresponding to an interactive operation B2 “detail page browsing”, and a task tower A3 corresponding to an interactive operation B3 “ordering”. As can be seen, the task nodes corresponding to the task towers can also represent an interactive operation, for example, a task node C1 corresponds to a task tower A1, that is, the task node C1 represents an interactive operation B1 “click”. Similarly, a task node C2 corresponds to a task tower A2, that is, the task node C2 represents an interactive operation B2 “detail page browsing”. A task node C3 corresponds to a task tower A3, that is, the task node C3 represents an interactive operation B3 “ordering”.

[0062] Based on this, there is continuity between the interactive operations represented by the task nodes, and since the continuity is used to describe that two interactive operations are adjacent and direct interactive operations, and there is no other interactive operation between the interactive operations. As can be seen, there is continuity between the represented interactive operations, that is, there is a connection relationship between the task nodes, there is a direct connection between the task nodes with a connection relationship, and the pointing direction of the directed edge is determined according to the order of the continuous interactive operations. Considering that there can be multiple interactive repetitions between two interactive operations in actual application, therefore, the directed edge between the task nodes can be a one-way directed edge, or a two-way directed edge, which is not limited here.

[0063] For example, in the object purchase scenario, the interaction operations performed to complete a one-time order purchase include: "clicking" and then "detail page browsing", and then "ordering". There is continuity between the interaction operation "clicking" and the interaction operation "detail page browsing", and there is continuity between the interaction operation "detail page browsing" and the interaction operation "ordering". Based on the foregoing example, it is known that there is a connection relationship between the task node C1 representing the interaction operation B1 "clicking" and the task node C2 representing the interaction operation B2 "detail page browsing". Since the pointing direction of the directed edge is determined according to the order of the continuous interaction operations, the directed edge between the task node C1 and the task node C2 is specifically: from the task node C1 to the task node C2. Similarly, there is a connection relationship between the task node C2 representing the interaction operation B2 "detail page browsing" and the task node C3 representing the interaction operation B3 "ordering". Similarly to the foregoing example, the directed edge between the task node C2 and the task node C3 is specifically: from the task node C2 to the task node C3.

[0064] On the contrary, since there is the interaction operation "detail page browsing" between the interaction operation "clicking" and the interaction operation "ordering", that is, the interaction operation "clicking" and the interaction operation "ordering" are not continuous interaction operations, that is, there is no continuity between the interaction operation "clicking" and the interaction operation "ordering". Therefore, there is no connection relationship between the task node C1 representing the interaction operation B1 "clicking" and the task node C3 representing the interaction operation B3 "ordering".

[0065] Specifically, the server obtains a directed graph representing the connection relationship between each task node of the target task. The directed graph obtained by the server can be a directed graph generated in real time by the server through the continuity of the interaction operation represented by each task node recorded in the historical operation record in the case of needing to perform conversion rate prediction. It can also be that the directed graph is generated through the historical operation record in the historical time period, and the generated directed graph is stored in the data storage system, so that the server obtains the directed graph from the data storage system with network connection relationship in the case of needing to perform conversion rate prediction. In this embodiment, the generation of the directed graph is not directly defined, nor is the directed graph obtained from the data storage system, or the directed graph is obtained in other ways.

[0066] For the convenience of understanding the directed graph, the interaction operations performed to complete a one-time order purchase include: the interaction operation "clicking", the interaction operation "detail page browsing", and the interaction operation "ordering", as shown in FIG. 2. Figure 4As shown, the directed graph includes a task node 402 representing the interaction operation "click", a task node 404 representing the interaction operation "detail page browsing", and a task node 406 representing the interaction operation "ordering". A directed edge 408 is directed from the task node 402 to the task node 404, and a directed edge 410 is directed from the task node 404 to the task node 406. The directed edge 408 can describe that the task node 402 representing the interaction operation "click" is connected to the task node 404 representing the interaction operation "detail page browsing", and the direction of the directed edge 408 can further represent that the operation order between the interaction operation "click" and the interaction operation "detail page browsing" is "click" first and then "detail page browsing". Similarly, the directed edge 410 can describe that the task node 404 representing the interaction operation "detail page browsing" is connected to the task node 406 representing the interaction operation "ordering", and the direction of the directed edge 410 can further represent that the operation order between the interaction operation "detail page browsing" and the interaction operation "ordering" is "detail page browsing" first and then "ordering".

[0067] It can be understood that, regardless of the direct or indirect way of obtaining the directed graph, the continuity of the interaction operations recorded by the historical operation record is required to generate the directed graph. Therefore, the following will introduce how to obtain the directed graph. In one embodiment, obtaining the directed graph representing the connection relationship between the task nodes of the target task includes: determining the continuity of the interaction operation represented by each task node and determining the operation order of the continuous operations between each task node based on the historical operation record; and connecting the task nodes in sequence based on the operation order to obtain the directed graph.

[0068] The historical operation record is used to record the continuous interaction operations in a historical time period, and the historical operation record can be collected for a target object or a target push object, which is not limited herein. For example, the historical operation record D1 is: the interaction operation "click", the interaction operation "detail page browsing", and the interaction operation "ordering". The historical operation record D2 is: the interaction operation "click", the interaction operation "detail page browsing", the interaction operation "jumping to a landing page", and the interaction operation "ordering". The historical operation record D3 is: the interaction operation "click", the interaction operation "detail page browsing", the interaction operation "jumping to a mobile program", and the interaction operation "ordering".

[0069] Therefore, the sequence of operations is used to describe the order of operations among multiple consecutive interactive operations recorded in the historical operation log. For ease of understanding, taking historical operation log D1 as an example, the sequence of operations is: interactive operation "click" → interactive operation "view details page" → interactive operation "place order". Similarly, taking historical operation log D2 as an example, the sequence of operations is: interactive operation "click" → interactive operation "view details page" → interactive operation "jump to landing page" → interactive operation "place order".

[0070] Specifically, the server uses historical operation records to track continuous interactive operations within a historical time period to determine the continuity of the interactive operations represented by each task node. This determines the sequential order of the interactive operations represented by each task node. Then, based on this sequential order, the matching task nodes are connected sequentially to obtain a directed graph. For ease of understanding, let's take the aforementioned example. Using historical operation record D1 as an example, the sequential order of operations is: "Click" → "View Details Page" → "Place Order". Task node C1 represents the "Click" interactive operation, and task node C2 represents the "View Details Page" interactive operation. Therefore, connecting task nodes C1 and C2 sequentially results in a directed edge pointing from task node C1 to task node C2. Similarly, task node C2 represents the "Place Order" interactive operation, and connecting task nodes C2 and C3 sequentially results in a directed edge pointing from task node C2 to task node C3.

[0071] The preceding example only illustrates the connection between task nodes for interactive operations recorded in historical operation record D1. However, there are usually multiple historical operation records. For further understanding, historical operation record D1, historical operation record D2, and historical operation record D3 are used as examples. Figure 5 As shown, for the interactive operations "click", "view details page", and "place order" recorded in the historical operation record D1, the interactive operation "click" is represented by task node 502, the interactive operation "view details page" is represented by task node 504, and the interactive operation "place order" is represented by task node 506. As can be seen from the order of operations in the historical operation record D1 in the previous example, task nodes 502, 504 and 506 are connected in sequence.

[0072] Similarly, for the interaction operations "click", "detail page browsing", "jump to landing page", and "order" recorded in the historical operation record D2, a task node 508 representing the interaction operation "jump to landing page" needs to be added, and then the task node 502, the task node 504, the task node 508, and the task node 506 are connected in sequence according to the operation sequence in the historical operation record D2 in the foregoing example. For the interaction operations "click", "detail page browsing", "jump to mobile program", and "order" recorded in the historical operation record D3, a task node 510 representing the interaction operation "jump to mobile program" needs to be added, and then the task node 502, the task node 504, the task node 510, and the task node 506 are connected in sequence according to the operation sequence in the historical operation record D3 in the foregoing example. Through the foregoing manner, the directed graph shown in FIG. 8 can be obtained. Figure 5

[0073] In step 304, the shared information in the target task for each task node is feature extracted to obtain the initial feature of each task node.

[0074] The target task is specifically a conversion rate prediction task, and the target task includes multiple sub-tasks for predicting the conversion rate of each interaction operation, that is, one sub-task has a corresponding task tower (that is, a task node). Secondly, the shared information is specifically shared features obtained by feature processing of object basic attribute information, media basic attribute information, and media pushing space information, that is, the features (that is, shared information) input for different task nodes are the same. Secondly, the directed graph includes multiple task nodes , and each task node , then the initial feature of each task node , is a feature vector.

[0075] Specifically, the server inputs the shared information for each task node, that is, the object basic attribute information, the media basic attribute information, and the media pushing space information are obtained from the task data of the target task, and then the obtained object basic attribute information, the media basic attribute information, and the media pushing space information are feature processed to obtain the shared information.

[0076] ​​Based on this, the following will specifically introduce how to obtain the sharing information: in a specific embodiment, the way of obtaining the sharing information further includes: obtaining object basic attribute information, media basic attribute information and media push space information from the task data of the target task; performing feature extraction on the object basic attribute information, the media basic attribute information and the media push space information respectively to obtain corresponding object features, media features and space features respectively; performing feature splicing processing on the object features, the media features and the space features to obtain spliced features; and performing feature cross processing on the spliced features to obtain the sharing information.

[0077] The task data of the target task is historical data and related object data collected for the target task. For example, the target task is a task of predicting the conversion rate of commodity 1, and the historical data of the target task can include objects that have interacted with commodity 1, push information corresponding to commodity 1, and a specific push space of commodity 1. Therefore, the task data at least includes object basic attribute information, media basic attribute information and media push space information.

[0078] Based on this, the object basic attribute information is at least used to describe: basic attributes of the pushed object, historical operation behavior information of the pushed object and interest information of the pushed object. The basic attributes of the pushed object at least include: object age, object gender, etc. The historical operation behavior information of the pushed object at least includes: media information (such as browsing advertisements) that has performed historical interactive operation, commodities (such as ordering commodities or browsing commodities, etc.) that have performed historical interactive operation. And the interest information of the pushed object can be interest features mined, such as media identifiers, commodity identifiers or commodity type identifiers that the pushed object is interested in, etc.

[0079] The media basic attribute information at least includes media basic information, media material information and media statistical information. The media basic information is media value and media delivery target. The media material information is information corresponding to the material in the delivered media. And the media statistical information can be historical conversion amount of the media obtained through historical data, etc. Secondly, the media push space information can also be referred to as context feature. The media push space information refers to the context information of the push position where the media push is performed, such as the push position of the media push, the surrounding environment of the push position of the media push, etc. Here, the object basic attribute information, the media basic attribute information and the media push space information are not specifically limited and should be flexibly determined based on different application scenarios and application requirements.

[0080] Specifically, the server takes the object basic attribute information, the media basic attribute information and the media push space information as the input of the prediction model for the target task, performs feature coding through the shared embedding layer to obtain object features corresponding to the object basic attribute information, media features corresponding to the media basic attribute information, and space features corresponding to the media push space information, takes the obtained object features, media features and space features as the input of the concatenation and flattening layer, that is, performs feature concatenation on the object features, media features and space features through the concatenation layer, and then performs feature dimension reduction processing on the features obtained after the feature concatenation to obtain the reduced concatenation features. Based on this, the reduced concatenation features are input into the deformable convolution layer for feature cross processing, and the cross features are obtained after the feature cross processing through the deformable convolution layer. Finally, the cross features are input into the dense and ReLU layer for dense connection processing and non-linear connection processing to obtain the shared features that can be shared by all task towers, that is, the shared information (features that can be shared by all task towers) is output through the Dense&ReLU layer.

[0081] Further, after obtaining the shared information through the foregoing manner, the shared information is input into different task nodes (i.e., task towers) for differentiated learning. Each task tower is composed of a linear layer and a nonlinear layer, so as to be able to output the initial features of each task node for the task optimization target of different task nodes. For ease of understanding, as shown in Figure 6 the object basic attribute information, the media basic attribute information and the media push space information are taken as the input of the shared embedding layer, the shared information is obtained through the foregoing manner, and the shared information is input into different task nodes, that is, the shared information is input into the task node 601, the task node 602, the task node 603 and the task node 604 respectively. Different task nodes output corresponding initial features through different task optimization targets. That is, the task node 601 outputs the initial features 605, the task node 602 outputs the initial features 606, the task node 603 outputs the initial features 607, and the task node 604 outputs the initial features 608.

[0082] In step 306, the initial features of each task node are respectively processed for feature optimization based on the directed graph to obtain the target features of each task node.

[0083] The feature optimization processing is specifically performed by a graph convolution network, and the feature optimization processing is specifically to optimize and update the initial features of the task nodes in view of the connection relationship between the task nodes in the directed graph. Therefore, the graph convolution network can be regarded as a feature extractor in this embodiment, that is, the initial features of the task nodes are taken as the node attribute features of the task nodes, and then the target features of the task nodes are extracted based on the connection relationship between the task nodes represented by the directed graph. Specifically, the server performs feature optimization processing on the initial features of each task node based on the connection relationship between the task nodes represented by the directed graph, thereby obtaining the target features of each task node. That is, the server inputs the directed graph and the initial features of each task node into the graph convolution network, so that the graph convolution network optimizes and predicts the node feature representation of the task nodes based on the connection relationship between the task nodes represented by the directed graph and the initial features of each task node as the node attribute features of the task nodes, and thereby outputs the target features of each task node after feature representation. Therefore, the connection relationship between different task nodes in the directed graph can be considered by the graph convolution network, and the receptive field is larger, so that more comprehensive global features can be extracted for each task node, so as to improve the accuracy and integrity of the obtained target features.

[0084] For ease of understanding, as shown in FIG. 7, Figure 7 The shared information is input into the task nodes 701, 702, 703 and 704, so that the corresponding initial features 705-708 are output by the task nodes 701-704, respectively. Based on this, the directed graph 709 representing the connection relationship between the task nodes 701, 702, 703 and 704 can also be obtained by the method introduced in the foregoing embodiments, and thereby the graph convolution network optimizes the feature representation of the initial features 705-708 corresponding to the task nodes 701-704 based on the connection relationship represented by the directed graph 709, to output the target features 710-713 corresponding to the task nodes 701-704, respectively.

[0085] In step 308, for each task node, the operation conversion rate of the interaction operation represented by the task node is predicted based on the target feature.

[0086] The target feature of each task node can be used to implement node classification, graph classification, edge prediction, and also obtain an embedding representation of the graph. In this embodiment, the target task is specifically a conversion rate prediction task, and therefore the target feature is specifically used to: predict the operation conversion rate of the interaction operation represented by each task node. Specifically, the server performs prediction of the operation conversion rate for each task node, that is, the server performs conversion rate prediction based on the target feature corresponding to each task node, to obtain the operation conversion rate of the interaction operation represented by each task node.

[0087] In one specific embodiment, for each task node, the operation conversion rate of the interaction operation represented by the task node is predicted based on the target feature, including: for each task node, performing dense connection processing and non-linear connection processing on the target feature to obtain the operation conversion rate of the interaction operation represented by the task node. Based on this, the server inputs the target feature corresponding to each task node into a dense & ReLU layer, at this time the dense & ReLU layer performs dense connection processing and non-linear connection processing on the target feature corresponding to each task node, thereby outputting the operation conversion rate of the interaction operation represented by each task node through the Dense & ReLU layer.

[0088] For ease of understanding, as shown in Figure 8 For each of the task nodes 801 to 804, the target feature corresponding to each of the task nodes 801 to 804 can be obtained in the manner introduced in the foregoing embodiments. Then, the target feature 805 of the target node 801 is input into a dense & ReLU layer to obtain the operation conversion rate 809 of the interaction operation represented by the target node 801. Similarly, the target feature 806 of the target node 802 is input into the dense & ReLU layer to obtain the operation conversion rate 810 of the interaction operation represented by the target node 802. The target feature 807 of the target node 803 is input into the dense & ReLU layer to obtain the operation conversion rate 811 of the interaction operation represented by the target node 803. The target feature 808 of the target node 804 is input into the dense & ReLU layer to obtain the operation conversion rate 812 of the interaction operation represented by the target node 804.

[0089] Further, in one optional embodiment, the method for conversion rate prediction further includes: performing splicing processing on the obtained operation conversion rate of the interaction operation represented by each task node to obtain a multi-task operation conversion rate prediction result for the target task.

[0090] Specifically, the server converts the target feature full connection layer & activation layer of each task node to output the operation conversion rate of the represented interactive operation through the foregoing embodiment mode, and then performs splicing processing on the operation conversion rate of the represented interactive operation to make the spliced operation conversion rate pass through the Sigmoid activation layer to obtain the final multi-task conversion rate prediction result. For ease of understanding, as shown in Figure 9 The foregoing embodiment introduces that the operation conversion rates 905 to 908 of the represented interactive operations of the task nodes 901 to 904 are obtained through the foregoing embodiment, and then the operation conversion rates 905 to 908 of the represented interactive operations of the task nodes 901 to 904 are spliced through the splicing layer, and then the spliced operation conversion rate output by the splicing processing is input to the Sigmoid activation layer, and the multi-task conversion rate prediction result 909 is output through the Sigmoid activation layer.

[0091] It should be understood that the corresponding examples in the embodiments of the present application are used to understand the present scheme, but should not be understood as a specific limitation of the present scheme.

[0092] In the above conversion rate prediction method, a directed graph representing the connection relationship between each task node of a target task is obtained, and the connection relationship is used to represent the continuity of the interactive operation for the task node, that is, there is a continuous interactive operation between the task nodes with the connection relationship. Then, the shared information for each task node in the target task is extracted to obtain the initial feature of each task node, and the initial feature at least includes the node attribute information of the task node. Based on the directed graph, the initial feature of each task node is respectively processed by feature optimization to obtain the target feature of each task node. The directed graph can describe the association between each task node, so that the target feature of each task node can consider the node features of the adjacent task nodes having the link relationship with the task node on the basis of the node features of the task node itself, so that the optimized target feature can consider the association between each interactive operation. Therefore, based on the target feature, the operation conversion rate of the interactive operation represented by each task node is predicted, which ensures that the obtained operation conversion rate considers the operation features of the interactive operation having the continuity with the interactive operation on the basis of considering the features of the interactive operation itself, thereby improving the reliability of the obtained operation conversion rate and further improving the accuracy of the conversion rate prediction.

[0093] The foregoing embodiment introduces that the operation conversion rate of the represented interactive operation through the foregoing embodiment, and then performs splicing processing on the operation conversion rate of the represented interactive operation to make the spliced operation conversion rate pass through the Sigmoid activation layer to obtain the final multi-task conversion rate prediction result. For ease of understanding, as shown in Figure 10As shown, based on the directed graph, feature optimization is performed on the initial features of each task node to obtain the target features of each task node, including:

[0094] Step 1002: Based on the connection relationships between task nodes in the directed graph, generate an adjacency matrix that matches the directed graph.

[0095] As described in the foregoing embodiments, feature optimization processing is specifically performed using the connection relationships between task nodes represented by the directed graph. Therefore, this embodiment specifically uses an adjacency matrix matching the directed graph to describe the connection relationships between task nodes. Thus, the adjacency matrix matching the directed graph is used to represent the connection relationships between task nodes in the directed graph. Therefore, the adjacency matrix can be a 0-1 matrix or a weight matrix, and the specific adjacency matrix matching the directed graph is... , This represents the number of nodes in the task node of the directed graph; therefore, the adjacency matrix that matches the directed graph is specifically... matrix.

[0096] Therefore, when the adjacency matrix is ​​a 0-1 matrix, meaning its elements contain only 0 or 1, a 1 indicates a connection between task nodes, while a 0 indicates no connection. Furthermore, a task node cannot be connected to itself; therefore, all diagonal elements of the adjacency matrix are 0. For ease of understanding, let's take a 0-1 adjacency matrix as an example, such as... Figure 11 As shown, based on Figure 11 The directed graph shown in Figure (A) can be obtained Figure 11 Figure (B) shows the adjacency matrix that matches the directed graph. This adjacency matrix is ​​a 0-1 matrix.

[0097] Specifically, the server determines the element values corresponding to the connection relationships between the task nodes in the directed graph based on the connection relationships between the task nodes in the directed graph, and determines the element values corresponding to the connection relationships between the task nodes as matrix elements, so as to form an adjacency matrix matched with the directed graph through the aforementioned matrix elements. It can be known that the construction of the adjacency matrix needs to consider the connection relationships between the task nodes, that is, the matrix elements in the adjacency matrix need to be determined through the connection relationships between the task nodes. The following will introduce the manner of determining the matrix elements and constructing the adjacency matrix. In one specific embodiment, the adjacency matrix matched with the directed graph is generated based on the connection relationships between the task nodes in the directed graph, and the method comprises: for each task node in the directed graph, determining a first type of task node having a connection relationship with the task node, and a second type of task node not having a connection relationship with the task node; determining a first element value representing a connection relationship type between the task node and the first type of task node, and a second element value representing a connection relationship type between the task node and the second type of task node; and constructing the adjacency matrix matched with the directed graph by taking the first element value and the second element value as matrix elements based on the first type of task node and the second type of task node corresponding to each task node.

[0098] In the directed graph, for each task node, there are the first type of task node and the second type of task node, the first type of task node has a connection relationship with the task node, and the second type of task node does not have a connection relationship with the task node. It can be understood that the task node itself does not have a connection relationship with the task node itself, and therefore the task node itself can be determined as the second type of task node of the task node.

[0099] Secondly, the first element value is used to represent the connection relationship type between the task node and the first type of task node, and the second element value is used to represent the connection relationship type between the task node and the second type of task node. It can be known that the first element value is used to represent that there is a connection relationship between the task node and the first type of task node, and the second element value is used to represent that there is no connection relationship between the task node and the first type of task node. If the adjacency matrix is a 0-1 matrix, the first element value can be determined as 1, and the second element value can be determined as 0.

[0100] Specifically, the server first determines, for each task node in the directed graph, the task nodes having a connection relationship with the task node as the first type of task node and the task nodes not having a connection relationship with the task node as the second type of task node through the connection relationships between each task node represented by the directed graph. For the sake of understanding, the following will take the task node A as an example. Figure 11For example, as shown in FIG. A, there are task node 1, task node 2, task node 3, task node 4 and task node 5. For task node 1, there is a connection relationship between task node 1 and task node 2, and there is no connection relationship between task node 1 and task node 3, task node 4 and task node 5. Therefore, the first type of task node of task node 1 is task node 2, and the second type of task node of task node 1 is task node 3, task node 4 and task node 5. In addition, task node 1 can also be determined as the second type of task node of task node 1 itself.

[0101] Similarly, for task node 2, there is a connection relationship between task node 2 and task node 1, task node 3, task node 4 and task node 5. Therefore, the first type of task node of task node 2 is task node 1, task node 3, task node 4 and task node 5. In this case, task node 2 can also be determined as the second type of task node of task node 2 itself. Similarly, the first type of task node of task node 3 is task node 2, task node 4 and task node 5, and the second type of task node of task node 3 is task node 1. In addition, task node 3 can also be determined as the second type of task node of task node 3 itself.

[0102] Similarly, the first type of task node of task node 4 is task node 2 and task node 3, and the second type of task node of task node 4 is task node 1 and task node 5. In addition, task node 4 can also be determined as the second type of task node of task node 4 itself. Similarly, the first type of task node of task node 5 is task node 2 and task node 3, and the second type of task node of task node 5 is task node 1 and task node 4. In addition, task node 5 can also be determined as the second type of task node of task node 5 itself.

[0103] Further, the server can determine the first type of task node and the second type of task node for each task node. Therefore, the server determines the first element value representing the type of connection relationship between the task node and the first type of task node, and the second element value representing the type of connection relationship between the task node and the second type of task node. That is, the first element value describes that there is a connection relationship between the task node and the first type of task node, and the second element value describes that there is no connection relationship between the task node and the second type of task node. For ease of understanding, the following table 1 shows the first element value and the second element value of the connection relationship between the task node and the first type of task node and the second type of task node. Figure 11As shown in the directed graph in FIG. (A) and in combination with the foregoing example, the first type of task node of task node 1 is task node 2, and the second type of task node of task node 1 is task node 3, task node 4, and task node 5. Task node 1 can also be determined as the second type of task node of task node 1 itself. Thus, the first element value is between task node 1 and task node 2, and the second element value is between task node 1 and task node 1, task node 3, task node 4, and task node 5.

[0104] Similarly, the first type of task node of task node 2 is task node 1, task node 3, task node 4, and task node 5. Task node 2 can also be determined as the second type of task node of task node 2 itself. Thus, the first element value is between task node 2 and task node 1, task node 3, task node 4, and task node 5, and the second element value is between task node 2 and task node 2. By analogy, the first element value is between task node 3 and task node 2, task node 4, and task node 5, and the second element value is between task node 3 and task node 1 and task node 3. Similarly, the first element value is between task node 4 and task node 2 and task node 3, and the second element value is between task node 4 and task node 1, task node 4, and task node 5. Also, the first element value is between task node 5 and task node 2 and task node 3, and the second element value is between task node 5 and task node 1, task node 4, and task node 5.

[0105] Further, the server constructs an adjacency matrix matched with the directed graph based on the first type of task node and the second type of task node corresponding to each task node, and the obtained first element value and second element value are matrix elements. That is, for each task node, a matrix row is constructed in the matrix, and the matrix column describes each task node. At this time, the obtained adjacency matrix is The matrix, and represents the number of nodes of the task nodes in the directed graph. That is, for each task node, the corresponding element value is filled in the matrix element position corresponding to the task node and other task nodes, thereby completing the construction of the adjacency matrix.

[0106] For ease of understanding, hereinafter, the directed graph shown in FIG. (A) is taken as an example, and the first element value is 1 and the second element value is 0. Specifically, as shown in FIG. (B), the first type of task node of task node 1 is task node 2, and the second type of task node of task node 1 is task node 3, task node 4, and task node 5. The first type of task node of task node 2 is task node 1, task node 3, task node 4, and task node 5. The first type of task node of task node 3 is task node 2, task node 4, and task node 5. The first type of task node of task node 4 is task node 2 and task node 3. The first type of task node of task node 5 is task node 2 and task node 3. Figure 11 Figure 12 ​As shown, according to the foregoing example, the first element value is between the task node 1 and the task node 2, and the second element value is between the task node 1 and the task node 1, the task node 3, the task node 4, and the task node 5. Therefore, in the matrix row 1201 corresponding to the task node 1, the matrix element position corresponding to the task node 1 and the task node 1 is filled with "0", the matrix element position corresponding to the task node 1 and the task node 2 is filled with "1", the matrix element position corresponding to the task node 1 and the task node 3 is filled with "0", the matrix element position corresponding to the task node 1 and the task node 4 is filled with "0", and the matrix element position corresponding to the task node 1 and the task node 5 is filled with "0".

[0107] Similarly, since the first element value is between the task node 2 and the task node 1, the task node 3, the task node 4, and the task node 5, and the second element value is between the task node 2 and the task node 2. Therefore, in the matrix row 1202 corresponding to the task node 2, the matrix element position corresponding to the task node 2 and the task node 1 is filled with "1", the matrix element position corresponding to the task node 2 and the task node 2 is filled with "0", the matrix element position corresponding to the task node 2 and the task node 3 is filled with "1", the matrix element position corresponding to the task node 2 and the task node 4 is filled with "1", and the matrix element position corresponding to the task node 2 and the task node 5 is filled with "1". The matrix rows corresponding to the remaining task nodes 3 to 5 are filled in turn, so that the adjacency matrix matched with the directed graph as shown in Figure 12 is obtained. Figure 11 The adjacency matrix matched with the directed graph as shown in the (B) of FIG. 1 is obtained.

[0108] In an optional embodiment, the adjacency matrix matched with the directed graph is generated based on the connection relationship between the task nodes in the directed graph, comprising: generating an initial adjacency matrix matched with the directed graph based on the connection relationship between the task nodes in the directed graph; and performing matrix splicing processing on the initial adjacency matrix and a unit matrix to obtain the adjacency matrix.

[0109] The unit matrix is a square matrix, and the elements on the diagonal line from the upper left corner to the lower right corner (referred to as the main diagonal line) are all 1, and all other elements are 0. The number of matrix columns and the number of matrix rows of the unit matrix are consistent with the initial adjacency matrix. That is, based on the initial adjacency matrix , the unit matrix is also , and the elements on the diagonal line from the upper left corner to the lower right corner are all 1, and all other elements are 0.

[0110] Specifically, the server determines, for each task node in the directed graph, a first type of task node having a connection relationship with the task node and a second type of task node not having a connection relationship with the task node, determines a first element value representing a connection relationship type between the task node and the first type of task node and a second element value representing a connection relationship type between the task node and the second type of task node, and constructs an initial adjacency matrix matching the directed graph based on the first type of task node and the second type of task node corresponding to each task node, with the first element value and the second element value as matrix elements.

[0111] Further, considering that the elements on the diagonal from the upper left corner to the lower right corner of the initial adjacency matrix are all 1, but the initial features of the task nodes should also be optimized to represent, a unit matrix is introduced to learn the initial features of the task nodes, and therefore, the server performs matrix splicing processing on the initial adjacency matrix and the unit matrix to obtain the adjacency matrix. That is, the server performs parameter multiplication processing on the unit matrix and a preset parameter, and then adds the unit matrix after the parameter multiplication processing to the initial adjacency matrix to obtain the adjacency matrix.

[0112] To facilitate understanding of the above-mentioned matrix splicing processing method, as formula (1):

[0113] ; (1)

[0114] Wherein, represents the adjacency matrix, represents the initial adjacency matrix, represents the unit matrix, represents a preset parameter.

[0115] Step 1004, based on the initial features of each task node and the adjacency matrix, performing feature optimization processing to obtain the target features of each task node.

[0116] Specifically, the server more completely represents the connection relationship between each task node through the adjacency matrix, and thus performs feature optimization processing on the initial features of each task node, thereby obtaining the target features of each task node. That is, the server performs feature optimization processing on the initial features based on the adjacency matrix through the graph convolution network, thereby outputting the target features of each task node through the graph convolution network.

[0117] It should be understood that the corresponding examples in the embodiments of the present application are used to understand the present scheme, but should not be understood as a specific limitation of the present scheme.

[0118] In this embodiment, the connection relationships between each task node represented by the directed graph are transformed into an adjacency matrix that provides a more comprehensive and clear understanding. Considering that the initial adjacency matrix obtained solely through connection relationships does not include the task itself, but the initial features of the task nodes should still be optimized, an identity matrix is ​​introduced to further learn the initial features of the task nodes. This results in the target features being able to more comprehensively and completely describe the features of the interactive operation of this application. Furthermore, considering the correlation between each interactive operation, the reliability of the operation conversion rate obtained based on the target features is improved, thereby enhancing the accuracy of the conversion rate prediction.

[0119] Considering that in one embodiment, such as Figure 13 As shown, conversion rate prediction methods also include:

[0120] Step 1302: Obtain the relation degree matrix through the connection relationship. The matrix elements of the relation degree matrix are used to represent the number of task nodes that have a connection relationship with each task node.

[0121] Specifically, the relation degree matrix is ​​as follows: , Represents task nodes Task nodes with connection relationships Therefore, the elements of the relation degree matrix are used to represent the number of task nodes that have connections with each task node. Since the number of task nodes with connections is determined for each task node, the matrix elements in the relation degree matrix are actually distributed along the diagonal from the top left to the bottom right. Except for the elements along the diagonal from the top left to the bottom right, all elements in the relation degree matrix are 0, and the values ​​of the elements along the diagonal from the top left to the bottom right represent the number of task nodes with connections. For ease of understanding, let's consider... Figure 11 Taking the directed graph shown in Figure (A) as an example, we can obtain the following: Figure 14 The relation degree matrix shown is shown.

[0122] Specifically, the server obtains a relation degree matrix matching the directed graph through connection relationships. That is, for each task node, the server first determines the number of task nodes connected to it through the connection relationships represented by the directed graph. Similar to the previous embodiment, for each task node in the directed graph, the server determines the first type of task nodes connected to it, and then counts the number of these first-type task nodes. The number of first-type task nodes is then the total number of task nodes connected to the task node. For example, further illustrating the previous example, since task node 1's first-type task node is task node 2, the number of first-type task nodes for task node 1 is 1. Similarly, since task node 2's first-type task nodes are: task node 1, task node 3, task node 4, and task node 5, the number of first-type task nodes for task node 2 is 4. And so on, task node 3's first-type task nodes are: task node 2, task node 4, and task node 5, so the number of first-type task nodes for task node 3 is 3. And task node 4's first-type task nodes are: task node 2 and task node 3, so the number of first-type task nodes for task node 4 is 2. Furthermore, the first type of task nodes for task node 5 are task node 2 and task node 3, therefore the number of first type task nodes for task node 5 is 2.

[0123] Furthermore, the server uses the number of task nodes with which a task node has a connection as matrix elements to generate a relation degree matrix that matches the directed graph. That is, it fills the corresponding matrix element positions (i.e., the number of task nodes with a connection) between each task node and the task node itself, thus completing the construction of the relation degree matrix. For ease of understanding, we will use... Figure 11 Taking the directed graph shown in Figure (A) as an example, the specific details are as follows: Figure 15 As shown in the previous example, the quantity of the first category of task node 1 is 1. Therefore, in matrix row 1501 corresponding to task node 1, the matrix element position 1502 corresponding to task node 1 is filled with the quantity of the first category of task node 1, "1". Similarly, since the quantity of the first category of task node 2 is 4, in matrix row 1503 corresponding to task node 2, the matrix element position 1504 corresponding to task node 2 is filled with the quantity of the first category of task node 1, "4". This process is repeated for the remaining matrix rows corresponding to task nodes 3 through 5, resulting in the following... Figure 15 The relation degree matrix shown is the one that matches the directed graph (also known as the relation degree matrix). Figure 14 The figure shows the relation degree matrix that matches the directed graph.

[0124] Based on this, feature optimization is performed on the initial features and adjacency matrix of each task node to obtain the target features of each task node, including:

[0125] In step 1304, the feature optimization processing is performed based on the initial feature of each task node, the adjacency matrix and the relationship degree matrix, to obtain the target feature of each task node.

[0126] Specifically, the server performs the feature optimization processing based on the initial feature of each task node, the adjacency matrix and the relationship degree matrix, to obtain the target feature of each task node, that is, the server performs the feature optimization processing on the initial feature of each task node based on the connection relationship between the task nodes represented by the adjacency matrix and the number of task nodes having the connection relationship with each task node represented by the relationship degree matrix through the graph convolution network, so as to output the target feature of each task node through the graph convolution network.

[0127] The specific manner of how to perform the feature optimization processing based on the initial feature, the adjacency matrix and the relationship degree matrix will be described in detail below. In one specific embodiment, the feature optimization processing is performed based on the initial feature of each task node, the adjacency matrix and the relationship degree matrix, to obtain the target feature of each task node, including: determining the node number of the task nodes included in the directed graph and the feature dimension of each initial feature; obtaining the node feature vector matrix based on the initial feature of each task node, the node number and the feature dimension; and performing the feature optimization processing through the node feature vector matrix, the adjacency matrix and the relationship degree matrix, to obtain the target feature of each task node.

[0128] Specifically, the node number of the task nodes included in the directed graph and the feature dimension of each initial feature are counted first. For example, the node number of the task nodes included in the directed graph is counted based on the initial feature of each task node, the node number and the feature dimension. Figure 11 For example, the directed graph shown in FIG. A includes task node 1, task node 2, task node 3, task node 4 and task node 5, so that the node number of the task nodes included in the directed graph is counted as 5. The feature dimension of the initial feature is flexibly determined based on the actual application requirement, which is not specifically limited here.

[0129] Further, the server obtains the node feature vector matrix based on the initial feature of each task node, the node number and the feature dimension, that is, the node feature vector matrix of each task node is obtained through the initial feature of each task node, the node number and the feature dimension. The construction manner of the node feature vector matrix of the task node is specifically as formula (2):

[0130] (2)

[0131] wherein, represents the node feature vector matrix of the task node, represents the node number, representing a feature dimension, representing an initial feature of each task node in the directed graph.

[0132] In the following, how to obtain the target feature by optimizing the feature in the foregoing manner will be detailed. In one specific embodiment, the feature optimization is performed by using the node feature vector matrix, the adjacency matrix, and the relationship degree matrix to obtain the target feature of each task node, including: determining the total adjacency feature of each task node by using the node feature vector matrix and the adjacency matrix, and determining the average adjacency feature of each task node by using each total adjacency feature and the relationship degree matrix; performing feature normalization based on the initial feature of each task node and the average adjacency feature of each task node to obtain the target feature of each task node.

[0133] Specifically, the server determines the total adjacency feature of each task node by using the node feature vector matrix and the adjacency matrix, that is, the server considers the adjacent task nodes having a connection relationship with each task node, and determines the node features of the adjacent task nodes having a connection relationship with each task node from the node feature vector matrix by using the adjacency matrix, thereby obtaining the total adjacency feature. The manner of obtaining the total adjacency feature will be described in detail below: in one specific embodiment, the total adjacency feature of each task node is determined by using the node feature vector matrix and the adjacency matrix, including: performing matrix multiplication on the node feature vector matrix and the adjacency matrix to obtain the total adjacency feature of each task node.

[0134] That is, the server determines the total adjacency feature of each task node by using the node feature vector matrix and the adjacency matrix to perform multiplication processing therebetween by using the graph convolution network. wherein, represents the node feature vector matrix of the task node, represents the adjacency matrix. The total adjacency feature It can be seen that the total adjacency feature includes the sum of the initial features corresponding to the adjacent task nodes (i.e., the first type of task node in the foregoing embodiment) having a connection relationship with the task node.

[0135] Further, the server further determines the average adjacency feature of each task node by using each total adjacency feature and the relationship degree matrix, that is, the server considers that there can be multiple task nodes (i.e., the first type of task node in the foregoing embodiment) having a connection relationship with each task node, and therefore needs to perform average processing on the sum of the initial features (i.e., the total adjacency feature) corresponding to the adjacent task nodes. The manner of determining the average adjacency feature of each task node can be: scaling the relationship degree matrix by row and column to determine the average adjacency feature, and scaling the relationship degree matrix by row to determine the average adjacency feature,

[0136] The two ways of obtaining the average adjacency feature will be introduced respectively. In an optional embodiment, the average adjacency feature of each task node is determined by each total adjacency feature and the relationship degree matrix, including: performing row and column scaling on the adjacency matrix, and obtaining the row and column scaled total adjacency feature of each task node through the row and column scaled adjacency matrix; performing feature averaging processing on each row and column scaled total adjacency feature according to the relationship degree matrix to determine the average adjacency feature of each task node.

[0137] Specifically, the server performs row and column scaling on the adjacency matrix, and obtains the row and column scaled total adjacency feature of each task node through the row and column scaled adjacency matrix. That is, the server first performs row and column scaling on the adjacency matrix, and then re-determines the total adjacency feature through the way introduced in the foregoing embodiment, that is, the server performs matrix multiplication processing on the node feature vector matrix and the row and column scaled adjacency matrix to obtain the row and column scaled total adjacency feature of each task node. Based on this, the server performs feature averaging processing on each row and column scaled total adjacency feature according to the relationship degree matrix to determine the average adjacency feature of each task node. At this time, the average adjacency feature is , denotes the inverse matrix of the arithmetic square root of the relationship degree matrix, denotes the total adjacency feature.

[0138] In an optional embodiment, the average adjacency feature of each task node is determined by each total adjacency feature and the relationship degree matrix, including: performing row scaling on the adjacency matrix, and obtaining the row scaled total adjacency feature of each task node through the row scaled adjacency matrix; performing feature averaging processing on each row scaled total adjacency feature according to the relationship degree matrix to determine the average adjacency feature of each task node.

[0139] Specifically, the server performs row scaling on the adjacency matrix, and obtains the row scaled total adjacency feature of each task node through the row scaled adjacency matrix. That is, the server first performs row scaling on the adjacency matrix, and then re-determines the total adjacency feature through the way introduced in the foregoing embodiment, that is, the server performs matrix multiplication processing on the node feature vector matrix and the row scaled adjacency matrix to obtain the row scaled total adjacency feature of each task node. Based on this, the server performs feature averaging processing on each row scaled total adjacency feature according to the relationship degree matrix to determine the average adjacency feature of each task node. At this time, the average adjacency feature is , denotes the inverse matrix of the relationship degree matrix, denotes the total adjacency feature.

[0140] Further, the server performs feature normalization processing on the initial feature of each task node and the average adjacency feature of each task node based on the initial feature of each task node and the average adjacency feature of each task node, to obtain the target feature of each task node. That is, the server performs feature normalization processing on the initial feature of each task node and the average adjacency feature of each task node through a normalization layer (Softmax) in the graph convolution network, so as to output the target feature of each task node through the normalization layer.

[0141] The normalization processing of the foregoing target feature is specifically as shown in formula (3):

[0142] (3)

[0143] wherein, the target feature of the task node, the node feature vector matrix, the initial adjacency matrix, the average adjacency feature.

[0144] It should be understood that the corresponding examples in the embodiments of the present application are used to understand the present scheme, but should not be understood as a specific limitation on the present scheme.

[0145] In the embodiment, the correlation between each interaction operation is further considered on the basis of considering the corresponding feature of each interaction operation, and a two-layer neural network is used to learn a more optimized feature representation, that is, the relationship degree matrix is considered and symmetrized to obtain the average adjacency matrix, so as to ensure that the optimized target feature obtained by performing feature normalization processing is more accurate, thereby improving the reliability of the operation conversion rate based on the target feature, and further improving the accuracy of the conversion rate prediction.

[0146] Based on the detailed introduction of the foregoing embodiment, the complete process of the method for predicting the conversion rate in the embodiments of the present application will be introduced below. In one embodiment, as shown in Figure 16 , a method for predicting the conversion rate is provided. Taking the server 104 in Figure 1 as an example for illustration, it should be understood that the method can also be applied to the terminal 102, and can also be applied to a system including the terminal 102 and the server 104, and is realized through the interaction of the terminal 102 and the server 104. In the embodiment, the method includes the following steps:

[0147] Step 1601, obtaining object basic attribute information, media basic attribute information and media push space information from task data of a target task.

[0148] Step 1602, feature extraction is performed on the object basic attribute information, the media basic attribute information and the media push space information respectively to obtain corresponding object features, media features and space features respectively.

[0149] Step 1603, feature splicing processing is performed on the object features, the media features and the space features to obtain spliced features.

[0150] Step 1604, feature intersection processing is performed on the spliced features to obtain shared information.

[0151] Step 1605, feature extraction is performed on the shared information for each task node in the target task to obtain initial features of each task node.

[0152] Step 1606, the continuity of the interactive operation represented by each task node is determined through historical operation records, and the operation sequence of the continuous operation between each task node is determined; the task nodes are connected in sequence through the operation sequence to obtain a directed graph.

[0153] Step 1607, based on the connection relationship between the task nodes in the directed graph, an adjacency matrix matching the directed graph is generated.

[0154] Step 1608, a relationship degree matrix is obtained through the connection relationship.

[0155] Step 1609, the number of nodes of the task nodes included in the directed graph and the feature dimension of each initial feature are determined.

[0156] Step 1610, based on the initial features of each task node, the number of nodes and the feature dimension, a node feature vector matrix is obtained.

[0157] Step 1611, matrix multiplication processing is performed on the node feature vector matrix and the adjacency matrix to obtain total adjacency features of each task node.

[0158] Step 1612, row and column scaling is performed on the adjacency matrix, and total adjacency features of each task node are obtained through the row and column scaled adjacency matrix; according to the relationship degree matrix, feature averaging processing is performed on each row and column scaled total adjacency feature to determine average adjacency features of each task node.

[0159] Step 1613, feature normalization processing is performed based on the initial features of each task node and the average adjacency features of each task node to obtain target features of each task node.

[0160] Step 1614, for each task node, dense connection processing and non-linear connection processing are performed on the target features to obtain operation conversion rates of the interactive operation represented by the task nodes.

[0161] Step 1615, the operation conversion rate of the obtained each task node representing the interactive operation is spliced to obtain a multi-task operation conversion rate prediction result for the target task.

[0162] It should be understood that the specific implementation of steps 1601 to 1615 is similar to the foregoing embodiments, and will not be described here.

[0163] It should be understood that although the steps in the flowchart involved in each of the above-described embodiments are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each of the above-described embodiments can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be alternately executed with at least part of other steps or steps or stages in other steps.

[0164] Based on the same inventive concept, the embodiments of the present application also provide a conversion rate prediction device for implementing the above-mentioned conversion rate prediction method. The problem-solving implementation scheme provided by the device is similar to the implementation scheme described in the above method, so the specific limitations in one or more conversion rate prediction device embodiments provided below can refer to the limitations of the conversion rate prediction method described above, and will not be described here.

[0165] In one embodiment, as shown in Figure 17 a conversion rate prediction device is provided, comprising: a directed graph acquisition module 1702, a feature extraction module 1704, a feature optimization module 1706 and a conversion rate prediction module 1708, wherein:

[0166] The directed graph acquisition module 1702 is configured to acquire a directed graph representing the connection relationship between the task nodes of the target task, and the connection relationship is used to represent the continuity of the interactive operation for the task node.

[0167] The feature extraction module 1704 is configured to perform feature extraction on the shared information for each task node in the target task to obtain the initial feature of each task node.

[0168] The feature optimization module 1706 is configured to perform feature optimization processing on the initial feature of each task node based on the directed graph to obtain the target feature of each task node.

[0169] The conversion rate prediction module 1708 is configured to predict, for each task node, an operation conversion rate of an interaction operation represented by the task node based on the target feature.

[0170] In an embodiment, the feature optimization module is specifically configured to generate an adjacency matrix matched with the directed graph based on the connection relationship between the task nodes in the directed graph; and perform feature optimization processing on each initial feature of each task node based on the adjacency matrix to obtain a target feature of each task node.

[0171] In an embodiment, the feature optimization module is specifically configured to determine, for each task node in the directed graph, a first type of task node having a connection relationship with the task node and a second type of task node not having a connection relationship with the task node; determine a first element value representing a connection relationship type between the task node and the first type of task node and a second element value representing a connection relationship type between the task node and the second type of task node; and construct an adjacency matrix matched with the directed graph based on the first type of task node and the second type of task node corresponding to each task node, and the first element value and the second element value as matrix elements.

[0172] In an embodiment, the feature optimization module is specifically configured to generate an initial adjacency matrix matched with the directed graph based on the connection relationship between the task nodes in the directed graph; and perform matrix splicing processing on the initial adjacency matrix and a unit matrix to obtain the adjacency matrix.

[0173] In an embodiment, the conversion rate prediction device further includes a relationship degree matrix construction module.

[0174] The relationship degree matrix construction module is configured to obtain a relationship degree matrix through the connection relationship, and a matrix element of the relationship degree matrix is used to represent a number of task nodes having a connection relationship with each task node.

[0175] The feature optimization module is specifically configured to perform feature optimization processing on each initial feature of each task node based on the adjacency matrix and the relationship degree matrix to obtain a target feature of each task node.

[0176] In an embodiment, the feature optimization module is specifically configured to determine a node number of the task nodes included in the directed graph and a feature dimension of each initial feature; obtain a node feature vector matrix based on each initial feature of each task node, the node number, and the feature dimension; and perform feature optimization processing on each initial feature of each task node through the node feature vector matrix, the adjacency matrix, and the relationship degree matrix to obtain a target feature of each task node.

[0177] In an embodiment, the feature optimization module is specifically configured to determine a respective total adjacency feature of each task node by performing matrix multiplication on the node feature vector matrix and the adjacency matrix, and determine a respective average adjacency feature of each task node by performing feature averaging on each total adjacency feature based on the relationship degree matrix.

[0178] In an embodiment, the feature optimization module is specifically configured to determine a respective total adjacency feature of each task node by performing matrix multiplication on the node feature vector matrix and the adjacency matrix.

[0179] In an embodiment, the feature optimization module is specifically configured to perform row and column scaling on the relationship degree matrix to obtain a row and column scaled relationship degree matrix.

[0180] In an embodiment, the feature optimization module is specifically configured to perform row and column scaling on the relationship degree matrix to obtain a row and column scaled relationship degree matrix.

[0181] In an embodiment, the feature optimization module is specifically configured to perform row scaling on the adjacency matrix, and obtain a respective row scaled total adjacency feature of each task node based on the row scaled adjacency matrix; and perform feature averaging on each row scaled total adjacency feature based on the relationship degree matrix to determine a respective average adjacency feature of each task node.

[0182] In an embodiment, the conversion rate prediction device further comprises a shared information acquisition module.

[0183] The shared information acquisition module is configured to acquire object basic attribute information, media basic attribute information, and media push space information from the task data of the target task; perform feature extraction on the object basic attribute information, the media basic attribute information, and the media push space information to obtain corresponding object features, media features, and space features, respectively; perform feature splicing processing on the object features, the media features, and the space features to obtain spliced features; and perform feature cross processing on the spliced features to obtain shared information.

[0184] In an embodiment, the directed graph acquisition module is specifically configured to determine the continuity of the interactive operation represented by each task node based on the historical operation records, and determine the operation sequence of the operations with continuity between each task node; and sequentially connect the task nodes based on the operation sequence to obtain a directed graph.

[0185] In an embodiment, the conversion rate prediction module is configured to perform dense connection processing and non-linear connection processing on the target features for each task node to obtain an operation conversion rate of the interactive operation represented by the task node.

[0186] The conversion rate prediction module is further configured to splice the operation conversion rates of the obtained interaction operations represented by each task node to obtain a multi-task operation conversion rate estimation result for the target task.

[0187] In one embodiment, a computer device is provided, which can be a server or a terminal. In this embodiment, the computer device is taken as a server as an example, and its internal structure diagram can be shown in FIG. 1. Figure 18 The computer device includes a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is configured to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for running the operating system and the computer program in the non-volatile storage medium. The database of the computer device is configured to store a directed graph and a conversion rate prediction model and other data related to the embodiments of the present application. The input / output interface of the computer device is configured to exchange information between the processor and external devices. The communication interface of the computer device is configured to communicate with external terminals through a network connection. The computer program is executed by the processor to implement a conversion rate prediction method.

[0188] Those skilled in the art can understand that Figure 18 The structure shown in FIG. 1 is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. Specifically, the computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0189] In one embodiment, a computer device is provided, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0190] In one embodiment, a computer readable storage medium is provided, which stores a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0191] In one embodiment, a computer program product is provided, which includes a computer program. The computer program is executed by a processor to implement the steps in the above method embodiments.

[0192] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of the country and region.

[0193] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing related hardware through a computer program, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. In the embodiments provided in the present application, any reference to memory, database or other medium can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (Read-Only Memory, ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (Magnetoresistive Random Access Memory, MRAM), ferroelectric memory (Ferroelectric Random Access Memory, FRAM), phase change memory (Phase Change Memory, PCM), graphene memory, etc. Volatile memory can include random access memory (Random Access Memory, RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (Static Random Access Memory, SRAM) or dynamic random access memory (Dynamic Random Access Memory, DRAM), etc. The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a block chain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0194] The technical feature information of the above embodiments can be combined arbitrarily. In order to make the description simple, all possible combinations of the technical feature information in the above embodiments are not described, but as long as the combination of these technical feature information does not exist contradictory, it should be considered as the scope of the present application.

[0195] The above-described embodiments are merely illustrative of several embodiments of the present application, and the description is relatively specific and detailed, but should not be understood as a limitation on the scope of the patent. It should be noted that for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of protection of the present application should be subject to the appended claims.

Claims

1. A method for predicting conversion rates, characterized in that, include: Obtain a directed graph representing the connection relationships between task nodes of the target task, wherein the connection relationships are used to represent the continuity of interactive operations for the task nodes; Feature extraction is performed on the shared information of each task node in the target task to obtain the initial features of each task node. Based on the directed graph, feature optimization processing is performed on the initial features of each task node to obtain the target features of each task node. For each task node, based on the target features, predict the operation conversion rate of the interactive operation represented by the task node.

2. The method according to claim 1, characterized in that, Based on the directed graph, feature optimization processing is performed on the initial features of each task node to obtain the target features of each task node, including: Based on the connection relationships between the task nodes in the directed graph, an adjacency matrix matching the directed graph is generated. Based on the initial features of each task node and the adjacency matrix, feature optimization processing is performed to obtain the target features of each task node.

3. The method according to claim 2, characterized in that, The step of generating an adjacency matrix matching the directed graph based on the connection relationships between the task nodes in the directed graph includes: For each task node in the directed graph, a first type of task node that has a connection relationship with the task node is determined, and a second type of task node that does not have a connection relationship with the task node is determined. Determine a first element value that represents the type of connection relationship between the task node and the first type of task node, and a second element value that represents the type of connection relationship between the task node and the second type of task node; Based on the first type of task node and the second type of task node corresponding to each of the task nodes, an adjacency matrix matching the directed graph is constructed using the first element value and the second element value as matrix elements.

4. The method according to claim 2, characterized in that, The step of generating an adjacency matrix matching the directed graph based on the connection relationships between the task nodes in the directed graph includes: Based on the connection relationships between the task nodes in the directed graph, an initial adjacency matrix matching the directed graph is generated; The initial adjacency matrix and the identity matrix are concatenated to obtain the adjacency matrix.

5. The method according to claim 2, characterized in that, The method further includes: A relation degree matrix is ​​obtained through the connection relationship, and the matrix elements of the relation degree matrix are used to represent: the number of task nodes that have a connection relationship with each of the task nodes; The feature optimization process based on the initial features of each task node and the adjacency matrix includes: Based on the initial features of each task node, the adjacency matrix, and the relation degree matrix, feature optimization processing is performed to obtain the target features of each task node.

6. The method according to claim 5, characterized in that, The feature optimization process, based on the initial features of each task node, the adjacency matrix, and the relation degree matrix, yields the target features for each task node, including: Determine the number of task nodes included in the directed graph, and the feature dimension of each initial feature; Based on the initial features of each task node, the number of nodes, and the feature dimension, a node feature vector matrix is ​​obtained; By performing feature optimization processing using the node feature vector matrix, the adjacency matrix, and the relation degree matrix, the target features of each task node are obtained.

7. The method according to claim 6, characterized in that, The feature optimization process, which uses the node feature vector matrix, the adjacency matrix, and the relation degree matrix to obtain the target features of each task node, includes: The total adjacency feature of each task node is determined by the node feature vector matrix and the adjacency matrix, and the average adjacency feature of each task node is determined by the total adjacency feature and the relation degree matrix. Based on the initial features of each task node and the average adjacency features of each task node, feature normalization processing is performed to obtain the target features of each task node.

8. The method according to claim 7, characterized in that, The step of determining the total adjacency features of each task node using the node feature vector matrix and the adjacency matrix includes: The node feature vector matrix and the adjacency matrix are multiplied to obtain the total adjacency features of each task node.

9. The method according to claim 7, characterized in that, The step of determining the average adjacency feature of each task node by using the total adjacency feature and the relation degree matrix includes: The adjacency matrix is ​​scaled in rows and columns, and the total adjacency features of each task node are obtained from the scaled adjacency matrix. Based on the relation degree matrix, feature averaging is performed on the total adjacency features of each row and column scaling to determine the average adjacency features of each task node.

10. The method according to claim 7, characterized in that, The step of determining the average adjacency feature of each task node by using the total adjacency feature and the relation degree matrix includes: The adjacency matrix is ​​row-scaled, and the row-scaled total adjacency features of each task node are obtained from the row-scaled adjacency matrix. Based on the relation degree matrix, the average adjacency feature of each row is scaled and averaged to determine the average adjacency feature of each task node.

11. The method according to claim 1, characterized in that, The method further includes: From the task data of the target task, obtain the basic attribute information of the object, the basic attribute information of the media, and the media push space information; Feature extraction is performed on the basic attribute information of the object, the basic attribute information of the media, and the spatial information of the media push, respectively, to obtain the corresponding object features, media features, and spatial features. The object features, the media features, and the spatial features are subjected to feature splicing processing to obtain spliced ​​features; The spliced ​​features are subjected to feature cross-processing to obtain the shared information.

12. The method according to claim 1, characterized in that, The process of obtaining the directed graph representing the connection relationships between task nodes of the target task includes: The continuity of the interactive operations represented by each task node is determined by the historical operation records, and the sequential order of operations with continuity between each task node is determined. The task nodes are connected sequentially by the order of operations to obtain a directed graph.

13. The method according to claim 1, characterized in that, For each task node, predicting the operation conversion rate of the interactive operation represented by the task node based on the target features includes: For each task node, dense connection processing and nonlinear connection processing are performed on the target features to obtain the operation conversion rate of the interactive operation represented by the task node. The method further includes: The operation conversion rates of the interactive operations represented by each task node are concatenated to obtain a multi-task operation conversion rate prediction result for the target task.

14. A conversion rate prediction device, characterized in that, The device includes: A directed graph acquisition module is used to acquire a directed graph representing the connection relationships between task nodes of a target task, wherein the connection relationships are used to represent the continuity of interactive operations for the task nodes. The feature extraction module is used to extract features from the shared information of each task node in the target task to obtain the initial features of each task node. The feature optimization module is used to perform feature optimization processing on the initial features of each task node based on the directed graph, so as to obtain the target features of each task node. The conversion rate prediction module is used to predict the operation conversion rate of the interactive operation represented by the task node for each task node based on the target features.

15. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 13.

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

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

Images