A method, apparatus, device, and storage medium for recommending information.

Abstract

This application discloses an information recommendation method based on artificial intelligence and cloud technology. Applicable fields include, but are not limited to, maps, navigation, vehicle-to-everything (V2X) communication, vehicle-road collaboration, and instant messaging. Application scenarios include various terminals such as mobile phones, computers, and in-vehicle terminals. This application includes acquiring target object information and candidate information; obtaining an estimated click-through rate (CTR) through a click-through rate prediction model; obtaining shallow conversion rate estimates and shallow-to-deep conversion rate estimates for candidate information through a conversion rate prediction model; for deep target information, determining the target revenue value based on the estimated CTR, shallow conversion rate estimates, shallow-to-deep conversion rate estimates, shallow bid, and deep bid; and recommending candidate information to the target object if the information recommendation conditions are met. This application also provides related devices, equipment, and media. This application can recommend targeted information content to specific objects, thereby improving the recommendation effect.

Description

Technical Field

[0001] This application relates to the field of machine learning technology, and in particular to an information recommendation method, apparatus, device, and storage medium. Background Technology

[0002] Driven by the development of big data technology, information recommendation has been widely applied in various fields as a way to improve user experience. For example, when users browse web pages, they can be recommended game information that they might be interested in. Another example is when users are filtering products on e-commerce platforms, they can be recommended product information that they might be interested in.

[0003] Personalized recommendation algorithms can help users find content they are interested in, which is beneficial for improving user online time and retention rates. In existing technologies, information that meets the criteria can be filtered based on information provided by channel providers and targeting conditions set by news content creators. This information is then sorted, and the sorting results are used as the basis for news recommendations.

[0004] The inventors discovered that existing technologies suffer from at least the following problems: the targeting criteria used in sorting information are often specific to a particular group, such as men aged 25 to 30. Therefore, untargeted groups may not be able to access this information, resulting in poor recommendation performance. Summary of the Invention

[0005] This application provides an information recommendation method, apparatus, device, and storage medium. By determining the target revenue value of information based on estimated click-through rate and estimated conversion rate, it is possible to recommend targeted information content to specific users, thereby improving recommendation effectiveness.

[0006] In view of this, this application provides an information recommendation method, including:

[0007] Obtain target object information and candidate information of candidate information;

[0008] Based on target object information and candidate information, the predicted click-through rate of the target object for the candidate information is obtained through a click-through rate prediction model;

[0009] Based on target information and candidate information, a conversion rate prediction model is used to obtain shallow conversion rate predictions and shallow-to-deep conversion rate predictions for candidate information.

[0010] If the candidate information belongs to deep target information, the target revenue value corresponding to the candidate information is obtained by calling the comprehensive function to calculate the estimated click-through rate, shallow conversion rate, shallow to deep conversion rate, shallow bid and deep bid. The deep target information has a pre-set shallow bid and deep bid.

[0011] If the candidate information meets the information recommendation criteria based on the target return value, then the candidate information is recommended to the target audience.

[0012] This application also provides an information recommendation device, comprising:

[0013] The acquisition module is used to acquire target object information and candidate information of candidate information;

[0014] The acquisition module is also used to obtain the estimated click-through rate of the target object for the candidate information based on the target object information and the candidate information through the click-through rate prediction model;

[0015] The acquisition module is also used to obtain shallow conversion rate estimates and shallow-to-deep conversion rate estimates of candidate information based on target object information and candidate information information through a conversion rate prediction model.

[0016] The determination module is used to calculate the target revenue value corresponding to the candidate information by calling a comprehensive function if the candidate information belongs to deep target information. The calculation is based on the estimated click-through rate, the estimated shallow conversion rate, the estimated shallow to deep conversion rate, the shallow bid, and the deep bid. The deep target information has a pre-set shallow bid and a deep bid.

[0017] The recommendation module is used to recommend candidate information to the target audience if the candidate information meets the information recommendation criteria based on the target return value.

[0018] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0019] The determination module is specifically used to determine the target deep conversion rate of candidate information based on the ratio of deep bids to shallow bids;

[0020] The conversion rate threshold is determined by multiplying the target deep conversion rate by the dynamic pricing factor. The dynamic pricing factor is related to the proportion of deep target information in the total amount of information.

[0021] The deep conversion rate estimate of the candidate information is determined by multiplying the shallow conversion rate estimate by the shallow to deep conversion rate estimate.

[0022] If the estimated conversion rate from shallow to deep layers is greater than the conversion rate threshold, then the target revenue value corresponding to the candidate information is calculated based on the estimated click-through rate, the estimated shallow conversion rate, the shallow bid, the dynamic price adjustment factor, the estimated deep conversion rate, and the deep bid.

[0023] If the estimated conversion rate from shallow to deep layers is less than or equal to the conversion rate threshold, then the target revenue value corresponding to the candidate information is determined as the minimum revenue value.

[0024] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0025] The determination module is also used to determine the target revenue value corresponding to the candidate information based on the estimated click-through rate, the estimated shallow conversion rate, and the shallow bid if the candidate information belongs to shallow target information. The shallow target information has a pre-set shallow bid.

[0026] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0027] The acquisition module is specifically used to acquire the target object information corresponding to the target object, wherein the target object information includes at least one of the target object's basic attribute information, behavioral information, and interest information;

[0028] Obtain the candidate information corresponding to the candidate information, wherein the candidate information includes at least one of the following: information identifier, main information identifier, information type, text information, and image information.

[0029] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0030] The acquisition module is specifically used to acquire P sets of object embedding vectors based on the target object information, wherein at least one set of object embedding vectors in the P sets of object embedding vectors is determined based on the embedding query relationship, and P is an integer greater than or equal to 1;

[0031] Q sets of information embedding vectors are obtained based on candidate information information, wherein at least one set of information embedding vectors in the Q sets of information embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1;

[0032] Add the elements at the same position in the embedding vector of the P groups of objects to obtain the object feature vector of the target object;

[0033] Add the elements at the same position in the Q-group information embedding vector to obtain the information feature vector of the candidate information;

[0034] The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector.

[0035] Based on the target splicing vector, shallow conversion rate predictions and shallow-to-deep conversion rate predictions of candidate information are obtained through a conversion rate prediction model.

[0036] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0037] The acquisition module is specifically used to perform feature processing on the target object information to obtain the object feature vector of the target object;

[0038] The candidate information is processed to obtain the information feature vector of the candidate information;

[0039] The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector.

[0040] Based on the target splicing vector, shallow conversion rate predictions and shallow-to-deep conversion rate predictions of candidate information are obtained through a conversion rate prediction model.

[0041] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0042] The acquisition module is specifically used to acquire object sub-information of the target object from the target object information, wherein the feature types included in the object sub-information are less than or equal to the feature types included in the target object information;

[0043] Obtain information sub-information of candidate information from candidate information information, wherein the feature types included in the information sub-information are less than or equal to the feature types included in the candidate information information;

[0044] The object sub-information is characterized to obtain the object sub-feature vector of the target object;

[0045] The information sub-information is characterized to obtain the information sub-feature vector of the candidate information;

[0046] Based on object sub-feature vectors and information sub-feature vectors, the predicted click-through rate of the target object for candidate information is obtained through a click-through rate prediction model.

[0047] In one possible design, in another implementation of another aspect of the embodiments of this application, the information recommendation device further includes a processing module and a training module;

[0048] The acquisition module is also used to acquire M sets of training sample data, wherein each set of training sample data includes object sub-sample information of training objects, information sub-sample information of training information, and labeled click tags of training objects, where M is an integer greater than or equal to 1.

[0049] The processing module is used to perform feature processing on the object sub-sample information for each set of training sample data to obtain the object sample sub-feature vector of the training object.

[0050] The processing module is also used to perform feature processing on the information sub-sample information for each group of training sample data to obtain the information sample sub-feature vector of the training information.

[0051] The acquisition module is also used to obtain the estimated click-through rate of the training object for the training information based on the object sample sub-feature vector and the information sample sub-feature vector for each group of training sample data, through the click-through rate prediction model to be trained.

[0052] The training module is used to update the model parameters of the click-through rate prediction model to be trained based on the predicted click-through rate of the training subjects for the training information and the labeled click tags of the training subjects, for M sets of training sample data, until the model training conditions are met, and the click-through rate prediction model is obtained.

[0053] In one possible design, in another implementation of another aspect of the embodiments of this application, the information recommendation device further includes an update module;

[0054] The update module is used to update the information sorting list according to the target profit value, in descending order of profit value. The information sorting list includes the mapping relationship between the information and the sorting results.

[0055] The determination module is also used to determine if the sorting result corresponding to the candidate information is among the top K sorting results in the information sorting list, and if so, that the candidate information satisfies the information recommendation conditions, where K is an integer greater than or equal to 1.

[0056] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0057] The determination module is also used to determine the group benefit value based on the basic attribute information included in the target object information and the information type included in the candidate information information;

[0058] The determination module is also used to determine the target comprehensive return value based on the target return value and the group return value;

[0059] The update module is also used to update the information sorting list according to the target comprehensive return value, in descending order of comprehensive return value. The information sorting list includes the mapping relationship between information and sorting results.

[0060] The determination module is also used to determine if the sorting result corresponding to the candidate information is among the top K sorting results in the information sorting list, and if so, that the candidate information satisfies the information recommendation conditions, where K is an integer greater than or equal to 1.

[0061] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0062] The acquisition module is also used to acquire N sets of training sample data, wherein each set of training sample data includes object sample information of the training object, information sample information of the training information and the transformation type of the training information, and N is an integer greater than or equal to 1.

[0063] The acquisition module is also used to acquire P sets of object sample embedding vectors for each set of training sample data based on object sample information. Among them, at least one set of object sample embedding vectors in the P sets of object sample embedding vectors is determined based on the embedding query relationship. The embedding query relationship includes the mapping relationship between information type and embedding vector, and P is an integer greater than or equal to 1.

[0064] The acquisition module is also used to acquire Q sets of information sample embedding vectors for each set of training sample data based on information sample information, wherein at least one set of information sample embedding vectors in the Q sets of information sample embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1;

[0065] The processing module is also used to add the elements corresponding to the same position in the embedding vector of the P groups of object samples for each group of training sample data to obtain the object sample feature vector of the training object.

[0066] The processing module is also used to add the elements corresponding to the same position in the embedding vector of the Q groups of information samples for each group of training sample data to obtain the information sample feature vector of the training information.

[0067] The processing module is also used to concatenate the feature vectors of the object samples and the feature vectors of the information samples for each set of training sample data to obtain the concatenated sample vector.

[0068] The acquisition module is also used to obtain the shallow conversion rate prediction value of the training information, the deep conversion rate prediction value of the training information, and the shallow to deep conversion rate prediction value of the training information for each group of training sample data, based on the sample concatenation vector and through the conversion rate prediction model to be trained.

[0069] The training module is also used to update the model parameters of the conversion rate prediction model to be trained based on the conversion type of the training information, the shallow conversion rate prediction value of the sample, and the shallow to deep conversion rate prediction value of the sample for N sets of training sample data, until the model training conditions are met, and the conversion rate prediction model is obtained.

[0070] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0071] The training module is also used to update the embedding vector in the embedded query relationship based on the conversion type of the training information, the shallow conversion rate estimate of the sample, and the shallow to deep conversion rate estimate of the sample for N sets of training sample data.

[0072] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0073] The training module is specifically used to determine the first loss value for N sets of training sample data, based on the conversion type of the training information and the estimated shallow conversion rate of the samples.

[0074] For N sets of training sample data, the second loss value is determined based on the conversion type of training information, the shallow conversion rate prediction of the sample, and the shallow to deep conversion rate prediction of the sample.

[0075] The model parameters of the conversion rate prediction model to be trained are updated based on the first loss value and the second loss value.

[0076] In one possible design, in another implementation of another aspect of the embodiments of this application,

[0077] The acquisition module is also used to acquire N sets of training sample data, wherein each set of training sample data includes object sample information of the training object, information sample information of the training information and the transformation type of the training information, and N is an integer greater than or equal to 1.

[0078] The processing module is also used to perform feature processing on the object sample information for each set of training sample data to obtain the object sample feature vector;

[0079] The processing module is also used to perform feature processing on the information sample information for each group of training sample data to obtain the information sample feature vector;

[0080] The processing module is also used to concatenate the feature vectors of the object samples and the feature vectors of the information samples for each set of training sample data to obtain the concatenated sample vector.

[0081] The acquisition module is also used to obtain the shallow conversion rate prediction value of the training information, the deep conversion rate prediction value of the training information, and the shallow to deep conversion rate prediction value of the training information for each group of training sample data, based on the sample concatenation vector and through the conversion rate prediction model to be trained.

[0082] The training module is also used to update the model parameters of the conversion rate prediction model to be trained based on the conversion type of the training information, the shallow conversion rate prediction value of the sample, and the shallow to deep conversion rate prediction value of the sample, until the model training conditions are met and the conversion rate prediction model is obtained.

[0083] This application also provides a computer device, including: a memory, a processor, and a bus system;

[0084] The memory is used to store programs;

[0085] The processor is used to execute programs in memory, and the processor is used to execute the methods mentioned above according to the instructions in the program code;

[0086] Bus systems are used to connect memory and processor to enable communication between them.

[0087] Another aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the methods described above.

[0088] Another aspect of this application provides a computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the above aspects.

[0089] As can be seen from the above technical solutions, the embodiments of this application have the following advantages:

[0090] This application provides an information recommendation method. First, it obtains target object information and candidate information for candidate information. Based on this, it uses a click-through rate (CTR) prediction model to obtain the predicted CTR of the target object for the candidate information, and a conversion rate prediction model to obtain the shallow conversion rate estimate and the shallow-to-deep conversion rate estimate of the candidate information. If the candidate information belongs to the deep target information category, it determines the target revenue value corresponding to the candidate information. Further, if the candidate information meets the information recommendation conditions based on the target revenue value, it is recommended to the target object. Through the above method, combining object association features and information association features, it predicts the predicted CTR, shallow conversion rate estimate, and shallow-to-deep conversion rate estimate of the candidate information. Using these estimates, it determines the target revenue value of the information; the higher the target revenue value, the higher the value of recommending the information to the object. Therefore, it is possible to recommend targeted information content to specific objects, thereby improving the recommendation effect. Attached Figure Description

[0091] Figure 1 This is a schematic diagram of an environment for implementing real-time information recommendation in an embodiment of this application;

[0092] Figure 2This is a schematic diagram of an environment for implementing non-real-time information recommendation in an embodiment of this application;

[0093] Figure 3 This is a flowchart illustrating the information recommendation method in an embodiment of this application;

[0094] Figure 4 This is a schematic diagram of an interface displaying information in an embodiment of this application;

[0095] Figure 5 This is a schematic diagram of the click-through rate prediction model in an embodiment of this application;

[0096] Figure 6 This is a schematic diagram of a process for recommending information based on a target benefit value in an embodiment of this application.

[0097] Figure 7 This is a schematic diagram of a process for recommending information based on comprehensive benefit value in an embodiment of this application;

[0098] Figure 8 This is a schematic diagram of the conversion rate prediction model in the embodiments of this application;

[0099] Figure 9 This is another structural diagram of the conversion rate prediction model in the embodiments of this application;

[0100] Figure 10 This is a schematic diagram of an information recommendation device in an embodiment of this application;

[0101] Figure 11 This is a schematic diagram of the structure of a computer device in an embodiment of this application. Detailed Implementation

[0102] This application provides an information recommendation method, apparatus, device, and storage medium. By determining the target revenue value of information based on estimated click-through rate and estimated conversion rate, it is possible to recommend targeted information content to specific users, thereby improving recommendation effectiveness.

[0103] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0104] In the field of information recommendation (e.g., advertising recommendation), it is often necessary to rank information. Recommending based on ranking not only reduces the likelihood of recommending the same information but also improves the user experience of the information recommendation system. When ranking information, the interests of users, information platforms, and information providers must be considered, and a balance must be struck among these three parties. For information providers, the goal is to drive more conversions (e.g., superficial and deep conversions). For information platforms, the goal is to maximize the exposure of their information. For users, the goal is to see information that interests them.

[0105] In news recommendation systems, whenever a user request arrives, the system needs to retrieve suitable candidate news from the candidate news set and deliver it to the target user. Personalized news recommendation systems, on the other hand, rank news based on the user's click-through rate and estimated conversion rate for each candidate news item and automatically bid on it. For example, news about app downloads involves different stages of conversion behavior, such as "download," "install," "activate," "register," "pay," and "day-two retention." Especially when news products use deeper conversion goals (e.g., "pay" or "day-two retention") as the performance indicator for delivery, they are prone to encountering situations where deep conversion behaviors are highly sparse and have high latency, leading to significant deviations in the model's conversion rate prediction results. For example, for shallow conversion behaviors, the conversion rate is relatively higher than that of deep conversion behaviors, and the news source's conversion data can be returned on the same day. However, for deep conversion behaviors, the conversion rate is much lower, and the latest data from the news source is delayed for a long time, usually more than a week. Therefore, in scenarios with high sparsity and high latency, predicting the conversion rate of deep behaviors is a challenging but potentially valuable task.

[0106] Based on this, this application proposes an information recommendation method that can predict shallow conversion rate estimates and shallow-to-deep conversion rate estimates through a model. Therefore, the deep conversion rate estimate can be obtained by multiplying the shallow conversion rate estimate and the shallow-to-deep conversion rate estimate. This method has higher accuracy compared to directly predicting the deep conversion rate estimate.

[0107] The information recommendation method provided in this application is applied to an information recommendation system, which includes a server and terminal devices. The client is deployed on the terminal device, which can run in the form of a browser or a standalone application (APP). The specific form of the client is not limited here. The server involved in this application can be a standalone physical server, a server cluster composed of multiple physical servers, or a distributed system. It can also be a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. The terminal device can be a smartphone, tablet, laptop, PDA, personal computer, smart TV, smartwatch, in-vehicle device, wearable device, etc., but is not limited to these. The terminal device and the server can be directly or indirectly connected via wired or wireless communication, which is not limited here. The number of servers and terminal devices is also not limited. The solution provided in this application can be completed independently by the terminal device, independently by the server, or jointly by the terminal device and the server. This application does not make any specific limitations on this.

[0108] For example, a server can access big data, such as extracting user profile information from a user database or information related to news from an information database. Big data refers to data sets that cannot be captured, managed, and processed within a certain timeframe using conventional software tools. It represents massive, rapidly growing, and diverse information assets that require new processing models to achieve stronger decision-making, insight discovery, and process optimization capabilities. With the advent of the cloud era, big data has attracted increasing attention. Big data requires specialized technologies to effectively process large amounts of data within a tolerable timeframe. Technologies suitable for big data include massively parallel processing databases, data mining, distributed file systems, distributed databases, cloud computing platforms, the Internet, and scalable storage systems.

[0109] The following will combine Figure 1 and Figure 2 This section introduces the processes of real-time and non-real-time news recommendation.

[0110] I. Real-time information recommendations;

[0111] Specifically, please refer to Figure 1 , Figure 1 This is a schematic diagram illustrating an environment for real-time information recommendation in this embodiment of the application. As shown in the figure, in step A1, when a user logs into the client, a login request is initiated, carrying a user identifier. In step A2, the terminal device sends a login request to the server, which can be an information platform server or an application server, etc. In step A3, the server retrieves the user's profile information from the user database based on the user identifier carried in the login request, and can also retrieve the information information of each candidate information from the information database. Based on this, the information sorting list can be updated in real time for the user. In step A4, the server selects the top K information items from the updated information sorting list as recommended information. In step A5, the server pushes the selected recommended information to the terminal device. In step A6, the terminal device displays the recommended information.

[0112] II. Information is not recommended in real time;

[0113] Specifically, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating an environment for implementing non-real-time information recommendation in this embodiment of the application. As shown in the figure, in step B1, when a user logs into the client for the first time, an initial login request is initiated, which carries a user identifier. In step B2, the terminal device sends the initial login request to the server, which can be an information platform server or an application server, etc. In step B3, the server retrieves the user's profile information from the user database based on the user identifier carried in the initial login request, and can also retrieve the information information of each candidate information from the information database. Based on this, the information sorting list can be updated for the user. In step B4, when the user logs into the client again, a subsequent login request is initiated, which carries a user identifier. In step B5, the terminal device sends a subsequent login request to the server. In step B6, the server selects the top K information items from the updated information sorting list as recommended information. In step B7, the server pushes the selected recommended information to the terminal device. In step B8, the terminal device displays the recommended information.

[0114] The technical solutions provided in this application involve Artificial Intelligence (AI), specifically Machine Learning (ML). ML is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and many other disciplines. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. ML is the core of AI and the fundamental way to enable computers to possess intelligence; its applications span all areas of AI. ML and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0115] AI (Artificial Intelligence) is a theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess the functions of perception, reasoning, and decision-making.

[0116] AI technology is a comprehensive discipline encompassing a wide range of fields, including both hardware and software technologies. Fundamental AI technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing, operating / interactive systems, and mechatronics. AI software technologies primarily include computer vision, speech processing, natural language processing, as well as machine learning / deep learning, autonomous driving, and intelligent transportation.

[0117] To facilitate understanding, the terminology used in this application will be introduced before the embodiments are described.

[0118] 1. Information exposure (or advertising exposure): The products that users observe being exposed to.

[0119] 2. News clicks (or, ad clicks): User clicks on content displayed.

[0120] 3. Superficial conversion of information (or, superficial conversion of advertising): Superficial conversion behavior of items after users click, such as application download, application activation, and form registration.

[0121] 4. Deep conversion of information (or deep conversion of advertising): Deep conversion behavior of users after clicking on items, such as payment and application retention the next day.

[0122] 5. Click-through rate (CTR): The ratio of information exposure to clicks.

[0123] 6. Conversion rate (CVR): The ratio of the number of conversions to the number of times the message is reached. In this application, CVR represents the ratio from a click on the message to a shallow conversion.

[0124] 7. Deep conversion rate (DeepCVR): The proportion of information converted from shallow to deep, which is the deep conversion rate divided by the shallow conversion rate.

[0125] 8. Shallow-targeted content: This refers to content with only one bidding objective, typically a shallow objective such as registration or activation. Shallow-targeted content requires estimation of click-through rate and conversion rate (CVR).

[0126] 9. Deep Targeting Information: This includes both shallow and deep targeting bids. Deep targeting bidding is a bidding model that involves bidding again based on depth metrics, which can be paid or retention metrics. Deep targeting information requires estimating click-through rate (CTR), conversion rate (CVR), and deep conversion rate (DeepCVR).

[0127] Based on the above introduction, the information recommendation method in this application will be described below. Please refer to [link / reference]. Figure 3 One embodiment of the information recommendation method in this application includes:

[0128] 110. Obtain target object information and candidate information of candidate information;

[0129] In one or more embodiments, the information recommendation device can obtain target object information related to a target object, wherein the target object may refer to a user. For example, the target object launches an application (e.g., a browser application) and logs in. After successful login, the target object information can be extracted from big data based on the login information. It is understood that the login information includes, but is not limited to, user account, mobile phone number, and email address. The information recommendation device can also obtain candidate information information corresponding to multiple candidate information items. It is understood that the candidate information can specifically be candidate advertisements, and the candidate information information is candidate advertisement information. For ease of explanation, this application uses the extraction of candidate information information corresponding to one candidate information item as an example; other candidate information items are processed in a similar manner.

[0130] It should be noted that the information recommendation device provided in this application can be deployed on a server, on a terminal device, or in a system composed of a server and a terminal device; no limitation is made here.

[0131] 120. Based on target object information and candidate information, obtain the estimated click-through rate of the target object for the candidate information through a click-through rate prediction model;

[0132] In one or more embodiments, the information recommendation device selects some or all of the information from the target object information and the candidate information, performs feature processing on the information, and then inputs these features into the click-through rate prediction model to predict the estimated click-through rate (ectr) of the target object for the candidate information.

[0133] 130. Based on target object information and candidate information, obtain the shallow conversion rate prediction value and the shallow to deep conversion rate prediction value of candidate information through the conversion rate prediction model.

[0134] In one or more embodiments, the information recommendation device may use a conversion rate prediction model to predict the shallow conversion rate estimate (ecvr1) of candidate information and the shallow-to-deep conversion rate estimate (e_deep_cvr) of candidate information.

[0135] Specifically, in one implementation, the conversion rate prediction model may include an embedding layer. This embedding layer is used to extract features from the input information. For example, after inputting target object information and candidate information into the embedding layer, the embedding layer can output embedded representations of the target object information and candidate information. These embedded representations are then input into the fully connected layer of the conversion rate prediction model. In another implementation, the conversion rate prediction model does not include an embedding layer. In this case, feature engineering can be used to feature the target object information and candidate information. The features corresponding to the target object information and the candidate information are then input into the fully connected layer of the conversion rate prediction model.

[0136] Understandably, the conversion rate prediction model includes two fully connected (FC) layers. One FC layer is used to output a shallow conversion rate prediction value (ecvr1) for candidate information, and the other FC layer is used to output a shallow to deep conversion rate prediction value (e_deep_cvr) for candidate information.

[0137] 140. If the candidate information belongs to deep target information, the target revenue value corresponding to the candidate information is obtained by calling the comprehensive function to calculate the estimated click-through rate, shallow conversion rate, shallow to deep conversion rate, shallow bid and deep bid. The deep target information has a pre-set shallow bid and deep bid.

[0138] In one or more embodiments, the information recommendation device can obtain the conversion type of candidate information. If the conversion type of the candidate information is a multi-target information type, then the candidate information belongs to deep-target information. The so-called "multi-target conversion type" refers to an information type involving multiple bidding targets. These multiple bids include shallow bids and deep bids set by the information provider. A shallow bid represents the information provider's bid for shallow conversions, for example, a bid of 10 yuan per download. A deep bid represents the bid for deep conversions, for example, a bid of 100 yuan per payment.

[0139] Specifically, based on the estimated click-through rate (ectr), the estimated shallow conversion rate (ecvr1), the estimated shallow-to-deep conversion rate (e_deep_cvr), the shallow bid (bid1), and the deep bid (bid2), the target revenue value corresponding to the candidate news can be calculated. In essence, the target revenue value can be expressed as the revenue generated per thousand impressions (effective cost per mille, ecpm). The higher the target revenue value, the more revenue the platform earns per impression of the news placement.

[0140] 150. If the candidate information meets the information recommendation criteria based on the target return value, then the candidate information is recommended to the target audience.

[0141] In one or more embodiments, the information recommendation device combines the target benefit value of candidate information to determine whether the candidate information meets the information recommendation criteria. If it does, the candidate information is recommended to the target object. If it does not meet the criteria, the candidate information does not need to be recommended to the target object. For example, if the target benefit value is greater than a benefit value threshold, the candidate information is considered to meet the information recommendation criteria. For example, if the target benefit value is greater than the benefit value of other information, the candidate information is considered to meet the information recommendation criteria. It is understood that the information recommendation criteria can also be flexibly set according to actual needs; an exhaustive list is not provided here.

[0142] For easier understanding, please refer to Figure 4 , Figure 4This is a schematic diagram of an interface displaying information in an embodiment of this application. As shown in the figure, taking a browser page as an example, assume the candidate information is about game application A, where the candidate information meets the information recommendation criteria. In one case, the target user enters the keyword "game A" in the "search bar," and based on this, information related to game application A can be displayed in the "information bar." In another case, the target user enters the keyword "rice cooker" in the "search bar," and based on this, information related to game application A can be displayed in the "business information bar."

[0143] This application provides an information recommendation method. By combining object association features and information association features, the method predicts the estimated click-through rate, shallow conversion rate, and shallow-to-deep conversion rate of candidate information. These estimates are then used to determine the target revenue value of the information; a higher target revenue value indicates a higher value for recommending the information to the target audience. Therefore, it is possible to recommend targeted information content to specific audiences, thereby improving the recommendation effect.

[0144] Optionally, in the above Figure 3 Based on the corresponding embodiments, in another optional embodiment provided by this application, the target revenue value corresponding to the candidate information is obtained by calling a comprehensive function to calculate the estimated click-through rate, the shallow conversion rate estimate, the shallow to deep conversion rate estimate, the shallow bid, and the deep bid. Specifically, this includes:

[0145] The target deep conversion rate for candidate information is determined based on the ratio of deep bids to shallow bids.

[0146] The conversion rate threshold is determined by multiplying the target deep conversion rate by the dynamic pricing factor. The dynamic pricing factor is related to the proportion of deep target information in the total amount of information.

[0147] The deep conversion rate estimate of the candidate information is determined by multiplying the shallow conversion rate estimate by the shallow to deep conversion rate estimate.

[0148] If the estimated conversion rate from shallow to deep layers is greater than the conversion rate threshold, then the target revenue value corresponding to the candidate information is calculated based on the estimated click-through rate, the estimated shallow conversion rate, the shallow bid, the dynamic price adjustment factor, the estimated deep conversion rate, and the deep bid.

[0149] If the estimated conversion rate from shallow to deep layers is less than or equal to the conversion rate threshold, then the target revenue value corresponding to the candidate information is determined as the minimum revenue value.

[0150] In one or more embodiments, a method for calculating target revenue values ​​for deep target information is described. As can be seen from the foregoing embodiments, the conversion rate prediction model can output a shallow conversion rate prediction value (ecvr1) and a shallow-to-deep conversion rate prediction value (e_deep_cvr) for candidate information. Therefore, the deep conversion rate prediction value (ecvr2) for candidate information can be determined by multiplying the shallow conversion rate prediction value (ecvr1) and the shallow-to-deep conversion rate prediction value (e_deep_cvr), i.e., ecvr2 = ecvr1 * e_deep_cvr.

[0151] Specifically, a dynamic pricing factor (α) can be set, with a value greater than 0 and less than or equal to 1. The dynamic pricing factor (α) is correlated with the proportion of deep-level target information in the total number of information items. That is, it determines whether to participate in deep bidding based on the estimated deep conversion rate, reducing the number of information items participating in deep bidding. If the number of information items participating in deep bidding is small, the dynamic pricing factor (α) is lowered; conversely, if the number of information items participating in deep bidding is large, the dynamic pricing factor (α) is increased. Furthermore, the target deep conversion rate (t_deep_cvr) of the candidate information can be determined based on the ratio of the deep bid (bid2) to the shallow bid (bid1), i.e., t_deep_cvr = bid2 / bid1. Therefore, the conversion rate threshold is determined by multiplying the target deep conversion rate (t_deep_cvr) by the dynamic pricing factor (α), i.e., α * t_deep_cvr.

[0152] It should be noted that the comprehensive function mainly consists of four parts: the first part is used to calculate the target deep conversion rate (t_deep_cvr), the second part is used to calculate the conversion rate threshold, the third part is used to calculate the deep conversion rate estimate (ecvr2), and the fourth part is used to calculate the target revenue value (ecpm).

[0153] For example, in one case, if the estimated conversion rate from shallow to deep layers (e_deep_cvr) is greater than the conversion rate threshold (α*t_deep_cvr), the target revenue value corresponding to the candidate information is calculated as follows:

[0154] ecpm=ectr*ecvr1*bid1*α+ectr*ecvr2*bid2;

[0155] Where ecpm represents the target revenue value, ectr represents the estimated click-through rate, ecvr1 represents the shallow conversion rate estimate, bid1 represents the shallow bid, α represents the dynamic price adjustment factor, ecvr2 represents the deep conversion rate estimate, and bid2 represents the deep bid.

[0156] For example, in another scenario, if the estimated conversion rate from shallow to deep layers (e_deep_cvr) is less than or equal to the conversion rate threshold (α*t_deep_cvr), then candidate information is directly filtered out. This is because if the estimated deep conversion rate is very low, there is no need to calculate the target revenue value and participate in the ranking. Therefore, the target revenue value corresponding to the candidate information can be determined as the minimum revenue value, where the minimum revenue value can be set to 0, or other negative numbers, etc., without limitation here.

[0157] Secondly, in this embodiment of the application, a method for calculating the target revenue value for deep target information is provided. In this method, if the estimated conversion rate from shallow to deep is less than or equal to the conversion rate threshold, it is considered that the deep conversion effect is not good. Therefore, the target revenue value is directly set to the minimum revenue value, that is, it does not need to participate in the information sorting (for example, it is placed directly at the end of the sorting), thereby reducing the amount of information to be calculated and saving computing resources.

[0158] Optionally, in the above Figure 3 Based on the corresponding embodiments, another optional embodiment provided in this application may further include:

[0159] If the candidate information belongs to shallow target information, the target revenue value corresponding to the candidate information is determined based on the estimated click-through rate, the estimated shallow conversion rate, and the shallow bid. The shallow target information has a pre-set shallow bid.

[0160] In one or more embodiments, a method for calculating target revenue values ​​for shallow target information is described. As can be seen from the foregoing embodiments, the conversion rate prediction model can output a shallow conversion rate prediction value (ecvr1) for candidate information.

[0161] Specifically, for shallow target information, the target return value corresponding to the candidate information can be calculated in the following way:

[0162] ecpm = ectr * ecvr1 * bid1;

[0163] Where ecpm represents the target revenue value, ectr represents the estimated click-through rate, ecvr1 represents the shallow conversion rate estimate, and bid1 represents the shallow bid.

[0164] Secondly, this application provides a method for calculating target revenue value for shallow target information. Through the above method, the target revenue value can also be calculated for shallow target information, thereby increasing the feasibility and operability of the solution.

[0165] Optionally, in the above Figure 3Based on the corresponding embodiments, in another optional embodiment provided by this application, obtaining the target object information of the target object and the candidate information of the candidate information specifically includes:

[0166] Obtain target object information corresponding to the target object, wherein the target object information includes at least one of the target object's basic attribute information, behavioral information, and interest information;

[0167] Obtain the candidate information corresponding to the candidate information, wherein the candidate information includes at least one of the following: information identifier, main information identifier, information type, text information, and image information.

[0168] In one or more embodiments, a method for obtaining target object information and candidate information is described. Target object information includes at least one of the target object's basic attribute information, behavioral information, and interest information. Candidate information includes at least one of the candidate information's information identifier, main information identifier, information type, text information, and image information. These information will be described below with examples.

[0169] I. Target object information;

[0170] (1) Basic attribute information: usually includes the user's age, gender, occupation and location.

[0171] (2) Behavioral information: This usually includes the user's frequency of use, usage time, usage duration, and registration time.

[0172] (3) Interest information: This usually includes the user's favorite sports, places they often visit, and items they often buy.

[0173] II. Candidate Information;

[0174] (1) Information Identifier: A unique identifier assigned to each piece of information. For example, the information identifier of a candidate piece of information is “10011”.

[0175] (2) Information owner identifier: A unique identifier assigned to each information owner. For example, if the information owner of a candidate information is "XXX Company", the information owner identifier is "1001".

[0176] (3) Information type: According to different information content, it can be divided into different information types, such as food information, game information and clothing information.

[0177] (4) Text information: Extract keywords as text information from news titles or text, or extract text information from news images or videos based on Optical Character Recognition (OCR) technology.

[0178] (5) Image information: Target images can be identified from news pictures or news videos based on image recognition technology, such as identifying "kitten" or "television".

[0179] It should be noted that in practical applications, target object information and candidate information include, but are not limited to, the examples above, and will not be exhaustively listed here.

[0180] Secondly, this application provides a method for obtaining target object information and candidate information. By using the above method, information on the target object and candidate information is extracted based on different dimensions, enriching the types of information and improving the accuracy of the model output results.

[0181] Optionally, in the above Figure 3 Based on the corresponding embodiments, in another optional embodiment provided by this application, based on target object information and candidate information, a shallow conversion rate prediction value and a shallow-to-deep conversion rate prediction value of candidate information are obtained through a conversion rate prediction model, specifically including:

[0182] Based on the target object information, obtain P sets of object embedding vectors, wherein at least one set of object embedding vectors in the P sets of object embedding vectors is determined based on the embedding query relationship, and P is an integer greater than or equal to 1;

[0183] Q sets of information embedding vectors are obtained based on candidate information information, wherein at least one set of information embedding vectors in the Q sets of information embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1;

[0184] Add the elements at the same position in the embedding vector of the P groups of objects to obtain the object feature vector of the target object;

[0185] Add the elements at the same position in the Q-group information embedding vector to obtain the information feature vector of the candidate information;

[0186] The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector.

[0187] Based on the target splicing vector, shallow conversion rate predictions and shallow-to-deep conversion rate predictions of candidate information are obtained through a conversion rate prediction model.

[0188] In one or more embodiments, a method for extracting object association features and information association features is described. As can be seen from the foregoing embodiments, feature extraction can be performed on target object information and candidate information. The feature extraction process requires the use of an embedding query relationship, which can be represented as an embedding feature table.

[0189] Specifically, taking the basic attribute information included in the target object information as an example, the basic attribute information may include the target object's age. For ease of understanding, please refer to Table 1, which is a schematic diagram of the embedded query relationship corresponding to the user's age.

[0190] Table 1

[0191] age Feature identifier Embedded vector 18 0 [0.1,0.3,0.7,0.4,0.5,0.1] 19 1 [0.3,0.2,0.9,0.1,0.2,0.4] 20 2 [0.1,0.5,0.1,0.8,0.4,0.5] 21 3 [0.1,0.2,0.8,0.7,0.2,0.3] 22 4 [0.7,0.3,0.2,0.5,0.9,0.4] 23 5 [0.6,0.1,0.2,0.9,0.3,0.4] 24 6 [0.2,0.1,0.6,0.6,0.1,0.4]

[0192] It is understood that the embedding vector shown in Table 1 has a dimension of 6. In practical applications, the dimension of the embedding vector can be set to 64 or 128, etc. This is only an illustration and should not be construed as a limitation of this application. The feature identifier is obtained by converting age using a hash function. The feature identifier is used as the key, and the corresponding embedding vector is stored as the value in the embedding query relation. Similarly, other information in the target object information can also have its corresponding embedding vector determined based on the corresponding embedding query relation, and this embedding vector is used as the object embedding vector.

[0193] Taking the information identifiers included in the candidate information as an example, please refer to Table 2 for easier understanding. Table 2 is a schematic diagram of the embedded query relationship corresponding to the information identifiers.

[0194] Table 2

[0195] Information label Feature identifier Embedded vector 15000 0 [0.2,0.1,0.2,0.5,0.9,0.1] 15001 1 [0.8,0.1,0.3,0.9,0.1,0.5] 15002 2 [0.3,0.2,0.2,0.8,0.5,0.4] 15003 3 [0.2,0.3,0.5,0.8,0.1,0.2] 15004 4 [0.3,0.4,0.5,0.1,0.2,0.2] 15005 5 [0.9,0.4,0.5,0.2,0.1,0.1] 15006 6 [0.7,0.7,0.6,0.4,0.4,0.5]

[0196] It is understood that the embedding vector shown in Table 2 has a dimension of 6. In practical applications, the dimension of the embedding vector can be set to 64 or 128, etc. This is only an illustration and should not be construed as a limitation of this application. The feature identifier is obtained by converting the information identifier using a hash function. The feature identifier is used as the key, and the corresponding embedding vector is stored as the value in the embedding query relationship. Similarly, the main information identifier and information type in the candidate information can also be determined based on the corresponding embedding query relationship, and the embedding vector is used as the information embedding vector. The text information and image information included in the candidate information can be extracted using a text encoder or an image encoder to extract their corresponding information embedding vectors.

[0197] Based on this, element-wise addition is performed on the P groups of object embedding vectors, that is, elements at the same position are added together. Assume the P groups of object embedding vectors include two sets of object embedding vectors, [0.1,0.5,0.1,0.8,0.4,0.5] and [0.5,0.3,0.9,0.1,0.1,0.2]. After element-wise addition, the object feature vector of the target object is obtained as [0.6,0.8,1,0.9,0.5,0.7]. Similarly, element-wise addition is performed on the Q groups of information embedding vectors, that is, elements at the same position are added together, to obtain the information feature vector of the candidate information. Next, the object feature vector and the information feature vector are concatenated to obtain the target concatenated vector. Assuming the object feature vector is n-dimensional and the information feature vector is n-dimensional, then the target concatenated vector is 2n-dimensional. Therefore, the target concatenation vector is input into the conversion rate prediction model, and the shallow conversion rate prediction value (ecvr1) and the shallow to deep conversion rate prediction value (e_deep_cvr) are output through the conversion rate prediction model.

[0198] Secondly, in this embodiment of the application, a method for extracting object association features and information association features is provided. In this way, considering that some information (e.g., information identifiers) has a large amount of data, an embedding vector with fewer feature dimensions can be obtained based on the trained embedding query relationship. Thus, on the one hand, feature dimensionality reduction can speed up the calculation, and on the other hand, it can avoid the situation of model overfitting due to too many features.

[0199] Optionally, in the above Figure 3 Based on the corresponding embodiments, in another optional embodiment provided by this application, based on target object information and candidate information, a shallow conversion rate prediction value and a shallow-to-deep conversion rate prediction value of candidate information are obtained through a conversion rate prediction model, specifically including:

[0200] The target object information is characterized to obtain the object feature vector of the target object;

[0201] The candidate information is processed to obtain the information feature vector of the candidate information;

[0202] The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector.

[0203] Based on the target splicing vector, shallow conversion rate predictions and shallow-to-deep conversion rate predictions of candidate information are obtained through a conversion rate prediction model.

[0204] In one or more embodiments, a method for extracting object association features and information association features is described. As can be seen from the foregoing embodiments, feature extraction can be performed on target object information and candidate information information, and feature engineering is required in the feature extraction process.

[0205] Specifically, taking the basic attribute information included in the target object information as an example, the basic attribute information may include the region of the target object. For example, the object embedding vector corresponding to "Shenzhen" is [0,0,0,0,0,0,0,1,0,0,0,0]. The basic attribute information may include the gender of the target object. For example, the object embedding vector corresponding to "female" is [0]. Taking the information type included in the candidate information as an example, for example, the object embedding vector corresponding to "game information" is [0,0,0,0,1,0,0,0,0].

[0206] Based on this, the features obtained after feature processing of the target object information are concatenated to obtain the object feature vector. Similarly, the features obtained after feature processing of the candidate information are concatenated to obtain the information feature vector. Next, the object feature vector and the information feature vector are concatenated to obtain the target concatenated vector. Assuming that the object feature vector is n-dimensional and the information feature vector is n-dimensional, then the target concatenated vector is 2n-dimensional. Therefore, the target concatenated vector is input into the conversion rate prediction model, which outputs a shallow conversion rate prediction value (ecvr1) and a shallow-to-deep conversion rate prediction value (e_deep_cvr).

[0207] Secondly, this application provides a method for extracting object association features and information association features. Through the above method, feature processing of target object information and candidate information can be performed directly based on feature engineering, without the need to train other feature extraction networks, thereby saving training resources.

[0208] Optionally, in the above Figure 3 Based on the corresponding embodiments, in another optional embodiment provided by this application, the predicted click-through rate of the target object for the candidate information is obtained through a click-through rate prediction model based on the target object information and candidate information, specifically including:

[0209] Obtain object sub-information of the target object from the target object information, wherein the feature types included in the object sub-information are less than or equal to the feature types included in the target object information;

[0210] Obtain information sub-information of candidate information from candidate information information, wherein the feature types included in the information sub-information are less than or equal to the feature types included in the candidate information information;

[0211] The object sub-information is characterized to obtain the object sub-feature vector of the target object;

[0212] The information sub-information is characterized to obtain the information sub-feature vector of the candidate information;

[0213] Based on object sub-feature vectors and information sub-feature vectors, the predicted click-through rate of the target object for candidate information is obtained through a click-through rate prediction model.

[0214] In one or more embodiments, a method for predicting and estimating click-through rates (CTR) is described. As can be seen from the foregoing embodiments, the CTR prediction model can output the predicted CTR (ectr). It should be noted that the CTR prediction model in this application includes, but is not limited to, Logistic Regression (LR) models, Gradient Boosting Decision Tree (GBDT) models, Factorization Machines (FM), Wide & Deep models, and Deep Interest Networks (DIN).

[0215] Specifically, and exemplarily, partial information can be extracted from the target object information as object sub-information. For example, only basic attribute information can be extracted, or the target object information can be directly used as object sub-information. Similarly, partial information can be extracted from candidate information information as information sub-information. For example, only information type can be extracted, or the candidate information information can be directly used as information sub-information. Based on this, feature processing can be performed on the object sub-information and the information sub-information of the candidate information, respectively, to obtain the object sub-feature vector of the target object and the information sub-feature vector of the candidate information.

[0216] For easier understanding, please refer to Figure 5 , Figure 5 This is a schematic diagram of a click-through rate (CTR) prediction model in an embodiment of this application. As shown in the figure, the object sub-feature vector and the information sub-feature vector are combined. For example, the FM and GBDT included in the CTR prediction model can be used to combine the features to obtain higher-order features. Then, the LR model included in the CTR prediction model is used to process the higher-order features and output the predicted click-through rate (ectr).

[0217] Secondly, this application provides a method for predicting and estimating click-through rates. Through this method, the predicted click-through rate can be directly output based on the trained click-through rate prediction model, thereby improving the feasibility and operability of the solution.

[0218] Optionally, in the above Figure 3Based on the corresponding embodiments, another optional embodiment provided in this application may further include:

[0219] Obtain M sets of training sample data, where each set of training sample data includes object sub-sample information of the training object, information sub-sample information of the training information, and labeled click tags of the training object, where M is an integer greater than or equal to 1;

[0220] For each set of training sample data, the object sub-sample information is characterized to obtain the object sample sub-feature vector of the training object;

[0221] For each set of training sample data, the information sub-sample information is characterized to obtain the information sample sub-feature vector of the training information;

[0222] For each set of training sample data, based on the object sample sub-feature vector and the information sample sub-feature vector, the predicted click rate of the training object for the training information is obtained through the click rate prediction model to be trained;

[0223] For M sets of training sample data, the model parameters of the click-through rate prediction model to be trained are updated based on the predicted click-through rate of the training subjects for the training information and the labeled click tags of the training subjects, until the model training conditions are met, and the click-through rate prediction model is obtained.

[0224] In one or more embodiments, a method for training a click-through rate (CTR) prediction model is described. As can be seen from the foregoing embodiments, the CTR prediction model needs to be trained, and thus, the CTR prediction model can output a predicted click-through rate (ectr).

[0225] Specifically, M sets of training sample data are used as a batch of training data, for example, M is 1000. Each set of training sample data includes object sub-sample information of the training object (e.g., age, gender, and interests), information sub-sample information of the training information (e.g., information type, information identifier, and main information identifier), and annotation click labels of the training object. When the training object clicks on the training information, the annotation click label can be represented as "1", and when the training object does not click on the training information, the annotation click label can be represented as "0".

[0226] Based on this, the object sub-sample information in each group of training sample data is characterized to obtain the object sample sub-feature vector of the training object. Similarly, the information sub-sample information in each group of training sample data is characterized to obtain the information sample sub-feature vector of the training information. After inputting the object sample sub-feature vector and the information sample sub-feature vector into the click-through rate (CTR) prediction model to be trained, the predicted CTR of the training object for the training information can be obtained. Using the cross-entropy loss function, with the predicted CTR as the predicted value and the labeled click value as the true value, a loss value can be calculated. Based on this loss value, the backpropagation gradient algorithm is used to update the model parameters of the CTR prediction model to be trained until the model training conditions are met, thus obtaining the CTR prediction model.

[0227] In one scenario, the model training condition can be considered satisfied when the number of model iterations is reached. In another scenario, the model training condition can be considered satisfied when the loss value converges.

[0228] Secondly, in this embodiment of the application, a method for training a click-through rate (CTR) prediction model is provided. Through the above method, a CTR prediction model for outputting the predicted CTR can be trained, thereby improving the feasibility and operability of the solution.

[0229] Optionally, in the above Figure 3 Based on the corresponding embodiments, another optional embodiment provided in this application may further include:

[0230] The information sorting list is updated according to the target profit value, in descending order of profit value. The information sorting list includes the mapping relationship between the information and the sorting results.

[0231] If the sorting result corresponding to the candidate information is among the top K sorting results in the information sorting list, then the candidate information is determined to meet the information recommendation conditions, where K is an integer greater than or equal to 1.

[0232] In one or more embodiments, a method for sorting information based on reward value is described. As can be seen from the foregoing embodiments, to achieve real-time recommendation, the information sorting list also needs to be updated. For example, the information can be sorted in descending order of reward value. If candidate information already exists in the information sorting list, its position in the list is directly adjusted according to the target reward value. If candidate information does not exist in the information sorting list, it can be added to the list based on the target reward value.

[0233] Specifically, for ease of understanding, please refer to Table 3, which is a schematic of the information sorting list before the update.

[0234] Table 3

[0235] Sorting results Information label Profitability 1 11023 15.7 2 32015 14.2 3 16044 12.0 4 98521 11.8 5 11258 8.7 6 33560 0 7 00461 0

[0236] The information ranking list shown in Table 3 is sorted in descending order of return value; however, this should not be construed as a limitation of this application. Assuming the target return value obtained after calculating the candidate information is "16.5", the information ranking list needs to be updated according to this target return value. For ease of understanding, please refer to Table 4, which is an illustration of the updated information ranking list.

[0237] Table 4

[0238] Sorting results Information label Profitability 1 66235 16.5 2 11023 15.7 3 32015 14.2 4 16044 12.0 5 98521 11.8 6 11258 8.7 7 33560 0 8 00461 0

[0239] As shown in Table 4, the ranking result corresponding to the candidate information is "1", meaning it is ranked first in the information ranking list. Assuming the preset K is 1, then the candidate information meets the criteria for being recommended to the target user, thus confirming that the candidate information meets the information recommendation conditions. It is understood that in practical applications, the value of K can be set according to the situation, and no limitation is made here.

[0240] Based on the above introduction, the following will combine... Figure 6 This section describes the process of recommending information based on target return values. Please refer to [link / reference]. Figure 6 , Figure 6 This is a schematic diagram of a process for recommending information based on a target benefit value in an embodiment of this application, as shown in the figure. Specifically:

[0241] In step C1, the target object initiates a login request to the server through a client running on the terminal device;

[0242] In step C2, the server retrieves and recalls some information from the information database. For example, it retrieves information that was launched in the past month as candidate information. If 20 pieces of information are recalled, then all of these pieces of information are candidate information.

[0243] In step C3, taking a candidate piece of information as an example, the candidate information information of the candidate information is extracted, and the target object information of the target object is extracted. Based on the candidate information information and the target object information, the estimated click-through rate of the candidate information is predicted by the click-through rate prediction model.

[0244] In step C4, based on candidate information and target object information, the shallow conversion rate estimate and the shallow to deep conversion rate estimate of candidate information are predicted by the conversion rate prediction model.

[0245] In step C5, for deep target information, a dynamic pricing factor is set, and the target revenue value of the candidate information is calculated based on indicators such as the estimated click-through rate, the estimated shallow conversion rate, and the estimated shallow-to-deep conversion rate of the candidate information.

[0246] In step C6, for shallow target information, the target revenue value of the candidate information is calculated based on indicators such as the estimated click-through rate and the estimated shallow conversion rate.

[0247] In step C7, the information in the information sorting list is reordered based on the target profit value;

[0248] In step C8, the top K pieces of information from the information sorting list are pushed to the client used by the target object.

[0249] Secondly, in this embodiment of the application, a method for sorting information based on benefit value is provided. By using the benefit value as the basis for sorting information, information with higher benefit value can be recommended first, thereby helping to improve the rationality of information recommendation and having a better information feedback effect.

[0250] Optionally, in the above Figure 3 Based on the corresponding embodiments, another optional embodiment provided in this application may further include:

[0251] The group benefit value is determined based on the basic attribute information included in the target object information and the information types included in the candidate information information.

[0252] Determine the target overall return value based on the target return value and the group return value;

[0253] The information sorting list is updated according to the target comprehensive return value, in descending order of comprehensive return value. The information sorting list includes the mapping relationship between information and sorting results.

[0254] If the sorting result corresponding to the candidate information is among the top K sorting results in the information sorting list, then the candidate information is determined to meet the information recommendation conditions, where K is an integer greater than or equal to 1.

[0255] In one or more embodiments, a method for sorting information based on return value and other factors is described. As can be seen from the foregoing embodiments, to achieve real-time recommendation, the information sorting list also needs to be updated. For example, the information can be sorted in descending order of comprehensive return value. If candidate information already exists in the information sorting list, its position in the list is directly adjusted based on the target comprehensive return value. If candidate information does not exist in the information sorting list, it can be added to the list based on the target comprehensive return value.

[0256] Specifically, the comprehensive return value can be calculated as follows:

[0257] Rank_Score = a*epcm + b*Q;

[0258] Wherein, Rank_Score represents the overall return value, a represents the first weight coefficient, b represents the second weight coefficient, epcm represents the target return value, and Q represents the relevant return value (e.g., the group return value).

[0259] Relevant revenue values ​​are typically related to ranking strategies such as diversity, frequency control, category control, and result-specific weighting. Diversity refers to the wide variety of categories of recommended information. Frequency control refers to controlling the exposure frequency of the same information on the same target. Category control refers to controlling the exposure frequency of information categories. Result-specific weighting involves artificially increasing the weight of specific information.

[0260] For example, taking the "specific result weighting" ranking strategy, assuming the target audience's basic attribute information is "youth," "female," and "Shanghai," if the candidate information's information type is "cosmetics information," then the group benefit value can be set to 10. If the candidate information's information type is "TV series information," then the group benefit value can be set to 3. It is evident that the group benefit value is related to the information's target audience. Therefore, multiplying the target benefit value by the first weighting coefficient and the group benefit value by the second weighting coefficient, and then summing the products, yields the comprehensive benefit value of the candidate information. For easier understanding, please refer to Table 5, which is a schematic of the information ranking list before the update.

[0261] Table 5

[0262] Sorting results Information label Profitability 1 11023 15.7 2 32015 14.2 3 16044 12.0 4 00461 0

[0263] The information ranking list shown in Table 5 is sorted in descending order of comprehensive return value; however, this should not be construed as a limitation of this application. Assuming the comprehensive return value of the candidate information is calculated to be "20.8", the information ranking list needs to be updated based on this comprehensive return value. For ease of understanding, please refer to Table 6, which is an illustration of the updated information ranking list.

[0264] Table 6

[0265] Sorting results Information label Profitability 1 66235 20.8 2 11023 15.7 3 32015 14.2 4 16044 12.0 5 00461 0

[0266] As shown in Table 6, the ranking result corresponding to the candidate information is "1", meaning it is ranked first in the information ranking list. Assuming the preset K is 1, then the candidate information meets the criteria for being recommended to the target user, thus confirming that the candidate information meets the information recommendation conditions. It is understood that in practical applications, the value of K can be set according to the situation, and no limitation is made here.

[0267] Based on the above introduction, the following will combine... Figure 7 This section describes the process of recommending information based on comprehensive return values. Please refer to [link / reference]. Figure 7 , Figure 7 The figure shows a flowchart of recommending information based on comprehensive benefit value in an embodiment of this application. It should be noted that steps D1 to D6 are similar to steps C1 to C6, so they will not be described in detail here.

[0268] In step D7, the information in the information ranking list is reordered based on the target benefit value and the ranking strategy (e.g., measurements related to diversity, frequency control, category control, and specific outcome weighting).

[0269] In step D8, the top K pieces of information from the information sorting list are pushed to the client used by the target object.

[0270] Secondly, in this embodiment of the application, a method for sorting information based on benefit value and other factors is provided. By using the above method, the benefit value and other factors are combined as the basis for sorting information. This not only prioritizes the recommendation of information with higher benefit value and achieves better information feedback, but also achieves the purpose of more comprehensively evaluating the value of information, thereby helping to improve the rationality of information recommendation.

[0271] Optionally, in the above Figure 3 Based on the corresponding embodiments, another optional embodiment provided in this application may further include:

[0272] Obtain N sets of training sample data, where each set of training sample data includes object sample information of the training object, information sample information of the training information, and transformation type of the training information, where N is an integer greater than or equal to 1.

[0273] For each set of training sample data, P sets of object sample embedding vectors are obtained based on the object sample information. Among them, at least one set of object sample embedding vectors is determined based on the embedding query relationship. The embedding query relationship includes the mapping relationship between information type and embedding vector. P is an integer greater than or equal to 1.

[0274] For each set of training sample data, Q sets of information sample embedding vectors are obtained based on the information sample information. Among them, at least one set of information sample embedding vectors in the Q sets of information sample embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1.

[0275] For each set of training sample data, the elements corresponding to the same position in the embedding vector of the P sets of object samples are added together to obtain the object sample feature vector of the training object.

[0276] For each set of training sample data, the elements corresponding to the same position in the embedding vector of the Q sets of information samples are added together to obtain the information sample feature vector of the training information.

[0277] For each set of training sample data, the feature vectors of the object samples and the feature vectors of the information samples are concatenated to obtain the concatenated sample vector.

[0278] For each set of training sample data, based on the sample concatenation vector, the shallow conversion rate prediction value of the training information sample, the deep conversion rate prediction value of the training information sample, and the shallow to deep conversion rate prediction value of the training information sample are obtained through the conversion rate prediction model to be trained.

[0279] For N sets of training sample data, the model parameters of the conversion rate prediction model to be trained are updated according to the conversion type of training information, the shallow conversion rate prediction value of the sample, and the shallow to deep conversion rate prediction value of the sample, until the model training conditions are met, and the conversion rate prediction model is obtained.

[0280] In one or more embodiments, a method for training a conversion rate prediction model is described. As can be seen from the foregoing embodiments, the conversion rate prediction model includes two tasks: one is the prediction task of the shallow conversion rate prediction value (ecvr1), and the other is the prediction task of the shallow conversion rate prediction value (ecvr1), wherein the two tasks can share the embedding vector in the embedded query relation.

[0281] Specifically, N sets of training sample data are used as a batch of training data, for example, N is 1000. Each set of training sample data includes object sample information of the training object (e.g., basic attribute information, behavioral information, and interest information), information sample information of the training information (e.g., information identifier, main information identifier, information type, text information, and image information), and transformation type of the training information. For shallow training information, the transformation type is "shallow transformation type", while for deep training information, the transformation type includes both "shallow transformation type" and "deep transformation type".

[0282] For easier understanding, please refer to Figure 8 , Figure 8The figure shows a schematic diagram of the conversion rate prediction model in this embodiment. Object information and information information (e.g., object sample information and information sample information) are used as input to the embedding layer. The embedding layer transforms the object sample information in each group of training sample data according to the embedding query relationship, thereby obtaining P groups of object sample embedding vectors. After element-wise addition of the P groups of object sample embedding vectors, the object sample feature vector is obtained. Similarly, the embedding layer transforms the information sample information in each group of training sample data according to the embedding query relationship, thereby obtaining Q groups of information sample embedding vectors. After element-wise addition of the Q groups of information sample embedding vectors, the information sample feature vector is obtained. The object sample feature vector and the information sample feature vector are concatenated to obtain the concatenated sample vector corresponding to each group of training sample data.

[0283] Fully connected layer 1 takes the concatenated vector of N samples as input, passes it through a multilayer perceptron, and then uses either sigmoid or softmax to obtain N shallow conversion rate predictions (ecvr1). Fully connected layer 2 takes the concatenated vector of N samples as input, passes it through a multilayer perceptron, and then uses either sigmoid or softmax to obtain N shallow-to-deep conversion rate predictions (e_deep_cvr). Based on this, the deep conversion rate prediction (ecvr2) is obtained by multiplying the shallow conversion rate predictions (ecvr1) and the shallow-to-deep conversion rate predictions (e_deep_cvr). Therefore, a loss function is constructed based on the N shallow conversion rate predictions (ecvr1) and the N shallow-to-deep conversion rate predictions (e_deep_cvr), and the loss value is calculated. Based on this loss value, the backpropagation gradient algorithm is used to update the model parameters of the conversion rate prediction model to be trained until the model training conditions are met, thus obtaining the conversion rate prediction model.

[0284] During the model prediction phase, the shallow conversion rate estimate (ecvr1) is output through fully connected layer 1, and the shallow-to-deep conversion rate estimate (e_deep_cvr) is output through fully connected layer 2. The deep conversion rate estimate (ecvr2) can be obtained by multiplying the shallow conversion rate estimate (ecvr1) and the shallow-to-deep conversion rate estimate (e_deep_cvr).

[0285] Secondly, this application provides a method for training a conversion rate prediction model. Through this method, a multi-task model is constructed, introducing a training task for shallow to deep conversion rate prediction as an auxiliary task, while simultaneously training both shallow and deep conversion rate targets. After the model training is complete, the shallow to deep conversion rate prediction can be used as an intermediate result output during online prediction, thereby enabling the calculation of revenue and improving the feasibility and operability of the solution.

[0286] Optionally, in the above Figure 3 Based on the corresponding embodiments, another optional embodiment provided in this application may further include:

[0287] For N sets of training sample data, the embedding vector in the embedded query relationship is updated based on the conversion type of the training information, the shallow conversion rate prediction of the sample, and the shallow to deep conversion rate prediction of the sample.

[0288] In one or more embodiments, a method for updating the embedding vector in an embedded query relation is described.

[0289] Specifically, based on the transformation type of each training information, a first set of training information belonging to the shallow transformation type and a second set of training information belonging to the deep transformation type are determined. Then, a first loss value is determined based on the transformation type of each training information and the estimated shallow transformation rate of each training sample in the N sets of training sample data. Furthermore, a second loss value is determined based on the first training information set, the second training information set, the estimated shallow transformation rate of each training sample in the N sets of training sample data, and the estimated shallow-to-deep transformation rate. Combining the first and second loss values, the backpropagation gradient is used to update each embedding vector in the embedded query relationship.

[0290] Understandably, before model training, the embedding vectors in the embedded query relations are randomly initialized values, so these embedding vectors need to be updated through model training.

[0291] Furthermore, in this embodiment of the application, a method for updating the embedding vector in the embedding query relationship is provided. Through the above method, the embedding vector in the embedding query relationship can be updated according to the inverse gradient, thereby achieving a better feature fitting effect. Based on the embedding query relationship, the transformation from high-dimensional sparse vector to low-dimensional dense feature vector can be completed, and gradient unification can be achieved.

[0292] Optionally, in the above Figure 3Based on the corresponding embodiments, in another optional embodiment provided by this application, for N sets of training sample data, the model parameters of the conversion rate prediction model to be trained are updated according to the conversion type of training information, the shallow conversion rate prediction value of the samples, and the shallow to deep conversion rate prediction value of the samples. Specifically, this includes:

[0293] For N sets of training sample data, the first loss value is determined based on the conversion type of the training information and the estimated shallow conversion rate of the samples.

[0294] For N sets of training sample data, the second loss value is determined based on the conversion type of training information, the shallow conversion rate prediction of the sample, and the shallow to deep conversion rate prediction of the sample.

[0295] The model parameters of the conversion rate prediction model to be trained are updated based on the first loss value and the second loss value.

[0296] In one or more embodiments, a method for updating the model parameters corresponding to the conversion rate prediction model is introduced. As can be seen from the foregoing embodiments, a loss function needs to be constructed based on the shallow conversion rate prediction values ​​(ecvr1) of N samples and the shallow-to-deep conversion rate prediction values ​​(e_deep_cvr) of N samples. Simultaneously, multiplying the shallow conversion rate prediction value (ecvr1) by the shallow-to-deep conversion rate prediction value (e_deep_cvr) yields the deep conversion rate prediction value (ecvr2). Therefore, the network structure has three sub-tasks, each used to output the aforementioned three results.

[0297] Specifically, the loss function mainly includes the task loss for the shallow conversion rate prediction (ecvr1) and the task loss for the shallow-to-deep conversion rate prediction (e_deep_cvr), while the deep conversion rate prediction (ecvr2) is learned as a latent variable. Based on this, the first loss function is constructed as follows:

[0298]

[0299] Where, L(θ) CVR () represents the first loss value. N represents the total number of training sample data. x i Let y represent the i-th training sample data. i This indicates training information that has undergone shallow transformation. θ CVR1 This represents the model parameters for predicting the shallow conversion rate of the sample (ecvr1). f(x) i ;θ CVR1 ) represents the predicted shallow conversion rate estimate (ecvr1) of the sample.

[0300] The second loss function is constructed as follows:

[0301]

[0302] Where, L(θ) DeepCVR ) represents the second loss value. N represents the total number of training sample data. x i Let y represent the i-th training sample data. i This indicates training information that has undergone shallow transformation. i This indicates training information that has undergone shallow transformation. θ DeepCVR This represents the model parameters for predicting the shallow-to-deep conversion rate (e_deep_cvr) of the sample. f(x) i ;θ CVR1 f(x) represents the predicted shallow conversion rate of the sample (ecvr1). i θ DeepCVR f(x) represents the predicted shallow-to-deep conversion rate (e_deep_cvr) of the sample. i ;θ CVR1 )*f(x i ;θ DeepCVR ) represents the predicted sample deep conversion rate estimate (ecvr2).

[0303] It should be noted that during the training process, training sample data that has not undergone shallow or deep transformation will be added as negative samples for training, thereby achieving a better fitting effect.

[0304] Furthermore, in this embodiment of the application, a method for updating the model parameters corresponding to the conversion rate prediction model is provided. Through the above method, the loss value can be used not only to learn the branch tasks, but also to learn the embedding vectors embedded in the query relationship. Thus, multiple branch tasks can share the underlying parameters, thereby alleviating the problem of data coefficients in the shallow to deep conversion rate prediction training branch tasks. It is equivalent to sharing the samples in the shallow conversion rate prediction training branch tasks with the shallow to deep conversion rate prediction training branch tasks.

[0305] Optionally, in the above Figure 3 Based on the corresponding embodiments, another optional embodiment provided in this application may further include:

[0306] Obtain N sets of training sample data, where each set of training sample data includes object sample information of the training object, information sample information of the training information, and transformation type of the training information, where N is an integer greater than or equal to 1.

[0307] For each set of training sample data, the object sample information is characterized to obtain the object sample feature vector;

[0308] For each set of training sample data, the information sample information is characterized to obtain the information sample feature vector;

[0309] For each set of training sample data, the feature vectors of the object samples and the feature vectors of the information samples are concatenated to obtain the concatenated sample vector.

[0310] For each set of training sample data, based on the sample concatenation vector, the shallow conversion rate prediction value of the training information sample, the deep conversion rate prediction value of the training information sample, and the shallow to deep conversion rate prediction value of the training information sample are obtained through the conversion rate prediction model to be trained.

[0311] Based on the conversion type of the training information, the shallow conversion rate prediction of the samples, and the shallow to deep conversion rate prediction of the samples, the model parameters of the conversion rate prediction model to be trained are updated until the model training conditions are met, and the conversion rate prediction model is obtained.

[0312] In one or more embodiments, another method for training a conversion rate prediction model is provided. As can be seen from the foregoing embodiments, the conversion rate prediction model includes two tasks: one is a shallow conversion rate prediction (ecvr1) prediction task, and the other is a shallow conversion rate prediction (ecvr1) prediction task, wherein these two tasks may share the embedding vector in the embedded query relation.

[0313] Specifically, N sets of training sample data are used as a batch of training data, for example, N is 1000. Each set of training sample data includes object sample information of the training object (e.g., basic attribute information, behavioral information, and interest information), information sample information of the training information (e.g., information identifier, main information identifier, information type, text information, and image information), and transformation type of the training information. For shallow training information, the transformation type is "shallow transformation type", while for deep training information, the transformation type includes both "shallow transformation type" and "deep transformation type".

[0314] For easier understanding, please refer to Figure 9 , Figure 9 This is another structural diagram of the conversion rate prediction model in this application embodiment. As shown in the figure, the object information and information information (e.g., object sample information and information sample information) are processed to obtain object feature vectors (e.g., object sample feature vectors) and information feature vectors (e.g., information sample feature vectors). The object sample feature vectors and information sample feature vectors are concatenated to obtain the sample concatenation vector corresponding to each set of training sample data.

[0315] Fully connected layer 1 takes the concatenated vector of N samples as input, passes it through a multilayer perceptron, and then uses either sigmoid or softmax to obtain N shallow conversion rate predictions (ecvr1). Fully connected layer 2 takes the concatenated vector of N samples as input, passes it through a multilayer perceptron, and then uses either sigmoid or softmax to obtain N shallow-to-deep conversion rate predictions (e_deep_cvr). Based on this, the deep conversion rate prediction (ecvr2) is obtained by multiplying the shallow conversion rate predictions (ecvr1) and the shallow-to-deep conversion rate predictions (e_deep_cvr). Therefore, a loss function is constructed based on the N shallow conversion rate predictions (ecvr1) and the N shallow-to-deep conversion rate predictions (e_deep_cvr), and the loss value is calculated. Based on this loss value, the backpropagation gradient algorithm is used to update the model parameters of the conversion rate prediction model to be trained until the model training conditions are met, thus obtaining the conversion rate prediction model.

[0316] During the model prediction phase, the shallow conversion rate estimate (ecvr1) is output through fully connected layer 1, and the shallow-to-deep conversion rate estimate (e_deep_cvr) is output through fully connected layer 2. The deep conversion rate estimate (ecvr2) can be obtained by multiplying the shallow conversion rate estimate (ecvr1) and the shallow-to-deep conversion rate estimate (e_deep_cvr).

[0317] Furthermore, this application provides another method for training the conversion rate prediction model. By introducing a training task for shallow-to-deep conversion rate prediction as an auxiliary task, both shallow and deep conversion rate targets are trained simultaneously. After model training is complete, the shallow-to-deep conversion rate prediction can be used as an intermediate output during online prediction, thereby enabling the calculation of revenue values ​​and improving the feasibility and operability of the solution. In addition, no additional embedding layer training is required, which helps improve model training efficiency.

[0318] The information recommendation device in this application is described in detail below. Please refer to [link / reference]. Figure 10 , Figure 10 This is a schematic diagram of one embodiment of the information recommendation device in this application. The information recommendation device 20 includes:

[0319] The acquisition module 210 is used to acquire target object information of the target object and candidate information of the candidate information;

[0320] The acquisition module 210 is also used to obtain the estimated click-through rate of the target object for the candidate information through a click-through rate prediction model based on the target object information and the candidate information.

[0321] The acquisition module 210 is also used to obtain the shallow conversion rate estimate and the shallow to deep conversion rate estimate of the candidate information based on the target object information and the candidate information information through the conversion rate prediction model.

[0322] The determination module 220 is used to calculate the estimated click-through rate, shallow conversion rate estimate, shallow to deep conversion rate estimate, shallow bid and deep bid by calling a comprehensive function if the candidate information belongs to deep target information, so as to obtain the target revenue value corresponding to the candidate information. The deep target information has a pre-set shallow bid and deep bid.

[0323] The recommendation module 230 is used to recommend candidate information to the target object if it is determined that the candidate information meets the information recommendation conditions based on the target benefit value.

[0324] This application provides an information recommendation device. Using this device, by combining object association features and information association features, the estimated click-through rate, shallow conversion rate, and shallow-to-deep conversion rate of candidate information are predicted. These estimates are then used to determine the target revenue value of the information; the higher the target revenue value, the higher the value of recommending the information to the target object. Therefore, it is possible to recommend targeted information content to specific objects, thereby improving the recommendation effect.

[0325] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0326] Module 220 is specifically used to determine the target deep conversion rate of candidate information based on the ratio of deep bid to shallow bid.

[0327] The conversion rate threshold is determined by multiplying the target deep conversion rate by the dynamic pricing factor. The dynamic pricing factor is related to the proportion of deep target information in the total amount of information.

[0328] The deep conversion rate estimate of the candidate information is determined by multiplying the shallow conversion rate estimate by the shallow to deep conversion rate estimate.

[0329] If the estimated conversion rate from shallow to deep layers is greater than the conversion rate threshold, then the target revenue value corresponding to the candidate information is calculated based on the estimated click-through rate, the estimated shallow conversion rate, the shallow bid, the dynamic price adjustment factor, the estimated deep conversion rate, and the deep bid.

[0330] If the estimated conversion rate from shallow to deep layers is less than or equal to the conversion rate threshold, then the target revenue value corresponding to the candidate information is determined as the minimum revenue value.

[0331] In this embodiment, an information recommendation device is provided. Using this device, if the estimated conversion rate from shallow to deep layers is less than or equal to a conversion rate threshold, the deep conversion effect is considered poor. Therefore, the target return value is directly set to the minimum return value, meaning it does not need to participate in information sorting, thereby reducing the amount of information computation and saving computational resources.

[0332] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0333] The determination module 220 is also used to determine the target revenue value corresponding to the candidate information based on the estimated click-through rate, the estimated shallow conversion rate, and the shallow bid if the candidate information belongs to shallow target information, wherein the shallow target information has a pre-set shallow bid.

[0334] This application provides an information recommendation device. Using this device, even for shallow target information, the corresponding target benefit value can be calculated, thereby increasing the feasibility and operability of the solution.

[0335] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0336] The acquisition module 210 is specifically used to acquire target object information corresponding to the target object, wherein the target object information includes at least one of the target object's basic attribute information, behavioral information, and interest information;

[0337] Obtain the candidate information corresponding to the candidate information, wherein the candidate information includes at least one of the following: information identifier, main information identifier, information type, text information, and image information.

[0338] This application provides an information recommendation device. Using this device, information about the target object and candidate information is extracted based on different dimensions, enriching the types of information and improving the accuracy of the model's output.

[0339] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0340] The acquisition module 210 is specifically used to acquire P sets of object embedding vectors based on the target object information, wherein at least one set of object embedding vectors in the P sets of object embedding vectors is determined based on the embedding query relationship, and P is an integer greater than or equal to 1.

[0341] Q sets of information embedding vectors are obtained based on candidate information information, wherein at least one set of information embedding vectors in the Q sets of information embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1;

[0342] Add the elements at the same position in the embedding vector of the P groups of objects to obtain the object feature vector of the target object;

[0343] Add the elements at the same position in the Q-group information embedding vector to obtain the information feature vector of the candidate information;

[0344] The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector.

[0345] Based on the target splicing vector, shallow conversion rate predictions and shallow-to-deep conversion rate predictions of candidate information are obtained through a conversion rate prediction model.

[0346] In this embodiment of the application, an information recommendation device is provided. Using the above device, considering that some information (e.g., information identifiers) has a large volume, embedding vectors with fewer feature dimensions can be obtained based on trained embedding query relationships. Therefore, on the one hand, feature dimensionality reduction can accelerate computation; on the other hand, it can avoid model overfitting due to too many features.

[0347] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0348] The acquisition module 210 is specifically used to perform feature processing on the target object information to obtain the object feature vector of the target object;

[0349] The candidate information is processed to obtain the information feature vector of the candidate information;

[0350] The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector.

[0351] Based on the target splicing vector, shallow conversion rate predictions and shallow-to-deep conversion rate predictions of candidate information are obtained through a conversion rate prediction model.

[0352] This application provides an information recommendation device. Using this device, feature processing of target object information and candidate information can be performed directly based on feature engineering, eliminating the need to train additional feature extraction networks and thus saving training resources.

[0353] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0354] The acquisition module 210 is specifically used to acquire object sub-information of the target object from the target object information, wherein the feature types included in the object sub-information are less than or equal to the feature types included in the target object information;

[0355] Obtain information sub-information of candidate information from candidate information information, wherein the feature types included in the information sub-information are less than or equal to the feature types included in the candidate information information;

[0356] The object sub-information is characterized to obtain the object sub-feature vector of the target object;

[0357] The information sub-information is characterized to obtain the information sub-feature vector of the candidate information;

[0358] Based on object sub-feature vectors and information sub-feature vectors, the predicted click-through rate of the target object for candidate information is obtained through a click-through rate prediction model.

[0359] This application provides an information recommendation device. Using this device, a pre-trained click-through rate (CTR) prediction model can directly output the predicted CTR, thereby improving the feasibility and operability of the solution.

[0360] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application, the information recommendation device 20 further includes a processing module 240 and a training module 250;

[0361] The acquisition module 210 is also used to acquire M sets of training sample data, wherein each set of training sample data includes object sub-sample information of the training object, information sub-sample information of the training information, and labeled click tags of the training object, and M is an integer greater than or equal to 1.

[0362] The processing module 240 is used to perform feature processing on the object sub-sample information for each group of training sample data to obtain the object sample sub-feature vector of the training object.

[0363] The processing module 240 is also used to perform feature processing on the information sub-sample information for each group of training sample data to obtain the information sample sub-feature vector of the training information.

[0364] The acquisition module 210 is also used to obtain the estimated click rate of the training object for the training information based on the object sample sub-feature vector and the information sample sub-feature vector for each group of training sample data through the click rate prediction model to be trained.

[0365] The training module 250 is used to update the model parameters of the click-through rate prediction model to be trained based on the predicted click-through rate of the training object for the training information and the labeled click tags of the training object for the M sets of training sample data, until the model training conditions are met, and the click-through rate prediction model is obtained.

[0366] This application provides an information recommendation device. Using this device, a click-through rate (CTR) prediction model can be trained to output the predicted CTR, thereby improving the feasibility and operability of the solution.

[0367] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application, the information recommendation device 20 further includes an update module 260;

[0368] The update module 260 is used to update the information sorting list according to the target profit value, in descending order of profit value. The information sorting list includes the mapping relationship between the information and the sorting results.

[0369] The determining module 220 is further configured to determine that the candidate information meets the information recommendation conditions if the sorting result corresponding to the candidate information is the first K sorting results in the information sorting list, where K is an integer greater than or equal to 1.

[0370] This application provides an information recommendation device. By using the above device and prioritizing information based on its reward value, information with higher reward values ​​can be recommended first, thereby improving the rationality of information recommendations and providing better information feedback.

[0371] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0372] The determination module 220 is also used to determine the group benefit value based on the basic attribute information included in the target object information and the information type included in the candidate information information;

[0373] The determination module 220 is also used to determine the target comprehensive return value based on the target return value and the group return value;

[0374] The update module 260 is also used to update the information sorting list according to the target comprehensive return value in descending order of comprehensive return value. The information sorting list includes the mapping relationship between information and sorting results.

[0375] The determining module 220 is further configured to determine that the candidate information meets the information recommendation conditions if the sorting result corresponding to the candidate information is the first K sorting results in the information sorting list, where K is an integer greater than or equal to 1.

[0376] This application provides an information recommendation device. By using this device, which integrates reward value and other factors as the basis for information ranking, it can not only prioritize recommending information with higher reward values ​​to achieve better information feedback, but also achieve a more comprehensive evaluation of information value, thereby helping to improve the rationality of information recommendations.

[0377] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0378] The acquisition module 210 is also used to acquire N sets of training sample data, wherein each set of training sample data includes object sample information of the training object, information sample information of the training information and transformation type of the training information, and N is an integer greater than or equal to 1.

[0379] The acquisition module 210 is also used to acquire P sets of object sample embedding vectors for each set of training sample data based on object sample information, wherein at least one set of object sample embedding vectors in the P sets of object sample embedding vectors is determined based on the embedding query relationship, the embedding query relationship includes the mapping relationship between information type and embedding vector, and P is an integer greater than or equal to 1.

[0380] The acquisition module 210 is also used to acquire Q sets of information sample embedding vectors for each set of training sample data based on information sample information, wherein at least one set of information sample embedding vectors in the Q sets of information sample embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1.

[0381] The processing module 240 is also used to add the elements corresponding to the same position in the embedding vector of the P groups of object samples for each group of training sample data to obtain the object sample feature vector of the training object.

[0382] The processing module 240 is also used to add the elements corresponding to the same position in the embedding vector of the Q groups of information samples for each group of training sample data to obtain the information sample feature vector of the training information.

[0383] The processing module 240 is also used to concatenate the object sample feature vector and the information sample feature vector for each set of training sample data to obtain a sample concatenation vector.

[0384] The acquisition module 210 is also used to obtain the shallow conversion rate prediction value of the training information, the deep conversion rate prediction value of the training information, and the shallow to deep conversion rate prediction value of the training information for each group of training sample data, based on the sample splicing vector and through the conversion rate prediction model to be trained.

[0385] The training module 250 is also used to update the model parameters of the conversion rate prediction model to be trained based on the conversion type of the training information, the shallow conversion rate prediction value of the sample, and the shallow to deep conversion rate prediction value of the sample for N sets of training sample data, until the model training conditions are met, and the conversion rate prediction model is obtained.

[0386] This application provides an information recommendation device. Using this device, a multi-task model is constructed, introducing a training task for shallow-to-deep conversion rate prediction as an auxiliary task, while simultaneously training both shallow and deep conversion rate targets. After model training is complete, the shallow-to-deep conversion rate prediction can be output as an intermediate result during online prediction, thereby enabling the calculation of revenue values ​​and improving the feasibility and operability of the solution.

[0387] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0388] The training module 250 is also used to update the embedding vector in the embedded query relationship based on the conversion type of the training information, the shallow conversion rate estimate of the sample, and the shallow to deep conversion rate estimate of the sample for N sets of training sample data.

[0389] This application provides an information recommendation device. Using this device, the embedding vector in the embedded query relationship can be updated according to the inverse gradient, thereby achieving better feature fitting results. Based on the embedded query relationship, the transformation from high-dimensional sparse vectors to low-dimensional dense feature vectors can be completed, achieving gradient unification.

[0390] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0391] Training module 250 is specifically used to determine the first loss value for N sets of training sample data based on the conversion type of training information and the shallow conversion rate estimate of the samples.

[0392] For N sets of training sample data, the second loss value is determined based on the conversion type of training information, the shallow conversion rate prediction of the sample, and the shallow to deep conversion rate prediction of the sample.

[0393] The model parameters of the conversion rate prediction model to be trained are updated based on the first loss value and the second loss value.

[0394] This application provides an information recommendation device. Using this device, not only can the loss value be used to learn about branch tasks, but also the embedding vectors embedded in the query relationship can be learned. Therefore, multiple branch tasks can share underlying parameters, thereby alleviating the data coefficient problem that occurs in the shallow-to-deep conversion rate prediction training branch task. This is equivalent to sharing the samples from the shallow conversion rate prediction training branch task with the shallow-to-deep conversion rate prediction training branch task.

[0395] Optionally, in the above Figure 10 Based on the corresponding embodiments, in another embodiment of the information recommendation device 20 provided in this application,

[0396] The acquisition module 210 is also used to acquire N sets of training sample data, wherein each set of training sample data includes object sample information of the training object, information sample information of the training information and transformation type of the training information, and N is an integer greater than or equal to 1.

[0397] The processing module 240 is also used to perform feature processing on the object sample information for each group of training sample data to obtain the object sample feature vector.

[0398] The processing module 240 is also used to perform feature processing on the information sample information for each group of training sample data to obtain the information sample feature vector;

[0399] The processing module 240 is also used to concatenate the object sample feature vector and the information sample feature vector for each set of training sample data to obtain a sample concatenation vector.

[0400] The acquisition module 210 is also used to obtain the shallow conversion rate prediction value of the training information, the deep conversion rate prediction value of the training information, and the shallow to deep conversion rate prediction value of the training information for each group of training sample data, based on the sample splicing vector and through the conversion rate prediction model to be trained.

[0401] The training module 250 is also used to update the model parameters of the conversion rate prediction model to be trained based on the conversion type of the training information, the shallow conversion rate prediction value of the sample, and the shallow to deep conversion rate prediction value of the sample, until the model training conditions are met and the conversion rate prediction model is obtained.

[0402] This application provides an information recommendation device. Using this device, a training task for shallow-to-deep conversion rate prediction is introduced as an auxiliary task, simultaneously training both shallow and deep conversion rate targets. After model training is complete, the shallow-to-deep conversion rate prediction can be output as an intermediate result during online prediction, thereby enabling the calculation of revenue values ​​and improving the feasibility and operability of the solution. Furthermore, no additional embedding layer training is required, which helps improve model training efficiency.

[0403] Figure 11 This is a schematic diagram of a computer device structure provided in an embodiment of this application. The computer device 300 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 322 (e.g., one or more processors) and a memory 332, and one or more storage media 330 (e.g., one or more mass storage devices) for storing application programs 342 or data 344. The memory 332 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the computer device. Furthermore, the CPU 322 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the computer device 300.

[0404] Computer device 300 may also include one or more power supplies 326, one or more wired or wireless network interfaces 350, one or more input / output interfaces 358, and / or one or more operating systems 341, such as Windows Server. TM Mac OS X TM Unix TM Linux TM FreeBSD TM etc.

[0405] The steps performed by the computer device in the above embodiments can be based on this Figure 11 The computer device structure shown.

[0406] This application also provides a computer-readable storage medium storing a computer program that, when run on a computer, causes the computer to perform the methods described in the foregoing embodiments.

[0407] This application also provides a computer program product including a program, which, when run on a computer, causes the computer to perform the methods described in the foregoing embodiments.

[0408] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0409] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0410] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0411] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0412] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0413] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An information recommendation method, characterized in that, include: Obtain target object information and candidate information of candidate information; Based on the target object information and the candidate information, the estimated click-through rate of the target object for the candidate information is obtained through a click-through rate prediction model; Based on the target object information and the candidate information, a shallow conversion rate estimate and a shallow-to-deep conversion rate estimate of the candidate information are obtained through a conversion rate prediction model. If the candidate information belongs to deep target information, then a comprehensive function is called to calculate the estimated click-through rate, the estimated shallow conversion rate, the estimated shallow-to-deep conversion rate, the shallow bid, and the deep bid to obtain the target revenue value corresponding to the candidate information. This includes: determining the target deep conversion rate of the candidate information based on the ratio of the deep bid to the shallow bid; and determining the conversion rate threshold based on the product of the target deep conversion rate and the dynamic pricing factor, wherein the dynamic pricing factor and the deep target information account for a certain percentage of the total number of information. The correlation is established; the deep conversion rate estimate of the candidate information is determined by multiplying the shallow conversion rate estimate by the shallow-to-deep conversion rate estimate; if the shallow-to-deep conversion rate estimate is greater than the conversion rate threshold, the target revenue value corresponding to the candidate information is calculated based on the estimated click-through rate, the shallow conversion rate estimate, the shallow bid, the dynamic pricing factor, the deep conversion rate estimate, and the deep bid, and the deep target information has the pre-set shallow bid and deep bid; If the candidate information is determined to meet the information recommendation criteria based on the target return value, then the candidate information is recommended to the target object.

2. The information recommendation method according to claim 1, characterized in that, The step of calculating the target revenue value corresponding to the candidate information by calling a comprehensive function to calculate the estimated click-through rate, the estimated shallow conversion rate, the estimated shallow-to-deep conversion rate, the shallow bid, and the deep bid, further includes: If the estimated conversion rate from shallow to deep layers is less than or equal to the conversion rate threshold, then the target revenue value corresponding to the candidate information is determined as the minimum revenue value.

3. The information recommendation method according to claim 1, characterized in that, The method further includes: If the candidate information belongs to shallow target information, then the target revenue value corresponding to the candidate information is determined based on the estimated click-through rate, the estimated shallow conversion rate, and the shallow bid, wherein the shallow target information has a pre-set shallow bid.

4. The information recommendation method according to claim 1, characterized in that, The acquisition of target object information and candidate information of candidate information includes: Obtain target object information corresponding to the target object, wherein the target object information includes at least one of the target object's basic attribute information, behavioral information, and interest information; Obtain the candidate information information corresponding to the candidate information, wherein the candidate information information includes at least one of the following: information identifier, main information identifier, information type, text information, and image information.

5. The information recommendation method according to claim 1, characterized in that, The step of obtaining shallow conversion rate estimates and shallow-to-deep conversion rate estimates of the candidate information based on the target object information and the candidate information through a conversion rate prediction model includes: Based on the target object information, P sets of object embedding vectors are obtained, wherein at least one set of object embedding vectors in the P sets of object embedding vectors is determined based on the embedding query relationship, and P is an integer greater than or equal to 1; Q sets of information embedding vectors are obtained based on the candidate information, wherein at least one set of information embedding vectors in the Q sets of information embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1; The elements corresponding to the same position in the embedding vectors of the P groups of objects are added together to obtain the object feature vector of the target object. The elements corresponding to the same position in the Q-group information embedding vector are added together to obtain the information feature vector of the candidate information; The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector. Based on the target splicing vector, the shallow conversion rate estimate and the shallow-to-deep conversion rate estimate of the candidate information are obtained through the conversion rate prediction model.

6. The information recommendation method according to claim 1, characterized in that, The step of obtaining shallow conversion rate estimates and shallow-to-deep conversion rate estimates of the candidate information based on the target object information and the candidate information through a conversion rate prediction model includes: The target object information is subjected to feature processing to obtain the object feature vector of the target object; The candidate information is subjected to feature processing to obtain the information feature vector of the candidate information; The object feature vector and the information feature vector are concatenated to obtain the target concatenated vector. Based on the target splicing vector, the shallow conversion rate estimate and the shallow-to-deep conversion rate estimate of the candidate information are obtained through the conversion rate prediction model.

7. The information recommendation method according to claim 1, characterized in that, The step of obtaining the estimated click-through rate of the target object for the candidate information through a click-through rate prediction model based on the target object information and the candidate information includes: Obtain object sub-information of the target object from the target object information, wherein the feature types included in the object sub-information are less than or equal to the feature types included in the target object information; Obtain information sub-information of the candidate information from the candidate information information, wherein the feature types included in the information sub-information are less than or equal to the feature types included in the candidate information information; The object sub-information is subjected to feature processing to obtain the object sub-feature vector of the target object; The information sub-information is subjected to feature processing to obtain the information sub-feature vector of the candidate information; Based on the object sub-feature vector and the information sub-feature vector, the predicted click-through rate of the target object for the candidate information is obtained through the click-through rate prediction model.

8. The information recommendation method according to claim 1, characterized in that, The method further includes: Obtain M sets of training sample data, wherein each set of training sample data includes object sub-sample information of the training object, information sub-sample information of the training information, and the labeled click tags of the training object, and M is an integer greater than or equal to 1; For each set of training sample data, the object sub-sample information is characterized to obtain the object sample sub-feature vector of the training object; For each set of training sample data, the information sub-sample information is characterized to obtain the information sample sub-feature vector of the training information; For each set of training sample data, based on the object sample sub-feature vector and the information sample sub-feature vector, the predicted click-through rate of the training object for the training information is obtained through the click-through rate prediction model to be trained; For the M sets of training sample data, based on the predicted click-through rate of the training object for the training information and the labeled click tags of the training object, the model parameters of the click-through rate prediction model to be trained are updated until the model training conditions are met, and the click-through rate prediction model is obtained.

9. The information recommendation method according to claim 1, characterized in that, The method further includes: The information sorting list is updated according to the target profit value, in descending order of profit value. The information sorting list includes the mapping relationship between information and sorting results. If the sorting result corresponding to the candidate information is one of the top K sorting results in the information sorting list, then the candidate information is determined to meet the information recommendation conditions, where K is an integer greater than or equal to 1.

10. The information recommendation method according to claim 1, characterized in that, The method further includes: The group benefit value is determined based on the basic attribute information included in the target object information and the information types included in the candidate information. Based on the target return value and the group return value, determine the target comprehensive return value; The information sorting list is updated according to the target comprehensive return value, in descending order of comprehensive return value. The information sorting list includes the mapping relationship between information and sorting results. If the sorting result corresponding to the candidate information is one of the top K sorting results in the information sorting list, then the candidate information is determined to meet the information recommendation conditions, where K is an integer greater than or equal to 1.

11. The information recommendation method according to any one of claims 1 to 10, characterized in that, The method further includes: Obtain N sets of training sample data, wherein each set of training sample data includes object sample information of the training object, information sample information of the training information, and the transformation type of the training information, and N is an integer greater than or equal to 1; For each set of training sample data, P sets of object sample embedding vectors are obtained based on the object sample information. Among the P sets of object sample embedding vectors, at least one set of object sample embedding vectors is determined based on an embedding query relationship. The embedding query relationship includes a mapping relationship between information type and embedding vector. P is an integer greater than or equal to 1. For each set of training sample data, Q sets of information sample embedding vectors are obtained based on the information sample information, wherein at least one set of information sample embedding vectors in the Q sets of information sample embedding vectors is determined based on the embedding query relationship, and Q is an integer greater than or equal to 1; For each set of training sample data, the elements corresponding to the same position in the embedding vector of the P sets of object samples are added together to obtain the object sample feature vector of the training object. For each set of training sample data, the elements corresponding to the same position in the embedding vector of the Q sets of information samples are added together to obtain the information sample feature vector of the training information. For each set of training sample data, the feature vector of the object sample and the feature vector of the information sample are concatenated to obtain a concatenated sample vector; For each set of training sample data, based on the sample concatenation vector, the shallow conversion rate prediction value of the training information, the deep conversion rate prediction value of the training information, and the shallow to deep conversion rate prediction value of the training information are obtained through the conversion rate prediction model to be trained. For the N sets of training sample data, the model parameters of the conversion rate prediction model to be trained are updated according to the conversion type of the training information, the shallow conversion rate prediction value of the sample, and the shallow to deep conversion rate prediction value of the sample, until the model training conditions are met, and the conversion rate prediction model is obtained.

12. The information recommendation method according to claim 11, characterized in that, The method further includes: For the N sets of training sample data, the embedding vector in the embedding query relationship is updated according to the conversion type of the training information, the shallow conversion rate estimate of the sample, and the shallow to deep conversion rate estimate of the sample.

13. The information recommendation method according to claim 11, characterized in that, The step of updating the model parameters of the conversion rate prediction model to be trained, based on the conversion type of the training information, the shallow conversion rate prediction value of the samples, and the shallow-to-deep conversion rate prediction value of the samples, for the N sets of training sample data, includes: For the N sets of training sample data, a first loss value is determined based on the conversion type of the training information and the estimated shallow conversion rate of the samples. For the N sets of training sample data, a second loss value is determined based on the conversion type of the training information, the shallow conversion rate estimate of the sample, and the shallow to deep conversion rate estimate of the sample. The model parameters of the conversion rate prediction model to be trained are updated based on the first loss value and the second loss value.

14. The information recommendation method according to any one of claims 1 to 10, characterized in that, The method further includes: Obtain N sets of training sample data, wherein each set of training sample data includes object sample information of the training object, information sample information of the training information, and the transformation type of the training information, and N is an integer greater than or equal to 1; For each set of training sample data, the object sample information is characterized to obtain an object sample feature vector. For each set of training sample data, the information sample information is characterized to obtain an information sample feature vector. For each set of training sample data, the feature vector of the object sample and the feature vector of the information sample are concatenated to obtain a concatenated sample vector; For each set of training sample data, based on the sample concatenation vector, the shallow conversion rate prediction value of the training information, the deep conversion rate prediction value of the training information, and the shallow to deep conversion rate prediction value of the training information are obtained through the conversion rate prediction model to be trained. Based on the conversion type of the training information, the shallow conversion rate prediction of the sample, and the shallow to deep conversion rate prediction of the sample, the model parameters of the conversion rate prediction model to be trained are updated until the model training conditions are met, and the conversion rate prediction model is obtained.

15. An information recommendation device, characterized in that, include: The acquisition module is used to acquire target object information and candidate information of candidate information; The acquisition module is further configured to obtain the estimated click-through rate of the target object for the candidate information based on the target object information and the candidate information, through a click-through rate prediction model; The acquisition module is further configured to obtain, based on the target object information and the candidate information information, a shallow conversion rate estimate and a shallow-to-deep conversion rate estimate of the candidate information through a conversion rate prediction model. The determination module is used to, if the candidate information belongs to deep target information, calculate the target revenue value corresponding to the candidate information by calling a comprehensive function to calculate the estimated click-through rate, the estimated shallow conversion rate, the estimated shallow-to-deep conversion rate, the shallow bid, and the deep bid. This includes: determining the target deep conversion rate of the candidate information based on the ratio of the deep bid to the shallow bid; and determining the conversion rate threshold based on the product of the target deep conversion rate and a dynamic pricing factor, wherein the dynamic pricing factor and the deep target information are related in terms of information... The proportion of the total is related; the deep conversion rate estimate of the candidate information is determined by multiplying the shallow conversion rate estimate by the shallow to deep conversion rate estimate; if the shallow to deep conversion rate estimate is greater than the conversion rate threshold, the target revenue value corresponding to the candidate information is calculated based on the estimated click-through rate, the shallow conversion rate estimate, the shallow bid, the dynamic pricing factor, the deep conversion rate estimate, and the deep bid, and the deep target information has the set shallow bid and the deep bid; The recommendation module is used to recommend the candidate information to the target object if the candidate information meets the information recommendation conditions based on the target benefit value.

16. A computer device, characterized in that, include: Memory, processor, and bus system; The memory is used to store programs; The processor is configured to execute a program in the memory, and the processor is configured to execute the information recommendation method according to any one of claims 1 to 14 according to the instructions in the program code; The bus system is used to connect the memory and the processor to enable communication between the memory and the processor.

17. A computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the information recommendation method as claimed in any one of claims 1 to 14.

18. A computer program product, comprising a computer program and instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the information recommendation method as described in any one of claims 1 to 14.

Images