Object recommendation method and apparatus, computer device, and storage medium

Abstract

The application relates to an object recommendation method and device, computer equipment and a storage medium. The method comprises the following steps: obtaining information features corresponding to a target user identifier and object features corresponding to candidate recommended objects of the target user identifier; performing shared mapping processing on the object features and the information features to obtain multi-level conversion task parameters matched with the candidate recommended objects; inputting the conversion task parameters of each level into a Bayesian network for mapping processing, obtaining each level conversion estimation value matched with the candidate recommended objects based on the mapping parameters output by the Bayesian network and the feature matching data of the object features and the information features, and the conversion estimation value corresponding to the conversion task indicated by the conversion task parameter; and performing object recommendation to a target user represented by the target user identifier based on the object recommendation probability corresponding to the conversion estimation value of each level. The method can improve the efficiency of object recommendation to users.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an object recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product. Background Technology

[0002] With the development of computer technology, object recommendation systems have emerged. Taking recommended advertising as an example, in an advertising recommendation system, when a user request arrives, the system sorts the ads and automatically bids based on the user's predicted click-through rate and conversion rate for each candidate ad, and then recalls suitable ads from the candidate ad set to deliver to the user. For example, for application download ads, this mainly involves the user's shallow conversion behavior and deep conversion behavior. Shallow conversion behavior includes downloading, installing, and activating, while deep conversion behavior includes paying and next-day retention.

[0003] However, when used as a performance indicator for advertising, deep conversion behaviors are often characterized by high sparsity and high latency, resulting in significant deviations in conversion rate predictions and consequently, low efficiency in recommending products to users. Summary of the Invention

[0004] Therefore, it is necessary to provide an object recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product that can improve the efficiency of recommending objects to users, in order to address the above-mentioned technical problems.

[0005] An object recommendation method, the method comprising:

[0006] Obtain the information features corresponding to the target user identifier, and the object features corresponding to the candidate recommendation objects for the target user identifier;

[0007] The object features and information features are shared and mapped to obtain multi-level conversion task parameters that match the candidate recommendation objects;

[0008] The conversion task parameters at each level are input into a Bayesian network for mapping processing. Based on the mapping parameters at each level output by the Bayesian network and the feature matching data between the object features and the information features, conversion prediction values ​​at each level that match the candidate recommendation object are obtained. The conversion prediction values ​​correspond one-to-one with the conversion tasks indicated by the conversion task parameters.

[0009] Based on the object recommendation probability corresponding to the conversion prediction value at each level, object recommendations are made to the target user represented by the target user identifier.

[0010] An object recommendation device, the device comprising:

[0011] The feature acquisition module is used to acquire the information features corresponding to the target user identifier and the object features corresponding to the candidate recommendation objects for the target user identifier;

[0012] The parameter determination module is used to perform a shared mapping process between the object features and the information features to obtain multi-level conversion task parameters that match the candidate recommendation object;

[0013] The computational processing module is used to input the conversion task parameters of each level into the Bayesian network for mapping processing. Based on the mapping parameters of each level output by the Bayesian network and the feature matching data of the object features and the information features, the conversion prediction value of each level that matches the candidate recommendation object is obtained. The conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters.

[0014] The object recommendation module is used to recommend objects to the target user represented by the target user identifier based on the object recommendation probability corresponding to the conversion prediction value at each level.

[0015] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0017] A computer program product includes a computer program that, when executed by a processor, implements the steps of the method described above.

[0018] The aforementioned object recommendation methods, apparatus, computer devices, computer-readable storage media, and computer program products, after acquiring the information features corresponding to the target user identifier and the object features corresponding to the candidate recommendation objects for the target user identifier, obtain multi-level conversion task parameters matching the candidate recommendation objects by sharing and mapping the object features and information features. This allows each level of conversion task to share underlying parameters, alleviating the data sparsity problem of higher-level conversion tasks and thus correcting data processing biases. By constructing a Bayesian network to process the multi-level conversion task parameters and obtaining conversion prediction values ​​at each level matching the candidate recommendation objects based on feature matching data of object features and information features, the multi-level conversion tasks can be correlated, improving the accuracy of the determined conversion prediction values ​​at each level. By recommending objects to the target user represented by the target user identifier based on the object recommendation probability corresponding to each level of conversion prediction value, the duplicate recommendation of similar objects can be effectively avoided, thereby improving object recommendation efficiency. Attached Figure Description

[0019] Figure 1 This is a diagram illustrating the application environment of an object recommendation method in one embodiment.

[0020] Figure 2 This is a flowchart illustrating an object recommendation method in one embodiment;

[0021] Figure 3 This is a schematic diagram of an object matching model and an object sorting model in one embodiment;

[0022] Figure 4 This is a flowchart illustrating an object recommendation method in a specific embodiment.

[0023] Figure 5 This is a schematic diagram of the ad matching model and ad ranking model in a specific embodiment;

[0024] Figure 6 This is a structural block diagram of an object recommendation device in one embodiment;

[0025] Figure 7 This is an internal structural diagram of a computer device in one embodiment;

[0026] Figure 8 This is a diagram of the internal structure of a computer device in another embodiment. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0028] In one embodiment, the object recommendation method provided in this application can be applied to, for example... Figure 1 In the illustrated application environment, both terminal 102 and server 104 may be involved. In other embodiments, the application environment may also involve terminal 106. Terminals 102 and 106 communicate with server 104 via a network, and a data storage system stores the data that server 104 needs to process. The data storage system may be integrated onto server 104 or located in the cloud or on another server.

[0029] Specifically, server 104 can obtain the target user identifier from terminal 102, thereby determining the information features corresponding to the target user identifier and the object features corresponding to the candidate recommended objects for the target user identifier. In server 104, object features and information features can be shared and mapped to obtain multi-level conversion task parameters matching the candidate recommended objects. Then, the conversion task parameters at each level are input into a Bayesian network for mapping processing. Based on the mapping parameters at each level output by the Bayesian network and the feature matching data of object features and information features, conversion prediction values ​​at each level matching the candidate recommended objects are obtained. Each conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters. Finally, server 104 calculates and determines the object recommendation probability corresponding to each level of conversion prediction value, and based on the object recommendation probability corresponding to each level of conversion prediction value, pushes the finally determined target recommended objects to terminals 102 and 106 to achieve object recommendation to the target user represented by the target user identifier.

[0030] In one embodiment, when the data processing capabilities of terminals 102 and 106 meet the data processing requirements, the object recommendation method provided in this application embodiment can be applied to an environment involving only terminal 102 or terminal 106. Specifically, terminal 102 or terminal 106 directly obtains the information features corresponding to the target user identifier and the object features corresponding to the candidate recommendation objects for the target user identifier. The object features and information features are then shared and mapped to obtain multi-level conversion task parameters matching the candidate recommendation objects. Then, the conversion task parameters at each level are input into a Bayesian network for mapping. Based on the mapping parameters at each level output by the Bayesian network and the feature matching data of object features and information features, conversion prediction values ​​at each level matching the candidate recommendation objects are obtained. Each conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters. Terminal 102 or terminal 106 displays the finally determined target recommendation object based on the object recommendation probability corresponding to each level of conversion prediction value, thus realizing object recommendation to the target user represented by the target user identifier.

[0031] Terminals 102 and 106 can be, but are not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart TVs, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, etc. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.

[0032] With the development of computer technology, object recommendation systems have emerged. Taking advertisements as an example, recommending ads to users is a key aspect. The effectiveness of an ad can be reflected in impressions, clicks, and conversions. When a user requests access, such as opening a webpage or application, the ad recommendation system needs to retrieve suitable ads from the candidate ad set and deliver them to the user. The system ranks and automatically bids ads based on the user's predicted click-through rate and conversion rate for each candidate ad. For example, for application download ads, different stages of conversion behavior are involved, including shallow conversion behaviors such as download, installation, and activation, and deep conversion behaviors such as payment and next-day retention. However, when used as a performance indicator for ad placement, deep conversion behaviors are often characterized by high sparsity and high latency, leading to significant deviations in conversion rate predictions. For instance, shallow conversion behaviors like activation have relatively higher conversion rates than deep conversion behaviors, and advertisers can submit their conversion data on the same day. However, deep conversion behaviors like payment have very low conversion rates, and the latest data from advertisers is delayed for a long time, typically more than a week.

[0033] In traditional techniques, the common practice for predicting ad click-through rates (CTR) and conversion rates is to use deep learning models to train separate CTR prediction models, shallow conversion rate prediction models, and deep conversion rate prediction models. Each time a user requests an ad, the CTR, shallow conversion rate, and deep conversion rate are predicted separately. However, because the shallow conversion rate prediction model is trained using samples from clicks to shallow conversions, and the deep conversion rate prediction model is trained using samples from shallow conversions to deep conversions, and both models use the displayed ad for each user request when predicting the conversion rate, the training and prediction sample distributions of the conversion rate prediction models become inconsistent, resulting in significant deviations in the conversion rate predictions.

[0034] To address the above problems, in one embodiment, such as Figure 2 As shown, an object recommendation method is provided, which is applied to... Figure 1 Taking server 104 as an example, the following steps are included:

[0035] Step S202: Obtain the information features corresponding to the target user identifier and the object features corresponding to the candidate recommendation objects for the target user identifier.

[0036] In this context, "target user" refers to the user who needs to be recommended content. The target user identifier is an identifier that can represent the target user, and its specific configuration can be set according to actual technical needs. For example, if the target user accesses a page through a terminal, the target user identifier could be the terminal identifier corresponding to the terminal used by the target user, or any other identifier that might represent the target user. "Candidate recommendation objects" refers to objects that need to be determined whether to recommend to the target user; multiple candidate recommendation objects are possible. The types of candidate recommendation objects include, but are not limited to, advertisements, news, videos, and applications.

[0037] In one embodiment, a candidate recommendation object dataset is pre-set, which includes multiple candidate recommendation objects. When an access request is received, a target user identifier is obtained. Then, based on the target user identifier, a candidate recommendation object corresponding to the target user identifier is determined from the candidate recommendation object dataset. For example, the candidate recommendation object dataset stores object information for multiple candidate recommendation objects, and the object information corresponds to a user identifier. After obtaining the target user identifier, the obtained target user identifier can be matched with the stored object information to determine the matching object information. By recalling the candidate recommendation objects corresponding to the matching object information, a candidate recommendation object corresponding to the target user identifier is determined. The matching and recall methods can be set according to actual technical needs and are not limited here.

[0038] In one embodiment, after obtaining the target user identifier and candidate recommended objects for the target user identifier, the information features corresponding to the target user identifier and the object features corresponding to the candidate recommended objects for the target user identifier are obtained. The information features corresponding to the target user identifier include, but are not limited to, the target user's basic attribute features and behavioral interest features. For example, basic attribute features include features such as name, gender, age, and city, while behavioral interest features include browsing behavior interest features and click behavior interest features. The object features corresponding to the candidate recommended objects include, but are not limited to, the recommended object identifier (id), the recommended object owner identifier (id), and the item features corresponding to the recommended object. Item features include, but are not limited to, the category to which the recommended object belongs and its semantic features and image features. In some embodiments, the object features also include context features, which include, but are not limited to, the context of the current access request and terminal device features. For example, context features include the semantics of the session control context accessed by the target user, the terminal device type, the terminal device address, and the text and image features of the currently accessed page.

[0039] Step S204: Perform shared mapping processing on object features and information features to obtain multi-level conversion task parameters that match the candidate recommendation objects.

[0040] In one embodiment, since some information features or object features may be high-dimensional, to improve data processing efficiency, before performing the shared mapping process between object features and information features, the method further includes mapping the acquired object features and information features to object features of a first preset dimension and information features of a second preset dimension, respectively. The first and second preset dimensions are low-dimensional and can be set according to actual technical needs; for example, the first and second preset dimensions can be set to be equal. The mapping method can also be set according to actual technical needs.

[0041] For example, if a feature mapping processing layer is used to perform the above mapping processing, and the feature mapping processing layer can be an embedding layer, then the object features obtained after processing are the embedding features of the first preset dimension, and the information features are the embedding features of the second preset dimension.

[0042] Specifically, such as Figure 3 The diagram shows the object matching model and object sorting model in an embodiment of this application. In the object sorting model, the above-mentioned mapping process can be performed to obtain multi-level transformation task parameters, which are then used for subsequent calculations. Figure 3 As shown in the feature layer of the object sorting model, the direction of the arrows in the feature layer is the direction of data transmission. That is, after obtaining the information features and object features, the feature mapping processing layer is first used to map the information features and object features, and then the parameters of the subsequent multi-level transformation tasks are determined.

[0043] In one embodiment, before making final object recommendations to the target user, to improve recommendation efficiency, it is necessary to first estimate the probability of various types of conversion behaviors corresponding to the candidate recommended objects. Each type of conversion behavior corresponds to a level of conversion task; that is, the candidate recommended objects have multiple levels of conversion tasks, specifically referred to as first-level conversion tasks, second-level conversion tasks, and so on. The number of levels can be determined according to actual technical needs, for example, based on the type corresponding to the candidate recommended objects.

[0044] For example, taking a recommended ad with three levels of conversion tasks, the conversion tasks specifically include: the first level of conversion task corresponding to ad impressions and clicks; the second level of conversion task corresponding to clicks and shallow conversions; and the third level of conversion task corresponding to shallow conversions and deeper conversions. It should be noted that if the ad also has updated deeper conversions, then the ad has a fourth level of conversion task corresponding to deeper conversions, which is not limited here.

[0045] In one embodiment, since higher-level transformations are performed based on lower-level transformations, the transformation task parameters of higher-level transformation tasks are more sparse and have higher latency compared to lower-level transformation tasks. Therefore, to alleviate the problem of sparse parameters and high latency in higher-level transformation tasks, the acquired object features and information features are shared and mapped to enable the sharing of underlying parameters among multi-level transformation tasks.

[0046] In one embodiment, shared mapping processing refers to mapping object features and information features through the same shared feature mapping processing layer to ultimately obtain multi-level conversion task parameters that match the candidate recommendation object. That is, each level of conversion prediction task can use these multi-level conversion task parameters. The shared feature mapping processing layer can be selected according to actual technical needs. For example, a shared multilayer perceptron (MLP) can be used for mapping processing. The training method of the multilayer perceptron involved in this application embodiment can be any feasible training method, including but not limited to backpropagation or pseudo-inverse learning training methods, and is not limited here.

[0047] In one embodiment, the shared mapping process between object features and information features to obtain multi-level conversion task parameters that match the candidate recommendation objects may include the following steps S2041 to S2042:

[0048] Step S2041: Perform a shared mapping process on the object features and information features to obtain the shared object features corresponding to the object features and the shared information features corresponding to the information features.

[0049] In one embodiment, a shared feature mapping processing layer performs shared mapping processing on object features and information features. The mapped object features are called shared object features, and the mapped information features are called shared information features. This yields the shared object features corresponding to the object features and the shared information features corresponding to the information features. Since object features and information features are embedding features, the shared object features and shared information features are also embedding features. Figure 3 The parameter sharing layer of the object ranking model is shown in the diagram. The arrows in the parameter sharing layer indicate the direction of data transmission. That is, information features and object features are connected to the shared feature processing layer through the feature mapping processing layer. After passing through the shared feature mapping processing layer, the shared object features corresponding to the object features and the shared information features corresponding to the information features are obtained.

[0050] Step S2042: According to the transformation tasks at each level, the shared object features and shared information features that match the transformation tasks are processed to obtain the transformation task parameters corresponding to each level.

[0051] In one embodiment, shared object features and shared information features matching the transformation tasks at each level are predetermined. Specifically, the shared object features and shared information features matching the transformation tasks at each level can be determined first, and then a series of calculations can be performed on these features to obtain the transformation task parameters corresponding to each level. The calculation method can be set according to actual technical needs.

[0052] In one embodiment, according to the transformation tasks at each level, the shared object features and shared information features that match the transformation tasks are processed to obtain the transformation task parameters corresponding to each level, including the following steps S20421 to S20422:

[0053] Step S20421: According to the transformation tasks at each level, the shared object features and shared information features that match the transformation tasks are concatenated to obtain concatenated features that match the transformation tasks.

[0054] In one embodiment, since there are multiple shared information features and shared object features, and different shared information features and shared object features are required when generating conversion task parameters at different levels, a shared gate structure is set to control the input shared object feature and shared information feature data. The number of shared gates can be set according to actual technical needs. Therefore, different levels of conversion tasks can control the input of different data by opening different shared gates.

[0055] Taking a conversion task with three levels as an example, such as Figure 3 As shown in the parameter sharing layer of the object sorting model, five shared gate structures are set. For the first-level transformation task, the input data when generating the transformation task parameters is the data of the first and second shared gates. For the second-level transformation task, the input data when generating the transformation task parameters is the data of the first, second, third, and fourth shared gates. For the third-level transformation task, the input data when generating the transformation task parameters is the data of the first, third, fourth, and fifth shared gates.

[0056] In one embodiment, shared object features and shared information features that match the transformation task can be concatenated to obtain concatenated features that match the transformation task. For example, assuming that the embedding lengths of both the shared object features and the shared information features are n, concatenating the shared object features and the shared information features will result in a concatenated feature with an embedding length of 2n.

[0057] Step S20422: Perform feature processing on each splicing feature to obtain the corresponding transformation task parameters for each level.

[0058] The feature processing method for each splicing feature can be set according to actual technical needs. In one embodiment, the feature processing method includes normalization processing and feature mapping processing. That is, each splicing feature is normalized to obtain processed splicing features, and the processed splicing features are then processed by feature mapping through the respective transformation task parameter processing model of each level to obtain the corresponding transformation task parameters for each level.

[0059] Specifically, the type and structure of the conversion task parameter processing model can be set according to actual technical needs. In one embodiment, the conversion task parameter processing model is selected as a multilayer perceptron (MLP). That is, the conversion task parameters corresponding to each conversion task are learned and determined through the conversion task parameter processing model of each layer.

[0060] Taking a conversion task with three levels as an example, such as Figure 3 As shown in the parameter sharing layer of the object sorting model, for each level of conversion task, after obtaining the splicing features that match the conversion task, the splicing features are normalized. Then, they are further processed through the first-level conversion task parameter processing model, the second-level conversion task parameter processing model, and the third-level conversion task parameter processing model that each level has, and finally output the corresponding conversion task parameters for each level.

[0061] Step S206: Input the conversion task parameters of each level into the Bayesian network for mapping processing. Based on the mapping parameters of each level output by the Bayesian network and the feature matching data of object features and information features, obtain the conversion prediction value of each level that matches the candidate recommendation object. The conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters.

[0062] A Bayesian network is a graphical network model used to describe uncertain causal relationships between variables. It consists of nodes, directed connections, and a node probability table, where directed connections represent causal dependencies between nodes. Therefore, to reflect the relationships between conversion tasks at different levels and improve the accuracy of predicting the results of these tasks—where the predicted result is called the conversion prediction value—a Bayesian network can be constructed to process the conversion task parameters at each level, ultimately yielding conversion prediction values ​​at each level that match the candidate recommendation objects.

[0063] In one embodiment, the type, number, and direction of directed connections in the Bayesian network can be determined according to actual technical needs. Specifically, the number of nodes is the same as the number of levels corresponding to the candidate recommendation objects, and the direction of the directed connections can be from one side to the other side being influenced. In the Bayesian network constructed in this embodiment, a node can correspond to the mapping processing model used for a level of conversion task, and the node of this Bayesian network is called a mapping node, that is, a mapping node in the Bayesian network is used to represent a level of conversion task mapping processing model. The conversion task mapping processing model can be set according to actual technical needs; in one embodiment, it can be selected as a multilayer perceptron neural network (MLP).

[0064] Furthermore, since higher-level transformation tasks are carried out on the basis of lower-level transformation tasks, the direction of the directed connection is from the mapping node of the lower-level transformation task to the mapping node of the higher-level transformation task. That is, there is a directed connection between the mapping node of the upper-level transformation task and the mapping node of the lower-level transformation task in the Bayesian network.

[0065] In one embodiment, after determining the conversion task parameters for each level, the conversion task parameters for each level can be input into a Bayesian network for mapping processing to obtain the mapping parameters for each level. Then, based on the mapping parameters for each level output by the Bayesian network and the feature matching data of object features and information features, the conversion prediction value for each level that matches the candidate recommendation object can be obtained. Specifically, this includes the following steps S2061 to S2062:

[0066] Step S2061: Map the transformation task parameters of each level to obtain the mapping parameters of each level; wherein, there is a directed connection between the mapping node of the transformation task of the previous level and the mapping node of the transformation task of the next level in the Bayesian network, and the input data of each mapping node includes the transformation task parameters of the current level and the mapping parameters of the previous level.

[0067] In one embodiment, within the Bayesian network, a separate transformation task mapping processing model is used for each level to map the transformation task parameters of each level. The resulting data is referred to as the mapping parameters. Specifically, the mapping parameters of each level can be represented as logit. n Where n represents the level, and n is greater than 0. For example, the mapping parameter of the first level can be represented as logit1, the mapping parameter of the second level can be represented as logit2, the mapping parameter of the third level can be represented as logit3, and so on.

[0068] It should be noted that, because there are directed connections between the mapping nodes of the upper-level transformation task and the mapping nodes of the lower-level transformation task in a Bayesian network, the input data of each mapping node includes the parameters of the current-level transformation task and the mapping parameters of the upper-level task. For example, if the candidate recommendation object has three levels of transformation tasks, for the first level, there is no upper-level task, meaning the input data of the mapping node for the first-level transformation task is the parameters of the first-level transformation task. For the second level, its upper-level task is the first level, so the input data of the mapping node for the second-level transformation task is the mapping parameters of the first level and the parameters of the second-level transformation task. The input data of the mapping node for the third-level transformation task is the mapping parameters of the second level and the parameters of the third-level transformation task, and so on.

[0069] In one embodiment, to effectively combine the two types of parameters—the current level conversion task parameters and the previous level mapping parameters—the mapping node of the previous level conversion task is also directed to the mapping node of the next level conversion task through the connection layer of the next level conversion task. Specifically, the directed connection between the mapping node of the previous level conversion task, the connection layer of the next level conversion task, and the mapping node of the next level conversion task is as follows: the mapping node of the previous level conversion task is directed to the connection layer of the next level conversion task and the mapping node of the next level conversion task; the connection layer of the next level conversion task is directed to the mapping node of the next level conversion task; and the conversion task parameters of the next level conversion task itself are input through the connection layer of the next level conversion task.

[0070] Taking a conversion task with three levels as an example, such as Figure 3The Bayesian network representing the object ranking model is shown below. The network contains three mapping nodes, with arrows indicating directed connections. Each level of the transformation task mapping model corresponds to one mapping node. The input data for the mapping node of the first-level transformation task includes the first-level transformation task parameters. The mapping node of the first-level transformation task has a directed connection layer to the second-level transformation task, and also contains mapping nodes for the second-level transformation task. The input data for the mapping node of the second-level transformation task includes the second-level transformation task parameters and the mapping parameters of the first level. The mapping node of the second-level transformation task has a directed connection layer to the third-level transformation task, and also contains mapping nodes for the third level. The input data for the mapping node of the third-level transformation task includes the third-level transformation task parameters and the mapping parameters of the second level. The input data of the mapping nodes of the first-level conversion task are mapped and processed by the first-level conversion task mapping processing model to obtain the first-level mapping parameters. The input data of the mapping nodes of the second-level conversion task are mapped and processed by the second-level conversion task mapping processing model to obtain the second-level mapping parameters. The input data of the mapping nodes of the third-level conversion task are mapped and processed by the third-level conversion task mapping processing model to obtain the third-level mapping parameters.

[0071] By adopting the above embodiment, it is equivalent to sharing the data of the first-level conversion task with the second-level conversion task, and sharing the data of the second-level conversion task with the third-level conversion task, which alleviates the data sparsity problem of high-level conversion tasks and corrects data processing deviations.

[0072] Step S2062: Based on the mapping parameters of each level and the feature matching data of object features and information features, obtain the conversion prediction value of each level that matches the candidate recommendation object.

[0073] In one embodiment, after determining the mapping parameters for each level, conversion prediction values ​​for each level that match the candidate recommendation objects can be obtained based on the mapping parameters for each level and the feature matching data of object features and information features. Each conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters. Therefore, subsequent object recommendations are performed based on the conversion prediction values ​​for each level. Specifically, the first-level conversion prediction value can be represented as CTR, and the other level conversion prediction values ​​can be represented as CVR. n-1 Where n represents the level, and n is greater than 1. For example, the conversion estimate for the second level is represented as cvr1, the conversion estimate for the third level is represented as cvr2, and so on.

[0074] Taking a conversion task with three levels as an example, such as Figure 3As shown in the object ranking model, the first-level mapping parameters, second-level mapping parameters, and third-level mapping parameters are obtained through a Bayesian network. A connection layer is then used to combine the mapping parameters of each level with feature matching data of object features and information features, ultimately obtaining the conversion prediction values ​​for each level that match the candidate recommended object, namely the first-level conversion prediction value, the second-level conversion prediction value, and the third-level conversion prediction value.

[0075] The feature matching data between object features and information features can be used to characterize the target user's interest in the candidate recommended object. In one embodiment, the method for determining the feature matching data between object features and information features includes the following steps S2063 to S2064:

[0076] Step S2063: Map the object features and information features respectively to obtain object mapping features and user mapping features.

[0077] In one embodiment, object features and information features are mapped separately, and the mapping method can be set according to actual technical needs. When determining the feature matching data between object features and information features, there is no hierarchical relationship. Therefore, to improve data processing efficiency, the object feature mapping model corresponding to the object features can be directly used to map the object features to obtain object-mapped features. The information feature mapping model corresponding to the information features is used to map the information features to obtain user-mapped features. The model type and model structure of the object feature mapping model and the information feature mapping model can be set according to actual technical needs. In one embodiment, a multilayer perceptron neural network (MLP) can be used.

[0078] It should be noted that user mapping features and object mapping features can be embedding features. That is, the user mapping features and object mapping features here are consistent with the feature types used in multi-level conversion tasks, thereby improving the accuracy of the final calculated conversion estimates for each level.

[0079] Step S2064: Perform a dot product operation on the object mapping features and the user mapping features to determine the feature matching data between the object features and the information features.

[0080] In one embodiment, feature matching data between object features and information features can be determined by performing a dot product operation on object mapping features and user mapping features. After performing the dot product operation, the result can be fed into a mapping processing model, and the final output of the mapping processing model is determined as the feature matching data between object features and information features. This mapping processing model introduces an activation function. The type of mapping processing model and its activation function can be set according to actual technical needs. In one embodiment, the mapping processing model can be a multilayer perceptron (MLP), and the activation function can be set to the sigmoid function.

[0081] Specifically, the feature matching data can be represented as logit0, and the calculation formula is as follows:

[0082] logit0 = P(match_score|x,H)

[0083] Where x represents information features and object features, and H represents the parameters of the mapping processing model.

[0084] By incorporating a nonlinear activation function into the mapping processing model using the above embodiments, the learning ability of the mapping processing model can be enhanced, thereby improving the accuracy of feature matching data between determined object features and information features.

[0085] Specifically, such as Figure 3 The diagram shows the object matching model and object ranking model in an embodiment of this application. In the object matching model, the dot product processing described above can be performed to obtain feature matching data between object features and information features. For example... Figure 3 As shown in the object matching model, its structure is a dual-tower structure. It takes object features and information features as input, and maps these features using their respective object feature mapping models and information feature mapping models to obtain object-mapped features and user-mapped features. The object-mapped features and user-mapped features are then multiplied by a dot product, and the result is fed into the mapping processing model to finally determine the feature matching data between the object features and information features.

[0086] In one embodiment, taking a conversion task with three levels as an example, the object ranking model is processed using a Bayesian network. The final conversion prediction values ​​for each level that match the candidate recommended objects are expressed as follows:

[0087] P(ctr,cvr1,cvr2|x,H,logit0)

[0088] =P(ctr|x,H,logit0)*P(cvr1|ctr,X,H,logit0)*P(cvr2|ctr,cvr1,X,H,logit0)

[0089] Where P(ctr|x, H, logit0) represents the first-level transformation prediction value, P(cvr1|ctr, X, H, logit0) represents the second-level transformation prediction value, P(cvr2|ctr, cvr1, X, H, logit0) represents the third-level transformation prediction value, x represents the input object features and information features, H represents the parameters of the model used in the Bayesian network for each level of transformation task, and logit0 is the feature matching data of information features and object features.

[0090] In one embodiment, taking the negative log-likelihood of the above formula, the loss function L(x, H) of the object ranking model is expressed as:

[0091] L(x,H)=-log(P(ctr,cvr1,cvr2x,H,logit0))

[0092] =-(log P(ctr|x,H,logit0)+log P(cvr1|ctr,X,H,logit0)+log P(cvr2|ctr,cvr1,X,H,logit0))

[0093] Where -log P(ctr|x, H, logit0) represents the loss function corresponding to the first-level transformation task, -log P(cvr1|ctr, X, H, logit0) represents the loss function corresponding to the second-level transformation task, and -log P(cvr2|ctr, cvr1, X, H, logit0) represents the loss function corresponding to the third-level transformation task.

[0094] In one embodiment, when considering the weights of the loss function, the loss function L(x, H) is expressed as:

[0095] L(x,H)=-(W1*log P(ctr|x,H,logit0)+W2*log P(cvr1|ctr,X,H,logit0)+W3*logP(cvr2|ctr,cvr1,X,H,logit0)]

[0096] Where W1, W2, and W3 are the weights of the loss functions corresponding to the first-level transformation task, the second-level transformation task, and the third-level transformation task, respectively.

[0097] It should be noted that the data processing process in the above embodiments of this application is the data processing process of the trained object matching model and object ranking model when in use. The data processing process of the object matching model and object ranking model during training corresponds to the data processing process described above when in use.

[0098] In one embodiment, after obtaining the conversion prediction values ​​at each level that match the candidate recommendation object, the object recommendation probability corresponding to each level of conversion prediction value can be calculated to determine whether to recommend the candidate recommendation object to the target user based on the object recommendation probability. Specifically, the method for determining the object recommendation probability corresponding to each level of conversion prediction value includes the following steps S2065 to S2066:

[0099] Step S2065: Determine the first-level conversion estimate among the conversion estimates of each level, and the maximum conversion estimate among the conversion estimates of each level excluding the first-level conversion estimate.

[0100] In one embodiment, since higher-level conversions are based on lower-level conversions, it is necessary to first determine the first-level conversion estimate among the conversion estimates for each level, and then determine the maximum conversion estimate among all levels excluding the first-level conversion estimate. For example, taking a conversion task with three levels as an example, after determining the first-level conversion estimate, a maximum conversion estimate is determined from the second-level and third-level conversion estimates for subsequent processing.

[0101] Step S2066: Determine the object recommendation probability of the candidate object based on the first-level conversion estimate, the maximum conversion estimate, and the recommendation benefit of the candidate object corresponding to the maximum conversion estimate.

[0102] In one embodiment, the recommendation probability of a candidate object is the product of the first-level conversion estimate, the maximum conversion estimate, and the recommendation benefit of the candidate object corresponding to the maximum conversion estimate. The recommendation benefit of a candidate object can refer to the bid corresponding to that subsequent recommended object. For example, taking a conversion task with three levels, if the maximum conversion estimate is the second-level conversion estimate, then the recommendation benefit of the candidate object corresponding to the maximum conversion estimate is the bid for the second-level conversion. If the maximum conversion estimate is the third-level conversion estimate, then the recommendation benefit of the candidate object corresponding to the maximum conversion estimate is the bid for the third-level conversion.

[0103] In one embodiment, the object recommendation probability is represented as ecpm, and its calculation formula is as follows:

[0104] ecpm = ctr * cvr max *bidmax

[0105] Where CTR represents the first-level conversion estimate, and CVR... max Bid represents the maximum conversion estimate excluding the first-level conversion estimate among all conversion estimates. max This represents the recommendation effectiveness of the candidate recommended object corresponding to the maximum conversion estimate.

[0106] Step S208: Based on the object recommendation probability corresponding to the conversion prediction value at each level, object recommendations are made to the target users represented by the target user identifier.

[0107] In one embodiment, after determining the object recommendation probability corresponding to each level of conversion prediction based on the conversion prediction value of the candidate recommendation object, object recommendations can be made to the target user represented by the target user identifier based on the object recommendation probability. Specifically, this includes the following steps S2081 to S2082:

[0108] Step S2081: Based on the object recommendation probability of each candidate recommended object, select the target recommended object whose object recommendation probability meets the recommendation conditions from the candidate recommended objects.

[0109] Specifically, based on the recommendation probability of each candidate object, target recommendation objects that meet the recommendation criteria can be selected from the candidate objects for recommendation to the target users. The recommendation criteria can be set according to actual technical needs.

[0110] For example, the recommendation criterion can be set to an object's recommendation probability being greater than or equal to a preset probability. In this case, one or more candidate objects with a recommendation probability greater than or equal to the preset probability can be selected from all candidate objects, and the selected object is designated as the target object. Alternatively, the recommendation criterion can be set to the highest probability among all object recommendations. In this case, the candidate object with the highest recommendation probability can be selected from all candidate objects, and the selected object is designated as the target object.

[0111] Step S2082: Recommend the target recommendation object to the target user represented by the target user identifier.

[0112] In one embodiment, any feasible method can be used to push the target recommended object to the target user represented by the target user identifier. For example, the determined target recommended object can be exposed on the page currently visited by the target user represented by the target user identifier to achieve object recommendation for that target user.

[0113] The aforementioned object recommendation method, after obtaining the information features corresponding to the target user identifier and the object features corresponding to the candidate recommendation objects for the target user identifier, obtains multi-level conversion task parameters matching the candidate recommendation objects by sharing and mapping object features and information features. This allows each level of conversion task to share underlying parameters, alleviating the data sparsity problem of higher-level conversion tasks and thus correcting data processing biases. By constructing a Bayesian network to process the multi-level conversion task parameters and obtaining conversion prediction values ​​at each level matching the candidate recommendation objects based on feature matching data of object features and information features, the method can associate multi-level conversion tasks and improve the accuracy of the determined conversion prediction values ​​at each level. By recommending objects to the target user represented by the target user identifier based on the object recommendation probability corresponding to each level of conversion prediction value, it can effectively avoid repeatedly recommending similar objects, thereby improving object recommendation efficiency and further enhancing the user experience.

[0114] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description, in conjunction with the accompanying drawings and a specific embodiment, will further illustrate this application. It should be understood that the specific embodiment described herein is merely illustrative and not intended to limit the scope of this application.

[0115] In one specific embodiment, we take recommending ads to users as an example. For instance, for application download ads, the benefits can be reflected in exposure, clicks, and conversions. These processes involve different stages of conversion behavior, mainly including shallow conversion behaviors such as downloading, installing, and activating, and deep conversion behaviors such as paying and next-day retention. For ease of description, the following specific embodiment uses an ad with three conversion tasks as an example. These conversion tasks specifically include: the first-level conversion task corresponding to ad exposure to clicks, the second-level conversion task corresponding to clicks to shallow conversions, and the third-level conversion task corresponding to shallow conversions to deep conversions. Figure 4 The diagram shows the flowchart of the object recommendation method. The steps of the object recommendation method are as follows:

[0116] If a user accesses a page on the server through a terminal, when the server receives the user's access request, it obtains the information features corresponding to the target user identifier, and matches and determines the candidate advertisements for the target user identifier from a pre-set candidate advertisement dataset based on the target user identifier; wherein, the target user identifier may be the terminal identifier corresponding to the terminal used by the target user, and the candidate advertisements include more than one.

[0117] For each candidate ad, perform the following processing:

[0118] The system acquires information features corresponding to the target user identifier and advertising features corresponding to candidate advertisements for the target user identifier. Information features mainly include the user's basic attributes and behavioral interest features; advertising features mainly include the advertisement ID, advertiser ID, item features corresponding to the advertisement, and contextual features; item features mainly include the category, semantic features, and image features of the items contained in the advertisement; contextual features mainly include the context of the current access request and terminal device features. For high-dimensional features such as the advertisement ID, an embedding layer can be used to map high-dimensional features to low-dimensional features. Specifically, each advertisement ID can be converted using a hash function, with the hash value (hash ID) used as the key and the corresponding feature embedding value used as the value, stored in an embedding table. In this embodiment, all embedding values ​​are initialized before model training, and the values ​​in the embedding table are updated based on the backpropagation gradient during model training.

[0119] After acquiring information features and advertising features, an advertising matching model is used to calculate feature matching data between the information features and advertising features. This feature matching data is used to characterize the target user's interest in the candidate advertisement. For example... Figure 5 As shown in the ad matching model, its structure is a dual-tower structure. It takes ad features and information features as input, and maps them using their respective ad feature mapping models and information feature mapping models to obtain ad-mapped features and user-mapped features. The ad-mapped features and user-mapped features are then multiplied by a dot product, and the result is fed into the mapping processing model to finally determine the feature matching data between ad features and information features. The ad feature mapping model, information feature mapping model, and mapping processing model can all employ a multilayer perceptron (MLP). The feature matching data can be represented as logit0, and the calculation formula is as follows:

[0120] logit0 = P(match_score|x, H)

[0121] Where x represents information features and advertising features, and H represents the parameters of the mapping processing model.

[0122] After acquiring information features and advertising features, based on the aforementioned feature matching data, an advertising ranking model is used to predict the conversion rates for the three levels of conversion tasks for the candidate advertisement. For example... Figure 5 As shown in the advertising ranking model, the advertising ranking model includes: a feature layer, a parameter sharing layer, and a Bayesian network.

[0123] In the feature layer of the ad ranking model, the input information features and ad features (including the features of the illustrated items and context features) are mapped by a feature mapping processing layer; the feature mapping processing layer can use an embedding layer, which corresponds to the embedding features.

[0124] Then, a shared feature mapping processing layer is connected to the feature mapping processing layer. The shared feature mapping processing layer performs shared mapping processing on the object features and information features to obtain the shared object features corresponding to the object features and the shared information features corresponding to the information features. The shared feature mapping processing layer can be a shared multilayer perceptron (MLP).

[0125] Then, according to the transformation tasks at each level, the shared object features and shared information features matching the transformation tasks are concatenated to obtain concatenated features matching the transformation tasks. Feature processing is then performed on each concatenated feature to obtain the transformation task parameters corresponding to each level. Specifically, in the parameter sharing layer, five shared gate structures are set up, denoted as Gate_1, Gate_2, Gate_3, Gate_4, and Gate_5. For the first-level transformation task, the input data when generating the transformation task parameters is the data of Gate_1 and Gate_2. For the second-level transformation task, the input data when generating the transformation task parameters is the data of Gate_1, Gate_2, Gate_3, and Gate_4. For the third-level transformation task, the input data when generating the transformation task parameters is the data of Gate_1, Gate_3, Gate_4, and Gate_5.

[0126] Then, the shared object features and shared information features that match the transformation task are concatenated to obtain the concatenated features that match the transformation task. Further processing is then performed using the first-level, second-level, and third-level transformation task parameter processing models for each level, ultimately outputting the transformation task parameters corresponding to each level. The transformation task parameter processing models for each level can be selected from multilayer perceptron neural networks (MLPs).

[0127] The transformation task parameters corresponding to each level are input into a Bayesian network for mapping processing. Specifically, mapping processing is performed on the transformation task parameters of each level to obtain the mapping parameters for that level. The mapping nodes of the previous and next level transformation tasks in the Bayesian network are connected by directed connection layers. The input data of each mapping node includes the transformation task parameters of the current level and the mapping parameters of the previous level. Specifically, the input data of the mapping node for the first level transformation task is the first level transformation task parameters; the input data of the mapping node for the second level transformation task is the mapping parameters of both the first and second levels; and the input data of the mapping node for the third level transformation task is the mapping parameters of both the second and third levels. The mapping parameters of the first level can be represented as logit1, the mapping parameters of the second level as logit2, and the mapping parameters of the third level as logit3.

[0128] Based on the mapping parameters at each level output by the Bayesian network and the feature matching data of object features and information features, conversion prediction values ​​at each level that match the candidate recommendation objects are obtained. Each conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters. Specifically, the first-level conversion prediction value is represented as CTR (click-through rate prediction), the second-level conversion prediction value is represented as CVR1 (shallow conversion rate prediction), and the third-level conversion prediction value is represented as CVR2 (deep conversion rate prediction). The final conversion prediction values ​​at each level that match the candidate ads output by the ad ranking model are represented as follows:

[0129] ctr = P(ctr|x, H, logit0)

[0130] cvr1=P(cvr1|ctr,X,H,logit0)

[0131] cvr2=P(cvr2|ctr,cvr1,X,H,logit0)

[0132] Based on the conversion estimates at each level of candidate ads, the ad recommendation probability corresponding to each candidate ad is determined. Taking a normal type of ad as an example, it only has two conversion tasks: the first-level conversion task from ad impression to click, and the second-level conversion task from click to shallow conversion. The corresponding ad recommendation probability ecpm is expressed as:

[0133] ecpm = ctr * cvr1 * bid

[0134] Taking multi-objective ads as an example, they have the three levels of conversion tasks mentioned above. When CVR1 > CVR2, the corresponding ad recommendation probability (ECPM) is expressed as:

[0135] ecpm = ctr * cvr1 * bid1

[0136] Otherwise, the corresponding ad recommendation probability ecpm is expressed as:

[0137] ecpm = ctr * cvr2 * bid2

[0138] Where, bid represents the maximum conversion estimate multiplied in the formula, excluding the first-level conversion estimate, and the corresponding level of recommendation benefit. Recommendation benefit can be the bid price. For example, bid1 represents the shallow bid of the candidate ad, and bid2 represents the deep bid of the candidate ad.

[0139] After calculating and determining the recommendation probability of each candidate advertisement, the target advertisement that meets the recommendation criteria is selected from the candidate advertisements. The recommendation criteria can be set according to actual technical needs. In one embodiment, the candidate advertisement with the highest recommendation probability is selected as the target advertisement.

[0140] The target ad is recommended and exposed to the target users represented by the target user identifier.

[0141] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0142] Based on the same inventive concept, this application also provides an object recommendation apparatus 600 for implementing the object recommendation method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method; therefore, the specific limitations of one or more embodiments of the object recommendation apparatus 600 provided below can be found in the limitations of the object recommendation method described above, and will not be repeated here.

[0143] In one embodiment, such as Figure 6 As shown, an object recommendation device 600 is provided, including: a feature acquisition module 610, a parameter determination module 620, a calculation and processing module 630, and an object recommendation module 640, wherein:

[0144] The feature acquisition module 610 is used to acquire the information features corresponding to the target user identifier and the object features corresponding to the candidate recommendation objects for the target user identifier.

[0145] The parameter determination module 620 is used to perform a shared mapping process between the object features and the information features to obtain multi-level conversion task parameters that match the candidate recommendation object.

[0146] The computational processing module 630 is used to input the conversion task parameters of each level into a Bayesian network for mapping processing. Based on the mapping parameters of each level output by the Bayesian network and the feature matching data of the object features and the information features, the conversion prediction value of each level that matches the candidate recommendation object is obtained. The conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters.

[0147] The object recommendation module 640 is used to recommend objects to the target user represented by the target user identifier based on the object recommendation probability corresponding to the conversion prediction value at each level.

[0148] In one embodiment, the parameter determination module 620 is used to perform shared mapping processing on the object features and the information features to obtain shared object features corresponding to the object features and shared information features corresponding to the information features; according to the transformation tasks at each level, the shared object features and the shared information features that match the transformation tasks are processed to obtain the transformation task parameters corresponding to each level.

[0149] In one embodiment, the parameter determination module 620 is used to concatenate the shared object features and the shared information features that match the conversion task according to the conversion task at each level, so as to obtain concatenated features that match the conversion task respectively; and to perform feature processing on each concatenated feature to obtain the conversion task parameters corresponding to each level.

[0150] In one embodiment, the object recommendation device 600 further includes: a feature matching data determination unit; the feature matching data determination unit is configured to perform mapping processing on the object features and the information features respectively to obtain object mapping features and user mapping features; and to perform a dot product operation on the object mapping features and the user mapping features to determine feature matching data between the object features and the information features.

[0151] In one embodiment, the computation processing module 630 is used to perform mapping processing on the conversion task parameters of each level to obtain the mapping parameters of each level; wherein, there is a directed connection between the mapping nodes of the previous level conversion task and the mapping nodes of the next level conversion task in the Bayesian network, and the input data of each mapping node includes the current level conversion task parameters and the previous level mapping parameters; based on the mapping parameters of each level and the feature matching data of the object features and the information features, the conversion prediction value of each level that matches the candidate recommendation object is obtained.

[0152] In one embodiment, the object recommendation device 600 further includes: an object recommendation probability determination unit; the object recommendation probability determination unit is configured to determine the first-level conversion estimate among the conversion estimate values ​​of each level, and the maximum conversion estimate value among the conversion estimate values ​​of each level other than the first-level conversion estimate value; and determine the object recommendation probability of the candidate recommendation object based on the first-level conversion estimate value, the maximum conversion estimate value, and the recommendation effectiveness of the candidate recommendation object corresponding to the maximum conversion estimate value.

[0153] In one embodiment, the object recommendation module 640 is configured to, based on the object recommendation probability of each of the candidate recommendation objects, select target recommendation objects whose object recommendation probability meets the recommendation conditions from the candidate recommendation objects; and recommend the target recommendation objects to the target user represented by the target user identifier.

[0154] Each module in the aforementioned object recommendation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0155] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores object recommendation data, such as information features, object features, multi-level transformation task parameters, feature matching data, transformation predictions at each level, object recommendation probabilities, etc. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements an object recommendation method.

[0156] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements an object recommendation method. The display unit of the computer device is used to form a visually visible image. It can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0157] Those skilled in the art will understand that Figure 7 and Figure 8The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0158] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the object recommendation method described above.

[0159] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the object recommendation method steps described above.

[0160] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the object recommendation method described above.

[0161] It should be noted that the user information (including but not limited to identification information, feature information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0162] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0163] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0164] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An object recommendation method, characterized in that, The method includes: Obtain the information features corresponding to the target user identifier, and the object features corresponding to the candidate recommendation objects for the target user identifier; The object features and the information features are shared and mapped to obtain the shared object features corresponding to the object features and the shared information features corresponding to the information features. According to the transformation tasks at each level, the shared object features and shared information features that match the transformation tasks are processed to obtain the transformation task parameters corresponding to each level. The conversion task parameters at each level are input into a Bayesian network for mapping processing. Based on the mapping parameters at each level output by the Bayesian network and the feature matching data between the object features and the information features, conversion prediction values ​​at each level that match the candidate recommendation object are obtained. The conversion prediction values ​​correspond one-to-one with the conversion tasks indicated by the conversion task parameters. The feature matching data is used to characterize the interest of the target user represented by the target user identifier in the candidate recommendation object. Based on the object recommendation probability corresponding to the conversion prediction value at each level, object recommendations are made to the target user represented by the target user identifier.

2. The method according to claim 1, characterized in that, The process involves performing calculations on the shared object features and shared information features that match the transformation tasks at each level, according to the transformation tasks at each level, to obtain the transformation task parameters corresponding to each level, including: According to the transformation tasks at each level, the shared object features that match the transformation tasks are concatenated with the shared information features to obtain concatenated features that match the transformation tasks. Feature processing is performed on each of the splicing features to obtain the corresponding transformation task parameters for each level.

3. The method according to claim 1, characterized in that, The method for determining the feature matching data between the object features and the information features includes: The object features and the information features are mapped to obtain object mapping features and user mapping features, respectively. Perform a dot product operation between the object mapping feature and the user mapping feature to determine the feature matching data between the object feature and the information feature.

4. The method according to claim 1, characterized in that, The process involves inputting the conversion task parameters at each level into a Bayesian network for mapping. Based on the mapping parameters output by the Bayesian network and the feature matching data between the object features and the information features, conversion prediction values ​​at each level that match the candidate recommendation object are obtained, including: The transformation task parameters at each level are mapped to obtain the mapping parameters at each level. In the Bayesian network, there is a directed connection between the mapping node of the transformation task at the previous level and the mapping node of the transformation task at the next level. The input data of each mapping node includes the transformation task parameters at the current level and the mapping parameters at the previous level. Based on the mapping parameters of each level and the feature matching data of the object features and the information features, the conversion prediction value of each level that matches the candidate recommendation object is obtained.

5. The method according to claim 1, characterized in that, The methods for determining the object recommendation probability corresponding to the conversion prediction value at each level include: Determine the first-level conversion estimate among the conversion estimates of each level, and the maximum conversion estimate among the conversion estimates of each level other than the first-level conversion estimate; Based on the first-level conversion estimate, the maximum conversion estimate, and the recommendation effectiveness of the candidate recommendation object corresponding to the maximum conversion estimate, the object recommendation probability of the candidate recommendation object is determined.

6. The method according to claim 5, characterized in that, The step of recommending objects to the target user represented by the target user identifier based on the object recommendation probability corresponding to the conversion prediction value at each level includes: Based on the object recommendation probability of each of the candidate recommendation objects, target recommendation objects whose object recommendation probability meets the recommendation conditions are selected from the candidate recommendation objects; The target recommendation object is recommended to the target user represented by the target user identifier.

7. An object recommendation device, characterized in that, The device includes: The feature acquisition module is used to acquire the information features corresponding to the target user identifier and the object features corresponding to the candidate recommendation objects for the target user identifier; The parameter determination module is used to perform shared mapping processing on the object features and the information features to obtain the shared object features corresponding to the object features and the shared information features corresponding to the information features; according to the transformation tasks at each level, the shared object features and the shared information features that match the transformation tasks are processed to obtain the transformation task parameters corresponding to each level. The computational processing module is used to input the conversion task parameters of each level into the Bayesian network for mapping processing. Based on the mapping parameters of each level output by the Bayesian network and the feature matching data of the object features and the information features, the conversion prediction value of each level that matches the candidate recommendation object is obtained. The conversion prediction value corresponds one-to-one with the conversion task indicated by the conversion task parameters. The object recommendation module is used to recommend objects to the target user represented by the target user identifier based on the object recommendation probability corresponding to the conversion prediction value at each level.

8. The object recommendation device according to claim 7, characterized in that, The parameter determination module is further configured to, according to the conversion tasks at each level, concatenate the shared object features that match the conversion tasks with the shared information features to obtain concatenated features that match the conversion tasks; and perform feature processing on each concatenated feature to obtain the conversion task parameters corresponding to each level.

9. The object recommendation device according to claim 7, characterized in that, The device further includes a feature matching data determination unit, which is used to perform mapping processing on the object features and the information features respectively to obtain object mapping features and user mapping features; and to perform dot product operation on the object mapping features and the user mapping features to determine feature matching data between the object features and the information features.

10. The object recommendation device according to claim 7, characterized in that, The computational processing module is further configured to perform mapping processing on the conversion task parameters of each level to obtain the mapping parameters of each level; wherein, there is a directed connection between the mapping nodes of the previous level conversion task and the mapping nodes of the next level conversion task in the Bayesian network, and the input data of each mapping node includes the current level conversion task parameters and the previous level mapping parameters; based on the mapping parameters of each level and the feature matching data of the object features and the information features, the conversion prediction values ​​of each level that match the candidate recommendation object are obtained.

11. The object recommendation device according to claim 7, characterized in that, The device further includes an object recommendation probability determination unit, which is used to determine the first-level conversion prediction value among the conversion prediction values ​​of each level, and the maximum conversion prediction value among the conversion prediction values ​​of each level other than the first-level conversion prediction value; and to determine the object recommendation probability of the candidate recommendation object based on the first-level conversion prediction value, the maximum conversion prediction value, and the recommendation effectiveness of the candidate recommendation object corresponding to the maximum conversion prediction value.

12. The object recommendation device according to claim 11, characterized in that, The object recommendation module is further configured to, based on the object recommendation probability of each of the candidate recommendation objects, select target recommendation objects whose object recommendation probability meets the recommendation conditions from the candidate recommendation objects; and recommend the target recommendation objects to the target user represented by the target user identifier.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

15. A computer program product comprising computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 6.

Images