Method and apparatus for training an item numbering model

Abstract

The embodiments of the present application disclose a method and apparatus for training an item numbering model. A specific embodiment of the method includes the steps of: obtaining an initial item numbering model and a training sample set; using sample user information of training samples in the training sample set as input for the initial item numbering model to obtain a probability of sample item number information corresponding to the input sample user information; maximizing the obtained probability of sample item number information corresponding to the input sample user information, adjusting the structural parameters of the initial item numbering model, and performing training to obtain an item numbering model, wherein the item numbering model represents the correspondence relationship between the input sample user information and the sample item number information and the correspondence relationship between the sample item information and the sample item number information. In this embodiment, item recommendations can be realized by the trained item numbering model, and the item number information can be used as an index to improve search efficiency.

Description

[Technical Field]

[0001] This application is based on and claims priority to a Chinese patent application bearing application number 202010613628.4, filed on June 30, 2020, and entitled "Method and Apparatus for Training an Item Numbering Model," the entire contents of which are incorporated herein by reference.

[0002] FIELD Embodiments of the present application relate to the field of computer technology, and more particularly to a method and apparatus for training an item numbering model. [Background technology]

[0003] With the development of Internet technology, various search systems are increasingly being applied. In search technology, the quality of recall information often has a significant impact on the ranking results of recommendation systems.

[0004] Existing recall models mainly include dot product models based on a "double-tower structure" such as the FBT (Full Binary Tree) algorithm and the HNSW (Hierarchical Navigable Small World graphs) algorithm, and deep tree structure models based on the TDM (Tree-based Deep Match) algorithm and the JTM (Joint Optimization of Tree-based Index and Deep Model) algorithm. Summary of the Invention

[0005] Embodiments of the present application provide a method and apparatus for training an item numbering model.

[0006] In a first aspect, an embodiment of the present application provides a method for training an item numbering model, the method including: obtaining a training sample set of training samples, including sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the corresponding sample item information; obtaining an initial item numbering model; using the sample user information of the training samples in the training sample set as input for the initial item numbering model to obtain a probability of the sample item number information corresponding to the input sample user information; maximizing the obtained probability of the sample item number information corresponding to the input sample user information, adjusting structural parameters of the initial item numbering model, and training to obtain an item numbering model; wherein the item numbering model is for representing the correspondence relationship between the input sample user information and the sample item number information and the correspondence relationship between the sample item information and the sample item number information.

[0007] In one embodiment, the step of obtaining the training sample set includes the steps of obtaining an initial training sample set of initial training samples including sample user information and sample item information corresponding to the sample user information; randomly generating sample item number information corresponding to the initial training samples in the initial training sample set; and combining the initial training samples and the randomly generated sample item number information to obtain training samples in the training sample set.

[0008] In one embodiment, the method further includes, in the process of adjusting the structural parameters of the initial item numbering model, a step of determining the degree of association between each sample item information and each item number information output by the initial item numbering model, and a step of adjusting the correspondence between the sample item information and the sample item number information based on the determined degree of association, wherein the degree of association is determined based on the probability of at least one item number information obtained by passing the input sample user information corresponding to the sample item information through the initial item numbering model.

[0009] In one embodiment, in the process of adjusting the structural parameters of the above-mentioned initial item numbering model, the step of identifying the relevance between each sample item information and each item number information output by the initial item numbering model includes the steps of: sequentially identifying the probability between the sample item information in each batch of training samples and each corresponding item number information in a chronological order, and using this as a sub-relevance; and combining the identified sub-relevance to generate a relevance between each sample item information and each item number information, and the weight of the sub-relevance in the relevance corresponds to the proximity between the current time and the training time.

[0010] In one embodiment, the step of adjusting the correspondence between the sample item information and the sample item number information according to a distribution consistent with the identified degree of relevance based on the above-mentioned identified degree of relevance includes adjusting the correspondence between the sample item information and the sample item number information according to a distribution consistent with the identified degree of relevance based on a predetermined penalty function, wherein the penalty function is positively correlated with the number of sample items corresponding to the item number information.

[0011] In one embodiment, the sample item information has a corresponding priority, and the step of adjusting the correspondence between the sample item information and the sample item number information based on the above-mentioned predetermined penalty function includes the steps of selecting a target number of sample item information from the sample item information corresponding to the same item number information based on the corresponding priority in response to determining that the number corresponding to the same item number information is greater than a predetermined threshold, and newly generating sample item number information corresponding to the selected target number of sample item information according to the identified degree of association.

[0012] In one embodiment, the numbering space corresponding to the sample item number information comprises a first number of layers, each layer in the first number of layers comprises a second number of dimensions, the first number being less than the second number.

[0013] In an embodiment, the step of using the sample user information of the training samples in the above training sample set as input to an initial item numbering model to obtain the probability of the sample item number information corresponding to the input sample user information includes: inputting the sample user information of the training samples in the training sample set into an initial item numbering model to obtain a probability distribution of at least one item number information corresponding to the input sample user information, wherein the at least one item number information includes the sample item number information corresponding to the input sample user information, and the probability in the probability distribution is obtained based on the sub-probabilities output by each layer of the numbering space, and the sub-probabilities are obtained based on the sample user information and the feature vectors output by the upper layer.

[0014] In a second aspect, an embodiment of the present application provides a method for generating recalled item information, the method including the steps of: acquiring user information, inputting the user information into an item numbering model pre-trained by any one of the methods of the first aspect to obtain item number information corresponding to the user information, and inputting the item number information and the user information into a pre-trained vector model to obtain recalled item information corresponding to the user information.

[0015] In one embodiment, the item numbering model and the vector model are jointly trained based on feature vectors corresponding to item number information.

[0016] In a third aspect, the present application provides an apparatus for training an item numbering model, the apparatus comprising: a sample obtaining means configured to obtain a training sample set of training samples including sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the corresponding sample item information; a model obtaining means configured to obtain an initial item numbering model; a generation means configured to obtain a probability of the sample item number information corresponding to the input sample user information using the sample user information of the training samples in the training sample set as an input of the initial item numbering model; and a training means configured to maximize the obtained probability of the sample item number information corresponding to the input sample user information, adjust structural parameters of the initial item numbering model, and perform training to obtain the item numbering model, wherein the item numbering model is for representing the correspondence relationship between the input sample user information and the sample item number information and the correspondence relationship between the sample item information and the sample item number information.

[0017] In one embodiment, the sample obtaining means comprises: an obtaining module configured to obtain an initial training sample set of initial training samples including sample user information and sample item information corresponding to the sample user information; a first generating module configured to randomly generate sample item number information corresponding to the initial training samples in the initial training sample set; and a combining module configured to combine the initial training samples and the randomly generated sample item number information into training samples in the training sample set.

[0018] In one embodiment, the device further includes an identification means configured to identify a degree of association between each sample item information and each item number information output by the initial item numbering model in the process of adjusting the structural parameters of the initial item numbering model, and an adjustment means configured to adjust the correspondence between the sample item information and the sample item number information based on the identified degree of association, wherein the degree of association is identified based on the probability of at least one item number information obtained by passing input sample user information corresponding to the sample item information through the initial item numbering model.

[0019] In one embodiment, the identification means includes an identification module configured to identify the probability between sample item information and corresponding item number information in the training samples of each batch in chronological order to generate sub-associations, and a fusion module configured to fuse the identified sub-associations to generate associativity between each sample item information and each item number information, and the weight of the sub-associations in the associativity corresponds to the proximity from the current time to the training time.

[0020] In one embodiment, the adjustment means is further configured to adjust the correspondence between the sample item information and the sample item number information according to a distribution consistent with the identified degree of association based on a predetermined penalty function, wherein the penalty function is positively correlated with the number of sample items corresponding to the item number information.

[0021] In one embodiment, the sample item information has a corresponding priority, and the adjustment means includes a selection module configured to select a target number of sample item information from the sample item information corresponding to the same item number information based on the corresponding priority in response to determining that the number corresponding to the same item number information is greater than a predetermined threshold, and a second generation module configured to newly generate sample item number information corresponding to the selected target number of sample item information according to the identified relevance.

[0022] In one embodiment, the numbering space corresponding to the sample item number information comprises a first number of layers, each layer in the first number of layers comprises a second number of dimensions, the first number being less than the second number.

[0023] In one embodiment, the generating means is further configured to input sample user information of the training samples in the training sample set into an initial item numbering model to obtain a probability distribution of at least one item number information corresponding to the input sample user information, wherein the at least one item number information includes sample item number information corresponding to the input sample user information, and the probabilities in the probability distribution are obtained based on the sub-probabilities output by each layer of the numbering space, and the sub-probabilities are obtained based on the sample user information and the feature vectors output by the upper layer.

[0024] In a fourth aspect, the present application provides an apparatus for generating recalled item information, the apparatus comprising: information acquiring means configured to acquire user information; numbering means configured to input the user information into an item numbering model pre-trained by any one of the methods of the first aspect to obtain item number information corresponding to the user information; and recalling means configured to input the item number information and the user information into a pre-trained vector model to obtain recalled item information corresponding to the user information.

[0025] In one embodiment, the item numbering model and the vector model are jointly trained based on feature vectors corresponding to item number information.

[0026] In a fifth aspect, the present application provides a server, the server comprising one or more processors and a storage device having one or more programs stored thereon, the one or more programs, when executed by the one or more processors, causing the one or more processors to implement a method according to any one of the implementation manners of the first aspect.

[0027] In a sixth aspect, the present application provides a computer-readable medium having stored thereon a computer program which, when executed by a processor, implements a method according to any one of the implementations of the first aspect.

[0028] The method and apparatus for training an item numbering model provided in the embodiments of the present application first obtain an initial item numbering model and a training sample set, where the training samples in the training sample set include sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the corresponding sample item information. Next, the sample user information of the training samples in the training sample set is used as input for the initial item numbering model to obtain the probability of the sample item number information corresponding to the input sample user information. Then, the probability of the obtained sample item number information corresponding to the input sample user information is maximized, and the structural parameters of the initial item numbering model are adjusted and training is performed to obtain an item numbering model, where the item numbering model represents the correspondence relationship between the input sample user information and the sample item number information, and the correspondence relationship between the sample item information and the sample item number information. In this way, item recommendations can be realized using the trained item numbering model, and search efficiency can be improved by using the item number information as an index. [Brief explanation of the drawings]

[0029] Other features, objects and advantages of the present application will become more apparent from the following non-limiting detailed description of the embodiments with reference to the drawings. [Figure 1] FIG. 1 is a diagram of an exemplary system configuration in which certain embodiments of the present application can be applied. [Figure 2a] 1 is a flowchart of an embodiment of a method for training an item numbering model according to the present application. [Figure 2b]FIG. 2 is a schematic diagram of a numbering space corresponding to item number information in an embodiment of a method for training an item numbering model according to the present application; [Figure 2c] FIG. 2 is a schematic flow chart illustrating identifying a probability distribution in an embodiment of a method for training an item numbering model according to the present application. [Figure 3] FIG. 1 is a schematic diagram of a usage scenario of a method for training an item numbering model in an embodiment of the present application; [Figure 4] 1 is a flowchart of an embodiment of a method for generating recalled item information according to the present application. [Figure 5] 1 is a block diagram of an embodiment of an apparatus for training an item numbering model according to the present application; [Figure 6] 1 is a schematic diagram of an embodiment of an apparatus for generating recalled item information according to the present application; [Figure 7] FIG. 1 is a schematic diagram illustrating the configuration of an electronic device suitable for implementing an embodiment of the present application. DETAILED DESCRIPTION OF THE INVENTION

[0030] The present application will now be described in more detail with reference to the drawings and examples. It should be understood that the specific examples described herein are merely illustrative of the relevant invention and are not intended to limit the scope of the present invention. For the sake of convenience, only parts relevant to the invention are shown in the drawings.

[0031] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments in the present application may be combined with each other. The present disclosure will now be described in detail with reference to the drawings and examples.

[0032] FIG. 1 illustrates an exemplary configuration 100 in which the method for training an item numbering model or the apparatus for training an item numbering model of the present application may be applied.

[0033] 1, system configuration 100 may include terminal devices 101, 102, 103, networks 104, 106, and servers 105, 107. Networks 104, 106 are for providing a medium for communication links between terminal devices 101, 102, 103 and server 105, and between server 105 and server 107. Networks 104, 106 may include various connection types, such as, for example, wired, wireless communication links, or fiber optic cables.

[0034] The terminal devices 101, 102, and 103 interact with a server 105 via a network 104 to receive or send messages, etc. Various communication client applications are installed on the terminal devices 101, 102, and 103, such as a web browser application, a shopping application, a search application, an instant messenger, an email client, social platform software, and a reading application.

[0035] The terminal devices 101, 102, and 103 may be hardware or software. If the terminal devices 101, 102, and 103 are hardware, they may be various electronic devices that have a display and support search, including, but not limited to, smartphones, tablets, e-book readers, laptops, and desktop computers. If the terminal devices 101, 102, and 103 are software, they are installed on the electronic devices listed above. They may be implemented as multiple software programs or software modules (e.g., software programs or software modules for providing distributed services) or as individual software programs or software modules. No specific limitations are provided here.

[0036] The server 105 may be a server that provides various services. For example, it may be a background server that provides support for web pages displayed on the terminal devices 101, 102, and 103. The server 107 may be a server for training an item numbering model. The background server 105 can obtain the trained item numbering model from the server 107. Then, it uses the obtained item numbering model to perform processing such as analysis on the user information received from the terminal device, generates processing results (e.g., sentences or product information that match the user information), and feeds them back to the terminal device.

[0037] It should be noted that the server 105 may be used to train the item numbering model, whereby the trained item numbering model may be directly stored locally on the server 105. The server 105 can directly extract the locally stored item numbering model. In this case, the network 104 and the server 107 may not be present.

[0038] The server may be hardware or software. If the server is hardware, it may be realized as a distributed server cluster consisting of multiple servers, or as an individual server. If the server is software, it may be realized as multiple software programs or software modules (e.g., software programs or software modules for providing distributed services), or as individual software programs or software modules. No specific limitations are imposed here.

[0039] It should be noted that the method for training an item numbering model provided in the embodiments of the present application is generally executed by the server 105 or 107. Correspondingly, the device for training the item numbering model is generally installed in the server 105 or 107.

[0040] It should be understood that the number of terminal devices, networks, and servers shown in Figure 1 is for illustrative purposes only, and any number of terminal devices, networks, and servers may be provided according to the needs of implementation.

[0041] 2a, there is shown a flow 200 of an embodiment of a method for training an item numbering model according to the present application. The method for training an item numbering model includes the following steps:

[0042] Step 201: Obtain a training sample set.

[0043] In this embodiment, the entity (server 105 shown in FIG. 1 ) that executes the method for training an item numbering model can acquire a training sample set via a wired or wireless connection. The training samples in the training sample set may include sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the corresponding sample item information. The item number information may represent a category to which the item belongs. The item number information may include various forms, such as numbers, letters, or character strings composed of these, but is not particularly limited here.

[0044] For example, the sample user information may include information related to the user acquired by the recommendation system, such as a user name or a user label. The sample product information corresponding to the sample user may include, for example, information about actual clicks by the user corresponding to the sample product information, or information that best matches the sample user information identified based on a predetermined recommendation algorithm, and is not particularly limited here. The sample product number information corresponding to the corresponding sample product information may be identified based on a predetermined numbering space. For example, if the numbering space is 1 to 1000, the sample product number information corresponding to the corresponding sample product information may be any number between 1 and 1000.

[0045] The correspondence between the sample item information and the sample item number information may be one-to-one, multiple-to-one, or one-to-multiple. It should be understood that in practical applications, an item often belongs to different categories when viewed from different classification perspectives. For example, "chocolate" may belong to either the "food" category or the "gift" category. A sentence about "university" may belong to either the "education" category or the "youth" category. Furthermore, due to the finite nature of the numbering space and the long tail effect in item classification, a method in which one sample item number information corresponds to multiple sample item information can improve representation efficiency.

[0046] In an alternative implementation of this embodiment, the execution entity may obtain the training sample set by the following steps:

[0047] First step: Obtain an initial training sample set.

[0048] In these implementations, the initial training samples in the initial training sample set may include sample user information and sample product information corresponding to the sample user information, and the sample user information and sample product information may be consistent with the above description.

[0049] Second step: randomly generate sample item number information corresponding to the initial training samples in the initial training sample set;

[0050] In these implementation methods, the executing entity may generate sample item number information corresponding to each initial training sample by randomly assigning the initial training samples in the initial training sample set to a numbering space corresponding to the sample item number information.

[0051] Third step: The initial training sample and the randomly generated sample item number information are combined as training samples in the training sample set.

[0052] Based on the above selective implementation method, the execution entity can make full use of the numbering space by random allocation, avoiding that a certain sample item number corresponds to many sample item information, and improves search efficiency in the form of partition by sample item number information.

[0053] In an alternative implementation of this embodiment, the numbering section (shown in FIG. 2b) corresponding to the sample item number information may have a first number of layers (d shown in FIG. 2b). Each layer in the first number of layers may have a second number of dimensions (k shown in FIG. 2b). The first number is generally smaller than the second number. For example, the first number may be 3 and the second number may be 1000. Therefore, the numbering space may include 10003 kinds of numbers. The sample item number information may be, for example, (36, 27, 20) or (18, 27, 35). The sample item number information may represent a position in the numbering space, for example, the 36th dimension in the first layer, the 27th dimension in the second layer, the 20th dimension in the third layer, and the 18th dimension in the first layer, the 27th dimension in the second layer, and the 35th dimension in the third layer. The 27th dimension in the second layer can be considered as the "common" node of the above two sample item number information.

[0054] Based on the above selective implementation method, the item number information is embodied by coding intervals represented by structural layers and dimensions, and the item number information can be associated by "common" nodes, as opposed to cluster centers in existing clustering techniques.

[0055] Step 202: Obtain an initial item numbering model.

[0056] In this embodiment, the executing entity can obtain the training sample set through a wired connection or a wireless connection. The initial item numbering model may include various classification artificial neural network (ANN) structures.

[0057] In this embodiment, the executing entity may obtain the initial item numbering model pre-stored locally, or may obtain the initial item numbering model from a communicatively connected electronic device, and there is no particular limitation here.

[0058] Step 203: Take the sample user information of the training sample in the training sample set as the input of the initial item numbering model, and obtain the probability of the sample item number information corresponding to the input sample user information.

[0059] In this embodiment, the executing entity can input sample user information of the training sample in the training sample set to an initial item numbering model, and obtain the probability of the sample item number information corresponding to the input sample user information. Generally, the initial item numbering model can output a probability distribution of at least one item number information corresponding to the input sample user information. The at least one item number information can generally include sample item number information corresponding to the input sample user information. The item number information is, in other words, a category classified in the classification model. The executing entity can obtain the probability of the sample item number information corresponding to the input sample user information from the obtained at least one item number information.

[0060] In an alternative implementation of this embodiment, the executing entity can also input sample user information of the training samples in the training sample set into the initial item numbering model based on the layer and dimension included in the numbering space corresponding to the sample item number information to obtain a probability distribution of at least one item number information corresponding to the input sample user information. The at least one item number information generally includes sample item number information corresponding to the input sample user information. The probabilities in the probability distribution are obtained based on the sub-probabilities output by each layer of the numbering space. The sub-probabilities are obtained based on the sample user information and the feature vectors output by the upper layer.

[0061] For example, with reference to FIG. 2c, the executing entity may use the initial item numbering model to determine a probability distribution corresponding to each dimension in the first layer based on the sample user information. Then, the executing entity may determine a probability distribution corresponding to each dimension in the second layer based on the sample user information and the generated feature vectors of the probability distributions corresponding to each dimension. Then, the executing entity may determine a probability distribution corresponding to each dimension in the second layer based on the sample user information and the generated feature vectors of the first and second sub-probabilities. Then, the executing entity may determine a probability distribution corresponding to each dimension in the third layer based on the sample user information and the generated feature vectors of the first and second sub-probabilities. Then, the executing entity may determine a probability corresponding to a dimension corresponding to a dimension in the third layer based on the sample user information and the generated feature vectors of the first and second sub-probabilities. Finally, the execution entity may use the fusion result (e.g., successive multiplication product) of the first sub-probability, the second sub-probability, and the third sub-probability as the probability of the sample item number information corresponding to the input sample user information.The execution entity may generate a probability distribution of at least one item number information corresponding to the input sample user information based on the probability distributions output by each layer, using a method consistent with the method for generating the probability of the sample item number information described above.

[0062] Based on the selective implementation method, the implementing entity can fully utilize the information transmission between each layer in the numbering space, and train the model to learn the relationship between the generated item number information, thereby improving the numbering effect and providing a technical basis for efficient searching, while significantly reducing the cost of manual classification and improving efficiency.

[0063] Step 204: Maximize the probability of the sample item number information corresponding to the obtained input sample user information, adjust the structural parameters of the initial item numbering model, and perform training to obtain the item numbering model.

[0064] In this embodiment, by maximizing the probability of the obtained sample item number information corresponding to the input sample user information, the executing entity can adjust the structural parameters of the initial item numbering model in various ways and perform training to obtain an item numbering model. The item numbering model represents the correspondence between the input sample user information and the sample item number information, and the correspondence between the input sample information and the sample item number information. For example, the executing entity can adjust the structural parameters of the initial item numbering model based on a trained loss function. The loss function may, for example, include the negative logarithm of the probability of the sample item number information corresponding to the input sample user information. Thus, the executing entity can adjust the structural parameters of the initial item numbering model using various adaptive machine learning algorithms to maximize the probability of the obtained sample item number information corresponding to the input sample user information.

[0065] In an alternative implementation of this embodiment, the execution entity may execute the following steps in succession:

[0066] First step: In the process of adjusting the structural parameters of the initial item numbering model, the degree of association between each sample item information and each item number information output by the initial item numbering model is identified.

[0067] In these implementations, the relevance can be determined based on the probability of at least one item number information obtained by passing input sample user information corresponding to the sample item information through the initial item numbering model. For example, for each training sample (e.g., training sample A, training sample B, training sample C) in each batch of training samples in the training sample set, the executing entity can obtain the probability of at least one item number information output by the initial item numbering model and corresponding to the sample user information of each training sample. Training sample A may include sample user information x1 and sample item information y1. Training sample B may include sample user information x2 and sample item information y2. Training sample C may include sample user information x3 and sample item information y1. For example, the probabilities that the item number information corresponding to training sample A is 009, 015, and 108 may be 0.7, 0.2, and 0.1, respectively. The probabilities that the item number information corresponding to training sample B is 009, 015, and 108 may be 0.3, 0.1, and 0.6, respectively. The probabilities that the item number information corresponding to training sample C is 009, 015, and 108 may be 0.8, 0.1, and 0.1, respectively. The executing entity may determine the degree of association between each sample item information and each item number information in various ways, such as a maximum value, a minimum value, or an average value, based on the probability of at least one item number information corresponding to the sample user information of each obtained training sample. For example, the executing entity may determine the degree of association between sample item information y1 and item number information 009, 015, and 108 as 0.75, 0.15, and 0.1, respectively, and the degree of association between sample item information y2 and item number information 009, 015, and 108 as 0.3, 0.1, and 0.6, respectively.

[0068] In the process of adjusting the structural parameters of the above initial item numbering model, multiple batches of training samples are often used, so it can be seen that the implementing entity can determine the degree of association between each sample item information and each item number information based on the probability of at least one item number information corresponding to each batch of training samples obtained.

[0069] Alternatively, the executing entity may identify the degree of association between each sample item information and each item number information according to the following steps.

[0070] S1: The probability of the sample item information in the training sample of each batch being associated with each corresponding item number information is determined sequentially in time series, and is used as the sub-relevance.

[0071] In these implementation methods, the execution body may determine the probability of the sample item information in each batch of training samples and the corresponding item number information in a chronological order, and use this as the sub-relevance. The method for determining the sub-relevance may be the same as the related description in the first step described above, so a redundant description will be omitted here.

[0072] S2: The identified sub-relevances are merged to generate a relevance between each piece of sample item information and each piece of item number information.

[0073] In these implementations, the executing entity may combine sub-relationships determined in various ways to generate a relevance between each sample item information and each item number information. The weight of the sub-relevance in the relevance generally corresponds to the proximity of the current time to the training time, i.e., the most recently trained sample has a high weight of the corresponding sub-relevance in the relevance. For example, the combining method may include, but is not limited to, a weighting method in which weights are preset, an exponential moving average method, etc.

[0074] For example, training samples D, E, and F in other batches trained after the training samples A, B, and C may include sample user information x1, sample product information y2, sample user information x3, sample product information y1, sample user information x2, and sample product information y3, respectively. The probabilities that the product number information corresponding to training samples D, E, and F is 009, 015, and 108 may be 0.2, 0.2, 0.6; 0.7, 0.1, 0.2; and 0.3, 0.2, and 0.5, respectively. Then, the executing entity can determine the sub-relevance between sample product information y1 and the product number information 009, 015, and 108 in the training samples of the batches of training samples D, E, and F as 0.7, 0.1, and 0.2, respectively. The executing entity determines the sample item information y1 based on the fact that the sub-relevance between the sample item information y1 and the item number information 009, 015, and 108 in the training samples of the batches of training samples A, B, and C for which the executing entity has been identified is 0.75, 0.15, and 0.1, respectively. y1 The sub-relevance (weighted at 0.4 and 0.6) between the sample product information 009, 015, and 108 can be determined as 0.72, 0.12, and 0.16, respectively. y1 The degree of association between the item number information 009 and the item number information 009 is calculated as 0.75×0.4+0.7×0.6=0.72.

[0075] In the above selective implementation, the training samples of each batch may belong to the same epoch or different epochs. Therefore, when the training samples in each epoch are not completely the same "stream" sample scene, giving more weight to the results of the samples that have recently participated in training can meet the time-sensitive needs of online learning and application, and contribute to the trained model having higher generalization ability.

[0076] Second step: Adjust the correspondence between the sample item information and the sample item number information based on the identified relevance.

[0077] In these implementation methods, the executing entity determines whether the correspondence between the adjusted sample item information and the sample item number information matches the correlation generated in the first step based on the identified correlation. do In this way, the correspondence between the sample item information and the sample item number information can be adjusted. y1 The correlation coefficients (weighted at 0.4 and 0.6) between the sample item information and the item number information 009, 015, and 108 are determined to be 0.72, 0.12, and 0.16, respectively. y1 If the sample product number information that originally corresponds to the sample product number is not "009", the execution body y1 The sample item number information corresponding to the item number "009" can be adjusted to the item number information "009" having the highest degree of association.

[0078] Optionally, the execution entity may store the adjusted correspondence between the sample item information and the sample item number information, so that the trained item numbering model and the adjusted correspondence between the sample item information and the sample item number information can be obtained through training the item numbering model for multiple epochs and alternating adjustment of the correspondence.

[0079] Based on the selective implementation method, the executing entity can adjust the correspondence between sample item information and sample item number information so as to optimize the map between each sample item information and item number information and improve search efficiency.

[0080] Optionally, the executing entity may adjust the correspondence between the sample item information and the sample item number information according to a distribution consistent with the identified relevance based on a predetermined penalty function. The penalty function may be positively correlated with the number of sample items corresponding to the item number information. For example, the penalty function may be proportional to the square of the number of sample items corresponding to the item number information.

[0081] Based on the above selective implementation method, the penalty function can be used to avoid that a certain sample product number information corresponds to many sample product information, which can contribute to reducing the burden on the service side in the model application process.

[0082] Optionally, the sample item information may have a corresponding priority, and the execution entity may adjust the correspondence between the sample item information and the sample item number information according to the corresponding priority according to the following steps:

[0083] S1: In response to determining that the number corresponding to the same item number information is greater than a predetermined threshold, select a target number of sample item information from the sample item information corresponding to the same item number information based on an appropriate priority.

[0084] In these implementations, the executing entity can select a target number of sample item information from the sample item information corresponding to the same item number information based on the corresponding priority in response to determining that the number corresponding to the same item number information is greater than a predetermined threshold. For example, assume that the item number information with the highest relevance matching corresponding to sample item information y2, y5, y8, and y9 is all 009. The predetermined threshold is, for example, 3. The executing entity can then select the sample item information with the lowest priority from the sample item information y2, y5, y8, and y9. Alternatively, the executing entity can select sample item information with a priority lower than a predetermined priority threshold from the sample item information y2, y5, y8, and y9, such as y2 and y5. The executing entity can then randomly select a target number of sample item information from the sample item information with a priority lower than the predetermined priority threshold.

[0085] S2: In accordance with the identified relevance, newly generate sample item number information corresponding to the selected target number of sample item information.

[0086] In these implementation methods, the execution entity can regenerate the target number of sample item information selected in step S1 into sample item number information that matches the relevance degree of each identified next relevance. For example, the execution entity can identify the item number information corresponding to the second highest relevance degree corresponding to the selected sample item information as the sample item number information corresponding to the sample item information.

[0087] Please continue to refer to FIG. 3. FIG. 3 is a schematic diagram of a usage scenario of the method for training an item numbering model of an embodiment of the present application. In the usage scenario of FIG. 3, a background server can obtain a training sample set 301 and an initial item numbering model 302. The training sample set may include "Zhang San abc" as sample user information, "mobile phone" as sample item information, and "015" as sample item number information. The background server can input "Zhang San abc" 3011 in the training sample into the initial item numbering model 302 to obtain a probability corresponding to "mobile phone", for example, 0.5 (shown as 303 in FIG. 3). The background server can adjust the model parameters of the initial item numbering model 302 using various algorithms (e.g., EM algorithm) to maximize the probability of outputting "mobile phone" corresponding to "Zhang San abc" 3011 in the training sample. This identifies the trained initial item numbering model 302 as the trained item numbering model.

[0088] Currently, one prior art approach maps commonly recalled information into a vector space and then uses an algorithm such as approximate nearest neighbor (ANN) to recall the closest information. The limitations of the vector space size and the distance used (typically dot product / Euclidean distance or cosine measure) prevent the full utilization of the model's capabilities. Furthermore, because the vector mapping model and the structure constructed by the ANN are relatively independent, joint training using ANN structure information is not possible during model training. Furthermore, the depth tree structure used in the prior art also makes it difficult to jointly train the tree's structural features with the model parameters due to the highly discretized tree structure. The method provided in the above-described embodiments of the present application introduces item number information as an intermediate variable between user information and item information, and through training, obtains an item numbering model that represents the correspondence between input sample user information and sample item number information, and the correspondence between sample item information and sample item number information. This allows the item numbering model to represent the correspondence between user information and item information, which can be used for item recommendation, and can also use the item number information as an index to improve search efficiency, and further realizes comprehensive model training.

[0089] Further, reference is made to Figure 4. Figure 4 shows a flow 400 of an embodiment of a method for generating recalled item information. The flow 400 of the method for generating recalled item information includes the following steps:

[0090] Step 401: User information is acquired.

[0091] In this embodiment, an entity executing the method for generating recalled item information (e.g., server 105 shown in FIG. 1 ) can acquire user information via a wired or wireless connection. The user information may include various information related to the user, such as a user name and a user label. The user information may further include user past click information. The past click information may include at least one of product information, text, video, and music, but is not limited to these.

[0092] Step 402: Input the user information into the pre-trained item numbering model to obtain item number information corresponding to the user information.

[0093] In this embodiment, the executing entity inputs the user information acquired in step 401 into a pre-trained item numbering model to obtain item number information corresponding to the user information. The item numbering model can be obtained by training according to the methods described in the above embodiments and optional implementation methods.

[0094] Step 403: Input the item number information and user information into a pre-trained vector model to obtain recalled item information corresponding to the user information.

[0095] In this embodiment, the executing entity inputs the item number information acquired in step 402 and the user information acquired in step 401 into a pre-trained vector model to obtain recalled item information corresponding to the user information. The vector model may include various machine learning models used for recall, such as a network structure incorporating a field-aware factorization machine (FFM) model and / or deep neural networks (DNN).

[0096] In one alternative implementation of this embodiment, the item numbering model and the vector model are jointly trained based on feature vectors that generally correspond to item number information.

[0097] Based on the above selective realization method, joint training is realized by using the feature vector corresponding to the item number information as an intermediate carrier between the above item numbering model and the above vector model, thereby making the connection between the two models closer and improving the matching degree of the generated recalled item information.

[0098] 4, the flow chart 400 of the method for generating recalled item information in this embodiment shows a step of inputting the output of the item numbering model into the vector model as a newly added dimension to obtain recalled item information corresponding to user information. In this way, the invention described in this embodiment can generate recalled item information not only by using user information but also by using the generated item number information, and can further improve the matching degree of the generated recalled item information.

[0099] Next, please refer to Figure 5. As an implementation of the methods shown in the above figures, the present application provides an embodiment of an apparatus for training an item numbering model. The embodiment of the apparatus corresponds to the embodiment of the method shown in Figure 2. The apparatus can be specifically applied to various electronic devices.

[0100] As shown in FIG. 5, the apparatus 500 for training an item numbering model provided by this embodiment includes a sample obtaining means 501, a model obtaining means 502, a generating means 503, and a training means 504. The sample obtaining means 501 is configured to obtain a training sample set of training samples including sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the corresponding sample item information. The model obtaining means 502 is configured to obtain an initial item numbering model. The generating means 503 is configured to use the sample user information of the training samples in the training sample set as input for the initial item numbering model to obtain the probability of the sample item number information corresponding to the input sample user information. The training means 504 is configured to maximize the probability of the sample item number information corresponding to the obtained input sample user information, adjust the structural parameters of the initial item numbering model, and perform training to obtain the item numbering model. The item numbering model is used to represent the correspondence between the input sample user information and sample item number information, and the correspondence between the input sample item information and sample item number information.

[0101] In this embodiment, in the apparatus 500 for training an item numbering model, the specific processing of the sample obtaining means 501, the model obtaining means 502, the generating means 503, and the training means 504 and the effects thereof can be referred to the relevant descriptions of steps 201, 202, 203, and 204 of the corresponding embodiment in Figure 2, respectively, and redundant descriptions will be omitted here.

[0102] In an alternative implementation of this embodiment, the sample acquiring means 501 may include an acquiring module (not shown), a first generating module (not shown), and a combining module (not shown). The acquiring module may be configured to acquire an initial training sample set of initial training samples including sample user information and sample item information corresponding to the sample user information. The first generating module may be configured to randomly generate sample item number information corresponding to the initial training samples in the initial training sample set. The combining module may be configured to combine the initial training samples and the randomly generated sample item number information to form training samples in the training sample set.

[0103] In an alternative implementation of this embodiment, the apparatus 500 for training the item numbering model may further include a determination means (not shown) and an adjustment means (not shown). The determination means may be configured to determine the degree of association between each sample item information and each item number information output by the initial item numbering model in the process of adjusting the structural parameters of the initial item numbering model. The degree of association may be determined based on the probability of at least one item number information obtained by passing input sample user information corresponding to the sample item information through the initial item numbering model. The adjustment means may be configured to adjust the correspondence between the sample item information and the sample item number information based on the determined degree of association.

[0104] In an alternative implementation of this embodiment, the identifying means may include an identifying module (not shown) and a fusion module (not shown). The identifying module may be configured to identify the probability between sample item information and each corresponding item number information in each batch of training samples in a time series order, and generate sub-associations. The fusion module may be configured to fuse the identified sub-associations to generate an association between each sample item information and each item number information. The weight of the sub-associations in the association may correspond to the proximity between the current time and the training time.

[0105] In an alternative implementation of this embodiment, the adjusting means may be further configured to adjust the correspondence between the sample item information and the sample item number information according to a distribution consistent with the determined relevance based on a predetermined penalty function, which is positively correlated with the number of sample items corresponding to the item number information.

[0106] In an alternative implementation of this embodiment, the sample item information may have a corresponding priority. The adjusting means may include a selection module (not shown) and a second generating module (not shown). The selection module may be configured to select a target number of sample item information from the sample item information corresponding to the same item number information based on the corresponding priority in response to determining that the number corresponding to the same item number information is greater than a predetermined threshold. The second generating module may be configured to newly generate sample item number information corresponding to the selected target number of sample item information according to the determined relevance.

[0107] In an alternative implementation of this embodiment, the numbering space corresponding to the sample item number information may have a first number of layers, and each layer in the first number of layers may have a second number of dimensions, with the first number generally being smaller than the second number.

[0108] In an alternative implementation of this embodiment, the generating means 503 may be further configured to input sample user information of the training samples in the training sample set into an initial item numbering model to obtain a probability distribution of at least one item number information corresponding to the input sample user information. The at least one item number information may include sample item number information corresponding to the input sample user information. The probabilities in the probability distribution are obtained based on the sub-probabilities output by each layer of the numbering space. The sub-probabilities are obtained based on the sample user information and the feature vectors output by the upper layer.

[0109] The device provided in the above embodiment of the present application acquires a training sample set using a sample acquisition means 501. The training samples in the training sample set include sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the corresponding sample item information. Next, a model acquisition means 502 acquires an initial item numbering model. Then, a generation means 503 uses the sample user information of the training samples in the training sample set as input for the initial item numbering model to obtain the probability of the sample item number information corresponding to the input sample user information. Finally, a training means 504 maximizes the probability of the obtained sample item number information corresponding to the input sample user information, adjusts the structural parameters of the initial item numbering model, and performs training to obtain an item numbering model. The item numbering model represents the correspondence between the input sample user information and the sample item number information, and the correspondence between the input sample user information and the sample item number information. This allows item recommendations to be realized using the trained item numbering model, and search efficiency can be improved using the item number information as an index.

[0110] Next, referring to Figure 6, as an implementation of the methods shown in the above figures, the present application provides an embodiment of an apparatus for generating recalled item information. The embodiment of the apparatus corresponds to the embodiment of the method shown in Figure 4. The apparatus can be specifically applied to various electronic devices.

[0111] 6, the apparatus 600 for generating recalled item information provided by this embodiment includes an information acquiring means 601, a numbering means 602, and a recalling means 603. The information acquiring means 601 is configured to acquire user information. The numbering means 602 is configured to input the user information into an item numbering model pre-trained by the method for training an item numbering model described above to obtain item number information corresponding to the user information. The recalling means 603 is configured to input the item number information and the user information into a pre-trained vector model to obtain recalled item information corresponding to the user information.

[0112] In this embodiment, in the device 600 for generating information on recalled items, the specific processing and effects of the information acquisition means 601, numbering means 602, and recall means 603 can be referred to the relevant explanations of steps 401, 402, and 403 of the corresponding embodiment in Figure 4, as well as their selective implementation methods, and redundant explanations will be omitted here.

[0113] In one alternative implementation of this embodiment, the item numbering model and the vector model are jointly trained based on feature vectors that generally correspond to item number information.

[0114] The device provided in the above embodiment of the present application first acquires user information through information acquisition means 601. Next, numbering means 602 inputs the user information into an item numbering model pre-trained by the above-mentioned method for training an item numbering model to obtain item number information corresponding to the user information. Recalling means 603 inputs the item number information and user information into a pre-trained vector model to obtain recalled item information corresponding to the user information. This not only utilizes user information, but also can use the generated item number information to generate recalled item information, further improving the matching degree of the generated recalled item information.

[0115] Next, reference is made to FIG. 7. FIG. 7 shows a schematic diagram of an electronic device (e.g., the server shown in FIG. 1) 700 for realizing an embodiment of the present disclosure. Terminal devices in the embodiment of the present disclosure include, but are not limited to, mobile terminals such as mobile phones, notebook computers, digital broadcast receivers, PDAs (personal digital assistants), PADs (tablets), PMPs (portable multimedia players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. The server shown in FIG. 7 is merely an example and does not limit the functions and scope of use of the embodiment of the present disclosure.

[0116] 7, electronic device 700 may include a processing unit (e.g., a central processing unit, a graphics processor, etc.) 701 that performs various appropriate operations and processes according to programs stored in read-only memory (ROM) 702 or loaded from storage device 708 into random access memory (RAM) 703. RAM 703 further stores various programs and data necessary for the operation of electronic device 700. The processing unit 701, ROM 702, and RAM 703 are interconnected by a bus 704. An input / output (I / O) interface 705 is also connected to bus 704.

[0117] Typically, the following devices can be connected to the input / output (I / O) interface 705: input devices 706, including, for example, a touch screen, touch panel, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 707, including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 708, including, for example, a magnetic tape, hard disk, etc.; and communication devices 709. The communication devices 709 can allow the electronic device 700 to exchange data with other devices via wireless or wired communication. While FIG. 7 illustrates the electronic device 700 with various devices, it should be understood that it is not required to implement or include all of the devices shown. Alternatively, the electronic device 700 may be implemented or include more or fewer devices. Each block illustrated in FIG. 7 may represent one device or multiple devices, if desired.

[0118] In particular, according to an embodiment of the present application, the processes described with reference to the flowcharts can be implemented as a computer software program. For example, an embodiment of the present application relates to a computer program product, which relates to a computer program stored on a computer-readable medium. The computer program includes program code for performing the method illustrated in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from the Internet via the communication device 709, or installed from the storage device 708 or from the ROM 702. When the computer program is executed by the processing device 701, the functions defined in the method of the embodiment of the present application are performed.

[0119] It should be noted that the computer-readable medium referred to in this application may be a computer-readable signal medium, a computer-readable storage medium, or any combination of the above. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of computer-readable storage media include, but are not limited to, an electrical connection having one or more leads, a portable computer magnetic disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a compact disc read-only memory (CD-ROM), an optical memory device, a magnetic memory device, or any combination of the above. In embodiments of this application, the computer-readable storage medium may be any tangible medium that contains or stores a program. The program may be used in or in combination with an instruction execution system, apparatus, or device. In embodiments of the present application, a computer-readable signal medium may include a propagated data signal, either in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take a variety of forms, including, but not limited to, electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium may also be any computer-readable medium other than a computer-readable storage medium. The computer-readable signal medium may transmit, propagate, or transmit a program for use in or in connection with an instruction execution system, apparatus, or device. The program code contained in the computer-readable medium may be transmitted by any suitable medium, including, but not limited to, electrical wires, optical tables, RF (Radio Frequency), or any combination of the above.

[0120] The computer-readable medium may be included in the server, or may exist independently of the server. The computer-readable medium has one or more programs stored therein, which, when executed by the server, cause the server to perform the following processes: Obtain a training sample set of training samples including sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the corresponding sample item information; Obtain an initial item numbering model; Use the sample user information of the training samples in the training sample set as input to the initial item numbering model to obtain the probability of the sample item number information corresponding to the input sample user information; Maximize the obtained probability of the sample item number information corresponding to the input sample user information, adjust the structural parameters of the initial item numbering model, and perform training to obtain the item numbering model; The item numbering model is intended to represent the correspondence between the input sample user information and the sample item number information, and the correspondence between the sample item information and the sample item number information.

[0121] Computer program code for carrying out operations of embodiments of the present application may be coded in one or more programming languages ​​or a combination thereof. Such programming languages ​​include object-oriented programming languages ​​such as Java, Smalltalk, C++, etc., and further include traditional procedural programming languages ​​such as "C" or similar programming languages. The program code may execute entirely on the user's computer, partially on the user's computer, as a standalone software package on the user's computer, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet Service Provider).

[0122] The flowcharts and block diagrams in the drawings illustrate system configurations, functions, and operations that can be implemented by systems, methods, and computer program products according to various embodiments of the present application. Each block in the flowcharts or block diagrams represents a module, program segment, or portion of code, which includes executable instructions for implementing one or more predetermined logical functions. Note that in some alternative implementations, the functions described in the blocks may occur out of the order described in the drawings. For example, two blocks shown in succession may actually be executed substantially simultaneously, or may sometimes be executed in the reverse order, depending on the functionality involved. Note that each block in the block diagrams and / or flowcharts, or a combination of blocks in the block diagrams and / or flowcharts, may be implemented by a dedicated hardware-based system that performs a predetermined function or operation, or by a combination of dedicated hardware and computer instructions.

[0123] The means according to the embodiments of the present application may be implemented by software or hardware. The described means may be installed in a processor, and may be described as a processor including, for example, a sample acquisition means, a model acquisition means, a generation means, and a training means. The names of these means may not necessarily be limiting. For example, the model acquisition means may be described as a "means for acquiring an initial item numbering model."

[0124] The above description is a description of the preferred embodiments and operational technical principles of the present application. Those skilled in the art should understand that the scope of the invention in the embodiments of the present application is not limited to the technical solution based on the specific combination of the above technical features. It should also include other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the spirit of the invention. For example, it should also include technical solutions formed by replacing the above features with technical features having similar functions disclosed in the embodiments of the present application (but not limited to these).

Claims

1. 1. A method for training an item numbering model, performed by an electronic device, comprising: obtaining a training sample set, wherein the training samples in the training sample set include sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the sample item information; obtaining an initial item numbering model; Using the sample user information of the training samples in the training sample set as input for the initial item numbering model, and obtaining a probability of sample item number information corresponding to the input sample user information; and training the initial item numbering model to maximize the obtained probability of the sample item number information corresponding to the input sample user information to obtain an item numbering model, the method comprising: In the process of adjusting the initial item numbering model, a step of determining a degree of association between each sample item information and each item number information output by the initial item numbering model, wherein the degree of association is determined based on a probability of at least one item number information obtained via the initial item numbering model from the input sample user information corresponding to the sample item information; and a step of adjusting the correspondence between the sample item information and the sample item number information based on the determined degree of association, the step including adjusting the correspondence between the sample item information and the sample item number information based on the determined degree of association so that the sample item number information corresponding to the sample item information becomes the sample item number information having the largest degree of association, The item numbering model is for representing a correspondence relationship between the input sample user information and the sample item number information, and a correspondence relationship between the input sample item information and the sample item number information. A method for training an item numbering model.

2. The step of obtaining the training sample set comprises: obtaining an initial training sample set of initial training samples including the sample user information and the sample item information corresponding to the sample user information; Randomly generating sample item number information corresponding to the initial training samples in the initial training sample set; combining the initial training samples with randomly generated sample item number information as training samples in the training sample set; The method of claim 1 , comprising:

3. In the process of adjusting the structural parameters of the initial item numbering model, the step of identifying the degree of association between each sample item information and each item number information output by the initial item numbering model includes: a step of sequentially determining the probability of each sample item information in the training sample of each batch and each item number information corresponding to the sample item information, and setting the result as a sub-relevance; a step of combining the identified sub-relevances to generate a relevance between each sample item information and each item number information; The weight of the sub-relevance to the generated relevance coincides with the proximity of the current time to the training time. The method of claim 1.

4. The step of adjusting the correspondence between the sample item information and the sample item number information based on the determined degree of association includes: adjusting the correspondence between the sample item information and the sample item number information according to a distribution that matches the determined relevance based on a predetermined penalty function; The penalty function is positively correlated with the number of sample items corresponding to the item number information. The method of claim 1.

5. The sample item information has a priority; adjusting the correspondence between the sample item information and the sample item number information according to a distribution that matches the identified relevance based on the predetermined penalty function; In response to determining that the number of sample item information corresponding to the same item number information is greater than a predetermined threshold, selecting a target number of sample item information from the sample item information corresponding to the same item number information based on the priority; generating new sample item number information corresponding to the selected target number of sample item information items according to the determined relevance; The method of claim 4, comprising:

6. a numbering space corresponding to the sample item number information having a first number of layers, each layer in the first number of layers having a second number of dimensions, the first number being less than the second number; 6. The method according to any one of claims 1 to 5.

7. The step of using sample user information of the training samples in the training sample set as input of the initial item numbering model to obtain a probability of sample item number information corresponding to the input sample user information includes: inputting sample user information of the training samples in the training sample set into the initial item numbering model to obtain a probability distribution of at least one item number information corresponding to the input sample user information; The at least one item number information includes sample item number information corresponding to the input sample user information, and the probability in the probability distribution is obtained based on sub-probabilities output by each layer of the numbering space, and the sub-probabilities are obtained based on the sample user information and feature vectors output by upper layers. The method of claim 6.

8. obtaining user information; inputting the user information into an item numbering model pre-trained by the method of any one of claims 1 to 7 to obtain item number information corresponding to the user information; inputting the item number information and the user information into a pre-trained vector model to obtain recalled item information corresponding to the user information; Including, The item numbering model and the vector model are trained using a feature vector corresponding to item number information as an intermediate carrier between the item numbering model and the vector model. A method for generating recalled item information.

9. a sample obtaining means configured to obtain a training sample set, wherein the training samples in the training sample set include sample user information, sample item information corresponding to the sample user information, and sample item number information corresponding to the sample item information; a model obtaining means configured to obtain an initial item numbering model; a generating means configured to use the sample user information of the training samples in the training sample set as input for the initial item numbering model to obtain a probability of sample item number information corresponding to the input sample user information; training means configured to adjust structural parameters of the initial item numbering model to maximize the probability that the sample item number information corresponds to the input sample user information, thereby obtaining an item numbering model; 1. An apparatus for training an item numbering model, comprising: A determination means for determining the degree of association between each sample item information and each item number information output by the initial item numbering model during the process of adjusting the structural parameters of the initial item numbering model, wherein the degree of association is determined based on the probability of at least one item number information obtained via the initial item numbering model from the input sample user information corresponding to the sample item information; and an adjusting means for adjusting the correspondence between the sample item information and the sample item number information based on the specified degree of association, so that the sample item number information corresponding to the sample item information becomes the sample item number information having the largest degree of association. The item numbering model is for representing a correspondence relationship between input sample user information and sample item number information, and a correspondence relationship between input sample item information and sample item number information. A device for training an item numbering model.

10. information obtaining means configured to obtain user information; a numbering means configured to input the user information into an item numbering model pre-trained by the method of any one of claims 1 to 8 to obtain item number information corresponding to the user information; a recall means configured to input the item number information and the user information into a pre-trained vector model to obtain recalled item information corresponding to the user information; Equipped with The item numbering model and the vector model are trained using a feature vector corresponding to item number information as an intermediate carrier between the item numbering model and the vector model. An apparatus for generating recalled item information.

11. one or more processors; a storage device having stored thereon one or more programs that cause the one or more processors to perform the method of any one of claims 1 to 8, server.

12. A computer readable storage medium having stored thereon a computer program for causing a processor to carry out the method of any one of claims 1 to 8.

Images