Model training method and member transition probability determination method

Abstract

The invention discloses a model training method and a member transition probability determination method, and the method comprises the steps: obtaining target data corresponding to a target product, the target data comprises function information of at least one historical version, user behavior data of a member user corresponding to each historical version and function information of a version to be released; inputting the target data into a member user feature prediction model to obtain at least one type of member user feature corresponding to the version to be released; and training a member user transition probability prediction model based on the at least one type of member user features and user behavior data of non-member users corresponding to at least one historical version corresponding to the target product.

Description

Technical Field

[0001] The present application belongs to the field of computer technology, and specifically relates to a model training method and a method for determining member conversion probability. Background Art

[0002] With the rapid development of mobile Internet, the role of cloud disk services in users' daily lives has become increasingly prominent. As a convenient file storage and sharing tool, how to attract and retain users, especially converting ordinary users into member (member) users, has become a challenge that cloud disk service providers must face. In related technologies, the prediction of ordinary users converting to member users adopts a decision tree strategy, and when processing large-scale data sets, the tree construction and prediction process may become very complex and run. Since the tree needs to traverse and compare each feature, and may need to build a large tree structure, it consumes a lot of computing resources and time when processing massive data. In addition, when processing data, the tree tends to split nodes based on a single feature, which makes it difficult to capture complex relationships and patterns in the data. In particular, when the data has highly complex local relationships or the data distribution is highly uneven, the tree may not be effectively constructed, resulting in low accuracy in predicting user conversion and high member churn. Summary of the invention

[0003] The embodiments of the present application provide a model training method and a method for determining member conversion probability, which can solve the problem of low prediction accuracy in traditional member conversion probability methods.

[0004] In a first aspect, an embodiment of the present application provides a model training method, the method comprising: obtaining target data corresponding to a target product, wherein the target data includes functional information of at least one historical version, user behavior data of member users corresponding to each of the historical versions, and functional information of a version to be released; obtaining at least one type of member user feature corresponding to the version to be released by inputting the target data into a member user feature prediction model; and training a member user conversion probability prediction model based on the at least one type of member user feature and user behavior data of non-member users corresponding to at least one historical version corresponding to the target product. In a second aspect, an embodiment of the present application provides a method for determining a member conversion probability, the method comprising: obtaining first behavior data of a user to be predicted; inputting the first behavior data into a member user conversion probability prediction model to obtain a conversion probability of the user to be predicted to be converted into each type of member user, wherein the member user conversion probability prediction model is obtained based on the model training method described in the first aspect; and outputting the conversion probability of the user to be predicted to be converted into each type of member user.

[0005] In a third aspect, an embodiment of the present application provides a model training device, which includes: an acquisition module for acquiring target data corresponding to a target product, wherein the target data includes functional information of at least one historical version, user behavior data of member users corresponding to each of the historical versions, and functional information of a version to be released; a processing module for obtaining at least one type of member user feature corresponding to the version to be released by inputting the target data into a member user feature prediction model; and a training module for training a member user conversion probability prediction model based on the at least one type of member user feature and user behavior data of non-member users corresponding to at least one historical version corresponding to the target product.

[0006] In a fourth aspect, an embodiment of the present application provides a device for determining member conversion probability, the device comprising: an acquisition module for acquiring first behavior data of a user to be predicted; a processing module for inputting the first behavior data into a member user conversion probability prediction model to obtain a conversion probability of the user to be predicted to be converted into each type of member user, wherein the member user conversion probability prediction model is obtained based on the model training method described in the first aspect; and an output module for outputting the conversion probability of the user to be predicted to be converted into each type of member user.

[0007] In a fifth aspect, an embodiment of the present application provides an electronic device, comprising a processor, a memory, and a program or instruction stored in the memory and executable on the processor, wherein the program or instruction, when executed by the processor, implements the steps of the method described in the first aspect, or implements the steps of the method described in the second aspect.

[0008] In a sixth aspect, an embodiment of the present application provides a computer-readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the method described in the first aspect are implemented, or the steps of the method described in the second aspect are implemented.

[0009] In the seventh aspect, an embodiment of the present application provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the method described in the first aspect, or to implement the steps of the method described in the second aspect.

[0010] In an eighth aspect, an embodiment of the present application provides a computer program product, which includes at least one computer program, and when the computer program is loaded and executed by a processor, it implements the method described in the first aspect, or implements the steps of the method described in the second aspect.

[0011] In an embodiment of the present application, by acquiring target data corresponding to a target product, wherein the target data includes functional information of at least one historical version, user behavior data of member users corresponding to each historical version, and functional information of the version to be released, and then by inputting the target data into a member user feature prediction model, at least one type of member user feature corresponding to the version to be released is obtained, and finally a member user conversion probability prediction model is trained based on at least one type of member user feature and user behavior data of non-member users corresponding to at least one historical version corresponding to the target product, that is, the user features of the member users are first predicted, and then the member user conversion probability prediction model is trained according to the predicted member user features and the user behavior data of the non-member users, and subsequently the probability of non-member users becoming different types of member users can be predicted by the trained member user conversion probability prediction model. In this way, since the feature form of the member users is not pre-defined, that is, the member user features are not pre-defined, the problem that the probability of non-member users becoming member users cannot be accurately predicted according to the pre-set member user features when the product functions or versions change can be avoided. In an embodiment of the present application, the prediction accuracy of the trained member user conversion probability prediction model is higher, and thus a more accurate marketing strategy can be formulated, thereby improving the user conversion rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is a flowchart of a model training method provided in an embodiment of the present application; Figure 2 It is a schematic diagram of the architecture of a member user feature prediction model provided in an embodiment of the present application; Figure 3 It is a schematic diagram of the architecture of a member user feature prediction model provided in an embodiment of the present application; Figure 4 is a schematic diagram of the architecture of an encoder provided in an embodiment of the present application; Figure 5 is a schematic diagram of the architecture of an encoder provided in an embodiment of the present application; Figure 6 is a schematic diagram of the architecture of an encoder provided in an embodiment of the present application; Figure 7a is a schematic diagram of the architecture of a first decoder provided in an embodiment of the present application; Figure 7b is a schematic diagram of the architecture of a first decoder provided in an embodiment of the present application; Figure 8 is a schematic diagram of the architecture of a first decoder provided in an embodiment of the present application; Fig. 9 is a schematic diagram of the architecture of a first decoder provided in an embodiment of the present application; Fig.10 is a schematic diagram of the architecture of a second decoder provided in an embodiment of the present application; Fig.11a is a schematic diagram of the architecture of a second decoder provided in an embodiment of the present application; Fig.11b is a schematic diagram of the architecture of a second decoder provided in an embodiment of the present application; Fig.12a is a schematic diagram of the architecture of a second decoder provided in an embodiment of the present application; Figure 12b It is a schematic diagram of the architecture of a cross-attention layer provided in an embodiment of the present application; Fig.12c is a schematic diagram of the architecture of a second decoder provided in an embodiment of the present application; Fig.13 It is a flowchart of a method for determining a member conversion probability provided by an embodiment of the present application; Fig.14 It is a structural schematic diagram of a model training device provided in an embodiment of the present application; Fig.15 It is a structural schematic diagram of a member conversion probability determination device provided in an embodiment of the present application; Fig.16 It is a schematic diagram of the structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0013] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0014] In conjunction with the accompanying drawings, a model training method and a method for determining a member conversion probability provided by an embodiment of the present application are described in detail through specific embodiments and their application scenarios.

[0015] Figure 1 A model training method provided by an embodiment of the present application is shown, and the method can be executed by an electronic device, and the electronic device may include: a server and / or a terminal device. In other words, the method can be executed by software or hardware installed in the electronic device, and the method includes the following steps: S110: Acquire target data corresponding to the target product.

[0016] The target data includes functional information of at least one historical version, user behavior data of member users corresponding to each historical version, and functional information of the version to be released.

[0017] Among them, the user behavior data refers to various operation records and interaction records generated by users in the process of using the target product. For example, taking the target product as a cloud disk, the user behavior data may include click data, access time, usage frequency, download and upload records, etc., wherein the click data may include data such as the function clicked by the user, and the access time may include the time the user stays on a specific page or application, etc. The usage frequency may include the frequency of users accessing or using a certain function, etc. The download and upload records may include the number and size of files downloaded or uploaded by users, etc.

[0018] S120: Obtain at least one type of member user feature corresponding to the version to be released by inputting the target data into a member user feature prediction model.

[0019] S130: Training a member user conversion probability prediction model based on the at least one type of member user characteristics and the user behavior data of non-member users corresponding to at least one historical version corresponding to the target product.

[0020] It can be understood that in the above S120 and S130, if the probability of an ordinary user becoming a member user is to be predicted, the member user characteristics of the member user need to be determined. In the related art, the member user characteristics of the target product are usually assumed to be fixed. Taking the cloud disk as an example, the member user characteristics of the cloud disk are usually assumed to include having a larger cloud storage space, having a faster upload and download speed, enjoying an exclusive customer service channel or priority technical support, enjoying some additional functions and privileges, such as customized personalized settings, having advanced data management functions, or being able to enjoy more file sharing and collaboration functions. However, in actual applications, with the continuous iteration and update of the functions of the target product, the permissions enjoyed by the member users of the target product will also change. That is, under target products with different functions / versions, the characteristic forms of the member users of the target product are different, which will change with the changes in the product functions or versions of the target product. Therefore, in this case, if the default member user characteristics in the related art are directly compared with the characteristics of ordinary users to predict the probability of ordinary users becoming member users, it is easy to lead to inaccurate prediction results.

[0021] In order to avoid this problem, in the embodiment of the present application, the user characteristics of the member user are first predicted, and then the predicted member user characteristics and the user characteristics of the ordinary user to be evaluated are used as the training data of the member user conversion probability prediction model for model training, and then the probability of an ordinary user becoming a member user is predicted by the trained member user conversion probability prediction model. In this way, since the characteristic form of the member user is not limited in advance, that is, the member user characteristics are not limited in advance, it can avoid the problem in the related art that the probability of an ordinary user becoming a member user cannot be accurately predicted according to the pre-set member user characteristics when the product function or version of the target product changes.

[0022] It should be noted that, in the embodiment of the present application, the target data may include at least one historical version, each historical version may correspond to an ordinary user and at least one member user, and each member user has different member user characteristics.

[0023] In an embodiment of the present application, by acquiring target data corresponding to a target product, wherein the target data includes functional information of at least one historical version, user behavior data of member users corresponding to each historical version, and functional information of the version to be released, and then by inputting the target data into a member user feature prediction model, at least one type of member user feature corresponding to the version to be released is obtained, and finally a member user conversion probability prediction model is trained based on at least one type of member user feature and user behavior data of non-member users corresponding to at least one historical version corresponding to the target product, that is, the user features of the member users are first predicted, and then the member user conversion probability prediction model is trained according to the predicted member user features and the user behavior data of the non-member users, and subsequently the probability of non-member users becoming different types of member users can be predicted by the trained member user conversion probability prediction model. In this way, since the feature form of the member users is not pre-defined, that is, the member user features are not pre-defined, the problem that the probability of non-member users becoming member users cannot be accurately predicted according to the pre-set member user features when the product functions or versions change can be avoided. In an embodiment of the present application, the prediction accuracy of the trained member user conversion probability prediction model is higher, and thus a more accurate marketing strategy can be formulated, thereby improving the user conversion rate.

[0024] In an exemplary embodiment, illustratively, Figure 2 As shown, the member user feature prediction model includes an encoder, a first decoder and a second decoder, such as Figure 3 As shown, it is another structural schematic diagram of the member user feature prediction model; by inputting the target data into the member user feature prediction model, obtaining at least one type of member user feature corresponding to the version to be released, which may include the following steps: S122: Perform feature characterization processing on the target data through the encoder to obtain a first sampling result set and a first target feature, wherein the first sampling result set includes N-1 first target sampling results, and N is an integer greater than or equal to 2.

[0025] It can be understood that the encoder is used to perform feature representation processing on the target data, that is, to convert the input data into a more compact and representative representation, and to retain the most meaningful features of the input data from the perspective of member user detection.

[0026] S124: Based on the first sampling result set and the first target feature, the first decoder performs a first reconstruction process to obtain a second sampling result set and a second target feature, wherein the second sampling result set includes N-1 second target sampling results.

[0027] The first reconstruction process refers to reconstructing the input data of the member user feature prediction model based on the first sampling result set and the first target feature, that is, reconstructing the target data.

[0028] It is understandable that, for the version to be released, since there is no user behavior data for this version, it is impossible to determine the characteristic form of the member user under this version, which makes it impossible to predict whether ordinary users will become member users under the version to be released. In the embodiment of the present application, after obtaining the more compact and representative coding features output by the encoder, the functional information of the target product under each version and the user behavior data of the user under each version can be further reconstructed through the first decoder.

[0029] S126: Based on the first sampling result set, the second sampling result set, the first target feature, and the second target feature, a second reconstruction process is performed by the second decoder to obtain at least one type of member user feature corresponding to the version to be released.

[0030] It can be understood that the second reconstruction process refers to predicting member user features based on the first sampling result set, the second sampling result set and the second feature data. That is, after obtaining the functional information of each reconstructed version and the user behavior data of the user under each historical version, the member user features corresponding to the version to be released can be predicted by the second encoder.

[0031] In an exemplary embodiment, if Figure 4As shown, the encoder includes N first sampling modules and N-1 residual modules connected in sequence, wherein the N-1 residual modules correspond one-to-one to the first N-1 first sampling modules, and each of the residual modules includes a different number of convolution filters; the encoder performs feature characterization processing on the target data to obtain a first sampling result set and a first target feature, including: encoding the target data through N first sampling modules in sequence to obtain the first target feature output by the Nth first sampling module and the first sampling result set corresponding to the N-1 residual modules, wherein the first sampling result set includes the first target sampling result output by each residual module.

[0032] It can be understood that the input data is first input to the first first sampling module, and the result output by the first first sampling module is simultaneously transmitted to the corresponding residual module and the next first sampling module until the Nth sampling module. The Nth first sampling module only outputs the first target feature, that is, the first sampling result set includes the first sampling results output by N-1 residual modules, and the first target feature is output by the Nth first sampling module. In an embodiment of the present application, different sampling modules can process different aspects or levels of the input data, i.e., the target data. For example, the first first sampling module can focus on the global features of the target data, while the second first sampling module focuses on the local details of the target data, etc. In this way, multiple levels of features can be captured through multi-scale processing, thereby obtaining a richer representation.

[0033] Among them, the residual module is used to solve the problem of gradient disappearance. It is understandable that during the model training process, if there is no residual processing, the target data may lose some key information after being processed by multiple first sampling modules. By collecting the input information of each first sampling module through the residual module, it can be ensured that the input information of each first sampling module will not be lost in the deep network, thereby solving the problem of gradient disappearance.

[0034] Further, in another exemplary embodiment, encoding the target data through N first sampling modules in sequence to obtain the first target feature output by the Nth first sampling module and the first sampling result set corresponding to N-1 residual modules may include the following steps: Step 1: For the first first sampling module, the target data is encoded to obtain the first first initial sampling result and the first first encoding feature output by the first first sampling module, and the first initial sampling result is processed by the residual module corresponding to the first first sampling module to obtain the first first target sampling result.

[0035] Step 2: For the Qth first sampling module, encode the Q-1th first coding feature output by the Q-1th first sampling module to obtain the Qth first initial sampling result and the Qth first coding feature output by the Qth first sampling module, and process the Qth first initial sampling result through the residual module corresponding to the Qth first sampling module to obtain the Qth first target sampling result, where Q is an integer greater than or equal to 2 and less than N.

[0036] Step 3: For the Nth first sampling module, encode the N-1th first coding feature output by the N-1th first sampling module to obtain the first target feature output by the Nth first sampling module.

[0037] It can be understood that the first first sampling module transmits the first first initial sampling result to the corresponding residual module for processing to obtain the first first target sampling result, and transmits the first first coding feature to the next first sampling module, that is, the second first sampling module. Any first sampling module from the second to the N-1th receives the first coding feature output by the previous sampling module, and obtains its own corresponding first initial sampling result and first coding based on the first coding feature output by the previous sampling module, and transmits the obtained first initial sampling result to the corresponding residual module for processing to obtain the corresponding first target sampling result, and transmits the obtained first coding feature to the next first sampling module. The Nth first sampling module receives the N-1th first coding feature output by the N-1th first sampling module, and processes the N-1th first coding feature to obtain the first target feature.

[0038] In an exemplary embodiment, illustratively, as Figure 5 As shown, the first first sampling module includes a first convolution layer, a first normalization layer, a first activation layer and a maximum pooling layer connected in sequence; the other first sampling modules except the first first sampling module among the N first sampling modules include the first convolution layer, the first normalization layer, the first activation layer, the second convolution layer, the second normalization layer, the first residual layer and the second activation layer connected in sequence.

[0039] based on Figure 5 In the illustrated embodiment, the specific processing processes of the 1st first sampling module, the 2nd first sampling module to the N-1th first sampling module, and the Nth first sampling module are schematically described below.

[0040] (1) In an exemplary embodiment, the encoding process is performed on the target data to obtain the first initial sampling result and the first encoding feature output by the first sampling module, including: Obtaining a first second encoding feature by inputting the target data into the first convolutional layer; By inputting the first second encoding feature into the first normalization layer, a second second encoding feature is obtained; By inputting the second second coding feature into the first activation layer, a third second coding feature and a first first initial sampling result are obtained; By inputting the third second encoding feature into the maximum pooling layer, the first first encoding feature is obtained.

[0041] (2) In an exemplary embodiment, the encoding process is performed on the Q-1th first coding feature output by the Q-1th first sampling module to obtain the Qth first initial sampling result and the Qth first coding feature output by the Qth first sampling module, including: Obtaining a first third encoding feature by inputting the Q-1th first encoding feature into the first convolutional layer; Obtaining a second third encoding feature by inputting the first third encoding feature into the first normalization layer; By inputting the second third coding feature into the first activation layer, a third third coding feature and a Qth first initial sampling result are obtained; By inputting the third third encoding feature into the second convolutional layer, a fourth third encoding feature is obtained; By inputting the fourth third encoding feature into the second normalization layer, a fifth third encoding feature is obtained; By inputting the fifth third coding feature and the Q-1th first coding feature into the first residual layer, a sixth third coding feature is obtained; By inputting the 6th third coding feature into the second activation layer, the Qth first coding feature is obtained.

[0042] (3) In an exemplary embodiment, the step of encoding the N-1th first encoding feature output by the N-1th first sampling module to obtain the first target feature output by the Nth first sampling module includes: Obtaining a first fourth coding feature by inputting the N-1th first coding feature into the first convolutional layer; Obtaining a second fourth coding feature by inputting the first fourth coding feature into the first normalization layer; By inputting the second fourth coding feature into the first activation layer, a third fourth coding feature and an Nth first initial sampling result are obtained; By inputting the third fourth encoding feature into the second convolutional layer, a fourth fourth encoding feature is obtained; By inputting the fourth fourth encoding feature into the second normalization layer, a fifth fourth encoding feature is obtained; Obtaining a sixth fourth coding feature by inputting the fifth fourth coding feature and the N-1th first coding feature into the first residual layer; The first target feature is obtained by inputting the sixth fourth coding feature into the second activation layer.

[0043] That is to say, when N is any integer greater than or equal to 2, the result of the output of the first activation layer of the first first sampling module is input into the corresponding residual module, and the result of the output of the maximum pooling layer of the first first sampling module is respectively input into the first convolution layer and the first residual layer of the next first sampling module; the result of the output of the second activation layer of any first sampling module from the 2nd to the N-1th is simultaneously input into the corresponding residual module and the first convolution layer and the first residual layer of the next first sampling module; the first convolution layer and the first residual layer of the Nth first sampling module receive the result of the output of the second activation layer of the previous first sampling module, and output the first target feature through the second activation layer. Exemplarily, assuming that N is 5, such as Figure 6 As shown, it is another structural schematic diagram of the encoder, wherein the result of the first activation layer output of the first first sampling module is input into the corresponding residual module, the result of the maximum pooling layer output of the first first sampling module is input into the first convolution layer and the first residual of the second first sampling module, the result of the second activation layer output of the second first sampling module is simultaneously input into the corresponding residual module and the first convolution layer and the first residual layer of the third first sampling module, the result of the second activation layer output of the third first sampling module is simultaneously input into the corresponding residual module and the first convolution layer and the first residual layer of the fourth first sampling module; the result of the second activation layer output of the fourth first sampling module is simultaneously input into the corresponding residual module and the first convolution layer and the first residual layer of the fifth first sampling module, and the result of the second activation layer output of the fifth first sampling module is the first target feature. Among them, in another exemplary embodiment, assuming that N is 5, the first residual module includes three staggered blocks, each residual block includes 64 convolution filters; the second residual module includes four residual blocks, each residual block includes 128 filters; the third residual module includes six residual blocks, each residual block includes 256 filters; the fourth residual module includes three residual blocks, each residual block includes 512 filters. Further, in order to reduce the spatial dimension of the input data, in the embodiment of the present application, the transition between each residual module uses a stepped convolution, so that the encoder can generate a first sampling result set based on the input data of the member user feature prediction model, and the first sampling result set has a reduced spatial dimension and a gradually complex representation:

[0044] in, represents the first sampling result set, represents the first initial sampling result output by the first residual module, represents the first initial sampling result output by the second residual module, represents the first initial sampling result output by the third residual module, Represents the first initial sampling result output by the 4th residual module.

[0045] In an exemplary embodiment, if Figure 7a As shown, the first decoder includes N second sampling modules connected in sequence, N-1 first-stage decoding modules and a first reconstruction module, wherein the N-1 first-stage decoding modules correspond one-to-one to the first N-1 second sampling modules, and each second sampling module includes the same number of convolution filters; based on the first sampling result set and the first target feature, the first reconstruction processing is performed by the first decoder to obtain the second sampling result set and the second target feature, including: sequentially decoding the first sampling result set and the first target feature through the N second sampling modules to obtain the Nth first decoding feature output by the Nth second sampling module and the second sampling result set corresponding to the N-1 first-stage decoding modules, wherein the second sampling result set includes the second target sampling result output by each first-stage decoding module; the first reconstruction processing is performed on the Nth first decoding feature through the first reconstruction module to obtain the second target feature.

[0046] It can be understood that the N second sampling modules are used for decoding processing, and the first reconstruction module is used for reconstructing the Nth first decoding feature output by the Nth second sampling module and re-projecting it onto the original domain.

[0047] Alternatively, if Figure 7bAs shown, the member user feature prediction model also includes a Dropout noise processing module, which is located between the encoder and the first decoder, that is, one end of the Dropout noise processing module is connected to the N-1 residual modules in the encoder, and the other end of the Dropout noise processing module is connected to the N-1 second sampling modules in the first decoder. It can be understood that the Dropout noise processing module enables the member user feature prediction model to better process unseen data by randomly discarding neurons, thereby enhancing the robustness of the model.

[0048] In the embodiment of the present application, the first stage decoding module has the same function as the residual module, and is also used to retain more information to avoid information loss, but the first stage decoding module is used in the decoding stage, and the residual module is used in the encoding stage. It can be understood that when mapping the target data to a low-dimensional representation in the encoding stage, a lot of information may have been compressed or lost, and the compressed or lost information needs to be restored in the decoding stage, and the first sampling result set output by the residual module can directly pass part of the information, so that the decoder can more easily reconstruct the details. At the same time, in the decoding stage, the input information of the corresponding second sampling module is also collected through each first-stage decoding module to ensure that the input information of each second sampling module will not be lost in the deep network to solve the problem of gradient disappearance.

[0049] Further, in another exemplary embodiment, the decoding process of the first sampling result set and the first target feature by the N second sampling modules in sequence to obtain the Nth first decoding feature output by the Nth second sampling module and the second sampling result set corresponding to the N-1 first-stage decoding modules may include the following steps: Step 1: For the first second sampling module, by decoding the first target feature and the N-1th first target sampling result in the first sampling result set, the first first decoding feature output by the first second sampling module and the first second target sampling result output by the first first stage decoding module are obtained.

[0050] Step 2: For the Mth second sampling module, decode the M-1th first decoding feature output by the M-1th second sampling module and the NMth first target sampling result in the first sampling result set to obtain the Mth first decoding feature output by the Mth second sampling module and the Mth second target sampling result output by the Mth first-stage decoding module, where M is an integer greater than or equal to 2 and less than N.

[0051] Step 3: For the Nth second sampling module, decode the N-1th first decoding feature output by the N-1th second sampling module to obtain the Nth first decoding feature output by the Nth second sampling module.

[0052] It can be understood that the first second sampling module transmits the first second initial sampling result to the corresponding first-stage decoding module for processing to obtain the first second target sampling result, and transmits the first second coding feature to the next second sampling module, that is, the second second sampling module. Any second sampling module from the second to the N-1th receives the second coding feature output by the previous second sampling module, and obtains its own corresponding second initial sampling result and first coding based on the second coding feature output by the previous sampling module, and transmits the obtained second initial sampling result to the corresponding first-stage decoding module for processing to obtain the corresponding second target sampling result, and transmits the obtained second coding feature to the next second sampling module. The Nth second sampling module receives the N-1th second coding feature output by the N-1th second sampling module, and processes the N-1th second coding feature to obtain the first target feature.

[0053] In an exemplary embodiment, if Figure 8 As shown, each of the second sampling modules includes a first upsampling layer and at least one first combination layer, and the first N-1 second sampling modules also include a first series layer, wherein the first upsampling layer in the first N-1 second sampling modules is connected to at least one first combination layer through the first series layer, and the first combination layer includes a third convolutional layer, a third normalization layer, and a third activation layer connected in sequence.

[0054] based on Figure 8 In the illustrated embodiment, the specific processing procedures of the 1st second sampling module, the 2nd second sampling module to the N-1th second sampling module, and the Nth second sampling module are schematically described below.

[0055] (1) In an exemplary embodiment, the decoding process of the first target feature and the N-1th first target sampling result in the first sampling result set to obtain the first first decoding feature output by the first second sampling module and the first second target sampling result output by the first first stage decoding module includes: Obtaining a first second decoding feature by inputting the first target feature into the first upsampling layer; By inputting the first second decoding feature and the N-1th first target sampling result into the first series connection layer, a second second decoding feature is obtained; The second second decoding feature is decoded in sequence through at least one of the first combination layers to obtain the first second target sampling result and the first first decoding feature.

[0056] (2) In an exemplary embodiment, the decoding process is performed on the M-1th first decoding feature output by the M-1th second sampling module and the NMth first target sampling result in the first sampling result set to obtain the Mth first decoding feature output by the Mth second sampling module and the Mth second target sampling result output by the Mth first-stage decoding module, including: By inputting the M-1th first decoding feature into the first upsampling layer, a first third decoding feature is obtained; By inputting the first third decoding feature and the NMth first target sampling result into the first series connection layer, a second third decoding feature is obtained; The second third decoding feature is decoded in sequence through at least one of the first combination layers to obtain the Mth second target sampling result and the Mth first decoding feature.

[0057] (3) In an exemplary embodiment, the decoding process is performed on the N-1th first decoding feature output by the N-1th second sampling module to obtain the Nth first decoding feature output by the Nth second sampling module, including: Obtaining a first fourth decoding feature by inputting the N-1th first decoding feature into the first upsampling layer; The first fourth decoding feature is decoded sequentially through at least one of the first combination layers to obtain the Nth first decoding feature.

[0058] For example, assuming N is 5, Fig. 9As shown, it is another structural schematic diagram of the first decoder, wherein the first target feature is first input into the first upsampling layer of the first second sampling module for upsampling processing to obtain the first second decoding feature, and then the first second decoding feature and the fourth first target sampling result in the first sampling result set are input into the first series connection layer to obtain the second second decoding feature, so as to selectively perturb the decoding feature, thereby improving the quality of the input data, and then the second second decoding feature is decoded through the two first combination layers in sequence to obtain the first second target sampling result and the first first decoding feature; and then the first first decoding feature is input into the first upsampling layer of the second second sampling module. , get the first third decoding feature, then input the first third decoding feature and the third first target sampling result in the first sampling result set into the first series layer, get the second third decoding feature, then decode the second third decoding feature through two first combination layers in turn, get the second second target sampling result and the second third decoding feature; the processing of the third second sampling module and the fourth second sampling module is the same as the second second sampling module, which will not be repeated here; the input of the first upsampling layer of the fifth second sampling module is the fourth third decoding feature, then decode the fourth third decoding feature through two first combination layers in turn, get the Nth first decoding feature. Among them, the second sampling result set can be expressed by the following formula:

[0059] in, represents the first sampling result set, represents the first target feature, represents the second initial sampling result corresponding to the fourth first-stage decoding module, represents the second initial sampling result corresponding to the third first-stage decoding module, represents the second initial sampling result corresponding to the second first-stage decoding module, Represents the second initial sampling result corresponding to the first first-stage decoding module.

[0060] In another exemplary embodiment, Fig.10 As shown, the first reconstruction module includes at least one second combination layer, a third combination layer, a second residual layer, a fourth activation layer, a fourth convolutional layer and a fifth activation layer connected in sequence, the second combination layer includes a fifth convolutional layer, a fourth normalization layer and a fifth activation layer, and the third combination layer includes a sixth convolutional layer and a fifth normalization layer; The performing a first reconstruction process on the Nth first decoding feature by the first reconstruction module to obtain the second target feature includes: Decoding the Nth first decoding feature through at least one of the second combination layers in sequence to obtain a first fifth decoding feature; Obtaining a second fifth decoding feature by inputting the first fifth decoding feature into the third combination layer; Obtaining a third fifth decoding feature by inputting the second fifth decoding feature into the second residual layer; Obtaining a fourth fifth decoding feature by inputting the third fifth decoding feature into the fourth activation layer; Obtaining a fifth fifth decoding feature by inputting the fourth fifth decoding feature into the fourth convolutional layer; The second target feature is obtained by inputting the 5th fifth decoding feature into the fifth activation layer.

[0061] It can be understood that the first decoding feature output by the Nth second sampling module continues to be input into the second combination layer of the first reconstruction module, and then is processed in sequence through the third combination layer, the second residual layer, the fourth activation layer, the fourth convolutional layer and the fifth activation layer, so as to obtain the first decoding feature and reproject it onto the original domain.

[0062] In an embodiment corresponding to the first decoder, each of the second sampling modules includes the same number of convolution filters to gradually rearrange the encoding features obtained from the encoder into a new vector representation. The first decoder in the embodiment of the present application can generate decoder features that are symmetrical with the encoder, which can then be used together by the second decoder segment in the next stage.

[0063] In an exemplary embodiment, if Fig.11a As shown, the second decoder includes N third sampling modules connected in sequence and a second reconstruction module, wherein the N-1 second-stage decoding modules correspond one-to-one to the first N-1 third sampling modules, and each of the third sampling modules includes the same number of convolution filters; The second reconstruction process is performed by the second decoder based on the first sampling result set, the second sampling result set, the first target feature, and the second target feature to obtain at least one type of member user feature corresponding to the version to be released, including: Decoding the first sampling result set, the second sampling result set, the first target feature, and the second target feature in sequence through N third sampling modules to obtain an Nth sixth decoding feature output by the Nth third sampling module; The second reconstruction module performs a second reconstruction process on the Nth sixth decoded feature to obtain at least one type of member user feature corresponding to the version to be released.

[0064] It is understandable that the second decoder is used to predict each type of member user feature based on the features extracted by the encoder and the first decoder. The second decoder enhances the second decoder by combining the high-level features of the encoder, i.e., each first target sampling result in the first sampling result set, with the fine-grained details of the first decoder, i.e., each second target sampling result in the second sampling result set, thereby ensuring that the second decoder can handle more complex and diverse member user feature detection conditions. Alternatively, if Fig.11b As shown, the member user feature prediction model also includes: a transfer module, which is located between the first decoder and the second decoder, that is, one end of the transfer module is connected to the N-1 first-stage decoding modules of the first decoder, and the other end of the transfer module is connected to the N-1 third sampling modules in the second decoder.

[0065] Further, in another exemplary embodiment, the first sampling result set, the second sampling result set, the first target feature, and the second target feature are decoded in sequence by N third sampling modules to obtain an Nth sixth decoding feature output by the Nth third sampling module, including: For the first third sampling module, by decoding the first target feature, the second target feature, the N-1th first target sampling result in the first sampling result set, and the N-1th second target sampling result in the second sampling result set, the first sixth decoding feature output by the first third sampling module is obtained; For the Lth third sampling module, the Lth sixth decoding feature output by the Lth third sampling module is obtained by decoding the L-1th sixth decoding feature output by the L-1th third sampling module, the NLth first target sampling result in the first sampling result set, and the NLth second target sampling result in the second sampling result set, where L is an integer greater than or equal to 2 and less than N; For the Nth third sampling module, the Nth sixth decoding feature output by the Nth third sampling module is obtained by decoding the N-1th sixth decoding feature output by the Nth third sampling module.

[0066] Figure 11a-Figure 11b The processing of the second decoder in the illustrated embodiment is similar to Figure 7a and 7b The processing process of the first decoder shown is similar, and can be referred to Figure 7a and 7b The illustrated embodiments will not be described in detail here.

[0067] In an exemplary embodiment, if Fig.12a As shown, each of the third sampling modules includes a second upsampling layer and at least one fourth combination layer, and the first N-1 third sampling modules also include a second series layer and a cross attention layer, wherein the second upsampling layers in the first N-1 second sampling modules are sequentially connected to at least one fourth combination layer through the second series layer and the cross attention layer, and the fourth combination layer includes a seventh convolutional layer, a sixth normalization layer and a sixth activation layer connected in sequence.

[0068] based on Fig.12a In the illustrated embodiment, the specific processing processes of the 1st third sampling module, the 2nd third sampling module to the N-1th third sampling module, and the Nth third sampling module are schematically described below.

[0069] (1) In an exemplary embodiment, the first sixth decoding feature output by the first third sampling module is obtained by decoding the first target feature, the N-1th first target sampling result in the first sampling result set, and the N-1th second target sampling result in the second sampling result set, including: By inputting the second target feature into the second upsampling layer, a first seventh decoding feature is obtained; By inputting the first seventh decoding feature, the N-1th first target sampling result and the N-1th second target sampling result into the second series connection layer, a second seventh decoding feature is obtained; By inputting the second seventh decoding feature into the cross attention layer, a third seventh decoding feature is obtained; The third of the seventh decoding features is decoded in sequence through at least one of the fourth combination layers to obtain the first of the sixth decoding features.

[0070] (2) In an exemplary embodiment, the decoding process of the L-1th sixth decoding feature output by the L-1th third sampling module, the NLth first target sampling result in the first sampling result set, and the NLth second target sampling result in the second sampling result set to obtain the Lth sixth decoding feature output by the Lth third sampling module includes: Obtaining a first eighth decoding feature by inputting the L-1th sixth decoding feature into the second upsampling layer; By inputting the first eighth decoding feature, the NLth first target sampling result and the NLth second target sampling result into the second series connection layer, a second eighth decoding feature is obtained; By inputting the second eighth decoding feature into the cross attention layer, a third eighth decoding feature is obtained; The third eighth decoding feature is decoded in sequence through at least one of the fourth combination layers to obtain the Lth sixth decoding feature.

[0071] (3) In an exemplary embodiment, the decoding process is performed on the N-1th sixth decoding feature output by the N-1th third sampling module to obtain the Nth sixth decoding feature output by the Nth third sampling module, including: By inputting the N-1th sixth decoding feature into the second upsampling layer, a first ninth decoding feature is obtained; The first of the ninth decoding features is decoded sequentially through at least one of the fourth combination layers to obtain the Nth of the sixth decoding features.

[0072] In the embodiment of the present application, the architecture of the second decoder is similar to that of the first decoder, both of which include five upsampling stages. However, the second decoder and the first decoder have the following differences: (1) The features learned by the first decoder are input into the concatenation layer (Cat) after being processed by the transport block.

[0073] (2) The second decoder integrates a cross-attention layer, allowing the second decoder to process different feature sequences to further enhance the robustness of member user feature prediction. For example, Figure 12b As shown in Figure 1, it is a schematic diagram of the network structure of the cross attention layer.

[0074] (4) A second decoder, further comprising a final convolution, identified as a second reconstruction module, to generate a responsive member user feature mask.

[0075] Exemplarily, the entire process of the second decoder can be expressed by the following formula:

[0076] In an exemplary embodiment, the second reconstruction module may include an eighth convolutional layer, exemplarily, such as Fig.12c shown.

[0077] In an exemplary embodiment, before obtaining at least one type of member user feature corresponding to the version to be released by inputting the target data into the member user feature prediction model, the method further includes: obtaining a target data set, wherein the target data set includes functional information of at least one of the historical versions corresponding to multiple products, user behavior data of at least one member user corresponding to each of the historical versions, and member user features of each member; and training the member user feature prediction model based on the target data set and a first loss function.

[0078] That is, before using the member user feature prediction model to perform member user feature prediction, the member user feature prediction model is first trained based on the target data set and the first loss function.

[0079] In an exemplary embodiment, the training of the member user conversion probability prediction model based on the at least one type of member user characteristics and the user behavior data of non-member users corresponding to at least one historical version of the target product includes: determining a training set based on the at least one type of member user characteristics and the user behavior data of non-member users corresponding to at least one historical version of the target product, wherein the training set includes a plurality of behavior sample data, each of the behavior sample data corresponds to at least one member label; and training the member user conversion probability prediction model based on the plurality of behavior sample data.

[0080] In this embodiment, the member user conversion probability prediction model is trained based on the training set determined based on the at least one type of member user feature and the user behavior data of non-member users corresponding to at least one historical version corresponding to the target product, so that the trained member user conversion probability prediction model can predict the probability of ordinary users becoming member users with higher accuracy. Because, since the feature form of member users is not limited in advance, that is, the member user features are not limited in advance, it can avoid the problem in the related art that the probability of ordinary users becoming member users cannot be accurately predicted according to the pre-set member user features when the product function or version of the target product changes.

[0081] Among them, in another exemplary embodiment, the training of the member user conversion probability prediction model based on the multiple behavior sample data includes: constructing member feature parameters corresponding to each type of the member user features in the member user conversion probability prediction model, and initializing the member feature parameters, wherein the member feature parameters include a weight matrix and a bias vector; based on the multiple behavior sample data and the second loss function, updating the weight matrix and bias vector corresponding to each type of the member user features.

[0082] The second loss function may be a cross entropy loss function.

[0083] In this embodiment, before model training, a target number of types of member user features can be preset, and each type of member user feature corresponds to a member feature parameter, which includes a weight matrix and a bias vector. During the training process, the weight matrix and the bias vector corresponding to each member user feature are updated through multiple behavior sample data and a second cross entropy loss function. During the training process, the member types can be dynamically expanded or reduced according to the training process.

[0084] Among them, in an exemplary embodiment, the weight matrix and bias vector corresponding to each of the member user features are updated based on the multiple behavior sample data and the second loss function, including: inputting the multiple behavior sample data into the member user conversion probability prediction model for iterative training; in each round of iteration, determining the loss value of the prediction result of the member user conversion probability prediction model and the member label corresponding to the current behavior sample data through the second loss function, and updating the weight matrix and bias vector corresponding to each of the member user features according to the loss value until the second loss function converges or reaches a preset number of iterations.

[0085] In this embodiment, by introducing the weight matrix and the bias vector, the model can dynamically adjust the model characteristics of each member user characteristic during the training process, thereby improving the accuracy, generalization ability and flexibility of the model.

[0086] In an exemplary embodiment, determining the prediction result of the member user conversion probability prediction model and the loss value of the member label corresponding to the current behavior sample data through the second loss function includes: obtaining the prediction probability vector of the current behavior sample data for different types of member user characteristics; determining the prediction result based on the prediction probability vector of the current behavior sample data for different types of member user characteristics; and determining the prediction result and the loss value of the member label through the second loss function.

[0087] In this embodiment, the difference between the predicted result and the actual member label is quantified by a loss function to evaluate the accuracy of the model prediction.

[0088] In an exemplary embodiment, obtaining the predicted probability vector of the current behavior sample data for different types of member user features includes: obtaining an implicit representation corresponding to the current behavior sample data by preprocessing the current behavior sample data; performing weighted summation on the implicit representation corresponding to the current behavior sample data using the weight matrix and the bias vector corresponding to different types of member user features, respectively, to obtain the focus weights of the current behavior sample data for different types of member user features; determining sub-feature vectors of the current behavior sample data for different types of member user features based on the focus weights of the current behavior sample data for different types of member user features; and merging the sub-feature vectors of the current behavior sample data for different types of member user features to obtain the predicted probability vector of the current behavior sample data for different types of member user features.

[0089] The above training process is schematically described below, which may include the following steps: Step 1: Initialize the weight matrix values ​​and bias item values ​​corresponding to each member user feature.

[0090] Exemplarily, the weight matrix value and bias item value corresponding to each member user feature of the member user conversion probability prediction model may be initialized in the following manner. The weight matrix value and bias item value corresponding to only one member user feature are used as an example for schematic description:

[0091]

[0092] in, The weight matrix representing the hidden layer of the member user conversion probability prediction model; They represent the attention branches, intermediate vectors, and weight matrices of output results in the member user conversion probability prediction model. Correspondingly, They respectively represent the attention branches in the member user conversion probability prediction model, as well as the intermediate vector and the bias vector of the output result.

[0093] Step 2: Input the training set into the member user conversion probability prediction model, wherein the training set includes multiple behavior sample data.

[0094] Step 3: Obtain the implicit representation corresponding to the current behavior sample data during the current iterative training process :

[0095] Wherein, i represents the i-th sub-behavior sample data corresponding to the current behavior sample data, and tanh is the hyperbolic tangent function.

[0096] It should be noted that each current behavior sample data corresponds to at least one implicit representation.

[0097] Step 4: For each sub-behavior sample data’s hidden representation, calculate the query as follows ,key ,value :

[0098]

[0099]

[0100] Step 5: Perform weighted summation on the implicit representation corresponding to each sub-behavior sample data through the weight matrix to obtain the focus weight of the current behavior sample data for each type of member characteristics:

[0101] Each behavior sample data is composed of at least one sub-behavior sample data.

[0102] It should be noted that, for the sake of ease of description, the target type member characteristics are used for exemplary explanation in the following, and the process of determining the predicted probability vector corresponding to other types of member characteristics is the same as that of the target member characteristics, and will not be repeated here.

[0103] Step 6: Based on the focus weight of the current behavior sample data for the target member characteristics, determine the sub-feature vector of the current behavior sample data for the target type member user characteristics:

[0104] in, .

[0105] Step 7: Merge each sub-feature vector to obtain the predicted probability vector :

[0106] Step 8: Merge the predicted probability vector Through a fully connected layer and sigmoid function, the final overall prediction result is obtained :

[0107]

[0108] Step 9: Based on the prediction results Based on the pre-annotated results, the weight matrix and bias vector corresponding to the member user features are updated based on the second loss function.

[0109] like Fig.13 As shown, the present application embodiment also provides a method for determining the member conversion probability, using the above Figure 1-12c The model obtained by training the model training method in any of the method embodiments shown is used for recognition, and the method comprises the following steps: S1310: Obtain the first behavior data of the user to be predicted.

[0110] In an exemplary embodiment, before obtaining the first behavior data of the user to be predicted, the method further includes: obtaining the third behavior data of the user to be predicted; and obtaining the first behavior data by preprocessing the third behavior data. The preprocessing may include the following steps: Step 1: Remove duplicate data samples.

[0111] Step 2: Remove missing values.

[0112] Check if there are any missing values ​​in the data and fill them with 0 if any.

[0113] Step 3: Use feature normalization to perform linear transformation on the value range of each feature and map all feature values ​​to between 0 and 1.

[0114] Step 4: Calculate the covariance matrix for the normalized data.

[0115] Step 5: Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues ​​and corresponding eigenvectors.

[0116] Step 6: Select the first 9000 eigenvectors as the first row of data in the order of eigenvalues, where 9000 is the target dimension after dimensionality reduction.

[0117] S1320: Input the first behavior data into a member user conversion probability prediction model to obtain the conversion probability of the predicted user to each type of member user, wherein the member user conversion probability prediction model is based on Figure 1-Figure 12c Obtained by training using the model training method in any of the method embodiments shown; S1330: Output the conversion probability of the user to be predicted to be converted into each type of member user.

[0118] In the embodiment of the present application, the first behavior data of the user to be predicted is first obtained, and then the first behavior data is input into the member user conversion probability prediction model to obtain the conversion probability of the user to be predicted to be converted into each type of member user, and finally the conversion probability of the user to be predicted to be converted into each type of member user is output. Figure 1-Figure 12c The model trained by the model training method provided in any of the illustrated embodiments has strong robustness, so that the conversion probability finally outputted is more accurate. Furthermore, when the conversion probability corresponding to a certain member meets the corresponding threshold, the introduction information of the membership rights that matches the preferences of the user to be predicted is pushed to enhance the user's interest and response rate, thereby improving the conversion rate.

[0119] In an exemplary embodiment, S1320 may include the following steps: S1322: Obtain the predicted probability vector of the first behavior data for different types of members.

[0120] Among them, the first behavior data includes the second behavior data of each unit time within a preset period, and the number of users to be predicted is at least one. Exemplarily, in units of days, the daily behavior data of the users to be predicted within the preset period is obtained from the user log system. If the number of users to be predicted is one, a one-dimensional array is constructed based on the first behavior data of the user to be predicted. If the number of users to be predicted is multiple, an n-dimensional array is constructed based on the first behavior data of all users to be predicted. Assuming that one user is a sample and the user's daily behavior data is a feature, each sample has 365 features, and each feature has 10,000 dimensions. Whether the user is a member user within the preset period, the member user label is encoded as an integer greater than 0, and the label encodings of different types of member users are different, and the non-member user label is encoded as 0. Based on this, the daily behavior features of each user can be represented as a 365×10,000 matrix, expressed as :

[0121] in, A 10,000-dimensional vector representing the daily behavior characteristics of user i on day t, where t ranges from 1 to 365. Then, by analogy, the behavior data of all users form an n-dimensional array X, whose size is n×365×10,000:

[0122] For example, suppose that the daily behavior feature of the user i to be predicted on the tth day is obtained, and the daily behavior feature is a 9000-dimensional vector, as shown in Table 1: Table 1

[0123] So, , ,

[0124] Finally, the user's membership label is represented as an n-dimensional vector y, where is an integer greater than 1, indicating that user i is a member user of a certain type. 0 means user i is not a VIP user:

[0125] The above S1322 may include the following steps: S13222: Obtain an implicit representation corresponding to each second behavior data by preprocessing each second behavior data.

[0126] For example, taking a type of member feature as an example, the weight matrix and bias vector corresponding to the member feature can be expressed as:

[0127]

[0128] in, The weight matrix representing the hidden layer of the member user conversion probability prediction model; They represent the attention branches, intermediate vectors, and weight matrices of output results in the member user conversion probability prediction model. Correspondingly, They represent the attention branches in the VIP user conversion probability prediction model, as well as the intermediate vector and the bias vector of the output result.

[0129] Assume that the first behavior data of the user to be predicted within the preset period is:

[0130] Then for each second line of data , the hidden representation can be obtained through a fully connected layer as follows :

[0131] Here, tanh is the hyperbolic tangent function.

[0132]

[0133]

[0134]

[0135] S13224: Perform weighted summation on the implicit representation corresponding to each second behavior data using the weight matrices and bias vectors corresponding to different types of members, to obtain the focus weight of each second behavior data for different types of members.

[0136] Continuing with the previous example, using hidden representation , calculate the query as follows ,key ,value :

[0137]

[0138]

[0139]

[0140]

[0141]

[0142]

[0143]

[0144]

[0145]

[0146]

[0147]

[0148] Then calculate the focus weight : , , is the dimension of focus weight, here is 3.

[0149]

[0150]

[0151]

[0152]

[0153]

[0154]

[0155] S13226: Determine sub-feature vectors of each piece of second behavior data for different types of members based on the focus weights of each piece of second behavior data for different types of members.

[0156] First, weighted sum of the values ​​according to the focus weights is performed to obtain the front and back block vectors :

[0157]

[0158]

[0159]

[0160] Then, the intermediate vector is obtained through a fully connected feedforward neural network :

[0161]

[0162]

[0163]

[0164] Then, Through a linear transformation and a sigmoid function, the sub-feature vector of each of the second behavior data for members of this type is obtained : , is the sigmoid function.

[0165]

[0166]

[0167]

[0168]

[0169] S13228: For different types of members, merge the sub-feature vectors of each of the second behavior data to obtain the predicted probability vector.

[0170] It can be understood that each second behavior data corresponds to a different sub-feature vector for different types of members. Therefore, according to the member type, each corresponding sub-feature vector is merged to obtain the prediction probability vector :

[0171]

[0172] S1324: Based on each of the prediction probability vectors, obtain a conversion probability of the user to be predicted to be converted into a member user of a corresponding type.

[0173] The combined predicted probability vector Through a fully connected layer and sigmoid function, the final overall prediction result is obtained :

[0174]

[0175]

[0176]

[0177] Finally, the probability that the user to be predicted will become this type of member user is 0.86881401.

[0178] Fig.14 A schematic diagram of the structure of a model training device provided in an embodiment of this specification is shown. Figure 3 As shown, the model training device 1400 may include: an acquisition module 1410 , a processing module 1420 and a training module 1430 .

[0179] In this embodiment, the acquisition module 1410 is used to acquire target data corresponding to the target product, wherein the target data includes function information of at least one historical version, user behavior data of member users corresponding to each of the historical versions, and function information of the version to be released; the processing module 1420 is used to obtain at least one type of member user feature corresponding to the version to be released by inputting the target data into a member user feature prediction model; the training module 1430 is used to train a member user conversion probability prediction model based on the at least one type of member user feature and the user behavior data of non-member users corresponding to at least one historical version corresponding to the target product.

[0180] In an exemplary embodiment, the member user feature prediction model includes an encoder, a first decoder and a second decoder; the processing module 1420 is specifically used to: perform feature characterization processing on the target data through the encoder to obtain a first sampling result set and a first target feature, wherein the first sampling result set includes N-1 first target sampling results, and N is an integer greater than or equal to 2; based on the first sampling result set and the first target feature, perform a first reconstruction processing through the first decoder to obtain a second sampling result set and a second target feature, wherein the second sampling result set includes N-1 second target sampling results; based on the first sampling result set, the second sampling result set, the first target feature and the second target feature, perform a second reconstruction processing through the second decoder to obtain at least one type of member user feature corresponding to the version to be released.

[0181] In an exemplary embodiment, the encoder includes N first sampling modules and N-1 residual modules connected in sequence, wherein the N-1 residual modules correspond one-to-one to the first N-1 first sampling modules, and each residual module includes a different number of convolution filters; the processing module 1420 is specifically used to: encode the target data through the N first sampling modules in sequence to obtain the first target feature output by the Nth first sampling module and the first sampling result set corresponding to the N-1 residual modules, wherein the first sampling result set includes the first target sampling result output by each residual module.

[0182] In an exemplary embodiment, the processing module 1420 is specifically used to: for the first first sampling module, encode the target data to obtain the first first initial sampling result and the first first encoding feature output by the first first sampling module, and process the first initial sampling result through the residual module corresponding to the first first sampling module to obtain the first first target sampling result; for the Qth first sampling module, encode the Q-1th first encoding feature output by the Q-1th first sampling module to obtain the Qth first initial sampling result and the Qth first encoding feature output by the Qth first sampling module, and process the Qth first initial sampling result through the residual module corresponding to the Qth first sampling module to obtain the Qth first target sampling result, wherein Q is an integer greater than or equal to 2 and less than N; for the Nth first sampling module, encode the N-1th first encoding feature output by the N-1th first sampling module to obtain the first target feature output by the Nth first sampling module.

[0183] In an exemplary embodiment, the first first sampling module includes a first convolution layer, a first normalization layer, a first activation layer and a maximum pooling layer connected in sequence; the other first sampling modules except the first first sampling module among the N first sampling modules include the first convolution layer, the first normalization layer, the first activation layer, the second convolution layer, the second normalization layer, the first residual layer and the second activation layer connected in sequence.

[0184] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first second encoded feature by inputting the target data into the first convolutional layer; obtain the second second encoded feature by inputting the first second encoded feature into the first normalization layer; obtain the third second encoded feature and the first first initial sampling result by inputting the second second encoded feature into the first activation layer; obtain the first first encoded feature by inputting the third second encoded feature into the maximum pooling layer.

[0185] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first third coding feature by inputting the Q-1th first coding feature into the first convolutional layer; obtain the second third coding feature by inputting the first third coding feature into the first normalization layer; obtain the third third coding feature and the Qth first initial sampling result by inputting the second third coding feature into the first activation layer; obtain the fourth third coding feature by inputting the third third coding feature into the second convolutional layer; obtain the fifth third coding feature by inputting the fourth third coding feature into the second normalization layer; obtain the sixth third coding feature by inputting the fifth third coding feature and the N-1th first coding feature into the first residual layer; obtain the Qth first coding feature by inputting the sixth third coding feature into the second activation layer.

[0186] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first fourth coding feature by inputting the N-1th first coding feature into the first convolutional layer; obtain the second fourth coding feature by inputting the first fourth coding feature into the first normalization layer; obtain the third fourth coding feature and the Nth first initial sampling result by inputting the second fourth coding feature into the first activation layer; obtain the fourth fourth coding feature by inputting the third fourth coding feature into the second convolutional layer; obtain the fifth fourth coding feature by inputting the fourth fourth coding feature into the second normalization layer; obtain the sixth fourth coding feature by inputting the fifth fourth coding feature and the N-1th first coding feature into the first residual layer; obtain the first target feature by inputting the sixth fourth coding feature into the second activation layer.

[0187] In an exemplary embodiment, the first decoder includes N second sampling modules connected in sequence, N-1 first-stage decoding modules and a first reconstruction module, wherein the N-1 first-stage decoding modules correspond one-to-one to the first N-1 second sampling modules, and each second sampling module includes the same number of convolution filters; the processing module 1420 is specifically used to: sequentially decode the first sampling result set and the first target feature through the N second sampling modules to obtain the Nth first decoding feature output by the Nth second sampling module and the second sampling result set corresponding to the N-1 first-stage decoding modules, wherein the second sampling result set includes the second target sampling result output by each first-stage decoding module; perform a first reconstruction process on the Nth first decoding feature through the first reconstruction module to obtain the second target feature.

[0188] In an exemplary embodiment, the processing module 1420 is specifically used to: for the first second sampling module, decode the first target feature and the N-1th first target sampling result in the first sampling result set to obtain the first first decoding feature output by the first second sampling module and the first second target sampling result output by the first first-stage decoding module; for the Mth second sampling module, decode the M-1th first decoding feature output by the M-1th second sampling module and the NMth first target sampling result in the first sampling result set to obtain the Mth first decoding feature output by the Mth second sampling module and the Mth second target sampling result output by the Mth first-stage decoding module, wherein M is an integer greater than or equal to 2 and less than N; for the Nth second sampling module, decode the N-1th first decoding feature output by the N-1th second sampling module to obtain the Nth first decoding feature output by the Nth second sampling module.

[0189] In an exemplary embodiment, each of the second sampling modules includes a first upsampling layer and at least one first combination layer, and the first N-1 second sampling modules also include a first series layer, wherein the first upsampling layers in the first N-1 second sampling modules are connected to at least one first combination layer through the first series layer, and the first combination layer includes a third convolution layer, a third normalization layer, and a third activation layer connected in sequence.

[0190] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first second decoding feature by inputting the first target feature into the first upsampling layer; obtain the second second decoding feature by inputting the first second decoding feature and the N-1th first target sampling result into the first series connection layer; and sequentially decode the second second decoding feature through at least one of the first combination layers to obtain the first second target sampling result and the first first decoding feature.

[0191] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first third decoding feature by inputting the M-1th first decoding feature into the first upsampling layer; obtain the second third decoding feature by inputting the first third decoding feature and the NMth first target sampling result into the first series connection layer; and decode the second third decoding feature through at least one of the first combination layers in turn to obtain the Mth second target sampling result and the Mth first decoding feature.

[0192] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first fourth decoding feature by inputting the N-1th first decoding feature into the first upsampling layer; and decode the first fourth decoding feature through at least one of the first combination layers in turn to obtain the Nth first decoding feature.

[0193] In an exemplary embodiment, the first reconstruction module includes at least a second combination layer, a third combination layer, a second residual layer, a fourth activation layer, a fourth convolution layer and a fifth activation layer connected in sequence, the second combination layer includes a fifth convolution layer, a fourth normalization layer and a fifth activation layer, and the third combination layer includes a sixth convolution layer and a fifth normalization layer; The processing module 1420 is specifically used to: decode the Nth first decoding feature through at least one of the second combination layers in sequence to obtain the first fifth decoding feature; obtain the second fifth decoding feature by inputting the first fifth decoding feature into the third combination layer; obtain the third fifth decoding feature by inputting the second fifth decoding feature into the second residual layer; obtain the fourth fifth decoding feature by inputting the third fifth decoding feature into the fourth activation layer; obtain the fifth fifth decoding feature by inputting the fourth fifth decoding feature into the fourth convolutional layer; obtain the second target feature by inputting the fifth fifth decoding feature into the fifth activation layer.

[0194] In an exemplary embodiment, the second decoder includes N third sampling modules connected in sequence, N-1 second-stage decoding modules, and a second reconstruction module, wherein the N-1 second-stage decoding modules correspond one-to-one to the first N-1 third sampling modules, and each of the third sampling modules includes the same number of convolution filters; the processing module 1420 is specifically used to: decode the first sampling result set, the second sampling result set, the first target feature, and the second target feature in sequence through the N third sampling modules to obtain the Nth sixth decoding feature output by the Nth third sampling module; perform a second reconstruction process on the Nth sixth decoding feature through the second reconstruction module to obtain at least one type of member user feature corresponding to the version to be released.

[0195] In an exemplary embodiment, the processing module 1420 is specifically used to: for the first third sampling module, by decoding the first target feature, the second target feature, the N-1th first target sampling result in the first sampling result set, and the N-1th second target sampling result in the second sampling result set, obtain the first sixth decoding feature output by the first third sampling module; for the Lth third sampling module, by decoding the L-1th sixth decoding feature output by the L-1th third sampling module, the NLth first target sampling result in the first sampling result set, and the NLth second target sampling result in the second sampling result set, obtain the Lth sixth decoding feature output by the Lth third sampling module, where L is an integer greater than or equal to 2 and less than N; for the Nth third sampling module, by decoding the N-1th sixth decoding feature output by the N-1th third sampling module, obtain the Nth sixth decoding feature output by the Nth third sampling module.

[0196] In an exemplary embodiment, each of the third sampling modules includes a second upsampling layer and at least one fourth combination layer, and the first N-1 third sampling modules also include a second series layer and a cross attention layer, wherein the second upsampling layer in the first N-1 second sampling modules is sequentially connected to at least one fourth combination layer through the second series layer and the cross attention layer, and the fourth combination layer includes a seventh convolutional layer, a sixth normalization layer and a sixth activation layer connected in sequence.

[0197] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first seventh decoding feature by inputting the second target feature into the second upsampling layer; obtain the second seventh decoding feature by inputting the first seventh decoding feature, the N-1th first target sampling result and the N-1th second target sampling result into the second series layer; obtain the third seventh decoding feature by inputting the second seventh decoding feature into the cross-attention layer; and obtain the first sixth decoding feature by decoding the third seventh decoding feature through at least one of the fourth combination layers in turn.

[0198] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first eighth decoding feature by inputting the L-1th sixth decoding feature into the second upsampling layer; obtain the second eighth decoding feature by inputting the first eighth decoding feature, the NLth first target sampling result and the NLth second target sampling result into the second series layer; obtain the third eighth decoding feature by inputting the second eighth decoding feature into the cross-attention layer; and obtain the Lth sixth decoding feature by decoding the third eighth decoding feature through at least one of the fourth combination layers in turn.

[0199] In an exemplary embodiment, the processing module 1420 is specifically used to: obtain the first ninth decoding feature by inputting the N-1th sixth decoding feature into the second upsampling layer; and decode the first ninth decoding feature through at least one of the fourth combination layers in turn to obtain the Nth sixth decoding feature.

[0200] In an exemplary embodiment, it also includes a member user feature prediction model training module, which is used to: obtain a target data set, wherein the target data set includes functional information of at least one historical version corresponding to multiple products, user behavior data of at least one member user corresponding to each historical version, and member user features of each member; train the member user feature prediction model based on the target data set and the first loss function.

[0201] In an exemplary embodiment, the training module 1430 is specifically used to: determine a training set based on the at least one type of member user characteristics and the user behavior data of non-member users corresponding to at least one historical version of the target product, wherein the training set includes multiple behavior sample data, each of the behavior sample data corresponds to at least one member label; and train the member user conversion probability prediction model based on the multiple behavior sample data.

[0202] In an exemplary embodiment, the training module 1430 is specifically used to: construct member feature parameters corresponding to each type of the member user feature in the member user conversion probability prediction model, and initialize the member feature parameters, wherein the member feature parameters include a weight matrix and a bias vector; based on the multiple behavior sample data and the second loss function, update the weight matrix and bias vector corresponding to each type of the member user feature.

[0203] In an exemplary embodiment, the training module 1430 is specifically used to: input the multiple behavior sample data into the member user conversion probability prediction model for iterative training; in each round of iteration, determine the loss value of the member label corresponding to the prediction result of the member user conversion probability prediction model and the current behavior sample data through the second loss function, and update the weight matrix and bias vector corresponding to each member user feature according to the loss value until the second loss function converges or reaches a preset number of iterations.

[0204] In an exemplary embodiment, the training module 1430 is specifically used to: obtain the predicted probability vector of the current behavior sample data for different types of member user characteristics; determine the predicted result based on the predicted probability vector of the current behavior sample data for different types of member user characteristics; and determine the loss value of the predicted result and the member label through the second loss function.

[0205] In an exemplary embodiment, the training module 1430 is specifically used to: obtain an implicit representation corresponding to the current behavior sample data by preprocessing the current behavior sample data; perform weighted summation on the implicit representation corresponding to the current behavior sample data using the weight matrices and the bias vectors corresponding to different types of member user features, respectively, to obtain the focus weights of the current behavior sample data for different types of member user features; determine sub-feature vectors of the current behavior sample data for different types of member user features based on the focus weights of the current behavior sample data for different types of member user features; and merge the sub-feature vectors of the current behavior sample data for different types of member user features to obtain prediction probability vectors of the current behavior sample data for different types of member user features.

[0206] A model training device provided in the embodiment of the present application can achieve Figure 1 To avoid repetition, the various processes implemented in the illustrated method embodiment will not be described again here.

[0207] Fig.15 A member conversion probability determination device provided in an embodiment of the present application is shown. The member conversion probability determination device 1500 may include: an acquisition module 1510 , a processing module 1520 , and an output module 1530 .

[0208] In this embodiment, the acquisition module 1510 is used to acquire the first behavior data of the user to be predicted; the processing module 1520 is used to input the first behavior data into the member user conversion probability prediction model to obtain the conversion probability of the user to be predicted to be converted into each type of member user, wherein the member user conversion probability prediction model is based on Figure 1-Figure 12c The output module 1530 is used to output the conversion probability of the user to be predicted to be converted into each type of member user.

[0209] In an exemplary embodiment, the processing module 1520 is specifically used to: obtain the prediction probability vector of the first behavior data for different types of members; and obtain the conversion probability of the user to be predicted to be converted into a member user of the corresponding type based on each of the prediction probability vectors.

[0210] In an exemplary embodiment, the first behavior data includes second behavior data for each unit time within a preset period; the processing module 1520 is specifically used to: obtain an implicit representation corresponding to each second behavior data by preprocessing each second behavior data; perform weighted summation on the implicit representation corresponding to each second behavior data using the weight matrices and bias vectors corresponding to different types of members, respectively, to obtain the focus weights of each second behavior data for different types of members; determine the sub-feature vectors of each second behavior data for different types of members based on the focus weights of each second behavior data for different types of members; and merge the sub-feature vectors of each second behavior data for different types of members to obtain the predicted probability vector.

[0211] The member conversion probability determination device provided in the embodiment of the present application can achieve Fig.13 To avoid repetition, the various processes implemented in the illustrated method embodiment will not be described again here.

[0212] A model training device and a member conversion probability determination device in the embodiment of the present application can be a device, or a component, an integrated circuit, or a chip in an electronic device. The embodiment of the present application is not specifically limited.

[0213] A model training device and a member conversion probability determination device in the embodiment of the present application may be a device with an operating system. The operating system may be an Android operating system, an iOS operating system, or other possible operating systems, which are not specifically limited in the embodiment of the present application.

[0214] Optional, such as Figure 5As shown, an embodiment of the present application also provides an electronic device 1600, including a processor 1610, a memory 1620, and a program or instruction stored in the memory 1620 and executable on the processor 1610. When the program or instruction is executed by the processor 1610, the various processes of the above-mentioned model training method embodiment are implemented, or the various processes of the above-mentioned member conversion probability determination method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0215] An embodiment of the present application also provides a computer-readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the various processes of the above-mentioned model training method embodiment, or the various processes of the above-mentioned member conversion probability determination method embodiment are implemented, and the same technical effect can be achieved. To avoid repetition, it will not be repeated here.

[0216] The processor is the processor in the electronic device described in the above embodiment. The readable storage medium includes a computer readable storage medium, such as a computer read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0217] An embodiment of the present application further provides a chip, which includes a processor and a communication interface, wherein the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the various processes of the above-mentioned model training method embodiment, or to implement the various processes of the above-mentioned member conversion probability determination method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0218] It should be understood that the chip mentioned in the embodiments of the present application can also be called a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0219] The embodiments of the present application also provide a computer program product, which includes at least one computer program. When the computer program is loaded and executed by a processor, it implements the various processes of the above-mentioned model training method embodiment, or implements the various processes of the above-mentioned member conversion probability determination method embodiment, and can achieve the same technical effect. To avoid repetition, it will not be repeated here.

[0220] It should be noted that, in this article, the terms "comprise", "include" or any other variant thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0221] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0222] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A model training method, characterized in that: The method comprises: Acquire target data corresponding to the target product, wherein the target data includes function information of at least one historical version, user behavior data of member users corresponding to each historical version, and function information of the version to be released; By inputting the target data into a member user feature prediction model, at least one type of member user feature corresponding to the version to be released is obtained; A member user conversion probability prediction model is trained based on the at least one type of member user characteristics and the user behavior data of non-member users corresponding to at least one historical version corresponding to the target product.

2. The method according to claim 1, characterized in that The member user feature prediction model includes an encoder, a first decoder and a second decoder; By inputting the target data into the member user feature prediction model, at least one type of member user feature corresponding to the version to be released is obtained, including: Performing feature characterization processing on the target data by the encoder to obtain a first sampling result set and a first target feature, wherein the first sampling result set includes N-1 first target sampling results, where N is an integer greater than or equal to 2; Based on the first sampling result set and the first target feature, performing a first reconstruction process by the first decoder to obtain a second sampling result set and a second target feature, wherein the second sampling result set includes N-1 second target sampling results; Based on the first sampling result set, the second sampling result set, the first target feature and the second target feature, a second reconstruction process is performed by the second decoder to obtain at least one type of member user feature corresponding to the version to be released.

3. The method according to claim 2, characterized in that The encoder includes N first sampling modules and N-1 residual modules connected in sequence, wherein the N-1 residual modules correspond one-to-one to the first N-1 first sampling modules, and each residual module includes a different number of convolution filters; The encoder performs feature characterization processing on the target data to obtain a first sampling result set and a first target feature, including: The target data is encoded and processed in sequence through N first sampling modules to obtain the first target feature output by the Nth first sampling module and the first sampling result set corresponding to N-1 residual modules, wherein the first sampling result set includes the first target sampling result output by each residual module.

4. The method according to claim 3, characterized in that The target data is encoded by N first sampling modules in sequence to obtain the first target feature output by the Nth first sampling module and the first sampling result set corresponding to N-1 residual modules, including: For the first first sampling module, encoding the target data to obtain a first first initial sampling result and a first first encoding feature output by the first first sampling module, and processing the first initial sampling result by the residual module corresponding to the first first sampling module to obtain a first first target sampling result; For the Qth first sampling module, encode the Q-1th first coding feature output by the Q-1th first sampling module to obtain the Qth first initial sampling result and the Qth first coding feature output by the Qth first sampling module, and process the Qth first initial sampling result by the residual module corresponding to the Qth first sampling module to obtain the Qth first target sampling result, where Q is an integer greater than or equal to 2 and less than N; For the Nth first sampling module, the first target feature output by the Nth first sampling module is obtained by encoding the N-1th first encoding feature output by the N-1th first sampling module.

5. The method according to claim 4, characterized in that The first first sampling module includes a first convolution layer, a first normalization layer, a first activation layer and a maximum pooling layer connected in sequence; the other first sampling modules except the first first sampling module among the N first sampling modules include the first convolution layer, the first normalization layer, the first activation layer, the second convolution layer, the second normalization layer, the first residual layer and the second activation layer connected in sequence.

6. The method according to claim 5, characterized in that The encoding process is performed on the target data to obtain the first initial sampling result and the first encoding feature output by the first sampling module, including: Obtaining a first second encoding feature by inputting the target data into the first convolutional layer; By inputting the first second encoding feature into the first normalization layer, a second second encoding feature is obtained; By inputting the second second coding feature into the first activation layer, a third second coding feature and a first first initial sampling result are obtained; By inputting the third second encoding feature into the maximum pooling layer, the first first encoding feature is obtained.

7. The method according to claim 5, characterized in that The method of encoding the Q-1th first coding feature output by the Q-1th first sampling module to obtain the Qth first initial sampling result and the Qth first coding feature output by the Qth first sampling module includes: Obtaining a first third encoding feature by inputting the Q-1th first encoding feature into the first convolutional layer; By inputting the first third encoding feature into the first normalization layer, a second third encoding feature is obtained; By inputting the second third coding feature into the first activation layer, a third third coding feature and a Qth first initial sampling result are obtained; By inputting the third third encoding feature into the second convolutional layer, a fourth third encoding feature is obtained; By inputting the fourth third encoding feature into the second normalization layer, a fifth third encoding feature is obtained; By inputting the fifth third coding feature and the N-1th first coding feature into the first residual layer, a sixth third coding feature is obtained; By inputting the 6th third coding feature into the second activation layer, the Qth first coding feature is obtained.

8. The method according to claim 5, characterized in that The step of encoding the N-1th first encoding feature output by the N-1th first sampling module to obtain the first target feature output by the Nth first sampling module includes: Obtaining a first fourth coding feature by inputting the N-1th first coding feature into the first convolutional layer; Obtaining a second fourth coding feature by inputting the first fourth coding feature into the first normalization layer; By inputting the second fourth coding feature into the first activation layer, a third fourth coding feature and an Nth first initial sampling result are obtained; By inputting the third fourth encoding feature into the second convolutional layer, a fourth fourth encoding feature is obtained; By inputting the fourth fourth encoding feature into the second normalization layer, a fifth fourth encoding feature is obtained; Obtaining a sixth fourth coding feature by inputting the fifth fourth coding feature and the N-1th first coding feature into the first residual layer; The first target feature is obtained by inputting the sixth fourth coding feature into the second activation layer.

9. The method according to claim 2, characterized in that: The first decoder includes N second sampling modules, N-1 first stage decoding modules and a first reconstruction module connected in sequence, wherein the N-1 first stage decoding modules correspond one to one with the first N-1 second sampling modules, and each second sampling module includes the same number of convolution filters; The performing a first reconstruction process based on the first sampling result set and the first target feature by the first decoder to obtain a second sampling result set and a second target feature includes: Decoding the first sampling result set and the first target feature through N second sampling modules in sequence to obtain the Nth first decoding feature output by the Nth second sampling module and the second sampling result set corresponding to N-1 first-stage decoding modules, wherein the second sampling result set includes the second target sampling result output by each first-stage decoding module; The first reconstruction module performs a first reconstruction process on the Nth first decoding feature to obtain the second target feature.

10. The method according to claim 9, characterized in that The first sampling result set and the first target feature are decoded in sequence by N second sampling modules to obtain the Nth first decoding feature output by the Nth second sampling module and the second sampling result set corresponding to N-1 first-stage decoding modules, including: For the first second sampling module, by decoding the first target feature and the N-1th first target sampling result in the first sampling result set, the first first decoding feature output by the first second sampling module and the first second target sampling result output by the first first-stage decoding module are obtained; For the Mth second sampling module, by decoding the M-1th first decoding feature output by the M-1th second sampling module and the NMth first target sampling result in the first sampling result set, the Mth first decoding feature output by the Mth second sampling module and the Mth second target sampling result output by the Mth first-stage decoding module are obtained, where M is an integer greater than or equal to 2 and less than N; For the Nth second sampling module, the Nth first decoding feature output by the Nth second sampling module is obtained by decoding the N-1th first decoding feature output by the Nth second sampling module.

11. The method according to claim 10, characterized in that Each of the second sampling modules includes a first upsampling layer and at least one first combination layer, and the first N-1 second sampling modules also include a first series layer, wherein the first upsampling layers in the first N-1 second sampling modules are connected to at least one first combination layer through the first series layer, and the first combination layer includes a third convolution layer, a third normalization layer, and a third activation layer connected in sequence.

12. The method according to claim 11, characterized in that The decoding process is performed on the first target feature and the N-1th first target sampling result in the first sampling result set to obtain the first first decoding feature output by the first second sampling module and the first second target sampling result output by the first first stage decoding module, including: Obtaining a first second decoding feature by inputting the first target feature into the first upsampling layer; By inputting the first second decoding feature and the N-1th first target sampling result into the first series connection layer, a second second decoding feature is obtained; The second second decoding feature is decoded in sequence through at least one of the first combination layers to obtain the first second target sampling result and the first first decoding feature.

13. The method according to claim 11, characterized in that The method of decoding the M-1th first decoding feature output by the M-1th second sampling module and the NMth first target sampling result in the first sampling result set to obtain the Mth first decoding feature output by the Mth second sampling module and the Mth second target sampling result output by the Mth first-stage decoding module includes: Obtaining a first third decoding feature by inputting the M-1th first decoding feature into the first upsampling layer; By inputting the first third decoding feature and the NMth first target sampling result into the first series connection layer, a second third decoding feature is obtained; The second third decoding feature is decoded in sequence through at least one of the first combination layers to obtain the Mth second target sampling result and the Mth first decoding feature.

14. The method according to claim 11, characterized in that The decoding process is performed on the N-1th first decoding feature output by the N-1th second sampling module to obtain the Nth first decoding feature output by the Nth second sampling module, including: Obtaining a first fourth decoding feature by inputting the N-1th first decoding feature into the first upsampling layer; The first fourth decoding feature is decoded sequentially through at least one of the first combination layers to obtain the Nth first decoding feature.

15. The method according to claim 9, characterized in that The first reconstruction module includes at least one second combination layer, a third combination layer, a second residual layer, a fourth activation layer, a fourth convolution layer and a fifth activation layer connected in sequence, the second combination layer includes a fifth convolution layer, a fourth normalization layer and a fifth activation layer, and the third combination layer includes a sixth convolution layer and a fifth normalization layer; The performing a first reconstruction process on the Nth first decoding feature by the first reconstruction module to obtain the second target feature includes: Decoding the Nth first decoding feature through at least one of the second combination layers in sequence to obtain a first fifth decoding feature; Obtaining a second fifth decoding feature by inputting the first fifth decoding feature into the third combination layer; Obtaining a third fifth decoding feature by inputting the second fifth decoding feature into the second residual layer; Obtaining a fourth fifth decoding feature by inputting the third fifth decoding feature into the fourth activation layer; Obtaining a fifth fifth decoding feature by inputting the fourth fifth decoding feature into the fourth convolutional layer; The second target feature is obtained by inputting the 5th fifth decoding feature into the fifth activation layer.

16. The method according to claim 2, characterized in that The second decoder comprises N third sampling modules, N-1 second stage decoding modules, and a second reconstruction module, wherein the N-1 second stage decoding modules correspond one-to-one to the first N-1 third sampling modules, and each third sampling module comprises the same number of convolution filters; The second reconstruction process is performed by the second decoder based on the first sampling result set, the second sampling result set, the first target feature, and the second target feature to obtain at least one type of member user feature corresponding to the version to be released, including: Decoding the first sampling result set, the second sampling result set, the first target feature, and the second target feature in sequence through N third sampling modules to obtain an Nth sixth decoding feature output by the Nth third sampling module; The second reconstruction module performs a second reconstruction process on the Nth sixth decoded feature to obtain at least one type of member user feature corresponding to the version to be released.

17. The method according to claim 1, characterized in that Before obtaining at least one type of member user feature corresponding to the version to be released by inputting the target data into the member user feature prediction model, the method further includes: Acquire a target data set, wherein the target data set includes function information of at least one of the historical versions corresponding to the multiple products, user behavior data of at least one member user corresponding to each of the historical versions, and member user characteristics of each member; The member user feature prediction model is trained based on the target data set and the first loss function.

18. The method according to claim 1, characterized in that The training of the member user conversion probability prediction model based on the at least one type of member user features and the user behavior data of non-member users corresponding to at least one historical version corresponding to the target product includes: Determine a training set based on the at least one type of member user feature and the user behavior data of non-member users corresponding to at least one historical version corresponding to the target product, wherein the training set includes a plurality of behavior sample data, each of the behavior sample data corresponds to at least one member label; The member user conversion probability prediction model is trained based on the multiple behavior sample data.

19. The method according to claim 18, characterized in that The training of the member user conversion probability prediction model based on the plurality of behavior sample data includes: Constructing member feature parameters corresponding to each type of member user feature in the member user conversion probability prediction model, and initializing the member feature parameters, wherein the member feature parameters include a weight matrix and a bias vector; Based on the multiple behavior sample data and the second loss function, the weight matrix and the bias vector corresponding to each member user feature are updated.

20. The method according to claim 19, characterized in that The updating of the weight matrix and the bias vector corresponding to each member user feature based on the plurality of behavior sample data and the second loss function includes: Inputting the plurality of behavior sample data into the member user conversion probability prediction model for iterative training; During each round of iteration, the second loss function is used to determine the loss value of the member label corresponding to the prediction result of the member user conversion probability prediction model and the current behavior sample data. According to the loss value, the weight matrix and bias vector corresponding to each member user feature are updated until the second loss function converges or reaches a preset number of iterations.

21. The method according to claim 20, characterized in that The determining, by the second loss function, the loss value of the member label corresponding to the prediction result of the member user conversion probability prediction model and the current behavior sample data includes: Obtaining a prediction probability vector of the current behavior sample data for different types of member user characteristics; Determining the prediction result according to the prediction probability vector of the current behavior sample data for different types of member user characteristics; The loss value of the prediction result and the member label is determined by the second loss function.

22. The method according to claim 21, characterized in that The obtaining of the prediction probability vector of the current behavior sample data for different types of member user characteristics includes: Obtaining an implicit representation corresponding to the current behavior sample data by preprocessing the current behavior sample data; By respectively performing weighted summation on the implicit representation corresponding to the current behavior sample data through the weight matrix and the bias vector corresponding to different types of member user features, the focus weight of the current behavior sample data for different types of member user features is obtained; Determining sub-feature vectors of the current behavior sample data for different types of member user features based on the focus weights of the current behavior sample data for different types of member user features; For different types of member user features, the sub-feature vectors of the current behavior sample data are merged to obtain a prediction probability vector of the current behavior sample data for different types of member user features.

23. A method for determining member conversion probability, characterized in that: include: Obtain the first behavior data of the user to be predicted; Inputting the first behavior data into a member user conversion probability prediction model to obtain a conversion probability of the user to be predicted to be converted into each type of member user, wherein the member user conversion probability prediction model is obtained based on the model training method according to any one of claims 1 to 24; The conversion probability of the user to be predicted to be converted into each type of member user is output.

24. The method according to claim 23, characterized in that The step of inputting the first behavior data into a member user conversion probability prediction model to obtain the conversion probability of the user to be predicted to be converted into each type of member user includes: Obtaining predicted probability vectors of the first behavior data for different types of members; Based on each of the prediction probability vectors, a conversion probability of the user to be predicted being converted into a member user of a corresponding type is obtained.

25. The method according to claim 24, characterized in that The first behavior data includes second behavior data of each unit time within a preset period; The obtaining of the predicted probability vectors of the first behavior data for different types of members includes: Preprocessing each of the second behavior data to obtain an implicit representation corresponding to each of the second behavior data; By using the weight matrices and bias vectors corresponding to different types of members, weighted summation is performed on the implicit representation corresponding to each second behavior data to obtain the focus weight of each second behavior data for different types of members; Determining sub-feature vectors of each of the second behavior data for different types of members based on the focus weights of each of the second behavior data for different types of members; For different types of members, the sub-feature vectors of each of the second behavior data are combined to obtain the predicted probability vector.

26. A model training device, characterized in that: include: An acquisition module, used to acquire target data corresponding to a target product, wherein the target data includes function information of at least one historical version, user behavior data of member users corresponding to each historical version, and function information of a version to be released; A processing module, configured to obtain at least one type of member user feature corresponding to the version to be released by inputting the target data into a member user feature prediction model; A training module is used to train a member user conversion probability prediction model based on the at least one type of member user characteristics and the user behavior data of non-member users corresponding to at least one historical version corresponding to the target product.

27. A device for determining member conversion probability, characterized in that: include: An acquisition module, used to acquire the first behavior data of the user to be predicted; a processing module, configured to input the first behavior data into a member user conversion probability prediction model to obtain a conversion probability of the user to be predicted to be converted into each type of member user, wherein the member user conversion probability prediction model is obtained based on the model training method according to any one of claims 1 to 22; The output module is used to output the conversion probability of the user to be predicted to be converted into each type of member user.

28. An electronic device, characterized in that: It includes a processor, a memory, and a program or instruction stored in the memory and executable on the processor. When the program or instruction is executed by the processor, it implements the steps of the model training method as described in any one of claims 1 to 22, or implements the steps of the method for determining the member conversion probability as described in any one of claims 23 to 25.

Images