Advertisement conversion rate prediction method and device, computer equipment and storage medium

Abstract

The invention provides an advertisement conversion rate prediction method and device, computer equipment and a storage medium. The advertisement conversion rate prediction method comprises the steps of obtaining behavior sequence data, object attribute data and target advertisement data of a target object, wherein the behavior sequence data comprises multiple pieces of advertisement operation behavior data of the target object sorted according to a time sequence; performing feature embedding on the object attribute data and the target advertisement data to obtain attribute features, and performing feature embedding on each piece of advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence; inputting the behavior feature sequence into at least two neural network models adopting different sequence feature extraction modes to obtain at least two sequence semantic features; splicing a fusion feature obtained by fusing the at least two sequence semantic features with the attribute feature to obtain a target feature; and predicting the advertisement conversion rate of the target object to the target advertisement based on the target feature. According to the embodiment of the invention, the accuracy of advertisement conversion rate prediction can be improved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of advertisement conversion rate prediction, and in particular to an advertisement conversion rate prediction method and device, a computer device, and a storage medium. BACKGROUND

[0002] The advertisement conversion rate prediction task refers to a task of predicting the probability of whether an object will purchase a product or other target behavior after seeing an advertisement. Advertisement conversion rate prediction can help advertisers make better advertisement placement strategies, thereby improving the effect and return on investment of advertisement placement.

[0003] Currently, when performing the advertisement conversion rate prediction task, a neural network model based on artificial intelligence is usually used to predict the advertisement conversion rate. However, the method of using the neural network model to predict the advertisement conversion rate has low accuracy of the predicted advertisement conversion rate. SUMMARY

[0004] The present disclosure provides an advertisement conversion rate prediction method, device, computer device, and storage medium, which can improve the accuracy of advertisement conversion rate prediction.

[0005] The first aspect of the present disclosure provides an advertisement conversion rate prediction method, comprising:

[0006] obtaining behavior sequence data of a target object, object attribute data of the target object, and target advertisement data, wherein the behavior sequence data comprises a plurality of advertisement operation behavior data of the target object sorted in chronological order;

[0007] performing feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and performing feature embedding processing on each of the advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence;

[0008] inputting the behavior feature sequence into at least two neural network models to extract sequence features, to obtain at least two sequence semantic features, wherein the at least two neural network models use different sequence feature extraction methods to extract sequence features, and the at least two neural network models are obtained by joint training;

[0009] performing feature fusion on the at least two sequence semantic features, and splicing the fused features with the attribute features to obtain target features;

[0010] predicting an advertisement conversion rate of the target object to the target advertisement based on the target features.

[0011] The second aspect of the present disclosure provides an advertisement conversion rate prediction device, comprising:

[0012] An acquisition unit is configured to acquire behavior sequence data of a target object, object attribute data of the target object, and target advertisement data, wherein the behavior sequence data comprises a plurality of advertisement operation behavior data of the target object in chronological order;

[0013] An embedding unit is configured to perform feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and perform feature embedding processing on each of the advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence;

[0014] An extraction unit is configured to input the behavior feature sequence into at least two neural network models to extract sequence features, to obtain at least two sequence semantic features, wherein the at least two neural network models use different sequence feature extraction methods to extract sequence features, and the at least two neural network models are obtained through joint training;

[0015] A fusion unit is configured to perform feature fusion on the at least two sequence semantic features, and concatenate the fused features with the attribute features to obtain target features;

[0016] A prediction unit is configured to predict an advertisement conversion rate of the target object to the target advertisement based on the target features.

[0017] Optionally, in some embodiments, the at least two neural network models comprise at least two first neural network models and a second neural network model, the first neural network model extracts sequence features based on the behavior feature sequence, and the second neural network model extracts sequence features based on the behavior feature sequence and advertisement features of the target advertisement data, and the extraction unit comprises:

[0018] A first extraction subunit is configured to input the behavior feature sequence into the at least two first neural network models to extract sequence features, to obtain at least two first sub-sequence semantic features;

[0019] A second extraction subunit is configured to perform feature embedding processing on the target advertisement data to obtain advertisement features, and input the advertisement features and the behavior feature sequence into the second neural network model to extract sequence features, to obtain a second sub-sequence semantic feature;

[0020] The fusion unit is further configured to:

[0021] perform feature fusion on the at least two first sub-sequence semantic features and the second sub-sequence semantic feature, and concatenate the fused features with the attribute features to obtain target features.

[0022] Optionally, in some embodiments, the at least two first neural network models comprise a global perception model, a nearby focus model, a time evolution model, and a self-attention model, the first extraction subunit comprises:

[0023] a global perception module configured to perform nonlinear transformation processing on a plurality of behavior features in the behavior feature sequence based on the global perception model, and perform mean pooling on the processed plurality of behavior features to obtain a first subsequence semantic feature; and

[0024] a nearby focus module configured to perform nonlinear transformation and summation processing on nearby behavior features in the behavior feature sequence based on the nearby focus model to obtain a first subsequence semantic feature, the nearby behavior features being a preset number of behavior features closest in time to a prediction time; and

[0025] a time evolution module configured to perform cyclic feature updating on a plurality of behavior features in the behavior feature sequence according to a time sequence based on the time evolution model to obtain a first subsequence semantic feature; and

[0026] a self-attention module configured to perform self-attention processing on a plurality of behavior features in the behavior feature sequence based on the self-attention model to obtain a first subsequence semantic feature.

[0027] Optionally, in some embodiments, the fusion unit comprises:

[0028] a first acquisition subunit configured to acquire weight coefficients corresponding to the at least two sequence semantic features;

[0029] a calculation subunit configured to perform weighted calculation on the at least two sequence semantic features based on the weight coefficients to obtain a fusion feature;

[0030] a first concatenation subunit configured to concatenate the fusion feature and the attribute feature to obtain a target feature.

[0031] Optionally, in some embodiments, the first acquisition subunit comprises:

[0032] a first acquisition module configured to acquire a preset weight matrix, the preset weight matrix being obtained during training of the at least two neural network models;

[0033] a first calculation module configured to calculate the weight coefficients corresponding to the at least two sequence semantic features based on the preset weight matrix and the at least two sequence semantic features.

[0034] Optionally, in some embodiments, the first calculation module comprises:

[0035] a concatenation submodule configured to concatenate the at least two sequence semantic features in a dimension expansion manner to obtain an expanded feature;

[0036] a calculation submodule configured to calculate a weight coefficient corresponding to the at least two sequence semantic features based on the preset weight matrix and the expanded feature.

[0037] Optionally, in some embodiments, the calculation submodule is specifically configured to:

[0038] calculate an attention parameter matrix based on the preset weight matrix and the expanded feature;

[0039] calculate the weight coefficient corresponding to the at least two sequence semantic features based on the attention parameter matrix.

[0040] Optionally, in some embodiments, the obtaining subunit comprises:

[0041] a second obtaining subunit configured to obtain advertisement environment data and advertisement context data corresponding to the target advertisement data;

[0042] a determination subunit configured to determine a weight coefficient corresponding to each sequence semantic feature based on the advertisement environment data and the advertisement context data.

[0043] Optionally, in some embodiments, the advertisement conversion rate prediction device provided by the present disclosure further comprises a training unit configured to train at least two neural network models, and the training unit comprises:

[0044] a third obtaining subunit configured to obtain training sample data, the training sample data comprising a plurality of sample data groups and a conversion label data corresponding to each sample data group, and each sample data group comprising sample object attribute data, sample behavior sequence data and sample advertisement data of a sample object;

[0045] an embedding subunit configured to perform feature embedding on the sample object attribute data and the sample advertisement data to obtain sample attribute features, and perform feature embedding on the sample behavior sequence data to obtain a sample behavior feature sequence;

[0046] a third extraction subunit configured to input the sample behavior feature sequence into the at least two neural network models respectively to perform sequence feature extraction, and perform feature fusion on the output at least two sample sequence semantic features to obtain sample fusion features;

[0047] a second concatenation subunit configured to concatenate the sample attribute features and the sample fusion features, and input the concatenated features into a third neural network model to perform conversion rate prediction, to obtain an output conversion rate prediction value;

[0048] The updating subunit is configured to calculate a loss value based on the conversion label data and the conversion rate prediction value, and update model parameters of the at least two neural network models based on the loss value.

[0049] Optionally, in some embodiments, the third extraction subunit comprises:

[0050] The splicing module is configured to perform dimension expansion splicing on the output at least two sample sequence semantic features to obtain sample dimension expansion features.

[0051] The second acquisition module is configured to acquire a weight matrix, parameters in the weight matrix being adjustable.

[0052] The second calculation module is configured to calculate sample weight coefficients of the at least two sample sequence semantic features based on the weight matrix and the sample dimension expansion features.

[0053] The fusion module is configured to fuse the at least two sample sequence semantic features to obtain sample fusion features according to the sample weight coefficients.

[0054] The updating subunit is further configured to:

[0055] update the model parameters of the at least two neural network models and the parameters in the weight matrix based on the loss value.

[0056] Optionally, in some embodiments, the prediction unit comprises:

[0057] The mapping subunit is configured to perform feature mapping on the target feature input trained full connection network model to obtain output features.

[0058] The normalization subunit is configured to perform normalization processing on the output features to obtain the advertisement conversion rate of the target object to the target advertisement.

[0059] The third aspect of the present disclosure provides a storage medium, the storage medium storing a computer program, the computer program being executed by a processor to implement the advertisement conversion rate prediction method of the first aspect or the advertisement conversion rate prediction method of the second aspect.

[0060] The fourth aspect of the present disclosure provides a computer device comprising a memory and a processor, the memory storing a computer program, the processor being executed to implement the advertisement conversion rate prediction method of the first aspect or the advertisement conversion rate prediction method of the second aspect.

[0061] The fifth aspect of the present disclosure provides a computer program product comprising a computer program, which is read and executed by a processor of a computer device, so that the computer device executes the advertisement conversion rate prediction method according to the first aspect, or implements the advertisement conversion rate prediction method according to the second aspect.

[0062] The advertisement conversion rate prediction method provided by the embodiments of the present disclosure comprises the following steps: obtaining behavior sequence data of a target object, object attribute data of the target object, and target advertisement data, wherein the behavior sequence data comprises a plurality of advertisement operation behavior data of the target object in chronological order; performing feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and performing feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence; inputting the behavior feature sequence into at least two neural network models to perform sequence feature extraction, so as to obtain at least two sequence semantic features, wherein the at least two neural network models perform sequence feature extraction in different sequence feature extraction manners, and the at least two neural network models are obtained through joint training; performing feature fusion on the at least two sequence semantic features, and splicing the fused features and the attribute features to obtain target features; and predicting an advertisement conversion rate of the target object to the target advertisement based on the target features.

[0063] The embodiments of the present disclosure can extract more accurate sequence semantic features of object behaviors from different dimensional perspectives by using at least two neural network models to perform sequence feature extraction on the behavior feature sequence of the target object in different sequence feature extraction manners, thereby avoiding the problem that a single model in the related art cannot sufficiently mine deep semantic features of the sequence features of the object behaviors, and thus the advertisement conversion rate prediction method provided by the present disclosure can improve the accuracy of advertisement conversion rate prediction.

[0064] Other features and advantages of the present disclosure will be described in the following description, and some will become apparent from the description, or will be learned through practice of the present disclosure. The purposes and other advantages of the present disclosure can be achieved and obtained by the structures specifically pointed out in the specification, claims, and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0065] The accompanying drawings are included to provide a further understanding of the technical solutions of the present disclosure, and constitute a part of the specification, and are used together with the embodiments of the present disclosure to explain the technical solutions of the present disclosure, and do not constitute a limitation on the technical solutions of the present disclosure.

[0066] Figure 1 A system architecture diagram applied to the advertisement conversion rate prediction method of the embodiments of the present disclosure;

[0067] Figure 2 A flowchart of the advertisement conversion rate prediction method provided by the present disclosure;

[0068] Figure 3 A first algorithm structure schematic diagram of the advertisement conversion rate prediction method provided by the present disclosure;

[0069] Figure 4 A second algorithm structure schematic diagram of the advertisement conversion rate prediction method provided by the present disclosure;

[0070] Figure 5 A third algorithm structure schematic diagram of the advertisement conversion rate prediction method provided by the present disclosure;

[0071] Figure 6 A model structure schematic diagram of the global perception model provided by the embodiment of the present disclosure;

[0072] Figure 7 A model structure schematic diagram of the proximity focus model provided by the embodiment of the present disclosure;

[0073] Figure 8 A model structure schematic diagram of the time evolution model provided by the embodiment of the present disclosure;

[0074] Figure 9 A model structure schematic diagram of the self-attention model provided by the embodiment of the present disclosure;

[0075] Figure 10 A model structure schematic diagram of the second neural network model provided by the embodiment of the present disclosure;

[0076] Figure 11 A specific flow schematic diagram of the advertisement conversion rate prediction method provided by the embodiment of the present disclosure;

[0077] Figure 12 A structure schematic diagram of the advertisement conversion rate prediction device provided by the embodiment of the present disclosure;

[0078] Figure 13 A terminal structure diagram for implementing each method according to an embodiment of the present disclosure;

[0079] Figure 14 A server structure diagram for implementing each method according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0080] In order to make the objectives, technical solutions and advantages of the present disclosure clearer, further detailed descriptions will be given to the present disclosure in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present disclosure, and are not used to limit the present disclosure.

[0081] Before the embodiments of the present disclosure are further described in detail, the terms and names involved in the embodiments of the present disclosure are explained, and the terms and names involved in the embodiments of the present disclosure are applicable to the following explanations:

[0082] Artificial Intelligence (AI): is to use digital computers or digital computer controlled machines to simulate, extend and expand human intelligence, perceive environment, acquire knowledge and use knowledge to obtain optimal 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 machines have the functions of perception, reasoning and decision-making.

[0083] Artificial intelligence technology is a comprehensive discipline, involving a wide range of fields, both hardware and software technologies. Artificial intelligence basic technologies generally include 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 various directions 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.

[0084] Machine Learning (ML): is a multi-disciplinary subject, involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, etc. It is a subject 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.

[0085] Neural network model: is a computational model inspired by biological neural networks, used in the field of machine learning and artificial intelligence. It consists of a hierarchical structure of multiple neurons, each connected to the neurons of the next layer. These connections have weights, and through these weights and activation functions, neural networks can learn complex patterns and relationships in input data. Neural network models are often used in image recognition, natural language processing, prediction, etc.

[0086] Linear rectified function (Rectified Linear Unit, Relu): a commonly used activation function for the hidden layer of neural networks. The definition of Relu function is: f(x) = max(0, x), that is, when the input value x is greater than or equal to zero, the output is x, otherwise the output is zero. Relu function is simple and efficient in calculation, which can effectively solve the problem of gradient disappearance, etc. In deep learning, Relu function is usually used as the activation function of hidden layer, which helps to learn the non-linear relationship of network.

[0087] The advertisement conversion rate refers to the proportion of the actual completion of the expected conversion behavior (such as registering an account, purchasing a product, downloading a product file, etc.) after the object clicks the advertisement in the process of advertising. The advertisement conversion rate can be used to measure the effectiveness and benefit of the advertisement, and a high conversion rate can indicate that the advertisement has a positive impact on the audience. The advertisement conversion rate prediction is to predict the probability of the actual completion of the expected conversion behavior after the object clicks the advertisement. The advertisement conversion rate prediction task refers to the task of predicting the probability of whether the object will purchase goods or other target behaviors after seeing the advertisement.

[0088] With the development of Internet technology and mobile terminal devices, advertisers usually can put advertisements through Internet advertising platforms to promote their products and expand sales in order to expand sales. For example, a clothing brand can choose to put advertisements of goods on different media platforms. Therefore, in different media platforms, the probability of object conversion to advertisements can be estimated through the pre-set advertisement conversion rate prediction model, which helps advertisers evaluate the effect of advertising strategies and make better advertising strategies, so as to improve the effect of advertising and return on investment.

[0089] Currently, when performing an advertisement conversion rate prediction task, the related art usually adopts an artificial intelligence-based neural network model to perform advertisement conversion rate prediction. For example, a method for extracting corresponding behavior sequence features based on a neural network model adopted by the related art is to directly take the mean or sum of the behavior node features in the input behavior sequence. Although this method is simple, it ignores the importance and difference of the behavior features at different positions in the sequence, resulting in low accuracy of the advertisement conversion rate predicted based on this method. Or, a method for extracting behavior sequence features based on an artificial intelligence-based neural network model adopted by the related art is to model multiple nodes in the order of the input behavior sequence, and perform advertisement conversion rate prediction according to the constructed model. Although this method can highlight the time sequence semantic information of the modeling sequence, its modeling efficiency is low, and it can only be processed in series according to the node order of the time sequence, resulting in low accuracy of the predicted advertisement conversion rate, and the like. That is, the related art usually adopts the above single model to extract sequence features. However, sequence features usually present in the form of a multi-value ordered list, and its length is often dozens or even hundreds, carrying a great information density. The single model of the related art cannot extract the most accurate features from the long behavior sequence, so that the conversion rate prediction model based on the single model for extracting sequence features has low accuracy of the predicted advertisement conversion rate.

[0090] To solve the problem of low accuracy of the advertisement conversion rate in the above scenario, the present disclosure provides an advertisement conversion rate prediction method to improve the accuracy of advertisement conversion rate prediction.

[0091] System architecture and scenario applied by embodiments of the present disclosure

[0092] Figure 1 It is a system architecture diagram applied by the advertisement conversion rate prediction method according to the embodiments of the present disclosure. It includes terminal 110, Internet 120, gateway 130, server 140, etc.

[0093] The terminal 110 includes desktop computers, laptop computers, PDAs (personal digital assistants), mobile phones, vehicle-mounted terminals, home theater terminals, special-purpose terminals, smart voice interaction devices, smart home appliances, or aircraft, etc. in various device forms. In addition, it can be a single device or a collection of multiple devices. For example, multiple devices are connected through a local area network, share a display device for collaborative work, and collectively constitute a terminal 110. The terminal 110 can communicate with the Internet 120 in a wired or wireless manner to exchange data.

[0094] The server 140 refers to a computer system capable of providing certain services to the terminal 110. Compared with the ordinary terminal 110, the server 140 has higher requirements in stability, security, performance, and the like. The server 140 can be a high-performance computer in a network platform, a cluster of multiple high-performance computers, a part of a high-performance computer (for example, a virtual machine), a combination of parts of multiple high-performance computers (for example, virtual machines), or the like. In the embodiments of the present disclosure, the server 140 specifically provides a storage function, that is, the server 140 here can be a database node of a distributed database, and the cluster of multiple servers 140 constitutes a distributed database in the present disclosure.

[0095] The gateway 130 is also called an inter-network connector or a protocol converter. The gateway realizes network interconnection at the transport layer and is a computer system or device acting as a conversion function. The gateway is a translator between two systems using different communication protocols, data formats or languages, or even having completely different architectures. Meanwhile, the gateway can also provide filtering and security functions. The messages sent by the terminal 110 to the server 140 are sent to the corresponding server 140 through the gateway 130. The messages sent by the server 140 to the terminal 110 are also sent to the corresponding terminal 110 through the gateway 130. In the embodiments of the present disclosure, the terminal 110 sends a data access request to the server 140 through the gateway 130, and the server 140 returns a data access result to the terminal 110 through the gateway 130.

[0096] The advertisement conversion rate prediction method provided by the embodiments of the present disclosure can be implemented in the terminal 110, or in the server 140, or partially in the terminal 110 and partially in the server 140.

[0097] When the advertisement conversion rate prediction method provided by the embodiments of the present disclosure is implemented in the terminal 110, the terminal 110 obtains behavior sequence data of a target object, object attribute data of the target object, and target advertisement data, and the behavior sequence data includes multiple advertisement operation behavior data of the target object sorted in time sequence. Then, the terminal 110 performs feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and performs feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence. Further, the terminal 110 inputs the behavior feature sequence into at least two neural network models to perform sequence feature extraction, to obtain at least two sequence semantic features, the at least two neural network models perform sequence feature extraction in different sequence feature extraction manners, and the at least two neural network models are obtained through joint training. Further, the terminal 110 performs feature fusion on the at least two sequence semantic features, and splices the fused features and the attribute features to obtain target features. Finally, the terminal 110 predicts an advertisement conversion rate of the target object to the target advertisement based on the target features.

[0098] When the advertisement conversion rate prediction method provided by the embodiment of the present disclosure is implemented in the server 140, the server 140 obtains the behavior sequence data of the target object, the object attribute data of the target object, and the target advertisement data, the behavior sequence data including a plurality of advertisement operation behavior data of the target object sorted in time sequence; then the server 140 performs feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and performs feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence; further, the server 140 inputs the behavior feature sequence into at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features, the at least two neural network models performing sequence feature extraction in different sequence feature extraction manners, the at least two neural network models being obtained by joint training; further, the server 140 performs feature fusion on the at least two sequence semantic features, and splices the fused feature and the attribute feature to obtain target features; finally, the server 140 predicts the advertisement conversion rate of the target object to the target advertisement based on the target features.

[0099] When the advertisement conversion rate prediction method provided by the embodiment of the present disclosure is partially implemented in the terminal 110 and partially implemented in the server 140, the terminal 110 obtains the behavior sequence data of the target object, the object attribute data of the target object, and the target advertisement data, the behavior sequence data including a plurality of advertisement operation behavior data of the target object sorted in time sequence; and performs feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and performs feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence; then, the terminal 110 can send the behavior feature sequence and the attribute features to the server 140, the server 140 inputs the behavior feature sequence into at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features, the at least two neural network models performing sequence feature extraction in different sequence feature extraction manners, the at least two neural network models being obtained by joint training; thereafter, the server 140 performs feature fusion on the at least two sequence semantic features, and splices the fused feature and the attribute feature to obtain target features; and predicts the advertisement conversion rate of the target object to the target advertisement based on the target features. Finally, the server 140 sends the advertisement conversion rate of the target advertisement to the terminal 110.

[0100] The advertisement conversion rate prediction method provided by the embodiment of the present disclosure can be applied to advertisement conversion rate prediction tasks in various Internet applications, and can be specifically applied to advertisement conversion rate prediction tasks in shopping applications, conversation applications, tourism recommendation applications, and various Internet applications.

[0101] For example, when the advertisement conversion rate prediction method provided by the embodiments of the present disclosure is applied to the advertisement conversion rate prediction task in the shopping application, the shopping behavior sequence data of the target object received by the shopping application, the object attribute data of the target object and the target advertisement data can be obtained, and the shopping behavior sequence data includes a plurality of target advertisement operation behavior data of the target object sorted in time sequence. Then, the attribute feature can be obtained by performing feature embedding processing on the object attribute data and the target advertisement data, and the behavior feature sequence can be obtained by performing feature embedding processing on each target advertisement operation behavior data in the shopping behavior sequence data. Further, the behavior feature sequence is input into at least two neural network models to perform sequence feature extraction, so as to obtain at least two sequence semantic features, the at least two neural network models perform sequence feature extraction by using different sequence feature extraction methods, and the at least two neural network models are obtained by joint training. Then, the at least two sequence semantic features are fused, and the fused feature obtained by fusion is spliced with the attribute feature to obtain the target feature. Finally, the advertisement conversion rate of the target object to the target advertisement is predicted based on the target feature. When the advertisement conversion rate is high, that is, the probability of actually completing the expected conversion behavior after the target object clicks the advertisement is high, advertisements similar to the target advertisement can be pushed and displayed to the target object in the shopping application more frequently. When the advertisement conversion rate is low, that is, the probability of actually completing the expected conversion behavior after the target object clicks the advertisement is low, advertisements similar to the target advertisement can not be pushed and displayed to the target object in the shopping application.

[0102] For example, when the advertisement conversion rate prediction method provided by the embodiments of the present disclosure is applied to the advertisement conversion rate prediction task in the conversation application, the conversation behavior sequence data of the target object received by the conversation application, the object attribute data of the target object, and the target advertisement data can be obtained, and the conversation behavior sequence data includes a plurality of target advertisement operation behavior data of the target object in time sequence order. Then, the object attribute data and the target advertisement data can be processed by feature embedding to obtain attribute features, and each target advertisement operation behavior data in the conversation behavior sequence data can be processed by feature embedding to obtain a behavior feature sequence. Further, the behavior feature sequence is input into at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features, the at least two neural network models use different sequence feature extraction methods for sequence feature extraction, and the at least two neural network models are obtained by joint training. Then, the at least two sequence semantic features are fused, and the fused features and the attribute features are spliced to obtain target features. Finally, the advertisement conversion rate of the target object to the target advertisement is predicted based on the target features. When the advertisement conversion rate is high, that is, the probability of the target object actually completing the expected conversion behavior after clicking the advertisement is high, advertisements similar to the target advertisement can be pushed and displayed to the target object in the conversation application more frequently. When the advertisement conversion rate is low, that is, the probability of the target object actually completing the expected conversion behavior after clicking the advertisement is low, advertisements similar to the target advertisement can not be pushed and displayed to the target object in the conversation application.

[0103] The above examples do not limit the scope of protection of the present case.

[0104] Overall description of embodiments of the present disclosure

[0105] According to one embodiment of the present disclosure, an advertisement conversion rate prediction method is provided. As shown in Figure 2 , a flowchart of the advertisement conversion rate prediction method provided by the present disclosure. The method can be applied to an advertisement conversion rate prediction device, which can be integrated in a computer device, which can be a terminal. The advertisement conversion rate prediction method can include:

[0106] Step 210, obtaining behavior sequence data of a target object, object attribute data of the target object, and target advertisement data.

[0107] Before the advertisement conversion rate prediction method provided by the present disclosure is introduced in detail, the background part of the advertisement conversion rate prediction method provided in the related art can be introduced in detail. First, the advertisement conversion rate prediction method can be applied to an advertisement recommendation system, and the advertisement recommendation system can be used to complete some processes such as person-goods matching, advertisement delivery, data processing. Among them, the person-goods matching refers to the process of matching the advertisement content with the target audience, so that the advertisement publisher can deliver the advertisement to the group of people who are most likely to be interested in its products or services, thereby improving the effect of advertisement delivery. The advertisement recommendation system can be composed of multiple modules such as recall, rough sorting, and fine sorting. Different recommendation models will be deployed in each module, and the advertisement conversion rate prediction model in the fine sorting link is one of the most important models. However, when performing the advertisement conversion rate prediction task, a neural network model based on artificial intelligence is usually used to predict the advertisement conversion rate, and the accuracy of the predicted advertisement conversion rate is not high.

[0108] To solve the problem of low accuracy of advertisement conversion rate in the above scenario, the present application provides an advertisement conversion rate prediction method to improve the accuracy of advertisement conversion rate prediction.

[0109] As previously disclosed, the advertisement conversion rate prediction method provided by the embodiments of the present disclosure can be applied to the advertisement conversion rate prediction task in various Internet applications to push different display types of target advertisements to target objects. The target object refers to an object that wants to view the pushed target advertisement. The target advertisement can be a video, an article, or other display types, and the content contained in the target advertisement can be a specific product, application software, etc. The behavior sequence data includes a series of multiple advertisement operation behavior data of the target object arranged in time sequence within a certain time period. The advertisement operation behavior data corresponding to the target object can include data corresponding to behavior operations such as clicking on an advertisement, watching a video advertisement, filling out a form, and downloading an application. The object attribute data refers to data describing the characteristics of the target object, such as geographic location, interest, etc. According to the object attribute data, the individual characteristics of the object can be more accurately understood, and the accuracy of the advertisement conversion rate prediction can be improved. The target advertisement data refers to the related information of the target advertisement to be predicted. The target advertisement data can include advertisement content, advertisement publisher information, advertisement type, advertisement budget, etc.

[0110] The object attribute has an impact on the advertisement conversion rate. For example, for the geographic location, if the target object is traveling in city A, it may tend to click on the advertisement related to city A; if the target object travels to city B, it may tend to click on the advertisement related to city B. For example, for the interest, if the target object likes extreme sports, it may tend to click on the advertisement related to the safety equipment of extreme sports; if the target object likes smart products, it may tend to click on the advertisement related to the latest smart device.

[0111] Note that the consent of the target object is obtained in advance when obtaining the behavior sequence data and the object attribute data of the target object. Moreover, the collection, use, and processing of the behavior sequence data and the object attribute data comply with relevant laws, regulations, and standards. When obtaining the consent of the target object, the individual permission or individual consent of the target object can be obtained through a pop-up window or by jumping to a confirmation page.

[0112] In any of the above scenarios, when the advertisement conversion rate needs to be predicted, the behavior sequence data of the target object, the object attribute data of the target object, and the target advertisement data can be obtained first. In the advertisement media flow, the object historical advertisement operation behavior data will be saved by the advertisement system log, such as advertisement clicks, advertisement likes, and advertisement comments. With the accumulation of time, these advertisement operation behavior data will constitute the behavior sequence data corresponding to the object. However, in fact, the interaction between the object and the advertisement is essentially sequence-dependent, that is, the interaction sequence between different historical objects and advertisements usually corresponds to different object preferences, and the object preferences will change over time. For example, an object plans to travel for a period of time, so he will click on advertisements related to travel. After he converts the hotel-related advertisement, he may be interested in the subsequent advertisements related to scenic spots and food. Therefore, obtaining the behavior sequence data of the target object and the object attribute data of the target object can help the subsequent model better understand the dynamic dependence relationship between the interaction sequence of the object and the advertisement, and improve the accuracy of the advertisement conversion rate prediction.

[0113] In step 220, the object attribute data and the target advertisement data are subjected to feature embedding processing to obtain attribute features, and each advertisement operation behavior data in the behavior sequence data is subjected to feature embedding processing to obtain a behavior feature sequence.

[0114] The feature embedding processing is a method of converting discrete variable data (such as object attribute data, target advertisement data or advertisement operation behavior data) into a continuous numerical vector, that is, high-dimensional sparse feature data can be mapped to a low-dimensional dense vector space. In machine learning and deep learning, feature embedding can be used to convert different types of feature data into continuous real number vectors to facilitate model learning and processing. A vector is an array composed of numerical values in different dimensions, which is a point in a multi-dimensional space. The line segment between the point and the origin in the multi-dimensional coordinate system has a size and a direction, which are the size and direction of the vector. Each numerical value in the vector is the point value of the point projected on the corresponding coordinate axis in the multi-dimensional coordinate system, that is, the vector element. The vector element can be a numerical value, a symbol or the like. The vector elements in the attribute feature and the behavior feature sequence of the embodiments of the present disclosure can be numerical values. In the advertisement conversion rate prediction process, for a plurality of advertisement operation behavior data, the feature embedding processing method can include Word2Vec (a word embedding model), Global Vectors for Word Representation (GloVe), Bidirectional Encoder Representations from Transformers (BERT), One-hot encoding, etc., to represent each advertisement operation behavior data as a fixed-length vector, and each element in the vector can represent the coordinate of the operation behavior in a certain latent semantic space.

[0115] The attribute feature refers to the vector form obtained after the object attribute data and the target advertisement data are subjected to feature embedding processing. After the feature embedding processing, the categorical variables in the object attribute data and the target advertisement data are all converted into numerical vectors, and the attribute feature can be regarded as a matrix, wherein each row represents a feature vector corresponding to an object attribute data or a target advertisement data, and each column represents a specific feature dimension.

[0116] The behavior feature sequence refers to the vector form obtained after each advertisement operation behavior data in the behavior sequence data is subjected to feature embedding processing. After the feature embedding processing, each advertisement operation behavior data is converted into a numerical vector, and then arranged in the order in the original behavior sequence data to form a behavior feature sequence. The behavior feature sequence can be regarded as a matrix, wherein each row represents a feature vector corresponding to an advertisement operation behavior data, and each column represents a specific feature dimension.

[0117] In some embodiments, the data used for ad conversion rate prediction can be divided into data with non-sequential features and data with sequential features. Non-sequential data refers to data without a clear sequential or time-series relationship, while sequential data refers to data with a clear sequential or time-series relationship. Therefore, non-sequential data can include object attribute data and target ad data. These data are typically independent, and each data point can represent an object attribute or ad attribute. For example, non-sequential data can include object interests, target ad content, and target ad publisher information. Sequential data can typically be arranged in a specific order, and the value of each feature may be influenced by the values ​​of previous features. Therefore, sequential data can be multiple ad action data items sorted in chronological order, and each ad action data item may be influenced by previously generated ad action data items. For example, if a target object is very interested in a target ad, then after liking it, they may add it to their favorites. In this case, the ad action data corresponding to the favorite action may be influenced by the ad action data corresponding to the previous like action.

[0118] like Figure 3 As shown, it is a schematic diagram of an algorithm structure for predicting the advertising conversion rate of a target object to a target advertisement provided by an embodiment of the present disclosure. As shown in the figure, the input data may include non-sequential feature data and sequential feature data. Among them, the non-sequential feature data may be object attribute data and target advertisement data, and may specifically include feature 1 to feature n1, where n1 represents the total number of features contained in the non-sequential feature data. The sequential feature data may be behavioral feature data x, where x = {x1, x2, ..., x j ,…,x n2}, that is, it can include n2 advertising operation behavior data, and x j Represents the j-th advertising operation behavior data. n2 represents the total number of features contained in the sequence feature data. n1 may be equal to n2 or not, and is not specifically limited here. Then, feature embedding processing is performed on the n1 non-sequence feature data respectively, and the n1 attribute sub-features obtained after the feature embedding processing are feature spliced ​​to obtain the attribute feature. Feature embedding processing is performed on the n2 advertising operation behavior data respectively, and the n2 behavior sub-features obtained after the feature embedding processing are feature spliced ​​to obtain the behavior feature sequence x. Therefore, after feature embedding processing of each advertising operation behavior data, an embedded vector representation is obtained, that is, the behavior feature x j For example, an advertisement action data is a user liking a video halfway through watching the video. After feature embedding, the corresponding behavior sub-features of the advertisement action data are obtained, for example, [12, 0, 1, 1, 5, 6].

[0119] It should be noted that since each behavior sub-feature and attribute sub-feature can be represented in the form of a vector, the attribute feature and the behavior feature sequence can be in the form of a matrix. For the attribute feature, each row represents an attribute sub-feature corresponding to an object attribute data, and each column represents a specific feature dimension. For the behavior feature sequence, each row represents a behavior sub-feature corresponding to an advertising operation behavior, and each column represents a specific feature dimension.

[0120] In step 230, the behavior feature sequence is input into at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features.

[0121] Among them, different neural network models in the at least two neural network models proposed by the embodiment of the disclosure can adopt different model structures to adopt different sequence feature extraction methods for sequence feature extraction of the behavior feature sequence. As shown in Figure 3 The neural network model 310 adopted by the embodiment of the disclosure can be a model constructed based on a recurrent neural network (RNN), a long short-term memory (LSTM), a gated recurrent unit (GRU), etc., which is not limited here. Therefore, the embodiment of the disclosure completes the modeling of the advertising conversion rate prediction method by adopting multiple modeling paradigms to describe the sequence feature representation semantics related to the target object from multiple angles, so as to fuse to obtain more comprehensive and accurate expression of the feature characteristics related to the target object, so as to obtain sequence features with higher depth and breadth. In this way, the accuracy of the advertising conversion rate prediction can be improved.

[0122] Among them, sequence feature extraction refers to the process of extracting useful feature information from the input behavior feature sequence. Comprehensive and accurate extraction of sequence features can improve the accuracy of advertising conversion rate prediction, thereby improving the accuracy of subsequent advertising pushing. A sequence semantic feature is used to represent the representation semantics extracted from the behavior feature sequence by a sequence feature extraction method from a perspective. That is, after sequence feature extraction by at least two neural network models, at least two sequence semantic features can be obtained. Each sequence semantic feature can be regarded as a matrix, wherein each row represents a sequence semantic feature vector, and each column represents a specific feature dimension.

[0123] It should be noted that the at least two neural network models can perform sequence feature extraction in different sequence feature extraction manners, and the at least two neural network models are obtained through joint training. The sequence feature extraction manner based on a single paradigm adopted by the related art cannot fully mine deep semantics in the behavior sequence features of the target object, and the embodiment of the present disclosure adopts at least two neural network models to perform sequence feature extraction on the behavior feature sequence of the target object in different sequence feature extraction manners, so that more accurate sequence semantic features of the object behavior can be extracted from different dimensional perspectives.

[0124] In some embodiments, as shown in FIG. 4, another algorithm structure diagram for predicting the advertisement conversion rate of the target object to the target advertisement is provided by the embodiment of the present disclosure. The at least two neural network models can include at least two first neural network models 410 and a second neural network model 420. The first neural network model 410 can perform sequence feature extraction based on the behavior feature sequence, and the second neural network model 420 can perform sequence feature extraction based on the behavior feature sequence and the advertisement features of the target advertisement data. That is, the first neural network model 410 is mainly used for sequence feature extraction based on the behavior feature sequence, and the obtained sequence semantic features can obtain the time sequence relationship between the behaviors of the target object, which is helpful to understand the interests and needs of the target object. The at least two first neural network models 410 can output at least two different sequence semantic features. The second neural network model 420 is mainly used for sequence feature extraction based on the behavior feature sequence and the advertisement features of the target advertisement data, and the obtained sequence semantic features not only obtain the time sequence relationship between the behaviors of the target object, but also combine the characteristics of the target advertisement, which is helpful to improve the accuracy of the advertisement conversion rate prediction. Figure 4 In some embodiments, the behavior feature sequence is input into the at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features, including:

[0125] The behavior feature sequence is input into the at least two first neural network models for sequence feature extraction to obtain at least two first sub-sequence semantic features;

[0126] The behavior feature sequence is input into the at least two first neural network models for sequence feature extraction to obtain at least two first sub-sequence semantic features;

[0127] The behavior feature sequence is input into the at least two first neural network models for sequence feature extraction to obtain at least two first sub-sequence semantic features;

[0128] The first subsequence semantic feature is a feature based on the output of the first neural network model, and the second subsequence semantic feature is a feature based on the output of the second neural network model. Sequence semantic features may include at least two first subsequence semantic features and second subsequence semantic features. The first subsequence semantic feature and the second subsequence semantic feature can each be represented by a matrix. The specific form of the matrix can be found in the matrix representation of each sequence semantic feature described above and will not be repeated here.

[0129] In other embodiments, the at least two neural network models may include at least two first neural network models. In this case, the behavioral feature sequence is input into the at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features. This specifically includes: inputting the behavioral feature sequence into the at least two first neural network models for sequence feature extraction to obtain at least two first subsequence semantic features. This not only allows for the extraction of more accurate sequence semantic features of object behavior from different dimensional perspectives, but also improves the efficiency of ad conversion rate prediction.

[0130] In some embodiments, as Figure 5 As shown, it is a schematic diagram of another algorithm structure for predicting the advertising conversion rate of the target object to the target advertisement provided by the embodiment of the present disclosure. Specifically, the at least two first neural network models may include models constructed by four paradigms: a global perception model 510, a proximity focus model 520, a temporal evolution model 530, and a self-attention model 540. The at least two first neural network models of the embodiment of the present disclosure may include the above-mentioned global perception model, proximity focus model, temporal evolution model, and self-attention model at the same time, or may include at least two of the above-mentioned four models, and may also introduce neural network models constructed based on other modeling paradigms, that is, they may be flexibly adjusted according to actual needs, which is not specifically limited here. Among them, the second neural network model may be an advertising heuristic model 550.

[0131] Based on this, the above-mentioned inputting of the behavioral feature sequence into at least two first neural network models to perform sequence feature extraction, and obtaining at least two first subsequence semantic features, specifically including:

[0132] Based on the global perception model, multiple behavioral features in the behavioral feature sequence are subjected to nonlinear transformation processing, and the processed multiple behavioral features are average-pooled to obtain a first subsequence semantic feature;

[0133] and performing nonlinear transformation and summation processing on adjacent behavior features in the behavior feature sequence based on the proximity focus model to obtain a first subsequence semantic feature, where the adjacent behavior features are a preset number of behavior features whose corresponding behavior time is closest to the predicted time;

[0134] and a first sub-sequence semantic feature is obtained by performing recurrent feature updating on the plurality of behavior features in the behavior feature sequence according to a time sequence based on a time sequence evolution model;

[0135] and a first sub-sequence semantic feature is obtained by performing self-attention processing on the plurality of behavior features in the behavior feature sequence based on a self-attention model.

[0136] Among them, the first modeling paradigm is a global perception-based modeling method. The global perception model can perceive all node information of the sequence and can fuse all node information at one time to obtain a global perception sequence semantic feature. The global perception model used in the embodiments of the present disclosure can be RNN, GRU, etc.

[0137] Specifically, as shown in Figure 6 , a model structure diagram of the global perception model 510 provided by the embodiments of the present disclosure. The global perception model 510 can include a fully connected layer (Dense layer), an activation layer, and a pooling layer. In the global perception model 510, the Dense layer can allow each neuron to be connected to all neurons of the previous layer, and each connection has a weight, which can learn the relationship between different features. The Dense layer performs linear transformation and feature learning by learning weights and biases, so that the corresponding neural network can perform complex nonlinear mapping on the input data. The activation layer can be a structure constructed according to the Relu function. After nonlinear mapping, the output of the Dense layer is nonlinearly converted by the Relu function, so that the corresponding neural network can learn and represent more complex patterns and features. Therefore, the plurality of behavior features in the behavior feature sequence are sequentially subjected to the Dense layer and the activation layer, which can perform nonlinear transformation processing on the plurality of behavior features x j in the behavior feature sequence x. Then, the plurality of behavior features processed by the activation layer are pooled to obtain a first sub-sequence semantic feature.

[0138] The pooling is a sampling method. The pooling can compress the dimension of the input features to speed up the operation of the model. The pooling can include average pooling and maximum pooling. After obtaining the processed multiple behavior features, the pooling layer of the first neural network model can perform average pooling on the processed multiple behavior features to obtain a first sub-sequence semantic feature; or perform maximum pooling on the processed multiple behavior features to obtain a first sub-sequence semantic feature; or perform average pooling and maximum pooling on the processed multiple behavior features to obtain maximum pooling features and average pooling features. Then, the maximum pooling features and the average pooling features are spliced to obtain a first sub-sequence semantic feature. The maximum pooling can retain the most significant features in the behavior feature sequence, and the average pooling can retain the overall and balanced features in the behavior feature sequence.

[0139] In an example, assuming that the size of the pooling kernel is 5x1 and the step is 1, if the processed multiple behavior features include {[70, 90, 80, 10, 40], [90, 80, 10, 40, 50], [80, 10, 40, 50, 30], [10, 40, 50, 30, 20], [40, 50, 30, 20, 80], [50, 30, 20, 80, 90], [30, 20, 80, 90, 100], [20, 80, 90, 100, 50]}, that is, the behavior feature sequence includes 8 behavior features.

[0140] If the average pooling is used, the values at the corresponding positions of the 8 behavior features can be calculated by average. Specifically, (70+90+80+10+40) / 5=58, (90+80+10+40+50) / 5=54, (80+10+40+50+30) / 5=42, (10+40+50+30+20) / 5=30, (40+50+30+20+80) / 5=44, (50+30+20+80+90) / 5=54, (30+20+80+90+100) / 5=64, (20+80+90+100+50) / 5]=68, and then a first sub-sequence semantic feature is “[58, 54, 42, 30, 44, 54, 64, 68]”.

[0141] If max pooling is adopted, the maximum value of the values in the positions corresponding to the 8 behavior features can be calculated. Specifically, the maximum values can be found from {[70, 90, 80, 10, 40], [90, 80, 10, 40, 50], [80, 10, 40, 50, 30], [10, 40, 50, 30, 20], [40, 50, 30, 20, 80], [50, 30, 20, 80, 90], [30, 20, 80, 90, 100], [20, 80, 90, 100, 50]} in turn, and 90, 90, 80, 50, 80, 90, 100, 100 are obtained. Then, the plurality of maximum values are spliced to obtain a first sub-sequence semantic feature of “[90, 90, 80, 50, 80, 90, 100, 100]”.

[0142] Based on this, if the global perception model of the embodiment of the present disclosure adopts the global mean pooling manner, the calculation process of the first sub-sequence semantic feature can be shown in the following formula 1:

[0143]

[0144] In formula 1, S (1) represents a first sub-sequence semantic feature based on the output of the global perception model, n represents the number of input behavior sequence data, that is, the number of advertisement operation behaviors, W (1) represents the corresponding trainable weight matrix of the Dense of the global perception model, b (1) represents the corresponding trainable bias parameter of the Dense layer of the global perception model, Relu(·) is the activation function corresponding to the activation layer, x j represents the jth behavior feature. Specifically, each behavior feature in the behavior feature sequence is first subjected to nonlinear transformation of Relu, and then the result of the nonlinear transformation is subjected to mean pooling of the corresponding dimension to obtain the output S (1) of the global perception sequence modeling paradigm.

[0145] Further, since the recent object behaviors can better reflect the current behavior interest of the target object, the information value provided is greater, and the global perception model does not perceive the time sequence semantics of the object behavior sequence. Thus, the second modeling paradigm can adopt a near focus model to specially focus on the most adjacent part nodes in the behavior sequence to obtain a sequence semantic feature that can highlight the current behavior interest of the target object. The near focus model adopted by the embodiment of the present disclosure can be RNN, GRU, etc.

[0146] Specifically, as Figure 7As shown, it is a model structure diagram of the proximity focusing model 520 provided by the embodiments of the present disclosure. The proximity focusing model 520 can include a filter, a Dense layer, an activation layer, and an output layer. In the proximity focusing model 520, the filter can be used to select proximity behavior features from the behavior feature sequence and mask the behavior features that are not the nearest neighbors, so as to perceive the time sequence semantics in the behavior feature sequence of the target object that can highlight the current behavior interest of the target object. The proximity behavior features are a preset number of behavior features corresponding to the behavior time closest to the prediction time. The behavior time is the time corresponding to the behavior feature used to determine whether it is the nearest behavior feature, and the prediction time is a time within the time interval of the proximity behavior feature. The proximity focusing model 520 used in the embodiments of the present disclosure can explicitly intercept the feature nodes within the interval [t-1, t-k], and mask the remaining feature nodes, where k is a preset number, t-1 represents the behavior time of the first behavior feature corresponding to the behavior time closest to the prediction time, and t-k represents the behavior time of the kth behavior feature corresponding to the behavior time closest to the prediction time. In this way, the embodiments of the present disclosure can start to determine the behavior time corresponding to the behavior feature closest to the prediction time, and select a preset number of behavior features closest to the prediction time. The Dense layer and the activation layer structure of the proximity focusing model 520 are the same as the corresponding structures in the global perception model, and will not be described here.

[0147] In the embodiments of the present disclosure, the size of the preset number can reflect the size of the interest attention time interval of the target object, for example, it can be 4, 6, 8, etc. The larger the preset number is, the more proximity behavior features of the target object that occur recently can be obtained, and the current behavior interest of the target object can be better reflected.

[0148] In some embodiments, the calculation process of the first subsequence semantic feature based on the proximity focusing model can be as shown in the following formula 2 and formula 3:

[0149]

[0150]

[0151] In formula 2, S (2) represents a first subsequence semantic feature output based on the proximity focusing model, n represents the number of input behavior sequence data, i.e., the number of advertisement operation behaviors, W (2) represents the weight matrix of the Dense layer in the proximity focusing model, b (2) represents the bias parameter of the Dense layer in the proximity focusing model, Relu(·) is the activation function corresponding to the activation layer, x j represents the jth behavior feature.

[0152] In Formula 3, δ represents the function corresponding to the filter. The filter selects the k nearest neighbor new behavioral features, that is, obtains a preset number of adjacent behavioral features of the corresponding advertising action after feature embedding. Then, through the Dense layer and activation layer, these adjacent behavioral features are nonlinearly transformed and summed to obtain a first subsequence semantic feature, which can reflect the current behavioral interests of the target object.

[0153] Furthermore, in addition to obtaining the interests of the target object at the most recent time point, if we can understand what the target object's interests were in the past period of time and how the past interests gradually changed, this will be helpful for a deeper understanding of the target object's interest evolution over time and improving the accuracy of advertising conversion rate prediction. In this way, the third modeling paradigm is to use a sequence modeling method based on temporal evolution to obtain sequence semantic features that can reflect the target object's interest evolution over time. The temporal evolution model used in the embodiment of the present disclosure can be RNN, LSTM, etc.

[0154] Specifically, if Figure 8 The figure shows a schematic diagram of a model structure of a time series evolution model provided by an embodiment of the present disclosure. The time series evolution model 530 includes an LSTM layer, and the LSTM layer can perform cyclic feature updates on multiple behavioral features in the behavioral feature sequence according to the time series. That is, the LSTM layer can update the behavioral feature x corresponding to the time with the largest time interval for the current advertising conversion rate prediction. n To begin, consider x n The behavior feature sequence is encoded in the time progressive order to the current most recent time x1. In this way, the output vector of the current most recent time x1 represents the semantic feature of the first subsequence of the behavior feature sequence, which is recorded as S (3) .

[0155] In some embodiments, the calculation process of calculating the semantic feature of the first subsequence based on the temporal evolution model 530 can be shown in the following formula 4 and formula 5:

[0156] S (3) =LSTM(h2,x1) (Formula 4)

[0157] h j =LSTM(h j+1 ,x j ) (Formula 5)

[0158] In the above formulas 4 and 5, h j is the hidden layer representation of the j-th node in the LSTM layer. At this time, the LSTM layer will use the behavioral feature x at the j-th moment jand the node x at the previous moment j+1 The output h of the LSTM layer at the jth moment j+1 The output h of the LSTM layer at the jth moment after processing by the LSTM layer j .

[0159] Further, since the advertising interest of the target object is not unique and is changeable, for example, the target object intends to travel for a period of time, he may click on the advertisement related to travel, and after he converts the advertisement related to the hotel, he may be interested in the advertisement related to the scenic spot and food, and then may click on the advertisement related to the scenic spot and food. Therefore, the behavior of the target object has very strong randomness and flexibility, thereby easily leading to a very messy and low-quality behavior feature sequence corresponding to the generated behavior, and the prediction accuracy of the advertisement conversion rate is low. Thus, the fourth modeling paradigm adopts a sequence modeling method of a self-attention mechanism, that is, the self-attention model constructed can capture the main interest of the target object from the sequence and reduce the interference of noise nodes.

[0160] Specifically, as shown in Figure 9 , a model structure diagram of a self-attention model provided by the embodiment of the disclosure. The self-attention model 540 is constructed based on a self-attention mechanism, which can find the key nodes in the behavior feature sequence by extracting the associated interaction semantics between the nodes corresponding to each behavior feature in the behavior feature sequence, and give the key nodes a larger attention weight. At the same time, the nodes with smaller influence degree screened out are given as small a weight as possible to weaken their interference.

[0161] In some embodiments, the calculation process of the first sub-sequence semantic feature based on the self-attention model can be as shown in the following formula 6 and formula 7:

[0162]

[0163] Q, K, V = {W (q) x, W (k) x, W (v) x} (formula 7)

[0164] In the above formula, Softmax(·) represents a normalized exponential function, Q, K, and V represent vectors corresponding to query (Query), key (Key), and value (Value) parameters in the self-attention model, respectively, W (q) represents a weight matrix corresponding to the query parameter in the self-attention model, W (k) represents a weight matrix corresponding to the key parameter in the self-attention model, and W (v) represents a weight matrix corresponding to the value parameter in the self-attention model, and d k is the dimension of the vector K corresponding to the key parameter.

[0165] wherein, according to formula 7, for each behavior feature x in the behavior feature sequence x j The Q, K and V in the self-attention operation can be respectively obtained through different parameter matrix transformations. Then, according to formula 6, the Q and K are mapped through the Softmax(·) function after the interaction operation of matrix multiplication, to obtain the attention weight of the node where the corresponding behavior feature is located, and then multiplied to the V to obtain the output feature of the node where each behavior feature is located. Then, the output features of the n nodes where the behavior features are located are summed to output a first sub-sequence semantic feature, denoted as S (4) .

[0166] In the self-attention model, W (q) , W (k) and W (v) are predefined weight matrices. In the self-attention mechanism, Q is a unit used for comparison with other units. For the input of each position, a corresponding query vector is calculated to measure the degree of association between the position and other positions. K is a unit used to represent the importance of other units. For the input of each position, a corresponding key vector is calculated to measure the degree of influence of other positions on the position. V is a unit used to represent the value of other units. For the input of each position, a corresponding numerical value vector is calculated to represent the value of the position. In this way, the importance between different positions can be dynamically adjusted according to the relationship between the query, key and numerical value, to obtain a better representation.

[0167] The normalization exponential function, i.e., the Softmax(·) function, is to compress a T-dimensional vector containing any real number into another T-dimensional vector, so that the range of each element is between [0, 1], and the sum of all elements is 1. Through the Softmax(·) function mapping, the disclosure embodiments can limit the attention weight of the node where the corresponding behavior feature is located to the range of [0, 1], so that the values of the attention weights corresponding to different behavior features are mapped in the same range, which are comparable.

[0168] In the above embodiments, the disclosure embodiments adopt four different sequence modeling paradigms to construct at least two first neural network models, to respectively perform sequence feature extraction on the behavior feature sequence of the target object in different sequence feature extraction manners, for example, the global perception model, the nearby focus model, the time sequence evolution model and the self-attention model can be respectively used for different sequence feature extraction. At least two sequence semantic features are obtained. In this way, the output results of different models are fused to realize complementary advantages, to increase the depth and breadth of sequence modeling semantic mining, thereby improving the accuracy of advertisement conversion rate prediction.

[0169] In some embodiments, the advertisement feature is a numerical vector converted from the target advertisement data after the feature embedding processing, which can be denoted as f. As shown in Figure 10 FIG. 5 is a schematic diagram of a model structure of an advertisement heuristic model 550 provided by an embodiment of the present disclosure.

[0170] In some embodiments, the advertisement feature and the behavior feature sequence are input into a second neural network model for sequence feature extraction to obtain a second sub-sequence semantic feature, which specifically includes:

[0171] Based on the advertisement heuristic model, the advertisement feature and the behavior feature sequence are subjected to nonlinear transformation processing, and the processed multiple advertisement transformation features are subjected to normalization processing.

[0172] Based on the processed multiple advertisement weights, the advertisement feature and the behavior feature sequence are subjected to weighted calculation to obtain the second sub-sequence semantic feature.

[0173] The process of nonlinear transformation processing of the advertisement feature and the behavior feature sequence based on the target advertisement heuristic model can be referred to the Dense layer and the activation layer structure described above, and will not be described here.

[0174] In some embodiments, the multiple advertisement weights of the embodiment of the present disclosure can be calculated based on the attention mechanism. Thus, the process of nonlinear transformation processing of the advertisement feature and the behavior feature sequence based on the target advertisement heuristic model, the normalization processing of the processed multiple advertisement transformation features, and the obtaining of the multiple advertisement weights can be shown in the following formula 8 and formula 9:

[0175]

[0176] F j = Relu(U (f) f + W (f) x j + b (f) ) (Formula 9)

[0177] In the above formula, U (j) represents the hidden layer parameter matrix related to the advertisement feature f, W (f) represents the hidden layer parameter matrix related to the behavior feature x j at the jth moment, and b (f) represents the bias parameter trainable in the advertisement heuristic model 550. Specifically, the advertisement feature f and the behavior feature x j at the jth moment are respectively multiplied by the hidden layer parameter matrix (i.e., including U (f) and W (f) ), and then are superimposed together, and then are subjected to Relu activation function transformation to obtain the importance measure F j of the behavior feature x jThe behavior feature x is then normalized by the attention mechanism to obtain a behavior feature x j The weight of the corresponding node The dimension of the attention weight vector is n x 1, which is one-to-one corresponding to the length of the behavior feature sequence. Therefore, the node corresponding to the behavior feature more relevant to the target advertisement will be assigned a higher weight.

[0178] In combination with the above embodiments, the process of weighting the advertisement features and the behavior feature sequence based on the processed multiple advertisement weight vectors to obtain the second sub-sequence semantic feature can be shown in the following formula 10:

[0179]

[0180] In formula 10, the behavior feature x at the jth moment is multiplied by the advertisement feature f bit by bit, and then the behavior feature x is multiplied by the result after multiplication. j j The weight of the corresponding node The n behavior features corresponding to the weighted vectors are summed up to obtain the second sub-sequence semantic feature S output by the target advertisement heuristic model. (5) .

[0181] According to at least two first neural network models and a second neural network model, the behavior feature sequence of the target object is extracted by different sequence feature extraction methods, which can extract more accurate sequence semantic features of the object behavior from different dimensional perspectives. For example, five sequence modeling paradigms can be constructed as shown in the above examples, which can include global perception model, near focus model, time sequence evolution model, self-attention model and target advertisement heuristic model. The five sequence modeling paradigms are calculated independently, which can avoid the problem that a single model cannot fully mine the deep semantic features of the object behavior sequence in related technologies. Therefore, the advertisement conversion rate prediction method provided by the present disclosure can improve the accuracy of advertisement conversion rate prediction.

[0182] In step 240, at least two sequence semantic features are fused, and the fused features are spliced with the attribute features to obtain target features.

[0183] After obtaining at least two first sub-sequence semantic features and second sub-sequence semantic features, the at least two sequence semantic features are fused, and the fused features are spliced with the attribute features to obtain target features, which specifically includes:

[0184] ​The at least two first sub-sequence semantic features and the second sub-sequence semantic feature are fused, and the fused feature is spliced with the attribute feature to obtain the target feature.

[0185] In some embodiments, the feature fusion of the at least two sequence semantic features refers to merging sequence semantic features extracted from different neural network models to generate a more comprehensive and representative feature representation. In this way, the feature fusion of the at least two sequence semantic features can be feature splicing of the at least two sequence semantic features to splice the at least two sequence semantic features along a certain dimension to form a fused feature.

[0186] In other embodiments, the feature fusion of the at least two sequence semantic features can also be element-wise addition of the at least two sequence semantic features with the same dimension to obtain the fused feature.

[0187] In other embodiments, the feature fusion of the at least two sequence semantic features can also be feature average value calculation of the at least two sequence semantic features with the same dimension to obtain the fused feature.

[0188] In addition, in some embodiments, the feature fusion of the at least two sequence semantic features and the splicing of the fused feature with the attribute feature to obtain the target feature specifically includes:

[0189] obtaining weight coefficients corresponding to the at least two sequence semantic features;

[0190] performing weighted calculation on the at least two sequence semantic features based on the weight coefficients to obtain the fused feature;

[0191] splicing the fused feature with the attribute feature to obtain the target feature.

[0192] The weight coefficient is used to measure the importance of each sequence semantic feature in the fused feature. The greater the value of the weight coefficient, the more important the corresponding sequence semantic feature in the fused feature. These weight coefficients can be obtained by various methods, such as manual setting, model-based training, reinforcement learning, etc.

[0193] In some embodiments, obtaining the weight coefficients corresponding to the at least two sequence semantic features includes:

[0194] obtaining advertisement environment data and advertisement context data corresponding to the target advertisement data;

[0195] determining the weight coefficient corresponding to each sequence semantic feature according to the advertisement environment data and the advertisement context data.

[0196] The advertisement environment data can include application information of advertisement publishing, network environment, geographical location, etc. The advertisement context data can include behavior history of the target object, search record, browsing record, and other information related to the advertisement. By obtaining the advertisement environment data and the advertisement context data corresponding to the target advertisement data, the relationship between the target object and the advertisement, and the behavior characteristics of the target object in different situations can be better understood.

[0197] It should be noted that when obtaining the advertisement environment data and the advertisement context data corresponding to the target advertisement data, the consent of the target object is required in advance. Moreover, the collection, use, and processing of the advertisement environment data and the advertisement context data will comply with relevant laws, regulations, and standards. When seeking the consent of the target object, the separate permission or separate consent of the target object can be obtained through a pop-up window or by jumping to a confirmation page.

[0198] In some embodiments, determining the weight coefficient corresponding to each sequence semantic feature according to the advertisement environment data and the advertisement context data can specifically include: determining a first score based on the advertisement environment data; determining a second score based on the advertisement context data; calculating a total score based on the first score and the second score, and determining the weight coefficient corresponding to each sequence semantic feature according to the total score.

[0199] Specifically, determining the first score based on the advertisement environment data can be in the form of a comparison table, which can specifically set a corresponding score for each sub-data of the application information of advertisement publishing, network environment, geographical location, etc. in the advertisement environment data.

[0200] Determining the second score based on the advertisement context data can also be in the form of a comparison table, which can specifically set a corresponding score for each sub-information of the behavior history of the target object, search record, browsing record, etc. related to the advertisement in the advertisement context data.

[0201] In this way, calculating the total score based on the first score and the second score can be to calculate the mean of the first score and the second score to obtain the total score, or to set corresponding weights for the advertisement environment data and the advertisement context data, and then to calculate the weighted average of the first score and the second score based on the weights to obtain the total score. The higher the total score is, the more important the corresponding sequence semantic feature is in the fused feature, and then a higher weight coefficient can be set for it.

[0202] The embodiments of the present disclosure can improve the accuracy of the obtained fusion features by determining the weight coefficients through fractions, thereby improving the accuracy of the advertisement conversion rate prediction. Moreover, the influence of the advertisement environment data and the advertisement context data on the calculation of the total score can be made the same by calculating the total score using the average method, thereby improving the fairness of determining the weight coefficients corresponding to each sequence semantic feature. The flexibility of determining the weight coefficients corresponding to each sequence semantic feature can be improved by flexibly setting different weights for the advertisement environment data and the advertisement context data according to actual needs by calculating the total score using the weighted average method.

[0203] In some other embodiments, after obtaining at least two sequence semantic features, the at least two sequence semantic features can be effectively fused through an attention mechanism to obtain the final fusion features of the sequence feature modeling. Thus, obtaining the weight coefficients corresponding to the at least two sequence semantic features can specifically include:

[0204] obtaining a preset weight matrix;

[0205] calculating the weight coefficients corresponding to the at least two sequence semantic features based on the preset weight matrix and the at least two sequence semantic features.

[0206] The preset weight matrix is obtained during the training of the at least two neural network models, and can be denoted as W (a) The preset weight matrix is a preset matrix, and the elements thereof represent the relative importance between different feature dimensions. The preset weight matrix after training is the weight matrix in actual application. At this time, calculating the weight coefficients corresponding to the at least two sequence semantic features based on the weight matrix and the at least two sequence semantic features can specifically include:

[0207] dimensionally expanding and splicing the at least two sequence semantic features to obtain an expanded feature;

[0208] calculating the weight coefficients corresponding to the at least two sequence semantic features based on the preset weight matrix and the expanded feature.

[0209] The dimension expansion and splicing refers to adding additional dimensions to each sequence semantic feature on the basis of maintaining the original feature dimensions, so as to obtain more information in subsequent calculations. The dimension expansion and splicing can be implemented by adding zero elements at the end of the vector of the sequence semantic feature, copying existing elements, or using an embedding layer, etc. Thus, the specific calculation process of the expanded feature vector obtained after the dimension expansion and splicing of the at least two sequence semantic features can be shown in the following formula 11:

[0210] E = Expand_dim ([S (1) ,S (2) ,S (3) ,S (4) ,S(5) ]) (Formula 11)

[0211] In Formula 11, E represents the expanded dimension feature, Expand_dim(·) represents an expansion function for expansion dimension splicing. Therefore, the matrix vector E is obtained by the expansion function Expand_dim(·), d is the dimension of the output vector of each paradigm model (i.e., the first neural network model), and 5 represents the total number of models of the first neural network model and the second neural network model adopted.

[0212] Further, because the at least two sequence semantic features can be effectively fused through the attention mechanism to obtain the final fused feature for sequence feature modeling, based on the preset weight matrix and the expanded dimension feature, the weight coefficients corresponding to the at least two sequence semantic features can be calculated, which can specifically include:

[0213] The attention parameter matrix is calculated according to the preset weight matrix and the expanded dimension feature.

[0214] The weight coefficients corresponding to the at least two sequence semantic features are calculated based on the attention parameter matrix.

[0215] Among them, the attention parameter matrix is calculated according to the preset weight matrix and the expanded dimension feature, which can mean that the output vectors of different first neural network models are mapped into an attention parameter matrix by using a nonlinear activation function.

[0216] In some embodiments, if a Tanh nonlinear activation function is used, the process of calculating the attention parameter matrix can be shown in the following Formula 12:

[0217] A = Tanh(W (a) E + b (a) ) (Formula 12)

[0218] In Formula 12, A represents the attention parameter matrix, Tanh(·) is a nonlinear activation function based on the hyperbolic tangent function, b (a) represents a trainable bias parameter in the fusion splicing process. After obtaining the expanded dimension feature E, E is subjected to nonlinear transformation by the Tanh(·) activation function to obtain the attention parameter matrix A, and the dimension of A is

[0219] Among them, the weight coefficients corresponding to the at least two sequence semantic features are calculated based on the attention parameter matrix, which can be specifically shown in the following Formula 13:

[0220]

[0221] In Formula 13, π i represents the weight coefficient corresponding to the i-th sequence semantic feature, and πi The value range is [0,1]. According to π i Higher weight coefficients can be assigned to the outputs of higher value neural network models, and as shown in the above formula, π i The calculation is dynamically determined based on the state of each neural network model.

[0222] Furthermore, after obtaining the weight coefficients corresponding to at least two sequential semantic features, a weighted calculation is performed on the at least two sequential semantic features based on the weight coefficients to obtain a fused feature. Specifically, each sequential semantic feature is multiplied by its corresponding weight coefficient, and then all weighted feature vectors are added together to obtain a fused feature. The process of calculating the fused feature is shown in Formula 14 below:

[0223]

[0224] In Formula 14, S represents the fusion feature, S (i) Represents the sequence semantic features output by the i-th neural network model. The disclosed embodiment can comprehensively consider at least two sequence semantic features from different neural network models and weight them according to their weight coefficients to fuse them to obtain a sequence feature with greater depth and breadth, thereby improving the accuracy of ad conversion rate prediction.

[0225] Furthermore, in some embodiments, Figure 3 As shown, the fusion features and attribute features are concatenated to obtain the target features. Specifically, the attribute features can be input into a feature representation extraction network to perform deep extraction of the attribute features after feature embedding, thereby obtaining a higher-level feature representation, which helps to better understand the intrinsic structural relationship of the data; then, the features output by the feature representation extraction network and the fusion features are concatenated to obtain the target features. The feature representation extraction network can be a convolutional neural network (CNN), RNN, autoencoder, etc., which is not specifically limited here.

[0226] The feature representation extraction network can also include nonlinear activation functions, which enables the network to learn nonlinear relationships in the data. Based on the feature representation extraction network, a more comprehensive and representative feature representation can be generated, further improving the quality and expressiveness of the features, thereby improving the accuracy of ad conversion rate prediction.

[0227] In other embodiments, Figure 3As shown, the target feature is obtained by splicing the fusion feature and the attribute feature, and specifically can include: splicing the fusion feature and the attribute feature to obtain initial splicing features, and inputting the initial splicing features into the feature representation extraction network to perform splicing feature extraction, to obtain the target feature. In this way, the feature representation extraction network can fuse the fusion feature and the attribute feature from different embedding layers to generate a more comprehensive and representative feature representation, thereby improving the accuracy of the advertisement conversion rate prediction.

[0228] At step 250, the advertisement conversion rate of the target object to the target advertisement is predicted based on the target feature.

[0229] The advertisement conversion rate of the target object to the target advertisement refers to the probability of completing the expected conversion after the target object clicks the target advertisement. After the target feature is determined, the advertisement conversion rate of the target object to the target advertisement can be predicted based on the target feature. Because the target feature of the present application is obtained by splicing the fused feature and the attribute feature from different perspectives through multiple different neural network models, it can fully mine the deep semantic problems of the object behavior sequence feature, and improve the accuracy of the advertisement conversion rate prediction.

[0230] In some embodiments, the advertisement conversion rate of the target object to the target advertisement is predicted based on the target feature, specifically including:

[0231] The target feature is input into the trained fully connected network model for feature mapping to obtain an output feature;

[0232] The output feature is normalized to obtain the advertisement conversion rate of the target object to the target advertisement.

[0233] The fully connected network model is a kind of deep learning model, in which the neurons of each layer are connected to all the neurons of the previous layer. The fully connected network model can learn the complex relationship between the input features. In the scenario of advertisement conversion rate prediction, the fully connected network model can be used to learn the mapping relationship between the target feature and the advertisement conversion rate. For example, Figure 3 As shown, after obtaining the target feature, the target feature is input into the trained fully connected network model for feature mapping to obtain an output feature.

[0234] The feature mapping refers to a series of linear and nonlinear transformations on the input target feature, which can be understood as mapping the target feature space to a higher-dimensional feature space. In this new space, the relationship between the advertisement conversion rate and other features can be linear. After processing by the layers of the fully connected network model, an output feature will be obtained. Each vector element of the output feature can represent a specific advertisement conversion rate prediction value, or a component of the advertisement conversion rate prediction value.

[0235] Furthermore, because the value range of the output features can be very large, directly using these values ​​as predictions for ad conversion rates may lead to unreasonable prediction results. Therefore, the output features can usually be normalized and converted to an appropriate range, such as [0, 1], to obtain the ad conversion rate of the target object for the target ad. Among them, normalization methods may include min-max scaling (Min-Max Scaling), Z-score normalization, etc. In this way, the ad conversion rate of the target object for the target ad obtained after normalization can be used to sort ads to select the best ads for display, or for other ad optimization strategies.

[0236] The disclosed embodiments employ at least two neural network models to extract sequence features from the target object's behavioral feature sequence using different sequence feature extraction methods, thereby extracting more accurate sequence semantic features of the object's behavior from different dimensional perspectives. Specifically, the disclosed embodiments propose a neural network model that employs at least two different sequence feature extraction methods to extract features from the target object's behavioral feature sequence. The extracted features are then weighted using different weight coefficients to adjust their importance, thereby weighting them to obtain more accurate sequence fusion features. The at least two neural network models include a global perception model, a proximity focus model, a temporal evolution model, a self-attention model, and an advertising heuristic model. Feature fusion is then performed on the at least two sequence semantic features, and the fused features are concatenated with attribute features to obtain target features, thereby improving the accuracy of ad conversion rate prediction. Compared to the single model employed in related technologies, the disclosed embodiments can fully exploit the deep semantics of the object's behavioral sequence features and can extract features from longer behavioral feature sequences to obtain sequence features with greater depth and breadth, thereby improving the accuracy of ad conversion rate prediction.

[0237] The embodiments of the present disclosure are described in detail in conjunction with specific application scenarios.

[0238] like Figure 11 The figure shows another flow chart of the advertising conversion rate prediction method provided by the present disclosure. This embodiment will use the advertising conversion rate prediction scenario as an example to introduce the advertising conversion rate prediction method in detail in combination with the execution subject of each step. The method may specifically include the following steps:

[0239] Step 1101: The terminal constructs an advertisement conversion rate prediction model and obtains training sample data.

[0240] In this embodiment, the advertisement conversion rate prediction method provided by the present disclosure will be introduced in detail taking the advertisement conversion rate prediction scenario as an example. The advertisement conversion rate prediction scenario can be triggered on various public networks, such as social platforms, video platforms, or e-commerce platforms, etc. The platforms provide the functions of clicking on advertisements and jumping to the purchase links of corresponding products, or viewing the complete content of the advertisements, etc. Therefore, accurately pushing the advertisements required by the objects and improving the advertisement conversion rate of the objects to the pushed advertisements can improve the use experience of the objects of the platform. The terminal in this embodiment can be any one of a personal computer, a mobile device, a wearable smart device, a vehicle-mounted device, or a server device, etc.

[0241] To improve the accuracy of advertisement conversion rate prediction, the present application provides an advertisement conversion rate prediction method based on a neural network model. The advertisement conversion rate prediction model used for advertisement conversion rate prediction can include an embedding layer, at least two sequence feature extraction layers, a feature fusion layer, a feature splicing layer, and a prediction layer. Each sequence feature extraction layer is a structure based on a neural network model. The structure of the advertisement conversion rate prediction model provided by the present application can refer to the model structure shown in Figure 5 The global perception model, the nearby focus model, the time sequence evolution model, the self-attention model, and the advertisement heuristic model can correspond to one sequence feature extraction layer, respectively. Compared with the single model used in the related art, the present embodiment adopts at least two sequence feature extraction layers with different sequence feature extraction methods, which can extract more accurate sequence semantic features of the object behavior from different dimensional perspectives, thereby improving the accuracy of advertisement conversion rate prediction.

[0242] After constructing the advertisement conversion rate prediction model with the structure shown in Figure 5 The training sample data for training the advertisement conversion rate prediction model can be further obtained. The training sample data includes a plurality of sample data groups and the conversion label data corresponding to each sample data group. The conversion label data can be a binary classification label, which can be 0 or 1, where 0 represents a conversion sample group and 1 represents a non-conversion sample group. The conversion label data can also be a sample conversion probability value, which can be any value in [0, 1]. The larger the sample conversion probability value, the greater the likelihood of conversion of the sample data group.

[0243] Each sample data group includes sample object attribute data, sample behavior sequence data, and sample advertisement data of a sample object. The sample object attribute data, sample behavior sequence data, and sample advertisement data have the same specific meanings as the object attribute data, behavior sequence data, and advertisement data in step 210, respectively. However, they are used as training samples here and will not be described again.

[0244] At step 1102, the terminal performs feature embedding on the sample object attribute data and the sample advertisement data based on an embedding layer of the advertisement conversion rate prediction model to obtain sample attribute features, and performs feature embedding on the sample behavior sequence data to obtain a sample behavior feature sequence.

[0245] Further, the sample attribute features and the sample behavior feature sequence obtained by performing feature embedding on the sample object attribute data and the sample advertisement data and on the sample behavior sequence data can preserve the feature independence of each data while ensuring their consistent dimensions, facilitating subsequent processing and analysis.

[0246] At step 1103, the terminal inputs the sample behavior feature sequence into at least two sequence feature extraction layers of the advertisement conversion rate prediction model for sequence feature extraction to obtain at least two sample sequence semantic features.

[0247] Further, the terminal inputs the sample behavior feature sequence into at least two sequence feature extraction layers of the advertisement conversion rate prediction model for sequence feature extraction, which can include a sequence feature extraction layer constructed based on a global perception model, a sequence feature extraction layer constructed based on a nearby focus model, a sequence feature extraction layer constructed based on a time sequence evolution model, a sequence feature extraction layer constructed based on a self-attention model, and a sequence feature extraction layer constructed based on an advertisement heuristic model. In this way, the terminal can obtain sample sequence semantic features output by each sequence feature extraction layer.

[0248] At step 1104, the terminal performs feature fusion on the at least two sample sequence semantic features based on a feature fusion layer of the advertisement conversion rate prediction model to obtain a sample fusion feature.

[0249] Further, the sample fusion feature can better obtain the correlation and interaction information between different sample sequence semantic features.

[0250] The process of performing feature fusion on the at least two output sample sequence semantic features to obtain the sample fusion feature can include the following steps.

[0251] Performing dimension expansion splicing on the at least two output sample sequence semantic features to obtain a sample dimension expansion feature;

[0252] Obtaining a weight matrix, and the parameters in the weight matrix are adjustable;

[0253] Calculating sample weight coefficients of the at least two sample sequence semantic features based on the weight matrix and the sample dimension expansion feature;

[0254] Fusing the at least two sample sequence semantic features according to the sample weight coefficients to obtain the sample fusion feature.

[0255] The specific process of the above steps can be referred to Formulas 11 to 14, which will not be described here again. The embodiments of the present disclosure can comprehensively consider at least two sequence semantic features from different sequence feature extraction layers, and weight them according to their weight coefficients, so as to fuse sequence features with higher depth and breadth, thereby improving the accuracy of the advertisement conversion rate prediction.

[0256] In step 1105, the terminal splices the sample attribute feature and the sample fusion feature based on the feature splicing layer of the advertisement conversion rate prediction model, to obtain a sample target feature.

[0257] Further, the process of splicing the sample attribute feature and the sample fusion feature to obtain the sample target feature can specifically include: inputting the sample attribute feature into the attribute feature extraction layer for attribute feature extraction, and splicing the feature after attribute feature extraction and the sample fusion feature to obtain the sample target feature. In this way, the key information inside the sample attribute feature can be deeply mined.

[0258] In another embodiment, the process of splicing the sample attribute feature and the sample fusion feature to obtain the sample target feature can specifically include: splicing the sample attribute feature and the sample fusion feature to obtain an initial sample splicing feature; inputting the initial sample splicing feature into the attribute feature extraction layer for attribute feature extraction to obtain the sample target feature. In this way, the interaction information between the sample attribute feature and the sample fusion feature can be more deeply mined.

[0259] In step 1106, the terminal performs conversion rate prediction on the spliced sample target feature based on the prediction layer of the advertisement conversion rate prediction model, to obtain an output conversion rate prediction value.

[0260] Further, the terminal can input the spliced sample target feature into the prediction layer of the advertisement conversion rate prediction model for conversion rate prediction, to obtain the conversion rate prediction value. At this time, the conversion rate prediction value can be a binary classification, i.e., 0 or 1, 0 indicating conversion success and 1 indicating conversion failure, or the conversion rate prediction value can also be a specific probability value, such as 96%, indicating that the conversion success is more likely. If the conversion label data is a binary classification label value, the corresponding conversion rate prediction value is also a label value, and if the conversion label data is a probability value, the corresponding conversion rate prediction value is also a probability value.

[0261] The prediction layer of the advertisement conversion rate prediction model can be a third neural network model, which can be specifically a CNN, a fully connected network, etc., which will not be described in detail here.

[0262] Step 1107, the terminal calculates a loss value according to the conversion label data and the conversion rate prediction value, and updates the model parameters of the at least two neural network models based on the loss value to obtain the trained advertisement conversion rate prediction model.

[0263] Further, the model parameters of the at least two neural network models of the advertisement conversion rate prediction model can be updated based on the difference between the conversion rate prediction value output by the prediction layer and the conversion label data of the sample data set. In addition, the model parameters in each layer of the advertisement conversion rate prediction model can also be updated based on the difference between the conversion rate prediction value output by the prediction layer and the conversion label data of the sample data set, which is not limited. Specifically, the loss value can be calculated according to the conversion rate prediction value and the conversion label data of the sample data set, and then the back propagation gradient is calculated based on the loss value, and the gradient back propagation processing is performed based on the back propagation gradient to update the model parameters and the parameters in the weight matrix of the at least two neural network models.

[0264] Further, a new batch of training sample data can be reacquired, and the advertisement conversion rate prediction model can be trained in a loop according to the above training method until the model parameters of the advertisement conversion rate prediction model converge, and the training of the advertisement conversion rate prediction model is completed. After the training of the advertisement conversion rate prediction model is completed, the advertisement conversion rate prediction model can be further deployed online in the terminal to perform the prediction task of the advertisement conversion rate.

[0265] Step 1108, the terminal acquires the behavior sequence data of the target object, the object attribute data of the target object, and the target advertisement data.

[0266] The behavior sequence data includes a plurality of advertisement operation behavior data of the target object in time sequence. In the terminal in which the trained advertisement conversion rate prediction model is deployed, the advertisement conversion rate prediction task of the target object to the target advertisement can be received. When the advertisement conversion rate prediction task is received, the behavior sequence data of the target object to be predicted, the object attribute data of the target object, and the target advertisement data can be acquired, and the behavior sequence data of the target object, the object attribute data of the target object, and the target advertisement data are input into the trained advertisement conversion rate prediction model to perform the advertisement conversion rate prediction task.

[0267] Step 1109, the terminal performs feature embedding processing on the object attribute data and the target advertisement data based on the embedding layer of the advertisement conversion rate prediction model to obtain attribute features, and performs feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence.

[0268] Further, the terminal can perform feature embedding processing on the object attribute data and the target advertisement data based on the embedding layer of the advertisement conversion rate prediction model to obtain attribute features, and perform feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence. At this time, the dimensions of the obtained attribute features and the behavior feature sequence are consistent with the feature dimensions obtained in the training process. The specific process has been described in detail in the above embodiment, and will not be described here.

[0269] In step 1110, the terminal inputs the behavior feature sequence into at least two sequence feature extraction layers of the advertisement conversion rate prediction model to perform sequence feature extraction to obtain at least two sequence semantic features.

[0270] Further, the behavior feature sequence is subjected to sequence feature extraction in the at least two sequence feature extraction layers of the advertisement conversion rate prediction model. Meanwhile, in one sequence feature extraction layer of the advertisement conversion rate prediction model, the interaction between the target advertisement data and the behavior feature sequence can be considered at the same time to obtain sequence semantic features from multiple perspectives.

[0271] In step 1111, the terminal performs feature fusion on the at least two sequence semantic features based on the feature fusion layer of the advertisement conversion rate prediction model to obtain fused features, and splices the attribute features and the fused features based on the feature splicing layer of the advertisement conversion rate prediction model to obtain target features.

[0272] Further, after obtaining the at least two sequence semantic features, the at least two sequence semantic features can be subjected to feature fusion to obtain fused features, and the attribute features and the fused features can be spliced to obtain target features, so as to facilitate the conversion rate prediction on the target features based on the prediction layer of the trained advertisement conversion rate prediction model.

[0273] In step 1112, the target features after splicing are subjected to conversion rate prediction based on the prediction layer of the advertisement conversion rate prediction model to obtain the advertisement conversion rate of the target object to the target advertisement.

[0274] Further, since the conversion label data corresponding to the sample data set when the advertisement conversion rate prediction model is trained is a probability value, the output advertisement conversion rate of the target object to the target advertisement is also a probability value, which is used to represent the possibility of the target object performing advertisement conversion on the target advertisement. Therefore, by adopting at least two neural network models to respectively perform sequence feature extraction on the behavior feature sequence of the target object in different sequence feature extraction manners, the advertisement conversion rate prediction method provided by the present disclosure can extract more accurate sequence semantic features of object behaviors from different dimensional perspectives, thereby avoiding the problem that a single model in the related art cannot sufficiently mine deep semantic features of object behavior sequences, and thus the advertisement conversion rate prediction method provided by the present disclosure can improve the accuracy of advertisement conversion rate prediction.

[0275] Apparatuses and devices of embodiments of the present disclosure

[0276] It can be understood that, although each step in each of the above flowcharts is displayed in sequence represented by an arrow, these steps are not necessarily executed in the order represented by the arrow. Unless otherwise specified in the embodiments, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, at least some of the steps in the above flowcharts can include multiple steps or multiple 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 executed alternately or alternately with at least part of other steps or steps or stages in other steps.

[0277] Figure 12 A structural schematic diagram of an advertisement conversion rate prediction apparatus 1200 provided by an embodiment of the present disclosure is provided. The advertisement conversion rate prediction apparatus 1200 comprises:

[0278] An acquisition unit 1210 is configured to acquire behavior sequence data of a target object, object attribute data of the target object, and target advertisement data, wherein the behavior sequence data comprises a plurality of advertisement operation behavior data of the target object in chronological order;

[0279] An embedding unit 1220 is configured to perform feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and perform feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence;

[0280] An extraction unit 1230 is configured to input the behavior feature sequence into at least two neural network models to perform sequence feature extraction, to obtain at least two sequence semantic features, wherein the at least two neural network models perform sequence feature extraction in different sequence feature extraction manners, and the at least two neural network models are obtained through joint training;

[0281] A fusion unit 1240 is configured to perform feature fusion on the at least two sequence semantic features, and splice the fused features and the attribute features to obtain target features;

[0282] A prediction unit 1250 is configured to predict an advertisement conversion rate of the target object to the target advertisement based on the target features.

[0283] Optionally, in some embodiments, the at least two neural network models comprise at least two first neural network models and a second neural network model, the first neural network model performs sequence feature extraction based on the behavior feature sequence, and the second neural network model performs sequence feature extraction based on the behavior feature sequence and advertisement features of the target advertisement data, and the extraction unit 1230 comprises:

[0284] a first extraction subunit (not shown) configured to input the behavior feature sequence into at least two first neural network models for sequence feature extraction to obtain at least two first subsequence semantic features;

[0285] a second extraction subunit (not shown) configured to perform feature embedding processing on the target advertisement data to obtain advertisement features, and input the advertisement features and the behavior feature sequence into a second neural network model for sequence feature extraction to obtain a second subsequence semantic feature;

[0286] the fusion unit 1240 is further configured to:

[0287] perform feature fusion on the at least two first subsequence semantic features and the second subsequence semantic feature, and concatenate the fused features and the attribute features to obtain the target features.

[0288] Optionally, in some embodiments, the at least two first neural network models include a global perception model, a nearby focus model, a time sequence evolution model, and a self-attention model, and the first extraction subunit (not shown) includes:

[0289] a global perception module (not shown) configured to perform nonlinear transformation processing on a plurality of behavior features in the behavior feature sequence based on the global perception model, and perform mean pooling on the processed plurality of behavior features to obtain a first subsequence semantic feature;

[0290] a nearby focus module (not shown) configured to perform nonlinear transformation and summation processing on nearby behavior features in the behavior feature sequence based on the nearby focus model to obtain a first subsequence semantic feature, the nearby behavior features being a preset number of behavior features closest to a corresponding behavior time and a prediction time;

[0291] a time sequence evolution module (not shown) configured to perform cyclic feature updating on a plurality of behavior features in the behavior feature sequence according to a time sequence based on the time sequence evolution model to obtain a first subsequence semantic feature;

[0292] a self-attention module (not shown) configured to perform self-attention processing on a plurality of behavior features in the behavior feature sequence based on the self-attention model to obtain a first subsequence semantic feature.

[0293] Optionally, in some embodiments, the fusion unit 1240 includes:

[0294] a first acquisition subunit (not shown) configured to acquire weight coefficients corresponding to the at least two sequence semantic features;

[0295] A calculation subunit (not shown) is configured to calculate at least two sequence semantic features based on the weight coefficients to obtain the fusion features.

[0296] A first splicing subunit (not shown) is configured to splice the fusion features and the attribute features to obtain the target features.

[0297] Optionally, in some embodiments, the first acquisition subunit (not shown) comprises:

[0298] A first acquisition module (not shown) is configured to acquire a preset weight matrix, the preset weight matrix being obtained during training of at least two neural network models.

[0299] A first calculation module (not shown) is configured to calculate the weight coefficients corresponding to the at least two sequence semantic features based on the preset weight matrix and the at least two sequence semantic features.

[0300] Optionally, in some embodiments, the first calculation module (not shown) comprises:

[0301] A splicing sub-module (not shown) is configured to perform dimension expansion splicing on the at least two sequence semantic features to obtain the dimension expansion features.

[0302] A calculation sub-module (not shown) is configured to calculate the weight coefficients corresponding to the at least two sequence semantic features based on the preset weight matrix and the dimension expansion features.

[0303] Optionally, in some embodiments, the calculation sub-module (not shown) is specifically configured to:

[0304] calculate an attention parameter matrix based on the preset weight matrix and the dimension expansion features;

[0305] calculate the weight coefficients corresponding to the at least two sequence semantic features based on the attention parameter matrix.

[0306] Optionally, in some embodiments, the first acquisition subunit (not shown) comprises:

[0307] A second acquisition module (not shown) is configured to acquire advertisement environment data and advertisement context data corresponding to the target advertisement data.

[0308] A determination module (not shown) is configured to determine the weight coefficients corresponding to each sequence semantic feature according to the advertisement environment data and the advertisement context data.

[0309] Optionally, in some embodiments, the advertisement conversion rate prediction device provided by the present disclosure further comprises a training unit (not shown), the training unit being configured to train the at least two neural network models, and the training unit comprising:

[0310] The second obtaining subunit (not shown) is configured to obtain training sample data, the training sample data comprising a plurality of sample data groups and conversion label data corresponding to each sample data group, and each sample data group comprising sample object attribute data, sample behavior sequence data and sample advertisement data of a sample object;

[0311] The embedding subunit (not shown) is configured to perform feature embedding on the sample object attribute data and the sample advertisement data to obtain sample attribute features, and perform feature embedding on the sample behavior sequence data to obtain a sample behavior feature sequence;

[0312] The third extracting subunit (not shown) is configured to input the sample behavior feature sequence into at least two neural network models respectively to perform sequence feature extraction, and perform feature fusion on the output at least two sample sequence semantic features to obtain sample fusion features;

[0313] The second splicing subunit (not shown) is configured to splice the sample attribute features and the sample fusion features, and input the spliced features into a third neural network model to perform conversion rate prediction, to obtain an output conversion rate prediction value;

[0314] The updating subunit (not shown) is configured to calculate a loss value based on the conversion label data and the conversion rate prediction value, and update model parameters of the at least two neural network models based on the loss value.

[0315] Optionally, in some embodiments, the third extracting subunit (not shown) comprises:

[0316] The splicing module (not shown) is configured to perform dimension expansion splicing on the output at least two sample sequence semantic features to obtain sample dimension expansion features;

[0317] The second obtaining module (not shown) is configured to obtain a weight matrix, and parameters in the weight matrix are adjustable;

[0318] The second calculating module (not shown) is configured to calculate sample weight coefficients of the at least two sample sequence semantic features based on the weight matrix and the sample dimension expansion features;

[0319] The fusion module (not shown) is configured to fuse the at least two sample sequence semantic features according to the sample weight coefficients to obtain the sample fusion features;

[0320] The updating subunit (not shown) is further configured to:

[0321] update the model parameters of the at least two neural network models and the parameters in the weight matrix based on the loss value.

[0322] Optionally, in some embodiments, the prediction unit 1250 comprises:

[0323] a mapping subunit (not shown) configured to input the target feature into the trained full connection network model to perform feature mapping, and obtain an output feature;

[0324] a normalization subunit (not shown) configured to perform normalization processing on the output feature, and obtain an advertisement conversion rate of the target object to the target advertisement.

[0325] In the embodiments of the present application, the term "module" or "unit" refers to a computer program or a part of a computer program with a predetermined function, and works together with other related parts to achieve a predetermined target, and can be implemented entirely or partially by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of the module or unit.

[0326] Reference Figure 13 , Figure 13 To implement the structure block diagram of the terminal 110 part of the advertisement conversion rate prediction method of the embodiments of the present application, the terminal 110 includes: radio frequency (RF) circuit 1310, memory 1315, input unit 1330, display unit 1340, sensor 1350, audio circuit 1360, wireless fidelity (WiFi) module 1370, processor 1380, and power supply 1390, and the like. Those skilled in the art can understand that the terminal 110 structure shown does not constitute a limitation on mobile phones or computers, and can include more or fewer components than shown, or combine certain components, or different component arrangements. Figure 13 The terminal 110 structure shown does not constitute a limitation on mobile phones or computers, and can include more or fewer components than shown, or combine certain components, or different component arrangements.

[0327] The RF circuit 1310 can be used for receiving and sending signals in the process of receiving information or calling, and in particular, receiving the downlink information of the base station and processing it by the processor 1380; in addition, sending the uplink data designed to the base station.

[0328] The memory 1315 can be used to store software programs and modules, and the processor 1380 executes various functions of the terminal by running the software programs and modules stored in the memory 1315.

[0329] The input unit 1330 can be used to receive input digital or character information, and generate key signal input related to the setting and function control of the terminal. Specifically, the input unit 1330 can include a touch panel 1331 and other input devices 1332.

[0330] The display unit 1340 can be used to display input information or provided information and various menus of the terminal. The display unit 1340 can include a display panel 1341.

[0331] The audio circuit 1360, the speaker 1361, and the microphone 1362 can provide an audio interface.

[0332] In this embodiment, the processor 1380 included in the terminal 110 can perform the advertisement conversion rate prediction method of the foregoing embodiments.

[0333] The terminal 110 of the embodiments of the present disclosure includes but is not limited to a mobile phone, a computer, a smart voice interactive device, a smart home appliance, a vehicle-mounted terminal, an aircraft, etc.

[0334] Figure 12 A structural block diagram of a part of the server 140 for implementing the advertisement conversion rate prediction method of the embodiments of the present disclosure. The server 140 can have a relatively large difference due to different configurations or performances, and can include one or more central processing units (CPUs) 1422 (for example, one or more processors) and a memory 1432, one or more storage media 1430 (for example, one or more mass storage devices) storing application programs 1442 or data 1444. Among them, the memory 1432 and the storage medium 1430 can be temporary storage or persistent storage. The programs stored in the storage medium 1430 can include one or more modules (not shown in the figure), each of which can include a series of instruction operations in the server 140. Further, the central processing unit 1422 can be configured to communicate with the storage medium 1430 and execute a series of instruction operations in the storage medium 1430 on the server 140.

[0335] The server 140 can also include one or more power supplies 1426, one or more wired or wireless network interfaces 1450, one or more input / output interfaces 1458, and / or one or more operating systems 1441, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.

[0336] The central processing unit 1422 in the server 140 can be used to execute the advertisement conversion rate prediction method of the embodiments of the present disclosure.

[0337] The embodiments of the present disclosure also provide a storage medium storing a computer program, which is executed by a processor to perform the advertisement conversion rate prediction method of the foregoing embodiments.

[0338] The embodiments of the present disclosure further provide a computer program product comprising a computer program. A processor of a computer device reads the computer program and executes, so that the computer device performs the advertisement conversion rate prediction method.

[0339] The terms "first", "second", "third", "fourth" and the like in the description of the present disclosure and the above drawings, if any, are used to distinguish similar objects, and are not necessarily used to describe a particular order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "comprise" and "include" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a list of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0340] It should be understood that in the present disclosure, "at least one" means one or more, and "multiple" means two or more. "And / or" is used to describe the association relationship of the associated objects, which means that there can be three relationships, for example, "A and / or B" can represent three cases of only A, only B and A and B existing at the same time, where A and B can be singular or plural. The character " / " generally represents that the associated objects before and after are in an "or" relationship. "At least one of the following" or similar expressions means any combination of these items, including any combination of single or multiple items. For example, at least one of a, b or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0341] It should be understood that in the description of the embodiments of the present disclosure, the meaning of multiple (or multiple items) is more than two, greater than, less than, more than, etc. are not included in the number, above, below, etc. are understood to include the number.

[0342] In several embodiments provided by the present disclosure, it should be understood that the disclosed system, apparatus and method can be implemented in other manners. For example, the apparatus embodiments described above are merely schematic. For example, the division of the units is merely logical function division. There can be another division manner for the actual implementation, for example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections between different units, can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0343] The units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purposes of the embodiments.

[0344] In addition, each functional unit in the various embodiments of the present disclosure can be integrated in one processing unit, or each unit can exist physically as a separate unit, or two or more units can be integrated in one unit. The integrated unit can be implemented in the form of hardware, or in the form of a software functional unit.

[0345] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on such an understanding, the technical solutions of the present disclosure essentially, or the part that contributes to the prior art, or all or a part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods in the various embodiments of the present disclosure. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, and various media that can store program codes.

[0346] It should also be understood that the various embodiments provided by the present disclosure can be combined in any manner to achieve different technical effects.

[0347] The above is a specific description of the embodiments of the present disclosure, but the present disclosure is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or replacements without departing from the spirit of the present disclosure, and these equivalent modifications or replacements are included in the scope defined by the claims of the present disclosure.

Claims

1. A method for predicting advertising conversion rate, characterized in that: The method comprises: Acquire behavior sequence data of a target object, object attribute data of the target object, and target advertisement data, wherein the behavior sequence data includes a plurality of advertisement operation behavior data of the target object sorted in chronological order; Performing feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and performing feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence; Inputting the behavior feature sequence into at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features, wherein the at least two neural network models use different sequence feature extraction methods for sequence feature extraction, and the at least two neural network models are obtained by joint training; Performing feature fusion on the at least two sequence semantic features, and concatenating the fused feature obtained by fusion with the attribute feature to obtain a target feature; The advertisement conversion rate of the target object to the target advertisement is predicted based on the target feature.

2. The method according to claim 1, characterized in that The at least two neural network models include at least two first neural network models and a second neural network model, the first neural network model performs sequence feature extraction based on the behavior feature sequence, and the second neural network model performs sequence feature extraction based on the behavior feature sequence and the advertisement features of the target advertisement data, and the behavior feature sequence is input into the at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features, including: Inputting the behavioral feature sequence into the at least two first neural network models to extract sequence features to obtain at least two first subsequence semantic features; Performing feature embedding processing on the target advertisement data to obtain advertisement features, and inputting the advertisement features and the behavioral feature sequence into the second neural network model to perform sequence feature extraction to obtain second subsequence semantic features; The step of fusing the at least two sequence semantic features and concatenating the fused feature obtained by fusion with the attribute feature to obtain the target feature includes: The at least two first subsequence semantic features and the second subsequence semantic features are subjected to feature fusion, and the fused feature obtained by fusion is concatenated with the attribute feature to obtain a target feature.

3. The method according to claim 2, characterized in that The at least two first neural network models include a global perception model, a proximity focus model, a temporal evolution model, and a self-attention model. The behavior feature sequence is input into the at least two first neural network models to extract sequence features, thereby obtaining at least two first subsequence semantic features, including: performing nonlinear transformation processing on multiple behavior features in the behavior feature sequence based on the global perception model, and performing mean pooling on the processed multiple behavior features to obtain a first subsequence semantic feature; and Performing nonlinear transformation and summation processing on adjacent behavior features in the behavior feature sequence based on the proximity focus model to obtain a first subsequence semantic feature, wherein the adjacent behavior features are a preset number of behavior features whose corresponding behavior times are closest to the predicted time; and Based on the temporal evolution model, cyclically updating the multiple behavior features in the behavior feature sequence according to the time series to obtain a first subsequence semantic feature; and Based on the self-attention model, self-attention processing is performed on multiple behavior features in the behavior feature sequence to obtain a first subsequence semantic feature.

4. The method according to claim 1, wherein The step of fusing the at least two sequence semantic features and concatenating the fused feature obtained by fusion with the attribute feature to obtain the target feature includes: Obtaining weight coefficients corresponding to the at least two sequence semantic features; Performing weighted calculation on the at least two sequence semantic features based on the weight coefficient to obtain a fusion feature; The fusion feature and the attribute feature are combined to obtain the target feature.

5. The method according to claim 4, characterized in that The obtaining of weight coefficients corresponding to the at least two sequence semantic features includes: Obtaining a preset weight matrix, where the preset weight matrix is ​​obtained during the training process of the at least two neural network models; The weight coefficients corresponding to the at least two sequence semantic features are calculated based on the preset weight matrix and the at least two sequence semantic features.

6. The method according to claim 5, characterized in that The calculating, based on the weight matrix and the at least two sequence semantic features, weight coefficients corresponding to the at least two sequence semantic features includes: Performing dimension expansion and concatenation on the at least two sequence semantic features to obtain an expanded dimension feature; The weight coefficients corresponding to the at least two sequence semantic features are calculated based on the preset weight matrix and the dimension-expanding feature.

7. The method according to claim 6, characterized in that The calculating weight coefficients corresponding to the at least two sequence semantic features based on the preset weight matrix and the dimension-expanding feature includes: An attention parameter matrix is ​​calculated based on the preset weight matrix and the dimension-expanding feature; The weight coefficients corresponding to the at least two sequence semantic features are calculated based on the attention parameter matrix.

8. The method according to claim 4, characterized in that The obtaining of weight coefficients corresponding to the at least two sequence semantic features includes: Acquiring advertising environment data and advertising context data corresponding to the target advertising data; A weight coefficient corresponding to each sequence semantic feature is determined according to the advertisement environment data and the advertisement context data.

9. The method according to any one of claims 1 to 8, characterized in that The training process of the at least two neural network models comprises the following steps: Acquire training sample data, the training sample data including a plurality of sample data groups and conversion label data corresponding to each sample data group, the sample data group including sample object attribute data, sample behavior sequence data, and sample advertisement data of a sample object; Performing feature embedding on the sample object attribute data and the sample advertisement data to obtain sample attribute features, and performing feature embedding on the sample behavior sequence data to obtain a sample behavior feature sequence; Inputting the sample behavior feature sequences into the at least two neural network models respectively to extract sequence features, and performing feature fusion on the output semantic features of the at least two sample sequences to obtain sample fusion features; Splicing the sample attribute features with the sample fusion features, and inputting the spliced ​​features into a third neural network model to perform conversion rate prediction, thereby obtaining an output conversion rate prediction value; A loss value is calculated based on the conversion label data and the conversion rate prediction value, and model parameters of the at least two neural network models are updated based on the loss value.

10. The method according to claim 9, characterized in that The step of fusing the semantic features of at least two output sample sequences to obtain a sample fusion feature includes: Perform dimension expansion and concatenation on the semantic features of at least two sample sequences output to obtain a sample dimension expansion feature; Obtaining a weight matrix, wherein parameters in the weight matrix are adjustable; Calculating sample weight coefficients of the semantic features of the at least two sample sequences based on the weight matrix and the sample dimension expansion feature; fusing the semantic features of the at least two sample sequences according to the sample weight coefficient to obtain a sample fusion feature; The updating of model parameters of the at least two neural network models based on the loss value includes: Model parameters of the at least two neural network models and parameters in the weight matrix are updated based on the loss value.

11. The method according to claim 1, wherein The predicting, based on the target feature, the advertisement conversion rate of the target object to the target advertisement includes: Inputting the target features into the trained fully connected network model for feature mapping to obtain output features; Normalization is performed on the output features to obtain an advertisement conversion rate of the target object to the target advertisement.

12. An advertisement conversion rate prediction device, characterized in that: The device comprises: an acquiring unit, configured to acquire behavior sequence data of a target object, object attribute data of the target object, and target advertisement data, wherein the behavior sequence data includes a plurality of advertisement operation behavior data of the target object sorted in chronological order; an embedding unit, configured to perform feature embedding processing on the object attribute data and the target advertisement data to obtain attribute features, and to perform feature embedding processing on each advertisement operation behavior data in the behavior sequence data to obtain a behavior feature sequence; an extraction unit, configured to input the behavioral feature sequence into at least two neural network models for sequence feature extraction to obtain at least two sequence semantic features, wherein the at least two neural network models use different sequence feature extraction methods for sequence feature extraction, and the at least two neural network models are obtained by joint training; A fusion unit, configured to perform feature fusion on the at least two sequence semantic features, and concatenate the fused feature with the attribute feature to obtain a target feature; A prediction unit is used to predict the advertisement conversion rate of the target object to the target advertisement based on the target feature.

13. A storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the advertising conversion rate prediction method according to any one of claims 1 to 11 is implemented.

14. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the advertising conversion rate prediction method according to any one of claims 1 to 11 is implemented.

15. A computer program product, comprising a computer program, wherein the computer program is read and executed by a processor of a computer device, so that the computer device executes the advertising conversion rate prediction method according to any one of claims 1 to 11.

Images