A content push method, device, equipment and storage medium

Abstract

An embodiment of the present application provides a content push method, device, equipment and medium. The content push method includes: obtaining a feature set of a target object; processing the feature set by using a visit conversion prediction model to obtain a visit conversion prediction value of the target object; if it is determined according to the visit conversion prediction value that the target object is an object to be pushed, pushing recommended content to the target object; wherein, the visit conversion prediction model is trained by using a conversion sample set and a visit sample set, the conversion sample set includes an un-converted sample subset and a converted sample subset, and the visit sample set includes a non-visited sample subset and a visited sample subset. By adopting the embodiment of the present application, objects to be pushed can be screened out for pushing, making the pushed objects more accurate, thereby improving the click-through rate and conversion rate of the recommended content.

Description

Technical Field

[0001] The present application relates to computer technology, and particularly to a content push method, a content push device, a terminal device, and a computer-readable storage medium. Background Art

[0002] With the development of information technology, recommended advertisements have become one of the main channels for various merchants to increase popularity and promote products. The effects of merchants' placed recommended advertisements are mainly measured by the click-through rate (CTR) and the conversion rate (CVR). The so-called click-through rate refers to the ratio of the number of clicks on a recommended advertisement to the number of times the recommended advertisement is displayed when browsing the recommended advertisement; the so-called conversion rate refers to the ratio of the number of transactions after entering the merchant's store.

[0003] Currently, in order to improve the click-through rate and conversion rate of recommended advertisements, the most commonly used method is to calculate the matching degree between the object label of a certain object and the product label, and recommend the recommended advertisement of the product with a higher matching degree to the object. Although the matching degree obtained by matching the object label and the product label can represent the relevance between the object's interest and the product, there are problems that the object label and the product label rely on manual tagging, are easily affected by personal subjectivity, cannot guarantee the accuracy of the label, resulting in poor recommendation effects and low conversion rates. Summary of the Invention

[0004] Embodiments of the present application provide a recommendation method, device, terminal device, and storage medium, which can screen out objects to be pushed for pushing, making the pushed objects more accurate, thereby improving the click-through rate and conversion rate of the recommended content.

[0005] In a first aspect, an embodiment of the present application provides a content push method, and the method includes:

[0006] Obtain a feature set of a target object;

[0007] Process the feature set by using a visit conversion prediction model to obtain a visit conversion prediction value of the target object;

[0008] If it is determined according to the visit conversion prediction value that the target object is an object to be pushed, then push recommended content to the target object;

[0009] Wherein, the visit conversion prediction model is trained by using a conversion sample set and a visit sample set, the conversion sample set includes an un-converted sample subset and a converted sample subset, and the visit sample set includes a non-visited sample subset and a visited sample subset.

[0010] Second aspect, embodiments of the present application provide a content push device, and the device includes:

[0011] An acquisition unit, configured to acquire a feature set of a target object;

[0012] A processing unit, configured to process the feature set by using a visit conversion prediction model to obtain a visit conversion prediction value of the target object;

[0013] A push unit, configured to push recommended content to the target object if it is determined according to the visit conversion prediction value that the target object is an object to be pushed;

[0014] Wherein, the visit conversion prediction model is trained by using a conversion sample set and a visit sample set, the conversion sample set includes an un-converted sample subset and a converted sample subset, and the visit sample set includes a non-visited sample subset and a visited sample subset.

[0015] In one implementation, the visit conversion prediction model includes a visit prediction module and a conversion prediction module, and the processing unit is specifically configured to:

[0016] Process the feature set by using the visit prediction module to obtain a visit prediction value of the target object;

[0017] Process the feature set by using the conversion prediction module to obtain a conversion prediction value of the target object;

[0018] Determine the visit conversion prediction value of the target object according to the visit prediction value and the conversion prediction value.

[0019] In one implementation, the device further includes:

[0020] The acquisition unit is further configured to acquire the conversion sample set and the visit sample set, each conversion sample in the conversion sample set includes a feature set of a first reference object and a conversion label, and each visit sample in the visit sample set includes a feature set of a second reference object and a visit label;

[0021] A determination unit, configured to train an initial visit conversion prediction model by using the conversion sample set and the visit sample set to determine target loss information;

[0022] An adjustment unit, configured to, when it is determined according to the target loss information that the training stop condition is not satisfied, adjust the model parameters of the initial visit conversion prediction model according to the target loss information, and train the adjusted visit conversion prediction model by using the conversion sample set and the visit sample set to obtain a trained visit conversion prediction model.

[0023] In one implementation, the above-mentioned determination unit is specifically configured to:

[0024] Train the initial incoming conversion prediction model by using the above-mentioned conversion sample set to determine the first loss information;

[0025] Train the initial incoming conversion prediction model by using the above-mentioned incoming sample set to determine the second loss information;

[0026] Determine the target loss information according to the above-mentioned first loss information and the above-mentioned second loss information.

[0027] In one implementation, if the conversion sample belongs to the converted sample subset, the conversion label of the above-mentioned conversion sample includes the conversion time, and the above-mentioned determination unit is specifically configured to:

[0028] For the first conversion sample, determine the conversion probability that the above-mentioned first conversion sample is converted at the conversion time corresponding to the above-mentioned first conversion sample, where the first conversion sample is any conversion sample in the above-mentioned converted sample subset;

[0029] For the second conversion sample, determine the conversion reference probability that the above-mentioned second conversion sample is not converted at the current training time, where the second conversion sample is any conversion sample in the above-mentioned unconverted sample subset;

[0030] Determine the first loss information according to the conversion probabilities of each first conversion sample and the reference probabilities of each second conversion sample.

[0031] In one implementation, the above-mentioned determination unit is specifically configured to:

[0032] Determine the probability that the above-mentioned second conversion sample will not be converted, and determine the probability that the above-mentioned second conversion sample will be converted after the current training time;

[0033] Determine the conversion reference probability that the above-mentioned second conversion sample is not converted at the current training time according to the probability that the above-mentioned second conversion sample will not be converted and the probability that the above-mentioned second conversion sample will be converted after the current training time.

[0034] In one implementation, if the incoming sample belongs to the incoming sample subset, the incoming label of the above-mentioned incoming sample includes the incoming time, and the above-mentioned determination unit is specifically configured to:

[0035] For the first incoming sample, determine the incoming probability that the above-mentioned first incoming sample is incoming at the incoming time corresponding to the above-mentioned first incoming sample, where the first incoming sample is any incoming sample in the above-mentioned incoming sample subset;

[0036] For the second visiting sample, determine the reference probability that the second visiting sample does not visit at the current training time, where the second visiting sample is any visiting sample in the subset of the future visiting samples;

[0037] Determine the second loss information according to the visiting probabilities of each first visiting sample and the reference probabilities of each second visiting sample.

[0038] In one implementation, the determining unit is specifically configured to:

[0039] Determine the probability that the second visiting sample will not visit, and determine the probability that the second visiting sample will visit after the current training time;

[0040] Determine the reference probability that the second visiting sample does not visit at the current training time according to the probability that the second visiting sample will not visit and the probability that the second visiting sample will visit after the current training time.

[0041] In one implementation, the apparatus further includes:

[0042] The obtaining unit is further configured to obtain the visiting conversion prediction value of each object in the pushable object set, where the target object is an object in the pushable object set;

[0043] The sorting unit is configured to sort the visiting conversion prediction values of each object in descending order of the visiting conversion prediction value;

[0044] The determining unit is further configured to determine that the target object is a to-be-pushed object if the visiting conversion prediction value of the target object is among the top M in the sorting result.

[0045] In a third aspect, an embodiment of the present application provides a content push device, including: a processor, a communication interface, and a memory. The processor, the communication interface, and the memory are interconnected. Wherein, the memory stores executable program code, and the processor is configured to call the executable program code to execute the content push method in the first aspect above.

[0046] In a fourth aspect, an embodiment of the present application further provides a computer-readable storage medium, where instructions are stored in the computer-readable storage medium, and when the instructions run on a computer, the computer is caused to execute the content push method in the first aspect above.

[0047] Fifth aspect, an embodiment of the present application further provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of a terminal device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the terminal device executes the content push method described in the first aspect above.

[0048] In an embodiment of the present application, by obtaining a feature set of a target object; processing the feature set by using a visit conversion prediction model to obtain a visit conversion prediction value of the target object; and then if it is determined according to the visit conversion prediction value that the target object is an object to be pushed, recommending content to the target object. It can be seen that the probability of predicting the visit and conversion of an object can be obtained based on multiple object features of the target object, and a visit conversion prediction value is obtained. The visit conversion prediction model not only considers visits and conversions, but also considers the behavior that the object will not visit immediately or convert immediately after receiving the recommended content, that is, it considers the behavior of delayed visits and delayed conversions of the object, so that the predicted visit conversion prediction value can be more accurate. Furthermore, screening the object according to the visit conversion prediction value can screen out the object with a greater probability of visiting and converting to push the recommended content, making the pushed object more accurate, thereby improving the click-through rate and conversion rate of the recommended content. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0050] Figure 1 is a schematic structural diagram of a content push system provided by an embodiment of the present application;

[0051] Figure 2 is a schematic flowchart of a content push method provided by an embodiment of the present application;

[0052] Figure 3 is a schematic structural diagram of a visit conversion prediction model provided by an embodiment of the present application;

[0053] Figure 4 is a schematic diagram of screening objects to be pushed provided by an embodiment of the present application;

[0054] Figure 5 is another schematic flowchart of a content push method provided by an embodiment of the present application;

[0055] Figure 6 It is a timing schematic diagram of a content push method provided by an embodiment of the present application;

[0056] Figure 7 It is a structural schematic diagram of a content push device provided by an embodiment of the present application;

[0057] Figure 8 It is a structural schematic diagram of an electronic device provided by an embodiment of the present application.

[0058] Specific implementation

[0059] Next, the technical solutions in the embodiments of the present application will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present application.

[0060] An embodiment of the present application proposes a content push solution, which can be applied to various push applications or push systems. The so-called push application or push system refers to an application or system with the function of pushing recommended content to an object. Among them, the object can be a user, and the user is the user who uses the above push application or push system; the object can also refer to the account, identifier, etc. used by the user in the above push application or push system. The present application does not make any limitations in this regard. For the convenience of description, the object is taken as an example of a user in the following description. Specifically, the push application or push system can determine the predicted value of the user's visit conversion based on the user's behavior characteristics, and then can screen out some users from all users as the users to be pushed, and then push the recommended content to the users to be pushed. It can be seen that the selected part of the users can be users with a relatively high probability of both visiting behavior (such as when the user receives the recommended content and clicks on the recommended content) and conversion behavior (such as when the user enters the page of the recommended content for transaction conversion), that is, screening out more accurate users to push the recommended content can have a better delivery effect, so as to achieve the purpose of improving the click-through rate and conversion rate of the recommended content.

[0061] The content push solution proposed by the embodiment of the present application involves technologies such as artificial intelligence and machine learning, among which:

[0062] Artificial Intelligence (AI) refers to the theory, methods, technologies, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to achieve the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines to enable machines to have the functions of perception, reasoning, and decision-making. Artificial intelligence technology is an interdisciplinary subject that involves a wide range of fields, including both hardware-level technologies and software-level technologies. The basic technologies of artificial intelligence generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operating (interaction) systems, and mechatronics. The software technologies of artificial intelligence mainly include several major directions such as computer vision technology, speech processing technology, natural language processing technology, and machine learning (deep learning).

[0063] Machine Learning (ML) is an interdisciplinary subject that involves multiple disciplines such as probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory. It specifically studies how computers simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize the existing knowledge structure to continuously improve their own performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent, and its applications cover all fields of artificial intelligence. Machine learning and deep learning (DL) usually include technologies such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and rote learning.

[0064] Based on the above description, please refer to Figure 1 , Figure 1 which is a schematic diagram of the architecture of a content push system provided by an embodiment of the present application. The content push system includes multiple terminal devices (such as the first terminal device 101, the second terminal device 102, and the third terminal device 103) and two electronic devices. Among them, the two electronic devices may include a content push device 104 and a model training device 105. The content push device 104 and the multiple terminal devices may be directly or indirectly connected by wired or wireless means, and the model training device 105 and the content push device 104 may be directly or indirectly connected by wired or wireless means. Optionally, the content push device 104 and the model training device 105 may be the same electronic device or two different electronic devices, and the present application does not make any limitations in this regard. It should be noted that Figure 1The number and form of the devices shown are for illustration purposes and do not constitute a limitation on the embodiments of the present application. In actual applications, the content push system may include more than three terminal devices and only include one electronic device. The embodiments of the present application will be described with three terminal devices (the first terminal device 101, the second terminal device 102, and the third terminal device 103) and one electronic device (i.e., the content push device 104 and the model training device 105 are the same electronic device).

[0065] As Figure 1 shown, the first terminal device 101, the second terminal device 102, and the third terminal device 103 may be terminal devices of three different users. The first terminal device 101, the second terminal device 102, and the third terminal device 103 may be used to receive the recommended content pushed. Furthermore, the user can click on the recommended content pushed through the corresponding terminal device to enter the page of the recommended content, that is, a visit occurs. A visit also refers to a visit behavior. Furthermore, after a visit occurs, a conversion can be made on the page of the pushed content through the terminal device, that is, a conversion occurs. Any one of the above-mentioned multiple terminal devices (such as the first terminal device 101, the second terminal device 102, and the third terminal device 103) may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc., but is not limited thereto.

[0066] The above-mentioned content push device 104 may be used to obtain the feature set of an object (such as a target user), and process the obtained feature set through the visit conversion prediction model to obtain the visit conversion prediction values of each user. Thus, the users to be pushed are determined according to the visit conversion prediction values, and recommended content is pushed to the users to be pushed. The above-mentioned model training device 105 may be used to construct and train the above-mentioned visit conversion prediction model. The above-mentioned content push device 104 and the above-mentioned model training device 105 may be a smart phone, a tablet computer, a notebook computer, a desktop computer, a smart speaker, a smart watch, etc.; the above-mentioned content push device 104 and the above-mentioned model training device 105 may also be a server. For example, it may be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It may also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (CDN), and big data and artificial intelligence platforms.

[0067] In a specific implementation, for example, in a shopping scenario, a target object (such as a target user) can install and run a shopping application with a push function on the first terminal device 101. Then, the target user can browse the content in the application, and the first terminal device 101 obtains a feature set by collecting the corresponding features of the target user. The content push device 104 can collect the feature set of the target user, process the feature set through a visit conversion prediction model to obtain a visit conversion prediction value of the target user. If the visit conversion prediction value of the target user meets the conditions of the user to be pushed, then the target user is determined as the user to be pushed, and further, the recommended content is pushed to the first terminal device 101 of the target user. Further, when the user receives the pushed recommended content through the first terminal device 101, the user can visit by clicking on the pushed content or the page of the recommended content, and then can complete a conversion on the page of the recommended content, resulting in a conversion.

[0068] In one implementation, both the feature set of the above-mentioned target object and the pushed recommended content can be stored in a blockchain, which can prevent the feature set of the target object and the pushed recommended content from being tampered with. Among them, the blockchain is a new application mode of computer technologies such as distributed data storage, peer-to-peer transmission, consensus mechanism, and encryption algorithms. In essence, it is a decentralized database, a string of data blocks generated by using cryptographic methods. Each data block contains information about a batch of network transactions, which is used to verify the validity of its information (anti-counterfeiting) and generate the next block.

[0069] Through the above content push system, a feature set of a target object is obtained; the feature set is processed by a visit conversion prediction model to obtain a visit conversion prediction value of the target object; and then, if the target object is determined as the object to be pushed according to the visit conversion prediction value, recommended content is pushed to the target object. It can be seen that the probability of predicting the visit and conversion of an object can be obtained based on multiple object features of the target object, and a visit conversion prediction value is obtained. The visit conversion prediction model not only considers the visit and conversion, but also considers the behavior that the object will not visit immediately or convert immediately after receiving the recommended content, that is, it considers the behavior of delayed visit and delayed conversion of the object, so that the predicted visit conversion prediction value can be more accurate. Then, the objects are screened according to the visit conversion prediction value, and the objects with a greater probability of visiting and converting can be screened out to push the recommended content, making the pushed objects more accurate, thereby improving the click-through rate and conversion rate of the recommended content.

[0070] It should be understood that the content push system described in the embodiments of the present application is for more clearly explaining the technical solutions of the embodiments of the present application, and does not constitute a limitation on the technical solutions provided by the embodiments of the present application. Those of ordinary skill in the art know that with the evolution of the system architecture and the emergence of new business scenarios, the technical solutions provided by the embodiments of the present application are equally applicable to similar technical problems.

[0071] Based on the above content push scheme and content push system, an embodiment of the present application provides a content push method. The content push method described above in the embodiments of the present application can be executed by an electronic device, and the electronic device can be Figure 1 the content push device 104 in the content push system shown. When the content push device 104 is a server, it can be a dedicated server or some Internet application servers. Through the Internet application servers, not only can the relevant steps of the embodiments of the present application be executed, but other services can also be provided. Please refer to Figure 2 , Figure 2 which is a schematic flowchart of a content push method provided by an embodiment of the present application. The content push method includes the following steps 201 to step 203:

[0072] Step 201, obtain the feature set of the target object.

[0073] Specifically, taking the target object as the target user as an example, the target user is any user who uses the above push application or push system. The feature set may include the portrait features of the user, the behavior features of the user, the item features of the user, and the scenario features. Among them, the portrait features of the user may be features such as the age, gender, age, and region where the user is located; the behavior features of the user may be the interaction behaviors of the user with the in-site content. For example, in a shopping scenario, the user's behaviors such as browsing products, collecting products, adding products to the shopping cart, sharing, and repurchasing; the item features of the user may be the item features of the user's interaction with the in-site content. For example, in the above shopping scenario, the product features browsed by the user, the product features collected by the user, the product features added to the shopping cart, the product features shared, and the product features repurchased, etc.; the scenario features can be understood as the features of the current scenario where the user is located. For example, in a shopping scenario, the scenario features may be the current popular trends. For another example, in the scenario of purchasing financial products, the scenario features may be the market conditions of the overall market, etc.

[0074] In a possible implementation, the target user is a user who can receive the pushed recommended content and is relevant to the currently pushed content. Exemplarily, the user can browse content through an application installed on the terminal device. When there is a visit, if the application has the function of purchasing goods, the user can also purchase goods through the application, resulting in a conversion. For the convenience of description, the following takes a shopping scenario as an example. The target user can receive the pushed recommended information through the application. For example, the target user has enabled the function of receiving push notifications of the application. Moreover, before pushing the recommended content, the business party, i.e., the publisher of the recommended content, can determine the recommended content and the whitelisted users who can be pushed. Among them, the users in the pushable whitelist can be users whose feature set includes one or several features, and the included features can be related to the recommended content that the business party wants to push. The content push device can screen out the target users from all users who use the application, i.e., all users, as the pushable users, and obtain a set of pushable users.

[0075] In the above process, the feature set of the target object (such as the target user) can characterize the preferences and interest characteristics of the target object. By obtaining the feature set of the target object, it provides effective and accurate data support for predicting the visit behavior and conversion behavior of the target object in the future.

[0076] Step 202: Use the visit conversion prediction model to process the above feature set to obtain the visit conversion prediction value of the above target object.

[0077] In a possible implementation, the visit conversion prediction model can be used to process the feature set of the target object to obtain the visit conversion prediction value of the target object. The above visit conversion prediction model can be a model trained based on historical push results and can be used to predict the visit and conversion prediction values of the target object after receiving the pushed recommended content. Among them, the visit conversion prediction value can be used to represent the probability that the target object (such as the target user) will be converted after visiting after receiving the pushed recommended content. Exemplarily, after receiving the pushed recommended content, the user can click on the recommended content to enter the page of the recommended content for browsing. This behavior of entering the page of the recommended content is considered a visit. The user can also enter the page of the recommended content through other means to have a visit behavior, which is not limited in this application. Further, if the user purchases goods on the page of the recommended content, it means that the user has been converted.

[0078] In scenarios such as shopping and advertisement pushing, after receiving the pushed recommended content, users may not click on the recommended content immediately, that is, there will be no visit at that moment; or after receiving the recommended content, users may click on the pushed content immediately to view it, but it is not certain that a conversion behavior will occur immediately after viewing; or a visit and conversion may occur some time after the recommended content is pushed. That is, there may be delayed visits and delayed conversions for users. In this application, in order to consider the delayed feedback of users, that is, the delayed visits and delayed conversions of users, when training the visit conversion prediction model, the training set may include a conversion sample set and a visit sample set. Among them, the conversion sample set may include an unconverted sample subset and a converted sample subset, and the visit sample set includes a non-visited sample subset and a visited sample subset.

[0079] For a visit and a conversion to occur, the samples of visits that occur during the period from when the user receives the pushed recommended content to when the model is trained (i.e., the waiting window) can be used as samples in the visited sample subset, and the samples of conversions that occur during this period can be used as samples in the converted sample subset. Moreover, the visit label of the samples in the visited sample subset also includes the visit time, and the visit label of the samples in the converted sample subset also includes the conversion time. For no visit and no conversion, the samples of no visit and no conversion that occur during the period from when the user receives the pushed recommended content to when the model is trained can be divided into two cases: one case is that the user is not interested in the recommended content and will not visit ultimately; the other case is that the user is interested in the recommended content, but the visit time is greater than the time of the waiting window, that is, no visit and no conversion occur when training this visit conversion prediction model. Thus, a visit conversion prediction model can be constructed more accurately, and samples can be characterized more accurately in the full sample space.

[0080] Among them, in order to consider both visits and conversions simultaneously, the visit conversion prediction model can be constructed as an Entire Space Delayed Feedback Model (ESDFM). A multi-task model refers to a model that can include at least two objectives, each corresponding to a sub-network in the multi-task model. Each sub-network processes the input data respectively and outputs the processing results of the learning tasks of each objective. Furthermore, the multi-task model can perform further processing based on the processing results of each sub-network to obtain the output result of the multi-task model. In the embodiments of the present application, the visit conversion prediction model can include two objectives. One objective can be the occurrence of a visit, and the other objective can be the occurrence of a conversion. A multi-objective delayed feedback model including two objectives is constructed, so that the link for user conversion is split into multiple objectives, and the possibility that the user conversion link is longer in more scenarios can be considered. In the present application, the multi-objective delayed feedback model can also use the model structures of other multi-objective learning models such as the Modeling Task Relationships in Multi-task Learning with Multi-gate Mixture-of-Experts (MMOE) and the Share bottom network. The multi-objective delayed feedback model can also add other objectives. For example, other situations of delayed conversion can be considered. Exemplarily, a delayed feedback task from user visit to user addition of self-selection can be added, so as to be able to screen out accurate users to be pushed. The present application is described by taking ESDFM as an example, and the structure of the multi-objective delayed feedback model is not limited in the present application.

[0081] Specifically, please refer to Figure 3 , Figure 3 which is a schematic structural diagram of a visit conversion prediction model provided by an embodiment of the present application. As Figure 3 shown, the visit conversion prediction model includes a visit prediction module and a conversion prediction module. The visit prediction module is used to predict whether an object (such as a user) will finally have a visit behavior. Further, the visit prediction module also needs to learn the distribution parameters of delayed visits to estimate the probability that a user will have a visit behavior at a certain future visit time (such as the nth day). Similarly, the conversion prediction module is used to predict whether an object (such as a user) will finally have a conversion behavior. Further, the visit prediction module also needs to learn the distribution parameters of delayed conversions to estimate the probability that a user will have a conversion behavior at a certain future conversion time (such as the mth day).

[0082] In a possible implementation manner, as Figure 3As shown, the user icon in the lower left corner and the data beside the user icon are represented as the input of the model, that is, the feature set of the target object (such as the target user). Then, the feature set is input into the embedding layer. Through the embedding layer, high-dimensional sparse features can be mapped to a low-dimensional semantic space. It should be noted that the portrait features, behavioral features, and scenario features of the user included in the feature set of the target object may all be datasets of several thousand dimensions, and the item features of the user may be datasets of several hundred dimensions. For the convenience of calculation and extraction of effective information, the data can be dimensionally reduced, that is, the feature set is input into the embedding layer (such as Figure 3 the squares in the dashed box shown) for dimensionality reduction processing to obtain the dimensionally reduced feature set. For example, the dataset can be dimensionally reduced to several hundred dimensions before processing. Optionally, the visit prediction module and the conversion prediction module can share the embedding layer, that is, the dimensionally reduced feature set is input into the multi-layer neural network in the visit prediction module for processing, and the dimensionally reduced feature set is input into the multi-layer neural network in the conversion prediction module for processing. Optionally, the visit prediction module and the conversion prediction module can have a sequential processing order or can be processed in parallel. This application does not make a limitation on this.

[0083] In a possible implementation manner, in the delayed feedback scenario, an exponential function can be selected as the parameter of the delayed visit distribution and the parameter of the delayed conversion distribution. This is because this function must always be positive, and generally, as the time when the pushed recommendation information reaches the user increases, for all the users who have been pushed, the number of visits and the number of conversions show an exponential decay trend.

[0084] Specifically, the distribution parameter of the delayed visit can be as shown in Formula 1:

[0085]

[0086] where λ v (x) represents the parameter estimation of the corresponding delayed visit of the object (such as the user), and this parameter estimation follows an exponential distribution. w vd is the model parameter related to the visit prediction module, x represents the feature set, and w vd ×x represents the parameter estimation of the exponential distribution.

[0087] In a possible implementation manner, for the prediction of the delayed visit, through the visit prediction module, it processes according to the feature set of the object (such as the user) to obtain the visit prediction value predicted by the visit prediction module, which is represented by pctr(x i ), and this pctr(x i) The value can be used to represent the probability that the user will eventually visit. It should be noted that this application is described with a day as the time unit, and other time units can also be selected. This application does not make any limitations in this regard. For the sake of description convenience, in the following text, the time of pushing the recommended content is taken as the first day, and the object is the user for description.

[0088] Further, under the condition that the user eventually visits, the probability that the visit time occurs on the vd i th day is shown in Formula 2,

[0089]

[0090] where Pr represents the conditional probability, VD represents the time interval between the occurrence of the marked event (such as pushing the recommended content) and the user's visit. Exemplarily, if the recommended content is pushed on the first day and the user visits on the vd i th day, then VD = vd i . V represents whether the user will eventually visit, V ∈ {0, 1}. If the user will eventually visit, then V = 1; conversely, V = 0 indicates that the user will not visit eventually. X represents the feature set, and x i represents the feature set of user i. Then the formula Pr(VD = d i |V = 1, X = x i ) represents the probability of VD = vd i under the condition of V = 1 and X = x i , that is, under the condition that user i will eventually visit and the feature set of user i is x i , the probability that user i visits on the vd i th day.

[0091] Therefore, the probability that the user's visit time is on the vd i th day can be shown in Formula 3:

[0092]

[0093] where Y V represents the current observed label, that is, whether the user has visited. Y C ∈ {0, 1}. If the user has visited, then Y V = 1; conversely, Y V = 0. VE represents the time when the marked time has passed. For example, if the number of days since the recommended content was pushed is ve i , then VE = ve iNote that the concepts of VD and VE need to be clearly distinguished. For example, assume that recommended content is pushed on the first day and a visit occurs on the 5th day, then VD = 5; VE can be used to represent the time for training the model, which has nothing to do with whether a user visits. If the above-mentioned visit conversion prediction model is trained on the 6th day, then VE = 6. Among them, since the elapsed time VE has nothing to do with whether a visit will ultimately occur, it can be removed during simplification. That is, the probability that a user visits on the vd i th day is the product of the probability that the visit time occurs on the vd i th day under the condition that the user ultimately visits and the probability pctr(x i ) indicating that the user will ultimately perform a visit behavior.

[0094] Furthermore, the probability that the visit prediction model predicts that a user will visit within n days can be as shown in Equation 4:

[0095] As shown in Equation 4:

[0096]

[0097] Among them, the purpose of the integration is to find the cumulative probability, that is, the probability of a visit within n days, which can be the sum of the probabilities that a user visits each day from the time of pushing the recommended content to the nth day.

[0098] Similarly, the distribution parameter of the delayed conversion can be as shown in Equation 5:

[0099]

[0100] Among them, λ(x) represents the parameter estimation of the user's corresponding delayed conversion, and this parameter estimation also follows an exponential distribution. w d is a model parameter related to the conversion prediction module, x represents the feature set, and w d ×x represents the parameter estimation of the exponential distribution.

[0101] Furthermore, the probability that the conversion time occurs on the d i th day under the condition that the user ultimately converts is as shown in Equation 6,

[0102]

[0103] Among them, D represents the time interval between the occurrence of a marked event (such as pushing recommended content) and the user's conversion. For example, if recommended content is pushed on the first day and the user converts on the d i th day, then D = d i . C represents whether the user will ultimately convert, C ∈ {0, 1}. If the user will ultimately convert, then C = 1; otherwise, C = 0, indicating that the user will not ultimately convert. X represents the feature set, xi represents the feature set of user i. Then the expression Pr(D = d i |C = 1, X = x i ) represents the probability of D = d occurring i under the condition that C = 1 and X = x i , that is, the probability of D = d occurring i when user i will eventually convert and the feature set of user i is x i , which is the probability of user i converting on the d

[0104] For the prediction of delayed conversion, through the conversion prediction module, it processes according to the feature set of the user to obtain the conversion prediction value predicted by the conversion prediction module, denoted by pcvr(x i ). The value of pcvr(x i ) can be used to represent the probability that the user will eventually convert after visiting.

[0105] After obtaining the visit prediction value pctr(x i ) and the conversion prediction value pcvr(x i ), the joint probability of the user converting after visiting can be obtained, that is, the visit conversion prediction value of the user. Among them, the joint probability of the user converting after visiting can be as shown in Formula 7:

[0106] pctcvr(x i ) = pctr(x i ) × pcvr(x i ) Formula 7

[0107] Among them, pctcvr(x i ) is the joint probability of the user converting after visiting, which is the product of the visit prediction value pctr(x i ) and the conversion prediction value pcvr(x i ).

[0108] Furthermore, the probability that the conversion prediction model predicts that the user will visit and convert within n days can be as shown in Formula 8:

[0109]

[0110] Similarly, the purpose of the integral is to find the cumulative probability, that is, the probability of visiting and converting within n days, which can be the sum of the probabilities of the user converting after visiting each day from the time of pushing the recommended content to the nth day.

[0111] Optionally, the visit conversion prediction model may determine the prediction score of a user based on the probability that the user visits and converts, that is, the visit conversion prediction value, or directly use the visit conversion prediction value as the prediction score. It can be understood that the visit conversion prediction value of a user is proportional to the prediction score of the user. Furthermore, the users to be pushed can be screened according to the prediction scores of each user, and recommended content can be pushed to them.

[0112] Step 203: If it is determined according to the above visit conversion prediction value that the above target object is an object to be pushed, then push recommended content to the above target object.

[0113] In a possible implementation, if an object (such as a user) receives the pushed recommended content through an application installed in a terminal device, the pushed recommended content may be the push information displayed in the notification bar interface after the pull-down status bar, or may be highlighted on the icon of the application, for example, a red dot is displayed on the icon of the application.

[0114] In another possible implementation, if an object (such as a user) receives the pushed recommended content through a small program integrated in an application, then the pushed recommended content may be displayed in the application, or may be highlighted on the icon of the small program in the application. Exemplarily, for example, if the shopping small program is integrated in an instant messaging application, then the pushed recommended content may be displayed in the push content of the instant messaging application, and may be highlighted in red on the icon displaying the shopping small program.

[0115] In a possible implementation, please refer to Figure 4 , Figure 4 which is a schematic diagram of screening objects to be pushed provided by an embodiment of the present application. As Figure 4 shown, taking all objects using the application as the full amount of objects as an example, the full amount of objects is the largest light-colored ellipse, and the objects that can be pushed are the objects selected from the full amount of objects that can receive the pushed recommended content and are related to the current pushed content. As Figure 4 shown, the objects that can be pushed are the smaller ellipses in the full amount of objects. Furthermore, the objects to be pushed are screened according to the above visit conversion prediction value, and recommended content is pushed to the screened objects to be pushed. Among them, the objects to be pushed may be high-potential objects screened before each delivery period. As Figure 4 shown, the objects to be pushed are the dark-filled ellipses in the ellipse of the objects that can be pushed. It can be understood that the sizes of the three ellipses in Figure 4 can be used to represent the number of objects respectively.

[0116] In a possible implementation, the content push device respectively obtains the visit conversion prediction values of each object in the set of pushable objects. The set of pushable objects may include multiple pushable objects, and the target object is an object in the set of pushable objects. The content push device may sort the visit conversion prediction values of each object in the set of pushable objects in descending order according to the visit conversion prediction values, and screen out the top M pushable objects in the sorting result as the objects to be pushed. For example, if the visit conversion prediction value of the target object is among the top M in the sorting result, the content push device may determine that the target object is an object to be pushed. Here, M is an integer greater than 1.

[0117] Optionally, the content push device may use, as the objects to be pushed, the objects corresponding to the visit conversion prediction values greater than a preset threshold among the visit conversion prediction values of each object in the recommendable set. Exemplarily, if the visit conversion prediction value of the target object is greater than the preset threshold, the target object may be used as a pushable object; otherwise, the target object is used as a non-pushable object.

[0118] In another possible implementation, the content push device respectively obtains the prediction scores corresponding to the visit conversion prediction values of each object in the set of pushable objects. Sort the prediction scores of each object in the set of pushable objects from largest to smallest, and screen out the top M pushable objects in the sorting result as the objects to be pushed. For example, if the visit conversion prediction value of the target object is among the top M in the sorting result, the content push device may determine that the target object is an object to be pushed. Here, M is an integer greater than 1.

[0119] Optionally, the content push device may screen out the prediction scores greater than a preset threshold among the prediction scores of each object in the set of pushable objects, and use the objects corresponding to the screened prediction scores greater than the preset threshold as the objects to be pushed.

[0120] Optionally, the content push device may also push the recommended content to the terminal devices of each object to be pushed by sending text messages.

[0121] Exemplarily, in a financial shopping scenario, for the objects to be pushed screened by the content push device using different prediction models, taking the object as a user as an example, the content push device respectively pushes the recommended content to the screened users to be pushed, and the push methods may be the way of pushing recommended content through the application and the way of sending text messages. The push effects of pushing content to new users can be as shown in Table 1:

[0122] Table 1

[0123] Model Push of application SMS push Deep neural network 5.42 1.33 Extreme gradient boosting tree 5.29 1.37 Delayed feedback model 5.41 1.36 ESDFM 5.69 1.38

[0124] Among them, new users can represent users in the feature set who do not have corresponding behavioral features. The numbers shown in Table 1 are the number of conversions after pushing recommended content 1,000 times. As shown in Table 1, when using the Deep Neural Networks (DNN) model to push 1,000 times through the application push method, the number of conversions is 5.42 people, and when pushing 1,000 times through the SMS push method, the number of conversions is 1.33 people; similarly, when using the eXtreme Gradient Boosting (XGBoost) model to push 1,000 times through the application push method, the number of conversions is 5.29 people, and when pushing 1,000 times through the SMS push method, the number of conversions is 1.37 people; the Delayed Feedback Model (DFM) has 5.41 people converted when pushing 1,000 times through the application push method, and 1.36 people converted when pushing 1,000 times through the SMS push method; and the multi-objective delayed feedback model ESDFM used in this application has 5.69 people converted when pushing 1,000 times through the application push method, and 1.38 people converted when pushing 1,000 times through the SMS push method. It can be seen that the effect of the multi-objective delayed feedback model used in this application is better than that of other several models.

[0125] In another exemplary case, in the financial shopping scenario, for the objects to be pushed screened by the content push device using different prediction models, taking the object as a user as an example, the content push device pushes recommended content to the screened users to be pushed respectively, and the push method can be the method of pushing recommended content through the application and the method of sending SMS. The push effects of pushing content to old users can be shown in Table 2:

[0126] Table 2

[0127] Model Push of application SMS push DNN 16.873 0.9579 XGBoost 16.852 0.8822 DFM 16.857 0.9616 Full-space multi-objective learning model 16.929 0.9571 ESDFM 16.954 0.9651

[0128] Among them, old users can represent users in the feature set who include corresponding behavioral features. The numbers shown in Table 2 are the number of people with a conversion amount greater than 1000 yuan after pushing the recommended content 1000 times. As shown in Table 2, when using the DNN model to push 1000 times through the application push method, the number of conversions with a conversion amount greater than 1000 yuan is 16.873 people, and when pushing 1000 times through the SMS push method, the number of conversions with a conversion amount greater than 1000 yuan is 0.9579 people; when using the XGBoost model to push 1000 times through the application push method, the number of conversions with a conversion amount greater than 1000 yuan is 16.852 people, and when pushing 1000 times through the SMS push method, the number of conversions with a conversion amount greater than 1000 yuan is 0.8822 people; when using DFM to push 1000 times through the application push method, the number of conversions with a conversion amount greater than 1000 yuan is 16.857 people, and when pushing 1000 times through the SMS push method, the number of conversions with a conversion amount greater than 1000 yuan is 0.9616 people; when using the Entire Space Multi-Task Model (ESMM) to push 1000 times through the application push method, the number of conversions with a conversion amount greater than 1000 yuan is 16.929 people, and when estimating to push 1000 times through the SMS push method, the number of conversions with a conversion amount greater than 1000 yuan is 0.9571 people; and when using the ESDFM used in this application to push 1000 times through the application push method, the number of conversions with a conversion amount greater than 1000 yuan is 16.954 people, and when pushing 1000 times through the SMS push method, the number of conversions is 0.9651 people. It can be seen that the ESDFM used in this application not only performs well in screening out users to be pushed (i.e., high-potential) among new users (total users) for pushing, but also outperforms other several models in screening out users to be pushed (i.e., high-potential) among old users.

[0129] In an embodiment of the present application, by obtaining a feature set of a target object; processing the feature set using a visit conversion prediction model to obtain a visit conversion prediction value of the target object; and then, if it is determined that the target object is a target to be pushed according to the visit conversion prediction value, pushing recommended content to the target object. It can be seen that the probability of predicting the visit and conversion of an object can be obtained based on multiple object features of the target object, and a visit conversion prediction value is obtained. This visit conversion prediction model not only considers visits and conversions, but also considers the behavior that the object will not visit immediately or convert immediately after receiving the recommended content, that is, it considers the behavior of delayed visits and delayed conversions of the object, so that the predicted visit conversion prediction value can be more accurate. Then, screening the objects according to the visit conversion prediction value can screen out the objects with a greater probability of visiting and converting and push the recommended content, making the pushed objects more accurate, thereby improving the click-through rate and conversion rate of the recommended content.

[0130] Please refer to Figure 5 , Figure 5 which is another flowchart of a content push method provided by an embodiment of the present application. This content push method can be executed by an electronic device, which can be Figure 1 the content push device 104 in the content push system shown in Figure 1 or the model training device 105 in the content push system shown in. Among them, the content push device 104 and the model training device 105 can be the same electronic device. This content push method includes steps 501-step 503:

[0131] 501. Obtain a conversion sample set and a visit sample set.

[0132] In a possible implementation manner, taking the content push device and the model training device as the execution subject of the same electronic device for explanation, the above content push device can construct an initial visit conversion prediction model and obtain a training sample set. Then, the initial visit conversion prediction model is trained by a supervised training method. Among them, the training sample set includes a conversion sample set and a visit sample set. The above conversion sample set includes an un-converted sample subset and a converted sample subset, and the above visit sample set includes a non-visited sample subset and a visited sample subset.

[0133] Specifically, each conversion sample in the conversion sample subset is a sample in which a first reference user has a conversion, which can be understood as a positive sample for training the visit conversion prediction model. The conversion sample subset includes a feature set of a first reference object and a conversion label. The first reference object is the object corresponding to a certain conversion sample in the conversion sample set. The conversion label is used to indicate whether the first reference object has a conversion. If there is a conversion, the conversion label is also used to indicate the time when the first reference object has a conversion.

[0134] Among them, each untransformed sample in the untransformed sample subset included in the above transformation sample set is a sample where each first reference object has not undergone transformation. The transformation label in each untransformed sample is used to indicate that the object corresponding to each untransformed sample has not undergone transformation, but it cannot be understood as a negative sample for training the visit conversion prediction model. This is because for the object of the sample that has not undergone transformation, there are two possible situations: (1) The first reference object is not interested in the recommended content, that is, no transformation will occur no matter how long it waits; (2) The first reference object has not undergone transformation currently, but the object will eventually undergo transformation, and the time of the object's transformation is after the current training time. For example, on the first day, recommended content is pushed to the first reference object, and on the 5th day, an initial visit conversion prediction model is constructed and trained on this initial visit conversion prediction model, but the first reference object undergoes transformation on the 10th day.

[0135] Similarly, each visit sample in the visit sample set includes a feature set of a second reference object and a visit label. The visit sample subset includes the feature set of the second reference object and the visit label, and the second reference object is the object corresponding to a certain visit sample in the visit sample set. The visit label is used to indicate whether the second reference object has visited, and when it is used to indicate a visit, the visit label is also used to indicate the time when the second reference object has visited.

[0136] Each unvisited sample in the unvisited sample subset included in the above visit sample set is a sample where the second reference object has not visited. The visit label in each unvisited sample is used to indicate that the second reference object has not visited, but it also cannot be understood as a negative sample for training the visit conversion prediction model.

[0137] It should be noted that the samples in the above visit sample set may be the same as the samples in the above transformation sample set, or may be different from the samples in the above visit sample set. In a possible implementation, the above visit sample set may include the samples in the above transformation sample set.

[0138] 502. Use the above transformation sample set and the above visit sample set to train the initial visit conversion prediction model to determine the target loss information.

[0139] In a possible implementation, use the transformation sample set to train the initial visit conversion prediction model to determine the first loss information; use the above visit sample set to train the initial visit conversion prediction model to determine the second loss information; determine the target loss information according to the above first loss information and the above second loss information. Among them, the visit conversion prediction model can be trained using the visit sample set to obtain the first loss information; the visit conversion prediction model can be trained using the transformation sample set to obtain the second loss information. Furthermore, the sum of the first loss information and the second loss information can be used as the target loss information of the visit conversion prediction model.

[0140] Exemplarily, for the first visited sample (i.e., the sample belonging to the visited sample subset), the visit probability that the second reference object (such as a user) visits on the vd i th day can be referred to the above formula 3; for the second visited sample (i.e., the sample belonging to the unvisited sample subset), the reference probability that the second reference object does not visit can be as shown in formula 9:

[0141]

[0142] where, pr(Y v =0|V = 0,X = x i ,E = e i )Pr(V = 0|X = x i ) represents the probability that the object will not visit, and Pr(Y v =0|V = 1,X = x i ,E = e i )Pr(V = 1|X = x i ) represents the probability that the conversion time of the object is after the current training time (i.e., VE = ve i ). The conversion prediction value pctr(x i ) represents the probability that the visit will finally occur. Therefore, 1 - pctr(x i ) represents the probability that the object will not be converted.

[0143] Furthermore, the second loss information of the visit prediction module can be determined according to the visit probabilities of each first visited sample and the above reference probabilities of each second visited sample. Among them, the second loss information can include the loss function of the visit prediction module, and the loss function of the visit prediction module can be as shown in formula 10:

[0144]

[0145] where, w v represents the model parameter related to visit prediction, w vd represents the model parameter related to the delayed visit time, and L(w v ,w vd ) represents the loss function of the visit prediction module. Specifically, the loss function of the visit prediction module can be obtained by maximum likelihood estimation. By taking the negative logarithm of the likelihood function in the form of a product, the optimization objective of maximizing the likelihood function for visit prediction (i.e., formula 3 and formula 9) is converted into minimizing the negative logarithm likelihood function, that is, the above formula 10 is obtained.

[0146] In another example, for the first transformed sample (i.e., the sample belonging to the transformed sample subset), the transformation probability that the first transformed sample is transformed at the corresponding transformation time of the first transformed sample can be determined, that is, pctcvr(x i ). The derivation of pctcvr(x i ) can refer to the above formula 7. For the second transformed sample (i.e., the sample belonging to the untransformed sample subset), the transformation reference probability that the second transformed sample is not transformed can include two parts. One part is the probability that the second transformed sample will not be transformed, and the second part is the probability that the second transformed sample is transformed after the current training time. Among them, the probability that the second transformed sample is not transformed at the current training time can be as shown in formula 11:

[0147]

[0148] Among them, Pr(Y = 0|C = 0,X = x i ,E = e i )Pr(C = 0|X = x i ) represents the probability that the object will not be transformed, and Pr(Y = 0|C = 1,X = x i ,E = e i )Pr(C = 1|X = x i ) represents the probability that the transformation time of the object is after the current training time (i.e., E = e i ). The transformation prediction value pcvr(x i ) represents the probability that the object will ultimately be transformed. Therefore, 1 - pctcvr(x i ) represents the probability that the object will not be transformed.

[0149] Since the probability that the transformation time occurs on the d i th day under the condition that the object is ultimately transformed can be as shown in formula 6, the probability that the object is transformed on the d i th day can be as shown in formula 12:

[0150]

[0151] Furthermore, the first loss information is determined according to the transformation probability of each first transformed sample and the transformation reference probability of each second transformed sample. Among them, the first loss information can include the loss function of the transformation prediction module, and the loss function of the transformation prediction module can be as shown in formula 13:

[0152]

[0153] Among them, w c represents the weight related to the visit transformation prediction value pctcvr(x i)The related model parameters, w d denote the model parameters related to the conversion prediction module. Specifically, the loss function of this conversion prediction module can also be obtained through maximum likelihood estimation. By taking the negative logarithm of the likelihood function in the form of a continuous product, the optimization method of maximizing the likelihood function for post-conversion prediction of the target (i.e., the above formulas 11 and 12) is converted into minimizing the negative logarithm likelihood function, that is, the above formula 13 is obtained.

[0154] Furthermore, the target loss information can be determined according to the first loss information and the above second loss information. The target loss information may include the loss function of this visit conversion prediction model (i.e., ESDFM). The loss function of this ESDFM can be as shown in formula 14:

[0155] L = L(w v , w vd ) + l(w c , w d ) Formula 14

[0156] where L represents the loss function of the visit conversion prediction model (i.e., ESDFM). The loss function of this ESDFM can be the sum of the loss function of the above visit prediction module and the loss function of the conversion prediction module.

[0157] It should be noted that in the visit conversion prediction model of this application, the conversion prediction value pcvr(x i ) after the object's visit is used as a latent variable for learning. Among them, the so-called latent variable is a variable that cannot be directly observed but has an impact on the system and the observable output. Using the conversion prediction value pcvr(x i ) after the object's visit as a latent variable for learning may simplify the observed variables and make the modeling problem simpler. Moreover, in the training sample set of the visit conversion prediction model, the currently un-converted samples are reasonably distinguished, avoiding confusing the samples with the intention to convert and the samples that do not convert, so as to depict the samples more precisely.

[0158] 503. When it is determined that the training stop condition is not satisfied according to the above target loss information, adjust the model parameters of the above initial visit conversion prediction model according to the above target loss information, and use the above conversion sample set and the above visit sample set to train the adjusted visit conversion prediction model to obtain the trained visit conversion prediction model.

[0159] In a possible implementation, when it is determined that the target loss information does not meet the training end condition, where the training end condition may be that the value of the loss function in the target loss information is within a preset threshold range, or the training end condition may also be that among the values of the loss function obtained through consecutive N iterative trainings, the number of times the difference between the values of the loss function obtained in two adjacent times is less than the preset difference threshold is greater than or equal to the preset number threshold. This application does not make a limitation, and it can be specifically determined according to the usage scenario.

[0160] Optionally, the training end condition may also be that the number of iterative trainings reaches the preset number threshold. And further, by using the above-mentioned transformed sample set and the above-mentioned visiting sample set, the visiting conversion prediction model after adjusting the model parameters is trained again until the training end condition is met, and the trained visiting conversion prediction model is obtained, that is, the above-mentioned ESDFM is obtained. Among them, in the process of each iterative training, the initial visiting conversion model may refer to the visiting conversion model after adjusting the model parameters after the previous training.

[0161] Optionally, the model parameters of the initial visiting conversion prediction model can be adjusted by the gradient descent method. When using the gradient descent method to update the model parameters, the gradient of the loss function is calculated, and the model parameters are iteratively updated according to this gradient, so as to gradually converge the initial visiting conversion prediction model to improve the prediction accuracy of the model.

[0162] Please refer to Figure 6 , Figure 6 which is a timing diagram of a content push method provided by an embodiment of this application. As Figure 6 shown, the timing of this content push method is as follows: the content push device collects the transformed sample set and the visiting sample set in the historical push; then the content push device trains the initial visiting conversion prediction model according to the transformed sample set and the visiting sample set, and then the content push device predicts the pushable object to obtain the users to be recommended, and pushes the recommended content to the objects to be pushed.

[0163] In the embodiments of the present application, by obtaining a feature set of a target object; using a visit conversion prediction model to process the feature set to obtain a visit conversion prediction value of the target object; and then, if it is determined according to the visit conversion prediction value that the target object is an object to be pushed, pushing recommended content to the target object. It can be seen that the probability of predicting the visit and conversion of an object can be obtained based on multiple object features of the target object, and a visit conversion prediction value is obtained. This visit conversion prediction model not only considers visits and conversions, but also considers the behavior that the object will not visit immediately or convert immediately after receiving the recommended content, that is, it considers the behavior of delayed visits and delayed conversions of the object, so that the predicted visit conversion prediction value can be more accurate. Furthermore, by screening objects according to the visit conversion prediction value, objects with a greater probability of visiting and converting can be screened out to push recommended content, making the pushed objects more accurate, thereby improving the click-through rate and conversion rate of the recommended content.

[0164] Please refer to Figure 7 , Figure 7 FIG. is a schematic structural diagram of a content push device according to an embodiment of the present application. The content push device 70 according to the embodiment of the present application can be set on an electronic device, and the electronic device can be the above-mentioned Figure 1 content push device. The above-mentioned content push device 70 includes the following units:

[0165] An acquisition unit 701, configured to acquire a feature set of a target object;

[0166] A processing unit 702, configured to use a visit conversion prediction model to process the above-mentioned feature set to obtain a visit conversion prediction value of the above-mentioned target object;

[0167] A push unit 703, configured to, if it is determined according to the above-mentioned visit conversion prediction value that the above-mentioned target object is an object to be pushed, push recommended content to the above-mentioned target object;

[0168] Among them, the above-mentioned visit conversion prediction model is trained using a conversion sample set and a visit sample set. The above-mentioned conversion sample set includes an un-converted sample subset and a converted sample subset, and the above-mentioned visit sample set includes a non-visited sample subset and a visited sample subset.

[0169] In one implementation, the above-mentioned visit conversion prediction model includes a visit prediction module and a conversion prediction module. The above-mentioned processing unit 702 is specifically configured to:

[0170] Use the above-mentioned visit prediction module to process the above-mentioned feature set to obtain a visit prediction value of the above-mentioned target object;

[0171] Use the above-mentioned conversion prediction module to process the above-mentioned feature set to obtain a conversion prediction value of the above-mentioned target object;

[0172] Based on the above visit prediction value and the above conversion prediction value, determine the visit conversion prediction value of the above target object.

[0173] In one implementation, the above device 70 further includes:

[0174] The above acquisition unit 701 is further configured to acquire the above conversion sample set and the above visit sample set. Each conversion sample in the above conversion sample set includes a feature set of a first reference object and a conversion label, and each visit sample in the above visit sample set includes a feature set of a second reference object and a visit label;

[0175] The determination unit 704 is configured to train an initial visit conversion prediction model by using the above conversion sample set and the above visit sample set, and determine target loss information;

[0176] The adjustment unit 705 is configured to, when it is determined according to the above target loss information that the training stop condition is not satisfied, adjust the model parameters of the above initial visit conversion prediction model according to the above target loss information, and train the adjusted visit conversion prediction model by using the above conversion sample set and the above visit sample set to obtain a trained visit conversion prediction model.

[0177] In one implementation, the above determination unit 704 is specifically configured to:

[0178] Train the initial visit conversion prediction model by using the above conversion sample set, and determine first loss information;

[0179] Train the above initial visit conversion prediction model by using the above visit sample set, and determine second loss information;

[0180] Determine the target loss information according to the above first loss information and the above second loss information.

[0181] In one implementation, if the conversion sample belongs to the converted sample subset, the conversion label of the above conversion sample includes the conversion time. The above determination unit 704 is specifically configured to:

[0182] For a first conversion sample, determine the conversion probability that the above first conversion sample is converted at the conversion time corresponding to the above first conversion sample. The above first conversion sample is any conversion sample in the above converted sample subset;

[0183] For a second conversion sample, determine the conversion reference probability that the above second conversion sample is not converted at the current training time. The above second conversion sample is any conversion sample in the above unconverted sample subset;

[0184] Determine the first loss information according to the conversion probabilities of the respective first conversion samples and the reference probabilities of the respective second conversion samples.

[0185] In one implementation, the determining unit 704 is specifically configured to:

[0186] Determine the probability that the above-mentioned second conversion sample will not undergo conversion, and determine the probability that the above-mentioned second conversion sample will undergo conversion after the current training time;

[0187] According to the probability that the above-mentioned second conversion sample will not undergo conversion and the probability that the above-mentioned second conversion sample will undergo conversion after the current training time, determine the reference probability that the above-mentioned second conversion sample did not undergo conversion at the above-mentioned current training time.

[0188] In one implementation, if the visiting sample belongs to the subset of visited samples, the visiting label of the above-mentioned visiting sample includes the visiting time, and the determining unit 704 is specifically configured to:

[0189] For the first visiting sample, determine the visiting probability that the above-mentioned first visiting sample visits at the visiting time corresponding to the above-mentioned first visiting sample, where the first visiting sample is any visiting sample in the above-mentioned subset of visited samples;

[0190] For the second visiting sample, determine the reference probability that the above-mentioned second visiting sample did not visit at the current training time, where the second visiting sample is any visiting sample in the above-mentioned subset of unvisited samples;

[0191] Determine the second loss information according to the visiting probabilities of the respective first visiting samples and the reference probabilities of the respective second visiting samples.

[0192] In one implementation, the determining unit 704 is specifically configured to:

[0193] Determine the probability that the above-mentioned second visiting sample will not visit, and determine the probability that the above-mentioned second visiting sample will visit after the current training time;

[0194] According to the probability that the above-mentioned second visiting sample will not visit and the probability that the above-mentioned second visiting sample will visit after the current training time, determine the reference probability that the above-mentioned second visiting sample did not visit at the above-mentioned current training time.

[0195] In one implementation, the device 70 further includes:

[0196] The obtaining unit 701 is further configured to obtain the visiting conversion prediction values of the respective objects in the set of objects to be pushed, where the target object is an object in the set of objects to be pushed;

[0197] A sorting unit 706 is configured to sort the predicted incoming conversion values of the above-mentioned respective objects in descending order according to the predicted incoming conversion values;

[0198] The above-mentioned determination unit 704 is further configured to determine the above-mentioned target object as an object to be pushed if the predicted incoming conversion value of the above-mentioned target object is among the top M in the sorting result.

[0199] According to an embodiment of the present application, Figure 2 Or Figure 5 Each step involved in the method shown can be executed by each unit in the Figure 7 content push device shown. For example, Figure 2 The step S201 shown is executed by the Figure 8 acquisition unit 701 shown in Figure 8 The step S202 is executed by the Figure 7 processing unit 702 shown in Figure 5 The step S501 shown is executed by the Figure 7 acquisition unit 701 shown in Figure 7 The step S502 is executed by the Figure 7 determination unit 704 shown in

[0200] According to another embodiment of the present application, Figure 7 Each unit in the content push device shown can be separately or all combined into one or several other units to form, or a certain (some) unit can be further split into multiple smaller units in terms of function to form, which can achieve the same operation without affecting the realization of the technical effects of the embodiments of the present application. The above units are divided based on logical functions. In practical applications, the function of one unit can also be realized by multiple units, or the functions of multiple units can be realized by one unit. In other embodiments of the present application, based on the content push device, other units can also be included. In practical applications, these functions can also be assisted by other units and can be realized by the cooperation of multiple units.

[0201] In the embodiments of the present application, by obtaining the feature set of the target object; using the visit conversion prediction model to process the feature set to obtain the visit conversion prediction value of the target object; and then if it is determined according to the visit conversion prediction value that the target object is a to-be-pushed object, push the recommended content to the target object. It can be seen that the probability of predicting the visit and conversion of the object can be obtained based on multiple object features of the target object, and the visit conversion prediction value is obtained. This visit conversion prediction model not only considers the visit and conversion, but also considers the behavior that the object will not visit immediately or convert immediately after receiving the recommended content, that is, it considers the behavior of the object's delayed visit and delayed conversion, so that the predicted visit conversion prediction value can be more accurate. Furthermore, by screening the objects according to the visit conversion prediction value, the objects with a greater probability of visiting and converting can be screened out to push the recommended content, making the pushed objects more accurate, thereby improving the click-through rate and conversion rate of the recommended content.

[0202] Based on the description of the above content push method embodiments, the embodiments of the present application also disclose an electronic device. Please refer to Figure 8 , the electronic device may at least include a processor 801, a communication interface 802, and a computer storage medium 803. Among them, the processor 801, the input interface 802, and the computer storage medium 803 in the electronic device can be connected through a bus or other means.

[0203] The above computer storage medium 803 is a memory device in the electronic device, used to store programs and data. It can be understood that the computer storage medium 803 here can include both the built-in storage medium of the electronic device and, of course, the extended storage medium supported by the electronic device. The computer storage medium 803 provides a storage space, and the operating system of the electronic device is stored in this storage space. And, one or more instructions suitable for being loaded and executed by the processor 801 are also stored in this storage space. These instructions can be one or more computer programs (including program codes). It should be noted that the computer storage medium here can be a high-speed RAM memory; optionally, it can also be at least one computer storage medium far from the aforementioned processor. The above processor can be called a Central Processing Unit (CPU), which is the core and control center of the electronic device, suitable for implementing one or more instructions, specifically loading and executing one or more instructions to implement the corresponding method flow or function.

[0204] In one implementation manner, one or more first instructions stored in the computer storage medium can be loaded and executed by the processor 801 to implement the corresponding steps of the method in the above content push method embodiments; in a specific implementation, one or more first instructions in the computer storage medium are loaded and executed by the processor 801 to perform the following operations:

[0205] Obtain the feature set of the target object;

[0206] Process the above feature set using the visit conversion prediction model to obtain the visit conversion prediction value of the above target object;

[0207] If it is determined according to the above visit conversion prediction value that the above target object is an object to be pushed, then push recommended content to the above target object;

[0208] Among them, the above visit conversion prediction model is trained using a conversion sample set and a visit sample set. The above conversion sample set includes an un-converted sample subset and a converted sample subset, and the above visit sample set includes a non-visited sample subset and a visited sample subset.

[0209] In one implementation, the above visit conversion prediction model includes a visit prediction module and a conversion prediction module. The above process of using the visit conversion prediction model to process the above feature set to obtain the visit conversion prediction value of the above target object includes:

[0210] Process the above feature set using the above visit prediction module to obtain the visit prediction value of the above target object;

[0211] Process the above feature set using the above conversion prediction module to obtain the conversion prediction value of the above target object;

[0212] Determine the visit conversion prediction value of the above target object according to the above visit prediction value and the above conversion prediction value.

[0213] In one implementation, one or more computer programs in the above computer storage medium are loaded and executed by the processor 801 to perform the following steps:

[0214] Obtain the above conversion sample set and the above visit sample set. Each conversion sample in the above conversion sample set includes the feature set and conversion label of the first reference object, and each visit sample in the above visit sample set includes the feature set and visit label of the second reference object;

[0215] Train the initial visit conversion prediction model using the above conversion sample set and the above visit sample set to determine the target loss information;

[0216] When it is determined according to the above loss information that the training stop condition is not satisfied, adjust the model parameters of the above initial visit conversion prediction model according to the above loss information, and train the adjusted visit conversion prediction model using the above conversion sample set and the above visit sample set to obtain the trained visit conversion prediction model.

[0217] In one implementation, training the initial visit conversion prediction model using the above-mentioned conversion sample set and the above-mentioned visit sample set to determine the target loss information includes:

[0218] Training the initial visit conversion prediction model using the above-mentioned conversion sample set to determine the first loss information;

[0219] Training the initial visit conversion prediction model using the above-mentioned visit sample set to determine the second loss information;

[0220] Determining the target loss information according to the above-mentioned first loss information and the above-mentioned second loss information.

[0221] In one implementation, if the conversion sample belongs to the converted sample subset, the conversion label of the above-mentioned conversion sample includes the conversion time, and training the initial visit conversion prediction model using the above-mentioned conversion sample set to determine the first loss information includes:

[0222] For the first conversion sample, determining the conversion probability that the above-mentioned first conversion sample is converted at the conversion time corresponding to the above-mentioned first conversion sample, where the first conversion sample is any conversion sample in the above-mentioned converted sample subset;

[0223] For the second conversion sample, determining the conversion reference probability that the above-mentioned second conversion sample is not converted at the current training time, where the second conversion sample is any conversion sample in the above-mentioned non-converted sample subset;

[0224] Determining the first loss information according to the conversion probabilities of each first conversion sample and the reference probabilities of each second conversion sample.

[0225] In one implementation, the above-mentioned determining the conversion reference probability that the above-mentioned second conversion sample is not converted at the current training time includes:

[0226] Determining the probability that the above-mentioned second conversion sample will not be converted, and determining the probability that the above-mentioned second conversion sample will be converted after the current training time;

[0227] Determining the conversion reference probability that the above-mentioned second conversion sample is not converted at the above-mentioned current training time according to the probability that the above-mentioned second conversion sample will not be converted and the probability that the above-mentioned second conversion sample will be converted after the current training time.

[0228] In one implementation, if the visit sample belongs to the visited sample subset, the visit label of the above-mentioned visit sample includes the visit time, and training the initial visit conversion prediction model using the above-mentioned visit sample set to determine the second loss information includes:

[0229] For the first visiting sample, determine the visiting probability that the first visiting sample visits at the visiting time corresponding to the first visiting sample, where the first visiting sample is any visiting sample in the subset of the visited samples;

[0230] For the second visiting sample, determine the reference probability that the second visiting sample does not visit at the current training time, where the second visiting sample is any visiting sample in the subset of the non-visited samples;

[0231] Determine the second loss information according to the visiting probabilities of the first visiting samples and the reference probabilities of the second visiting samples.

[0232] In one implementation, determining the reference probability that the second visiting sample does not visit at the current training time includes:

[0233] Determine the probability that the second visiting sample will not visit, and determine the probability that the second visiting sample will visit after the current training time;

[0234] Determine the reference probability that the second visiting sample does not visit at the current training time according to the probability that the second visiting sample will not visit and the probability that the second visiting sample will visit after the current training time.

[0235] In one implementation, one or more computer programs in the computer storage medium are loaded and executed by the processor 801 to perform the following steps:

[0236] Obtain the visiting conversion prediction values of the objects in the set of objects to be pushed, where the target object is an object in the set of objects to be pushed;

[0237] Sort the visiting conversion prediction values of the objects in descending order according to the visiting conversion prediction values;

[0238] If the visiting conversion prediction value of the target object is among the top M in the sorting result, determine that the target object is an object to be pushed.

[0239] For the specific implementation of each step executed by the processor 801 in the embodiments of the present application, reference may be made to the description of the relevant content in the foregoing embodiments, and the same technical effects can be achieved, which will not be elaborated here.

[0240] The embodiments of the present application also provide a computer-readable storage medium, in which a computer program is stored, and the processor runs the computer program to enable the electronic device to execute the method provided in the foregoing embodiments.

[0241] An embodiment of the present application also provides a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the electronic device executes the method provided in the foregoing embodiment.

[0242] The steps in the method of the embodiment of the present application can be adjusted, combined, and deleted according to actual needs.

[0243] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above methods. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0244] The foregoing disclosure is only a preferred embodiment of the present application. Of course, the scope of rights of the present application cannot be limited thereby. Those of ordinary skill in the art can understand all or part of the processes of the above embodiments, and the equivalent changes made according to the claims of the present application still fall within the scope covered by the invention.

Claims

1. A content push method, characterized in that, the method includes: obtaining a feature set of a target object; processing the feature set by using a visit conversion prediction model to obtain a visit conversion prediction value of the target object; if it is determined according to the visit conversion prediction value that the target object is an object to be pushed, then pushing recommended content to the target object; wherein, the visit conversion prediction model is obtained by adjusting model parameters of an initial visit conversion prediction model according to target loss information, and the target loss information is determined according to first loss information and second loss information; the first loss information is determined according to the conversion probability that each first conversion sample is converted at the conversion time corresponding to the first conversion sample and the reference probability that each second conversion sample is not converted at the current training time, and the reference probability is determined according to the probability that the second conversion sample will not be converted and the probability that the second conversion sample is converted after the current training time; the first conversion sample is any conversion sample in the subset of converted samples, the second conversion sample is any conversion sample in the subset of unconverted samples, and the subset of converted samples and the subset of unconverted samples are included in the conversion sample set; the second loss information is determined by training the initial visit conversion prediction model by using a visit sample set, and the visit sample set includes a subset of non-visited samples and a subset of visited samples.

2. The method according to claim 1, characterized in that, the visit conversion prediction model includes a visit prediction module and a conversion prediction module, and the processing of the feature set by using the visit conversion prediction model to obtain the visit conversion prediction value of the target object includes: processing the feature set by using the visit prediction module to obtain a visit prediction value of the target object; processing the feature set by using the conversion prediction module to obtain a conversion prediction value of the target object; determining the visit conversion prediction value of the target object according to the visit prediction value and the conversion prediction value.

3. The method according to claim 1 or 2, characterized in that, the method further includes: obtaining the conversion sample set and the visit sample set, where each conversion sample in the conversion sample set includes a feature set of a first reference object and a conversion label, and each visit sample in the visit sample set includes a feature set of a second reference object and a visit label; when it is determined according to the target loss information that the training stop condition is not satisfied, adjusting the model parameters of the initial visit conversion prediction model according to the target loss information, and training the adjusted visit conversion prediction model by using the conversion sample set and the visit sample set to obtain a trained visit conversion prediction model.

4. The method according to claim 3, characterized in that, if the conversion sample belongs to the subset of converted samples, the conversion label of the conversion sample includes the conversion time.

5. The method according to claim 3, characterized in that, If the visited sample belongs to the subset of visited samples, the visit label of the visited sample includes the visit time, and the initial visit conversion prediction model is trained using the visited sample set to determine the second loss information, including: For the first visited sample, determine the visit probability that the first visited sample visits at the visit time corresponding to the first visited sample, where the first visited sample is any visited sample in the subset of visited samples; For the second visited sample, determine the reference probability that the second visited sample does not visit at the current training time, where the second visited sample is any visited sample in the subset of unvisited samples; Determine the second loss information according to the visit probabilities of each first visited sample and the reference probabilities of each second visited sample.

6. The method according to claim 5, wherein, the determining the reference probability that the second visited sample does not visit at the current training time includes: determining the probability that the second visited sample will not visit, and determining the probability that the second visited sample will visit after the current training time; determining the reference probability that the second visited sample does not visit at the current training time according to the probability that the second visited sample will not visit and the probability that the second visited sample will visit after the current training time.

7. The method according to claim 1 or 2, wherein, the method further includes: obtaining the visit conversion prediction values of each object in the set of objects to be pushed, where the target object is an object in the set of objects to be pushed; sorting the visit conversion prediction values of each object in descending order according to the visit conversion prediction values; if the visit conversion prediction value of the target object is among the top M in the sorting result, determining that the target object is an object to be pushed.

8. A content push device, wherein, the device includes: an acquisition unit, configured to acquire the feature set of the target object; a processing unit, configured to process the feature set using the visit conversion prediction model to obtain the visit conversion prediction value of the target object; a push unit, configured to push recommended content to the target object if it is determined that the target object is an object to be pushed according to the visit conversion prediction value; Among them, the visiting conversion prediction model is obtained by adjusting the model parameters of the initial visiting conversion prediction model according to the target loss information, and the target loss information is determined according to the first loss information and the second loss information; the first loss information is determined according to the conversion probability that each first conversion sample is converted at the conversion time corresponding to the first conversion sample and the reference probability that each second conversion sample is not converted at the current training time, and the reference probability is determined according to the probability that the second conversion sample will not be converted and the probability that the second conversion sample is converted after the current training time; the first conversion sample is any conversion sample in the converted sample subset, the second conversion sample is any conversion sample in the unconverted sample subset, and the converted sample subset and the unconverted sample subset are included in the conversion sample set; the second loss information is determined by training the initial visiting conversion prediction model using the visiting sample set, and the visiting sample set includes a non-visiting sample subset and a visited sample subset.

9. A content push device, characterized in that, it includes a processor, a communication interface and a memory, the processor, the communication interface and the memory are connected to each other, wherein, the memory stores executable program code, and the processor is used to call the executable program code to execute the content push method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it executes the content push method according to any one of claims 1-7.

11. A computer program product, characterized in that, the computer program product includes a computer program, the computer program is stored in a computer-readable storage medium, the processor reads the computer program from the computer-readable storage medium, and the processor executes the computer program to execute the content push method according to any one of claims 1-7.

Images