Training method of model, material recall method and device

Abstract

The application discloses a model training method, a material recall method and device. The training method comprises the following steps: obtaining a plurality of user feature sets corresponding to a target user within a first preset time, and a material set composed of materials clicked by the user within the first preset time, wherein the user feature set is a set of user features corresponding to the target user each time the target user clicks a material; determining a feature vector corresponding to the user feature set based on the user feature set; iteratively training a target model to be trained through a plurality of feature vectors, inputting the feature vector into the target model to be trained each time the iteration is trained, and obtaining an interest preference vector corresponding to the target user output by the target model to be trained; determining an estimated click probability of a target material clicked by the target user within the first preset time based on the interest preference vector and a material vector matrix constructed according to the material set; and obtaining a target model trained in a case where the estimated click probability reaches a target value.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information recall, and in particular to a model training method, a material recall method and device. BACKGROUND

[0002] In a social scenario, a personalized recommendation system is responsible for pushing the micro-blog posts (i.e. materials) of each blogger to interested users in the form of message push, so that the users can obtain the interested information in time. The personalized recommendation system includes a recall module, which applies various recall algorithms to filter out the posts that the user may be interested in from a million-level candidate post library, to form a recall post library. Since the user collaborative filtering algorithm can calculate the similarity of users through the historical behavior of the users, and recommend the materials that the similar users like, the user collaborative filtering algorithm has been widely applied in the recall module of various personalized recommendation systems.

[0003] In the related art, the calculation steps of the user collaborative filtering algorithm are: constructing a behavior matrix based on the user identity (ID) and the material ID, then decomposing the behavior matrix into a user matrix representing the user interest preference and a material matrix representing the material attribute, in which case the interest preference of each user can be represented by the vector corresponding to the user in the user matrix, then grouping the users based on the vectors corresponding to the users, and finally counting the materials that the users in the same group like, and recalling the materials for other users in the same group who have not distributed the materials. However, the above scheme determines the user interest preference based on the user ID and the material ID, and has the problem of low accuracy of representing the user interest preference. SUMMARY

[0004] The present application discloses a model training method, a material recall method and device to solve the problem of low accuracy of representing the user interest preference in the related art.

[0005] To solve the above problems, the present application adopts the following technical solutions:

[0006] In a first aspect, the embodiments of the present application provide a model training method, comprising: obtaining a plurality of user feature sets corresponding to a target user within a first preset time, and a material set composed of materials clicked by the user within the first preset time, wherein the user feature set is a set of user features corresponding to the target user each time the target user clicks a material; determining a feature vector corresponding to the user feature set based on the user feature set; iteratively training a target model to be trained by a plurality of the feature vectors, wherein each iteration of training inputs the feature vector into the target model to be trained, and obtains an interest preference vector corresponding to the target user output by the target model to be trained; determining an estimated click probability of a target material clicked by the target user within the first preset time based on the interest preference vector and a material vector matrix constructed according to the material set; and obtaining a trained target model when the estimated click probability reaches a target value.

[0007] In a second aspect, the embodiments of the present application provide a material recall method, comprising: obtaining a user feature set corresponding to each of a plurality of users respectively; determining a feature vector corresponding to the user feature set based on the user feature set; inputting each of a plurality of the feature vectors into a target model respectively, and obtaining a plurality of interest preference vectors corresponding to the users output by the target model; determining at least one user group by clustering a plurality of the interest preference vectors; obtaining a click number of a plurality of materials by a user in a target user group within a second preset time respectively, wherein the target user group is any one of the at least one user group; determining a hot material list corresponding to the target user group based on the click number; and recalling materials for the users in the target user group based on the hot material list, wherein the target model is trained according to the model training method of the first aspect.

[0008] In a third aspect, an embodiment of the present application provides a model training apparatus, comprising: an acquisition module configured to acquire a plurality of user feature sets corresponding to a target user within a first preset time, and a material set composed of materials clicked by a user within the first preset time, wherein the user feature set is a set of user features corresponding to the target user each time the target user clicks a material; a first determination module configured to determine a feature vector corresponding to the user feature set based on the user feature set; a training module configured to iteratively train a target model to be trained by using a plurality of the feature vectors, wherein each iteration of the training inputs the feature vector into the target model to be trained, and acquires an interest preference vector corresponding to the target user output by the target model to be trained; a second determination module configured to determine an estimated click probability of a target material clicked by the target user within the first preset time based on the interest preference vector and a material vector matrix constructed according to the material set; and a obtaining module configured to obtain the target model trained when the estimated click probability reaches a target value.

[0009] In a fourth aspect, an embodiment of the present application provides a material recall apparatus, comprising: a first acquisition module configured to acquire a user feature set corresponding to each of a plurality of users; a first determination module configured to determine a feature vector corresponding to the user feature set based on the user feature set; a second acquisition module configured to input each of the feature vectors into a target model, and acquire a plurality of interest preference vectors corresponding to the users output by the target model; a second determination module configured to determine at least one user group by clustering the interest preference vectors; a third acquisition module configured to acquire a number of clicks on a plurality of materials by a user in a target user group within a second preset time, wherein the target user group is any one of the at least one user group; a third determination module configured to determine a hot material list corresponding to the target user group based on the number of clicks; and a recall module configured to recall materials for the users in the target user group based on the hot material list, wherein the target model is trained according to the model training method of the first aspect.

[0010] In a fifth aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory, wherein the memory stores programs or instructions executable on the processor, and the programs or instructions are executed by the processor to implement the steps of the method of the first aspect or the second aspect.

[0011] In a sixth aspect, an embodiment of the present application provides a readable storage medium, wherein the readable storage medium stores programs or instructions, and the programs or instructions are executed by a processor to implement the steps of the method of the first aspect or the second aspect.

[0012] In a seventh aspect, an embodiment of the present application provides a chip, which comprises a processor and a communication interface, the communication interface is coupled with the processor, and the processor is configured to run programs or instructions to implement the method according to the first aspect or the second aspect.

[0013] In an eighth aspect, an embodiment of the present application provides a computer program product stored in a storage medium, which is executed by at least one processor to implement the method according to the first aspect or the second aspect.

[0014] The embodiment of the present application provides a training method of a model, a plurality of user feature sets corresponding to a target user within a first preset time are acquired, and a material set composed of materials clicked by the user within the first preset time is acquired, a feature vector corresponding to the user feature set is determined based on the user feature set, a target model to be trained is iteratively trained through the plurality of feature vectors, and then a target material clicked by the target user within the first preset time is determined based on an interest preference vector corresponding to the target user output by the target model to be trained and a material vector matrix constructed according to the material set, an estimated click probability of the target material is determined, in the case that the estimated click probability reaches a target value, a target model trained is obtained, the target model trained can output an interest preference vector corresponding to a user based on a user feature set of the user, and the representation accuracy of the user interest preference is relatively high.

[0015] The embodiment of the present application provides a recall method of a material, a user feature set corresponding to a user is acquired respectively, a feature vector corresponding to the user feature set is determined based on the user feature set, each feature vector in the plurality of feature vectors is input into a target model respectively, a plurality of interest preference vectors corresponding to the user output by the target model are acquired, at least one user group is determined by clustering the plurality of interest preference vectors, the number of clicks of a plurality of materials by a user of the target user group within a second preset time is acquired respectively, a hot material list corresponding to the target user group is determined based on the number of clicks of the plurality of materials by the user of the target user group, and the materials are recalled for the user in the target user group based on the hot material list. The materials recalled for the user by the material recall method disclosed in the present application have a relatively high relevance to the interest preference of the user, and the plurality of users are clustered into at least one user group, and then the materials are recalled for the user in the user group based on the interest preference of each user group, so that the calculation amount of online services can be reduced. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 A flowchart of a training method of a model disclosed in an embodiment of the present application is shown;

[0017] Figure 2 A flowchart of a recall method of a material disclosed in an embodiment of the present application is shown.

[0018] Figure 3 A structural schematic diagram of a model training device disclosed by an embodiment of the application;

[0019] Figure 4 A structural schematic diagram of a material recall device disclosed by an embodiment of the application;

[0020] Figure 5 A structural schematic diagram of an electronic device disclosed by an embodiment of the application. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the application will be clearly described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some, but not all of the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of the application.

[0022] The terms "first", "second", and the like in the specification and claims of the application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually a category and do not limit the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims indicates at least one of the connected objects, and the character " / ", generally indicates that the front and rear associated objects are in an "or" relationship.

[0023] The model training method, the material recall method and the device provided by the embodiments of the application will be described in detail below with reference to the drawings, specific embodiments and application scenarios.

[0024] The embodiments of the application provide a model training method, Figure 1 A flowchart of a model training method disclosed by an embodiment of the application is shown in FIG. 1. As shown in the figure, the method comprises the following steps. Figure 1

[0025] S110, acquiring a plurality of user feature sets corresponding to a target user within a first preset time, and a material set composed of materials clicked by the user within the first preset time, wherein the user feature set is a set of user features corresponding to the target user each time the material is clicked.

[0026] For example, the first preset time can be 10 days, 20 days, 30 days, etc., which is not limited by the application.

[0027] ​It should be noted that the user corresponds to multiple user features each time the user clicks the material, and the multiple user features corresponding to the user constitute a user feature set corresponding to the user. The user feature set corresponding to the user each time the user clicks the material can be different. For example, the label included in the user feature set corresponding to the user when the user clicks the first material is geography, the label included in the user feature set corresponding to the user when the user clicks the second material is entertainment, and the label included in the user feature set corresponding to the user when the user clicks the third material is sports, and the like.

[0028] In the present application, the material set can include all materials clicked by all users within a first preset time. For example, the first preset time is 10 days, all users are 10 users, and each user clicks 10 materials per day. Therefore, the material set includes 1000 materials.

[0029] In S120, a feature vector corresponding to the user feature set is determined based on the user feature set.

[0030] In S130, the target model to be trained is iteratively trained by using multiple feature vectors. In each iteration, the feature vector is input into the target model to be trained, and an interest preference vector corresponding to the target user output by the target model to be trained is obtained.

[0031] After obtaining multiple feature vectors corresponding to the target user, the feature vectors are input into the target model to be trained for iterative training. The target model after training can output an interest preference vector corresponding to the user according to the input feature vector corresponding to the user.

[0032] In the present application, the target model can be a three-layer deep neural network (DNN) model. The number of neurons in the three layers of the DNN model can be 1024, 512, and 256, respectively. The third layer of neurons outputs an interest preference vector corresponding to the user. In this case, the performance of the target model is better.

[0033] In an implementation manner, the target model to be trained can be trained by using an Adam optimizer.

[0034] In S140, an estimated click probability of a target material clicked by the target user within the first preset time is determined based on the interest preference vector and a material vector matrix constructed according to the material set.

[0035] It should be noted that the dimension of the material vector matrix constructed according to the material set is the same as the number of neurons in the third layer of the DNN model. For example, when the number of neurons in the third layer of the DNN model is 256, the material vector matrix V constructed according to the material set can be a |V|*256 vector matrix, where |V| represents the number of materials included in the material set, and one material corresponds to one row in the material vector matrix V.

[0036] In a case where the estimated click probability reaches a target value, a trained target model is obtained.

[0037] The learning task of the target model requires determining, based on the interest preference vector corresponding to the target user output by the target model and the material vector matrix constructed according to the material set, that the estimated click probability of the target material clicked by the target user in the first preset time is maximum, and the estimated click probability of other materials not clicked by the target user in the first preset time is smaller. Based on this, in a case where the estimated click probability of the target material clicked by the target user in the first preset time reaches a maximum value, a trained target model is obtained.

[0038] The embodiment of the present application provides a model training method. A plurality of user feature sets corresponding to a target user in a first preset time and a material set composed of materials clicked by the user in the first preset time are obtained, and a feature vector corresponding to each user feature set is determined based on the user feature set. The target model to be trained is iteratively trained through a plurality of feature vectors. Then, based on the interest preference vector corresponding to the target user output by the target model to be trained and the material vector matrix constructed according to the material set, the estimated click probability of the target material clicked by the target user in the first preset time is determined. In a case where the estimated click probability reaches a target value, a trained target model is obtained. The trained target model can output an interest preference vector corresponding to a user based on the user feature set of the user, and the representation accuracy of the user interest preference is high.

[0039] In the embodiment of the present application, the determination of the estimated click probability of the target material clicked by the target user in the first preset time based on the interest preference vector and the material vector matrix constructed according to the material set can include: performing vector inner product operation on the interest preference vector and a target material vector, and performing normalization processing to determine the estimated click probability of the target material clicked by the target user in the first preset time, wherein the target material vector is a vector corresponding to the target material in the material vector matrix constructed according to the material set.

[0040] In an implementation manner, the vector inner product operation of the interest preference vector and the target material vector is performed, and normalization processing is performed to determine the estimated click probability of the target material clicked by the target user in the first preset time, which can include: determining the estimated click probability of the target material clicked by the target user in the first preset time through the following formula: Wherein, P(v i |u) is the estimated click probability of the target material clicked by the target user in the first preset time, v i is the target material vector, u is the interest preference vector, V is the material vector matrix, v j is any material vector in the material vector matrix.

[0041] In this application, the feature vector corresponding to the user feature set is determined based on the user feature set, which can include: respectively encoding multiple user features in the user feature set into vectors; obtaining the feature vector corresponding to the user feature set by splicing multiple vectors. Exemplarily, the multiple user features in the user feature set can be embedded into low dimension respectively, and then concatenated to obtain the feature vector corresponding to the user feature set, and then the obtained feature vector is input into the target model.

[0042] In the embodiment of the present application, the user features in the user feature set can include at least one of the user's basic information, the user's interest portrait and the user's behavior characteristics. In this application, the user's basic information can include at least one of the user's gender, the user's age, the user's model of using electronic equipment and the user's region. The user's interest portrait can include the user's multi-level interest label, which can include first-level interest label, second-level interest label and third-level interest label. Exemplarily, the first-level interest label can be sports, the second-level interest label can be basketball, and the third-level interest label can be basketball player A, or the first-level interest label can be entertainment, the second-level interest label can be A area entertainment star, and the third-level interest label can be A area entertainment star B, etc. The user's behavior characteristics can include but are not limited to attention to bloggers, search words, etc.

[0043] The embodiment of the present application also provides a material recall method, Figure 2 is a flowchart of a material recall method disclosed by the embodiment of the present application. As Figure 2 shown, the method includes the following steps.

[0044] S210, respectively acquiring multiple user corresponding user feature sets.

[0045] Exemplarily, the user feature set can include the gender, age, model of the electronic device used, region, first-level interest label, second-level interest label, third-level interest label, followed blogger, search word, and the like of the user.

[0046] S220, determining a feature vector corresponding to the user feature set based on the user feature set.

[0047] In the present application, a plurality of user features in the user feature set can be encoded into vectors respectively, and a feature vector corresponding to the user feature set can be obtained by concatenating the plurality of vectors. Exemplarily, a plurality of user features in the user feature set can be embedded into low dimensions respectively, and concatenated to obtain a feature vector corresponding to the user feature set.

[0048] S230, inputting each of the plurality of feature vectors into a target model respectively to obtain a plurality of interest preference vectors corresponding to the user output by the target model.

[0049] It should be noted that the target model is trained according to the training method of the model described above.

[0050] S240, determining at least one user group by clustering the plurality of interest preference vectors.

[0051] It should be noted that if the interest preference vectors of two users are similar, the interest preferences of the two users are similar, and the users in each user group are users with similar interest preferences.

[0052] In the present application, a plurality of interest preference vectors are clustered, and then users with similar interest preference vectors are attributed to the same user group, and users with large differences in interest preference vectors are attributed to different user groups. Each user is attributed to and only attributed to one user group.

[0053] In an implementation manner, the clustering can be performed by a clustering algorithm Kmeans.

[0054] In addition, the number of user groups can be controlled to balance the heat of hot materials and the degree of personalization in the user group. In the case of a small number of user groups, the number of users in each user group is relatively large, and the heat of hot materials in the user group is higher and the degree of personalization is lower. In the case of a large number of user groups, the number of users in each user group is relatively small, and the heat of hot materials in the user group is lower and the degree of personalization is higher.

[0055] S250, respectively obtaining the number of clicks of a plurality of materials by users in a target user group within a second preset time, wherein the target user group is any user group in the at least one user group.

[0056] In the present application, the second preset time can be the current time, that is, the number of clicks of the users in the target user group on the multiple materials can be obtained by real-time statistics of the clicks of the user group on the pushed materials.

[0057] In addition, the second preset time can also be a time period, which is not limited in the present application.

[0058] S260, determining a hot material list corresponding to the target user group based on the number of clicks.

[0059] In the present application, the hot material list corresponding to the target user group can be determined by arranging the number of clicks of the users in the target user group on the multiple materials in descending order.

[0060] For example, the users in the target user group click on material A 300 times, material B 500 times, material C 600 times, material D 200 times, material E 20 times, and material F 5 times, and the hot material list corresponding to the target user group is as follows:

[0061] Material Clicks Material C 600 times Material B 500 times Material A 300 times Material D 200 times Material E 20 times Material F 5 times

[0062] S270, recalling materials for the users in the target user group based on the hot material list.

[0063] When pushing to a user in the target user group, the top N unread materials in the hot material list corresponding to the target user group can be taken as a recall material set in real time, and the materials are recalled for the user. It should be noted that N is an integer greater than 0.

[0064] The embodiment of the present application provides a material recall method, by respectively acquiring user corresponding user feature sets, determining feature vectors corresponding to the user feature sets based on the user feature sets, respectively inputting each feature vector in the multiple feature vectors into a target model, acquiring multiple interest preference vectors corresponding to the users output by the target model, and then clustering the multiple interest preference vectors to determine at least one user group, respectively acquiring the number of clicks of the users in the target user group on the multiple materials within a second preset time, determining a hot material list corresponding to the target user group based on the number of clicks of the users in the target user group on the multiple materials, and recalling materials for the users in the target user group based on the hot material list. The materials recalled for the user by the material recall method disclosed in the present application have a high degree of relevance to the user's interest preference, and by clustering multiple users into at least one user group and then recalling materials for the users in the user group based on the interest preference of each user group, the calculation amount of online services can be reduced.

[0065] The model training method provided in the embodiments of the present application can be executed by a model training device. The model training method executed by the model training device is taken as an example in the embodiments of the present application to describe the model training device provided in the embodiments of the present application.

[0066] Figure 3 FIG. 1 shows a structural schematic diagram of a model training device disclosed in the embodiments of the present application. As shown in FIG. 1, the model training device 300 includes an acquisition module 310, a first determination module 320, a training module 330, a second determination module 340, and an obtaining module 350. Figure 3

[0067] In the present application, the acquisition module 310 is configured to acquire a plurality of user feature sets corresponding to a target user within a first preset time, and a material set composed of materials clicked by a user within the first preset time, wherein the user feature set is a set of user features corresponding to the target user each time the target user clicks a material; the first determination module 320 is configured to determine a feature vector corresponding to the user feature set based on the user feature set; the training module 330 is configured to iteratively train a target model to be trained by using a plurality of the feature vectors, wherein each time the feature vector is input into the target model to be trained, an interest preference vector corresponding to the target user output by the target model to be trained is acquired; the second determination module 340 is configured to determine an estimated click probability of a target material clicked by the target user within the first preset time based on the interest preference vector and a material vector matrix constructed according to the material set; and the obtaining module 350 is configured to obtain the target model trained completely in a case where the estimated click probability reaches a target value.

[0068] In an implementation manner, the second determination module 340 determines the estimated click probability of the target material clicked by the target user within the first preset time based on the interest preference vector and the material vector matrix constructed according to the material set, including: performing vector inner product operation on the interest preference vector and a target material vector, and performing normalization processing to determine the estimated click probability of the target material clicked by the target user within the first preset time, wherein the target material vector is a vector corresponding to the target material in the material vector matrix constructed according to the material set.

[0069] In an implementation manner, the second determination module 340 performs vector inner product operation on the interest preference vector and a target material vector, and performs normalization processing to determine the estimated click probability of the target material clicked by the target user within the first preset time, including: determining the estimated click probability of the target material clicked by the target user within the first preset time by using the following formula: ​wherein, P(v i is a predicted click probability of the target material clicked by the target user in the first preset time, v i is the target material vector, u is the interest preference vector, V is the material vector matrix, v j is any material vector in the material vector matrix.

[0070] In an implementation manner, the first determination module 320 determines the feature vector corresponding to the user feature set based on the user feature set, including: respectively encoding a plurality of user features in the user feature set into vectors; and obtaining the feature vector corresponding to the user feature set by splicing a plurality of the vectors.

[0071] In an implementation manner, the user features in the user feature set at least include one of basic information of the user, an interest portrait of the user, and a behavior feature of the user.

[0072] The model training apparatus in the embodiments of the present application can be an electronic device, or a component in an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal, or other devices other than the terminal.

[0073] The model training apparatus provided in the embodiments of the present application can implement all processes of the model training method, and thus the details are not repeated here.

[0074] The execution subject of the material recall method provided in the embodiments of the present application can be a material recall apparatus. The embodiments of the present application take the material recall apparatus as an example to describe the material recall apparatus provided in the embodiments of the present application.

[0075] Figure 4 FIG. 4 is a structural schematic diagram of a material recall apparatus according to an embodiment of the present application. As shown in FIG. 4, the material recall apparatus 400 includes a first acquisition module 410, a first determination module 420, a second acquisition module 430, a second determination module 440, a third acquisition module 450, a third determination module 460, and a recall module 470. Figure 4

[0076] ​In the present application, the first obtaining module 410 is configured to obtain a user feature set corresponding to each of a plurality of users respectively; the first determining module 420 is configured to determine a feature vector corresponding to the user feature set based on the user feature set; the second obtaining module 430 is configured to input each of a plurality of feature vectors into a target model respectively, and obtain a plurality of interest preference vectors corresponding to the users output by the target model; the second determining module 440 is configured to determine at least one user group by clustering the plurality of interest preference vectors; the third obtaining module 450 is configured to obtain a number of clicks of a plurality of materials by a user in a target user group within a second preset time respectively, wherein the target user group is any one of the at least one user group; the third determining module 460 is configured to determine a hot material list corresponding to the target user group based on the number of clicks; and the recall module 470 is configured to recall materials for the user in the target user group based on the hot material list, wherein the target model is trained according to the training method of the model.

[0077] The recall device of the material in the embodiments of the present application can be an electronic device, or a component in the electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal, or other devices other than the terminal.

[0078] The recall device of the material provided in the embodiments of the present application can implement all the processes of the recall method of the material, and thus the details are not repeated here.

[0079] Optionally, as shown in Figure 5 The program or instructions stored on the storage 502 can be executed by the processor 501 to implement each step of the training method of the model or the recall method of the material, and achieve the same technical effects. Thus, the details are not repeated here.

[0080] It should be noted that the electronic device in the embodiments of the present application includes the mobile electronic device and the non-mobile electronic device.

[0081] The readable storage medium stores a program or instructions, which are executed by a processor to implement each process of the training method of the model or the recall method of the material, and achieve the same technical effects. Thus, the details are not repeated here.

[0082] The processor is the processor in the electronic device described in the above embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0083] The chip provided in the embodiments of the present application includes a processor and a communication interface, the communication interface is coupled with the processor, the processor is used to run programs or instructions to realize each process of the training method of the above model or the recall method of the material, and the same technical effects can be achieved. To avoid repetition, details are not described here.

[0084] It should be understood that the chip mentioned in the embodiments of the present application can also be referred to as a system-level chip, a system chip, a chip system or a system-on-chip chip, etc.

[0085] The embodiments of the present application provide a computer program product stored in a storage medium, which is executed by at least one processor to realize each process of the training method of the above model or the recall method of the material, and the same technical effects can be achieved. To avoid repetition, details are not described here.

[0086] It should be noted that in this paper, the term "includes", "contains" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, method, article or device. Without more limitations, the element defined by the statement "includes a" does not exclude the presence of another identical element in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiments of the present application is not limited to the order of functions shown or discussed, but can also include functions performed in a substantially simultaneous manner or in the opposite order, for example, the described method can be performed in an order different from that described, and various steps can also be added, omitted or combined. In addition, the features described with reference to some examples can be combined in other examples.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned example methods can be realized by means of software and a necessary general hardware platform, and of course, can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a computer software product in essence or in the form of a part that contributes to the prior art, which is stored in a storage medium (such as a ROM / RAM, a magnetic disk, or an optical disk) and includes a plurality of instructions for causing a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application.

[0088] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative and not restrictive. Those skilled in the art can make many forms under the inspiration of the present application without departing from the scope of the present application and the scope protected by the claims.

Claims

1. A method for training a model, characterized in that, include: Acquire multiple user feature sets corresponding to the target user within a first preset time period, and a material set consisting of materials clicked by the user within the first preset time period, wherein the user feature set is a set of user features corresponding to the target user each time the target user clicks on a material; Based on the user feature set, determine the feature vector corresponding to the user feature set; The target model to be trained is iteratively trained using multiple feature vectors, wherein in each iteration, the feature vectors are input into the target model to be trained, and the interest preference vector output by the target model to be trained, corresponding to the target user, is obtained. Based on the interest preference vector and the material vector matrix constructed according to the material set, the estimated click probability of the target material clicked by the target user within the first preset time period is determined. If the estimated click probability reaches the target value, the trained target model is obtained.

2. The training method according to claim 1, characterized in that, The step of determining the estimated click probability of the target material clicked by the target user within the first preset time period based on the interest preference vector and the material vector matrix constructed according to the material set includes: Perform a vector inner product operation on the interest preference vector and the target material vector, and perform normalization processing to determine the estimated click probability of the target material clicked by the target user within the first preset time period, wherein the target material vector is the vector corresponding to the target material in the material vector matrix constructed based on the material set.

3. The training method according to claim 2, characterized in that, The step of performing a vector dot product operation on the interest preference vector and the target material vector, and then performing normalization processing to determine the estimated click probability of the target material clicked by the target user within the first preset time period includes: The estimated click probability of the target material clicked by the target user within the first preset time period is determined by the following formula; Wherein, P(v i |u) represents the estimated click probability of the target material clicked by the target user within the first preset time period, v i Let u be the target material vector, v be the interest preference vector, and V be the material vector matrix. j Let be any material vector in the material vector matrix.

4. The training method according to claim 1, characterized in that, The step of determining the feature vector corresponding to the user feature set based on the user feature set includes: Encode each of the multiple user features in the user feature set into a vector; By concatenating multiple vectors, a feature vector corresponding to the user feature set is obtained.

5. The training method according to claim 1, characterized in that, The user features in the user feature set include at least one of the user's basic information, user interest profile, and user behavioral features.

6. A method for recalling materials, characterized in that, include: Obtain user feature sets for multiple users respectively; Based on the user feature set, determine the feature vector corresponding to the user feature set; Each of the multiple feature vectors is input into the target model to obtain multiple interest preference vectors output by the target model that correspond to the user. By clustering multiple interest preference vectors, at least one user group can be identified; The number of clicks on multiple materials by users in the target user group within a second preset time period is obtained respectively, wherein the target user group is any one of the at least one user group; Based on the number of clicks, a list of trending materials corresponding to the target user group is determined; Based on the list of trending materials, materials are recalled for users in the target user group; The target model is trained according to the model training method described in any one of claims 1 to 5.

7. A training device for a model, characterized in that, include: The acquisition module is used to acquire multiple user feature sets corresponding to the target user within a first preset time period, and a material set consisting of materials clicked by the user within the first preset time period, wherein the user feature set is a set of user features corresponding to the target user each time the target user clicks on a material; The first determining module is used to determine the feature vector corresponding to the user feature set based on the user feature set; The training module is used to iteratively train the target model to be trained using multiple feature vectors. In each iteration, the feature vectors are input into the target model to be trained, and the interest preference vector output by the target model to be trained is obtained, which corresponds to the target user. The second determining module is used to determine the estimated click probability of the target material clicked by the target user within the first preset time period based on the interest preference vector and the material vector matrix constructed according to the material set. The module is used to obtain the trained target model when the estimated click probability reaches the target value.

8. A material recall device, characterized in that, include: The first acquisition module is used to acquire user feature sets corresponding to multiple users respectively; The first determining module is used to determine the feature vector corresponding to the user feature set based on the user feature set; The second acquisition module is used to input each of the multiple feature vectors into the target model to acquire multiple interest preference vectors output by the target model corresponding to the user. The second determining module is used to determine at least one user group by clustering multiple interest preference vectors; The third acquisition module is used to acquire the number of times users in the target user group click on multiple materials within a second preset time period, wherein the target user group is any one of the at least one user group. The third determining module is used to determine a list of hot materials corresponding to the target user group based on the number of clicks. The recall module is used to recall materials for users in the target user group based on the hot material list; The target model is trained according to the model training method described in any one of claims 1 to 5.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a program or instructions that can run on the processor, the program or instructions being executed by the processor to implement the steps of the training method of the model as described in any one of claims 1-5, or the steps of the material recall method as described in claim 6.

10. A readable storage medium, characterized in that, The readable storage medium stores a program or instructions that, when executed by a processor, implement the steps of the training method for the model as described in any one of claims 1-5, or, when executed by the processor, implement the steps of the material recall method as described in claim 6.

Images