Information recommendation training method and system, storage medium and server

Abstract

The embodiment of the invention discloses an information recommendation training method and system, a storage medium and a server, and is applied to the technical field of information processing. In the process of training the information recommendation model, an original training sample is adopted to train initial information recommendation, then multimedia information associated with positive sample users is expanded through the initial information recommendation model, initial recommendation information is expanded for each positive sample user, the number of training samples used for training is increased, and the user experience is improved. The accuracy of the information recommendation model obtained by training can be ensured, and the number does not need to be obtained from other business application systems except the current business application system to serve as training samples, so that the training of the information recommendation model is not influenced by different business application systems; the problem that the information recommendation model is inaccurate due to different service application systems is avoided, and the accuracy of the information recommendation model obtained through training is improved.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a training method, system, storage medium and server for information recommendation. Background Art

[0002] When recommending information to user terminals, the information recommendation system can recommend relevant multimedia information to the user terminals based on the user information. In order to improve the accuracy and efficiency of information recommendation, the information recommendation system can generally use information recommendation models to recommend information to users. This method can be applied to a variety of business application systems, such as news systems, multimedia systems, etc.

[0003] Among them, the information recommendation model is a machine learning model that can be trained based on training samples. The larger the number of training samples, the more accurate the information recommended by the trained information recommendation model. However, in some business application systems, due to factors such as the small number of users, the number of training samples collected from the business application system is relatively small.

[0004] In the prior art, in order to improve the accuracy of the information recommendation model, data can generally be collected from other business application systems outside the current business application system as training samples. However, due to the differences between business application systems, the information recommendation model trained by using training samples obtained from other systems is not very effective in actual information recommendation. Summary of the invention

[0005] The embodiments of the present invention provide a training method, system, storage medium and server for information recommendation, which improve the accuracy of the trained information recommendation model without the need to obtain data from outside the current business application system.

[0006] An embodiment of the present invention provides, on one hand, a training method for information recommendation, comprising:

[0007] Acquire original training samples, where the original training samples include associated sample information corresponding to a plurality of positive sample users;

[0008] An initial information recommendation model is obtained by training according to the original training samples;

[0009] Determining the initial recommended information for the positive sample user according to the initial information recommendation model;

[0010] Fusion the associated sample information and the initial recommendation information of the positive sample user to obtain a fused original training sample;

[0011] The final information recommendation model is obtained by training the fused original training samples, and the information recommendation model is used to obtain corresponding recommended information according to user-related information.

[0012] Another aspect of the present invention provides a training system for information recommendation, including:

[0013] An original sample unit, used to obtain an original training sample, wherein the original training sample includes associated sample information corresponding to a plurality of positive sample users;

[0014] An initial training unit, used for training an initial information recommendation model according to the original training samples;

[0015] An initial recommendation unit, configured to determine the initial recommendation information of the positive sample user according to the initial information recommendation model;

[0016] A fusion unit, used to fuse the associated sample information and the initial recommendation information of the positive sample user to obtain a fused original training sample;

[0017] The final training unit is used to train the final information recommendation model according to the fused original training samples, and the information recommendation model is used to obtain corresponding recommendation information according to the user-related information.

[0018] Another aspect of the embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a plurality of computer programs, wherein the computer programs are suitable for being loaded by a processor and executing the information recommendation training method as described in one aspect of the embodiment of the present invention.

[0019] Another aspect of the embodiment of the present invention further provides a server, including a processor and a memory;

[0020] The memory is used to store multiple computer programs, and the computer programs are used to be loaded by the processor and execute the information recommendation training method as described in one aspect of an embodiment of the present invention; the processor is used to implement each computer program in the multiple computer programs.

[0021] It can be seen that in the training process of the information recommendation model of this embodiment, real data will be collected from the business application system as the original training sample, and the original training sample will be used for the initial training to obtain the initial information recommendation model. Then, the initial information recommendation model determines the initial recommendation information for the positive sample user in the original training sample, and fuses it with the associated sample information in the original training sample to obtain the fused original training sample, and then retrains according to the fused original training sample to obtain the final information recommendation model. In this process, the multimedia information associated with the positive sample user can be expanded by the initial information recommendation model, and the number of training samples used for training is increased, which can ensure the accuracy of the information recommendation model obtained by training, and there is no need to obtain the number from other business application systems outside the current business application system as training samples, so that the training of the information recommendation model will not be affected by different business application systems, avoiding the problem of inaccurate information recommendation model due to different business application systems, and improving the accuracy of the information recommendation model obtained by training. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0023] Figure 1 is a schematic diagram of a training method for information recommendation provided by an embodiment of the present invention;

[0024] Figure 2 is a flow chart of a training method for information recommendation provided by an embodiment of the present invention;

[0025] Figure 3 is a schematic diagram of an initial information recommendation model in one embodiment of the present invention;

[0026] Figure 4 is a flow chart of training an initial information recommendation model in one embodiment of the present invention;

[0027] Figure 5a is a flow chart of a training method for information recommendation provided in an application embodiment of the present invention;

[0028] Figure 5b is a schematic diagram of a training method for information recommendation provided in an application embodiment of the present invention;

[0029] Figure 6 It is a logical structure diagram of an information recommendation training system provided by an embodiment of the present invention;

[0030] Figure 7 It is a schematic diagram of the logical structure of a server provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can, for example, be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0033] The embodiment of the present invention provides a training method for information recommendation, which is mainly used to train an information recommendation model applied to any business application system (especially a business application system based on a small amount of user data), and in the training process, the data collected in the business application system is used as a training sample, such as Figure 1 As shown, the information recommendation model can be trained by the information recommendation training system according to the following steps:

[0034] Obtain original training samples, wherein the original training samples include associated sample information corresponding to multiple positive sample users; train an initial information recommendation model based on the original training samples; determine the initial recommendation information of the positive sample users based on the initial information recommendation model; fuse the associated sample information of the positive sample users with the initial recommendation information to obtain a fused original training sample; train a final information recommendation model based on the fused original training samples, wherein the information recommendation model is used to obtain corresponding recommendation information based on user-related information.

[0035] When training is performed based on the fused original training samples, training may be continued based on the above initial information recommendation model, or a new information recommendation model may be determined and the final information recommendation model may be obtained by training the new information recommendation model.

[0036] The above information recommendation model is a machine learning model based on artificial intelligence. Here, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines so that machines have the functions of perception, reasoning and decision-making.

[0037] Artificial intelligence technology is a comprehensive discipline that covers a wide range of fields, including both hardware-level and software-level technologies. The basic technologies of artificial intelligence generally include sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics and other technologies. Artificial intelligence software technology mainly includes computer vision technology, speech processing technology, natural language processing technology, machine learning and / deep learning.

[0038] Machine Learning (ML) is a multi-disciplinary subject that involves probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications are spread across all areas of artificial intelligence. Machine learning and deep learning usually include artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and teaching learning.

[0039] It can be seen that in the process of training the information recommendation model, the multimedia information associated with the positive sample user can be expanded through the initial information recommendation model, which increases the number of training samples used in training, and can ensure the accuracy of the trained information recommendation model. There is no need to obtain a number of training samples from other business application systems outside the current business application system, so that the training of the information recommendation model will not be affected by different business application systems, avoiding the problem of inaccurate information recommendation models due to different business application systems, and improving the accuracy of the trained information recommendation model.

[0040] In one embodiment of the present invention, a training method for information recommendation is provided, which is mainly a method executed by a training system for information recommendation. The flow chart is as follows: Figure 2 As shown, including:

[0041] Step 101: obtaining original training samples, wherein the original training samples include associated sample information corresponding to a plurality of positive sample users.

[0042] It can be understood that if it is necessary to train an information recommendation model in a business application system (especially a business application system with less user-based data), data can be collected from the business application system as original training samples, which can include the attribute information of positive sample users and related sample information, and can also include the attribute information of negative sample users and non-related sample information, which are all real data generated by the business application system during actual operation. Among them:

[0043] The attribute information of each sample user includes the attributes of each sample user, such as the user's age, gender, location, interests and hobbies; the associated sample information of the positive sample user refers to the multimedia information in which the positive sample user performs certain operations through the user terminal, such as multimedia information in which the user views details, likes, comments, and forwards. The non-associated sample information of the negative sample user refers to the multimedia information in which the negative sample user does not perform any operations through the user terminal, such as multimedia information in which the user does not perform any operations, that is, sample information that is not associated with the negative sample user. The multimedia information refers to the multimedia information provided in the above-mentioned business application system, such as news, self-media content, etc.

[0044] Specifically, when obtaining the associated sample information of the positive sample user, multiple multimedia information in which the positive sample user performs any operation within a period of time before the current moment can be obtained from the target business application system as the associated sample information.

[0045] Furthermore, if the trained information recommendation model is to increase the frequency of users in the target business application system performing a certain type of operation on multimedia information, when obtaining the original training samples, specifically, multiple multimedia information in which positive sample users perform specific types of operations (such as clicking, viewing, commenting or forwarding) are obtained from the target business application as the associated sample information corresponding to the positive sample users.

[0046] Step 102: An initial information recommendation model is obtained by training according to the original training samples.

[0047] Specifically, training the initial information recommendation model is a process of continuously adjusting the parameter values ​​of the parameters (such as weight values, etc.) in the initial information recommendation model according to the original training samples. In this process, a training target based on the parameter value will be set (generally a calculated loss function). The adjustment of the parameter value of the parameter needs to ultimately reach the training target based on the parameter value. After the final adjusted parameter value is obtained, the initial information recommendation model is trained.

[0048] Among them, the trained initial information recommendation model is specifically used to obtain the first feature information of any multimedia information, and obtain the second feature information related to the user (including the user's own attributes and the multimedia information of the user's operation, etc.), and calculate the similarity between the first feature information and the second feature information. If the calculated similarity is greater than the preset similarity, the corresponding multimedia information will be used as the user's information to be recommended.

[0049] The initial information recommendation model can be a similar recommendation (i2i) recall model of any structure, such as a deep structured semantic model (DSSM), etc. The specific structure of the initial information recommendation model is not limited here. Figure 3 As shown, the initial information recommendation model consists of two towers. The user tower (usertower) is used to obtain the second feature information related to the user, and the information tower (item tower) is used to obtain the first feature information of the multimedia information provided by the business application system, and then calculate the similarity between the first feature information and the second feature information. Then, based on the calculated similarity, multimedia information with a similarity greater than a preset similarity is recommended to the user.

[0050] Step 103: determine the initial recommended information for the positive sample users according to the initial information recommendation model.

[0051] The associated sample information of the positive sample users obtained from the original training samples in the above step 101 is directly obtained from the target business application system.

[0052] When determining the initial recommendation information in step 103, in one case, the initial information recommendation model is used to directly obtain the initial information recommendation, specifically:

[0053] The relevant information of the positive sample user is input into the initial information recommendation model trained in the above step 102, and the initial information recommendation model can output the initial recommendation information recommended for the positive sample user. The relevant information of the positive sample user may include the attribute information of the positive sample user, and may also include the multimedia information that the positive sample user has operated historically, such as the associated sample information of the positive sample user.

[0054] In a specific application, in order to determine the initial recommendation information, the initial information recommendation model may first determine multiple recommendation information based on the relevant information of the positive sample user, and then select at least one recommendation information from the multiple recommendation information as the initial recommendation information. When selecting at least one recommendation information, specifically: determine the first similarity between the relevant information of the positive sample user and the associated sample information, and determine the second similarity between the relevant information of the positive sample user and each recommendation information, and then select at least one recommendation information whose second similarity is greater than the first similarity from the multiple recommendation information.

[0055] In other embodiments, the number of recommended information selected for a positive sample user may be limited to not exceed a preset value, for example, the number of recommended information selected for a positive sample user is limited to 5 or less.

[0056] In this step 103, when determining the initial recommendation information, in another case, the initial recommendation information is not completely obtained by the initial information recommendation model. The initial information recommendation model is only used to obtain the first feature information of the multimedia information. Specifically:

[0057] The above-mentioned initial information recommendation model is used to obtain the first feature information of each multimedia information provided in the target business application system, and then through the search function in the information recommendation training system, such as the k-NearestNeighbor (KNN) algorithm, the second feature information related to the positive sample user is used as the search key information, combined with the first feature information of each multimedia information obtained by the initial information recommendation model, to search for multiple multimedia information corresponding to the first feature information that is most similar to the second feature information, and the multiple multimedia information obtained by the search is used as the initial recommendation information.

[0058] Step 104: fuse the associated sample information of the positive sample user with the initial recommendation information to obtain a fused original training sample.

[0059] Specifically, the initial recommendation information of the positive sample user is directly added to the associated sample information of the corresponding positive sample user to obtain the fused original training sample.

[0060] Step 105, a final information recommendation model is obtained by training according to the fused original training samples, and the information recommendation model is used to obtain corresponding recommended information according to the relevant information of the user.

[0061] Specifically, in one case, the initial information recommendation model can be directly trained according to the fused original training samples, and the parameter values ​​in the initial information recommendation model can be fine-tuned to obtain the final information recommendation model.

[0062] Alternatively, in another case, a final information recommendation model can be retrained based on the original training samples of the fused original training samples, that is, a new information recommendation model is determined, and the parameter values ​​in the new information recommendation model are adjusted to obtain the final information recommendation model. Here, the structure of the new information recommendation model can be the same as that of the initial information recommendation model, except that the parameter values ​​of the parameters are different.

[0063] The method for training the final information recommendation model here is similar to the method for training the initial information recommendation model in step 102, except that the number of training samples used in training the final information recommendation model will increase. Specifically, in order to ensure that the process of training the information recommendation model is mainly trained on real training samples to improve the accuracy of the trained information recommendation model, in the process of executing step 105, weight values ​​can be set for the initial recommendation information and the associated sample information of the positive sample user, and the weight value of the initial recommendation information can be set to be less than the preset weight value, specifically, Figure 4 As shown, in the case of continuing to train based on the initial information recommendation model, this step 105 can be performed through the following steps: a business model canvas drawing method,

[0064] Step 201 , obtaining a first similarity between a positive sample user and associated sample information in a fused original training sample according to an initial information recommendation model, and obtaining a second similarity between a positive sample user and initial recommendation information.

[0065] It should be noted that if the original training samples also include non-associated sample information corresponding to negative sample users, since the non-associated sample information is multimedia information on which the negative sample users have not performed any operations, it is not necessary to perform the fusion in the above steps 103 to 104. In the process of training the initial information recommendation model, for the negative sample users, the non-associated sample information obtained in the above step 101 can continue to be used. In this way, the third similarity between the negative sample users and the non-associated sample information can also be obtained according to the initial information recommendation model.

[0066] Step 202: Calculate a loss function related to the initial information recommendation model according to the first similarity and the second similarity.

[0067] Specifically, the information recommendation training system will first calculate the loss function related to the initial information recommendation model based on the first similarity and the second similarity of each positive sample user obtained in the above step 201. The loss function is used to indicate the error between the first similarity and the second similarity of each positive sample user obtained by the initial information recommendation model and the actual situation of each positive sample user, such as the cross entropy loss function. Among them, the first similarity and the second similarity of any positive sample user need to be greater than the preset similarity to meet the actual situation of the positive sample user, that is, the positive sample user is associated with the associated sample information and the initial recommended information.

[0068] Furthermore, if the third similarity between each negative sample user and the non-associated sample information is obtained in the above step 201, then when calculating the above loss function, it is necessary to calculate the loss function related to the initial information recommendation model based on the first similarity and the second similarity of each positive sample user obtained in the above step 201, and the third similarity of each negative sample user. This loss function is not only used to indicate the error between the first similarity and the second similarity of each positive sample user obtained by the initial information recommendation model and the actual situation of each positive sample user, but also used to indicate the error between the third similarity of each negative sample user obtained by the initial information recommendation model and the actual situation of each negative sample user. Among them, the third similarity of any negative sample user needs to be no greater than the preset similarity to meet the actual situation of the negative sample user, that is, the negative sample user is not associated with the non-associated sample information.

[0069] In this embodiment, in order to ensure that the process of training the information recommendation model is mainly trained on real training samples to improve the accuracy of the trained information recommendation model, during the execution of step 202:

[0070] The first weight value and the second weight value are respectively set for the first similarity and the second similarity corresponding to the positive sample user, and the first weight value and the second weight value are respectively added to the first similarity and the second similarity to obtain the weighted first similarity and the weighted second similarity, and the loss function is calculated according to the weighted first similarity and the weighted second similarity. Among them, the second weight value needs to be less than a preset weight value, and the first weight value needs to be greater than another preset weight value, and the other preset weight value needs to be greater than or equal to a preset weight value.

[0071] In order to further improve the accuracy of the information recommendation model, since each associated sample information corresponding to the positive sample user is multimedia information on which the positive sample user performs certain operations, these operations may include click, view, forward, comment and other operations. In order to increase the frequency of users' operations based on a certain type of operation on the multimedia information provided in the target business application system, when setting the first weight value of the first similarity corresponding to each positive sample user, different first weight values ​​may be set for the first similarities obtained from the associated sample information of different types of operations. For example, a higher first weight value may be set for the first similarity obtained from the associated sample information of deep operations such as forwarding or commenting, and a lower first weight may be set for the first similarity obtained from the associated sample information of shallow operations such as clicking or viewing.

[0072] Step 203: adjust the initial information recommendation model according to the loss function to obtain a final information recommendation model.

[0073] Specifically, the training process of the information recommendation model needs to minimize the above error, that is, the value of the loss function. The training process continuously optimizes the parameter values ​​of the parameters in the above initial information recommendation model through a series of mathematical optimization methods such as back propagation derivation and gradient descent, and minimizes the calculated value of the above loss function.

[0074] It should be noted that the above steps 201 to 203 are a one-time adjustment of the parameter values ​​in the initial information recommendation model, and in actual applications, it is necessary to continuously loop through the above steps 201 to 203 until the adjustment of the parameter values ​​meets certain stop conditions.

[0075] Therefore, after executing steps 201 to 203 of the above embodiment, the information recommendation model still needs to determine whether the current adjustment of the parameter value meets the preset stop condition. If it does, the process ends and the parameter value obtained after adjustment in step 203 is used as the parameter value of the final information recommendation model; if it does not meet the condition, the initial information recommendation model after the parameter value is adjusted returns to execute the above steps 201 to 203, that is, a new batch of training samples is used to continue adjusting the parameter value in the initial information recommendation model. The preset stop condition includes but is not limited to any one of the following conditions: the difference between the current adjusted parameter value and the last adjusted parameter value is less than a threshold value, that is, the adjusted parameter value reaches convergence; and the number of parameter value adjustments is equal to the preset number of times, etc.

[0076] It should be noted that in the training process of the information recommendation model in this embodiment, when the amount of user-based data in the target business application system is currently small and insufficient to train the information recommendation model, the above-mentioned methods from steps 101 to 105 can be used to train the information recommendation model, and the training samples used to train the information recommendation model can be expanded in the target business application system. Therefore, before the method of the specific application embodiment is used, it can also be determined whether the total number of users in the target business application system is less than the preset value, or whether the online time of the target business application system is less than the preset time. If the total number of users is less than the preset value or the online time is less than the preset time, it means that the amount of user-based data in the target business application system is small, and the above-mentioned methods from steps 101 to 105 need to be used to train the information recommendation model.

[0077] It can be seen that in the training process of the information recommendation model of this embodiment, real data will first be collected from the business application system as the original training sample, and the original training sample will be used for the initial training to obtain the initial information recommendation model, and then the initial information recommendation model will determine the initial recommendation information for the positive sample user in the original training sample, and fuse it with the associated sample information in the original training sample to obtain the fused original training sample, and then train again according to the fused original training sample to obtain the final information recommendation model. In this process, the multimedia information associated with the positive sample user can be expanded by the initial information recommendation model, and the number of training samples used for training is increased, which can ensure the accuracy of the information recommendation model obtained by training, and there is no need to obtain the number from other business application systems outside the current business application system as training samples, so that the training of the information recommendation model will not be affected by different business application systems, avoiding the problem of inaccurate information recommendation model due to different business application systems, and improving the accuracy of the information recommendation model obtained by training.

[0078] The following is a specific application example to illustrate the information recommendation training method of the present invention. In this embodiment, the target business application system mainly targeted is a news system on a smart device, such as a news system such as Kankan. The amount of user-based data in the news system is relatively small, such as Figure 5a and Figure 5b As shown, the following method is needed to train the information recommendation model in the news system:

[0079] Step 301, collects the self-attribute information and related sample information of each user as a positive sample user from the news system, and obtains the self-attribute information and non-related sample information of each user as a negative sample user, so that the original training sample includes: the self-attribute information and related sample information of the positive sample users, and the self-attribute information and non-related sample information of the negative sample users.

[0080] Step 302: an initial information recommendation model is obtained by training according to the original training samples. In this embodiment, the initial information recommendation model may specifically be a DSSM.

[0081] Step 303, using the above-mentioned initial information recommendation model to obtain the first feature information of each multimedia information provided in the news system, and then through the search function in the information recommendation training system, such as KNN, using the second feature information related to the positive sample user as the search key information, combined with the first feature information of each multimedia information obtained by the initial information recommendation model, search for multiple multimedia information corresponding to the first feature information that is most similar to the second feature information, and use the multiple multimedia information obtained by the search as the initial recommendation information for the positive sample user.

[0082] Step 304, the associated sample information and the initial recommendation information of the positive sample user are merged to obtain a merged original training sample, which includes the initial recommendation information and associated sample information corresponding to the positive sample user, and the non-associated sample information corresponding to the negative sample user.

[0083] Step 305, obtaining the first similarity between the positive sample user and the associated sample information in the fused original training sample according to the initial information recommendation model, obtaining the second similarity between the positive sample user and the initial recommendation information, and obtaining the third similarity between the negative sample user and the non-associated sample information.

[0084] Step 306, calculating a loss function related to the initial information recommendation model according to the first similarity, the second similarity and the third similarity, and adjusting parameter values ​​in the initial information recommendation model according to the loss function to obtain a final trained information recommendation model.

[0085] In which, when calculating the loss function, corresponding weight values ​​can be set for each similarity, wherein the weight value set for the above-mentioned second similarity needs to be less than a preset weight value, and the weight value set for the first similarity is greater than another preset weight value. This can ensure that the process of training the information recommendation model is mainly trained on real training samples to improve the accuracy of the trained information recommendation model.

[0086] It should be noted that the above steps 305 and 306 are the process of retraining the information recommendation model based on the initial information recommendation model. In other embodiments, a new information recommendation model can be determined, and the new information recommendation model can be trained using the fused original training samples to obtain the final information recommendation model.

[0087] In the specific practice, after respectively using the methods in the above steps 301 to 306 and using the existing methods to train the information recommendation model 1 and the information recommendation model 2, and running the two information recommendation models in the news system, after the information recommendation model 1 trained by the method of this embodiment recommends information to the user, the user's video fast-swiping rate (the proportion of videos with a playback time of less than 5 seconds) is reduced by 5.345% compared with the video fast-swiping rate corresponding to the information recommendation model 2, and the video completion rate is increased by 3.043% compared with the video completion rate corresponding to the information recommendation model 2, and the per capita video playback volume is increased by 1.761% compared with the video playback volume corresponding to the information recommendation model 2. It can be seen that the accuracy rate of the information recommendation model trained by the training method of this embodiment is greatly improved.

[0088] It can be seen that in the training method of this embodiment, there is no need to use the numbers in other business application systems other than the news system as training samples. Instead, based on the associated sample information corresponding to the current positive sample user, other data with higher confidence, namely the initial recommendation information, are mined to perform data enhancement on the original training samples, thereby improving the generalization ability of the information recommendation model, expanding the field of view of the information recommendation model, solving the problem of insufficient data volume, and improving the accuracy of the information recommendation model.

[0089] The embodiment of the present invention also provides an information recommendation training system, the structural diagram of which is shown in FIG. Figure 6 As shown, it may specifically include:

[0090] The original sample unit 10 is used to obtain original training samples, wherein the original training samples include associated sample information corresponding to a plurality of positive sample users.

[0091] The original sample unit 10 is specifically used to obtain multiple multimedia information of the positive sample user performing a specific type of operation from the target business application system as the associated sample information corresponding to the positive sample user.

[0092] The initial training unit 11 is used to train an initial information recommendation model according to the original training samples obtained by the original sample unit 10.

[0093] The initial recommendation unit 12 is used to determine the initial recommendation information of the positive sample user according to the initial information recommendation model obtained by the initial training unit 11.

[0094] The initial recommendation unit 12 is specifically used for the initial information recommendation model to determine multiple recommendation information according to the relevant information of the positive sample user; and select at least one recommendation information from the multiple recommendation information as the initial recommendation information. Wherein, when selecting at least one recommendation information from the multiple recommendation information as the initial recommendation information, the initial recommendation unit 12 is specifically used to determine the first similarity between the relevant information of the positive sample user and the associated sample information, and determine the second similarity between the relevant information of the positive sample user and the recommendation information; and select at least one recommendation information from the multiple recommendation information whose second similarity is greater than the first similarity.

[0095] The initial recommendation unit 12 is specifically configured to execute the step of determining the initial recommendation information only when it is determined that the total number of users in the target business application system is less than a preset value, or the online duration of the target business application system is less than a preset duration.

[0096] The fusion unit 13 is used to fuse the associated sample information of the positive sample user and the initial recommendation information determined by the initial recommendation unit 12 to obtain a fused original training sample.

[0097] The final training unit 14 is used to train the final information recommendation model according to the fused original training samples obtained by the fusion unit 13, and the information recommendation model is used to obtain corresponding recommendation information according to the user-related information.

[0098] The final training unit 14 is specifically used to obtain, according to the initial information recommendation model, a first similarity between a positive sample user and associated sample information in the fused original training sample, and to obtain a second similarity between the positive sample user and the initial recommendation information; calculate a loss function related to the initial information recommendation model according to the first similarity and the second similarity; and adjust the initial information recommendation model according to the loss function to obtain a final information recommendation model.

[0099] Among them, when calculating the loss function related to the initial information recommendation model based on the first similarity and the second similarity, the final training unit 14 is specifically used to set corresponding first weight values ​​and second weight values ​​for the first similarity and the second similarity corresponding to the positive sample user; the second weight value is less than the preset weight value; the first weight value and the second weight value are added to the first similarity and the second similarity respectively to obtain a weighted first similarity and a weighted second similarity; and the loss function is calculated based on the weighted first similarity and the weighted second similarity.

[0100] It can be seen that in the process of training the information recommendation model, the system of this embodiment can expand the multimedia information associated with the positive sample user through the initial information recommendation model, thereby increasing the number of training samples used for training, and ensuring the accuracy of the trained information recommendation model. There is no need to obtain a number of training samples from other business application systems outside the current business application system, so that the training of the information recommendation model will not be affected by different business application systems, avoiding the problem of inaccurate information recommendation models due to differences in business application systems, and improving the accuracy of the trained information recommendation model.

[0101] The embodiment of the present invention further provides a server, the structural diagram of which is shown in Figure 7 As shown, the server may have relatively large differences due to different configurations or performances, and may include one or more central processing units (CPU) 20 (for example, one or more processors) and memory 21, and one or more storage media 22 (for example, one or more mass storage devices) storing application programs 221 or data 222. Among them, the memory 21 and the storage medium 22 can be short-term storage or permanent storage. The program stored in the storage medium 22 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations in the server. Furthermore, the central processing unit 20 can be configured to communicate with the storage medium 22 and execute a series of instruction operations in the storage medium 22 on the server.

[0102] Specifically, the application 221 stored in the storage medium 22 includes an application for information recommendation training, and the program may include the original sample unit 10, the initial training unit 11, the initial recommendation unit 12, the fusion unit 13 and the final training unit 14 in the above-mentioned information recommendation training system, which will not be described in detail here. Furthermore, the central processor 20 can be configured to communicate with the storage medium 22 and execute a series of operations corresponding to the application for information recommendation training stored in the storage medium 22 on the server.

[0103] The server may also include one or more power supplies 23, one or more wired or wireless network interfaces 24, one or more input and output interfaces 25, and / or one or more operating systems 223, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, FreeBSD TM, etc.

[0104] The steps performed by the information recommendation training system described in the above method embodiment can be based on the Figure 7 The structure of the server is shown.

[0105] Furthermore, another aspect of the embodiment of the present invention also provides a computer-readable storage medium, which stores multiple computer programs, and the computer programs are suitable for being loaded by a processor and executing the information recommendation training method performed by the above-mentioned information recommendation training system.

[0106] Another aspect of the embodiment of the present invention further provides a server, including a processor and a memory;

[0107] The memory is used to store multiple computer programs, and the computer programs are used to be loaded by the processor and execute the information recommendation training method performed by the above-mentioned information recommendation training system; the processor is used to implement each computer program among the multiple computer programs.

[0108] In addition, according to one aspect of the present application, a computer program product or a computer program is provided, the computer program product or the computer program includes computer instructions, the computer instructions are stored in a computer-readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device executes the information recommendation training method provided in the above various optional implementations.

[0109] A person skilled in the art may understand that all or part of the steps in the various methods of the above-mentioned embodiments may be completed by instructing the relevant hardware through a program, and the program may be stored in a computer-readable storage medium, and the storage medium may include: a read-only memory (ROM), a random access memory (RAM), a disk or an optical disk, etc.

[0110] The above is a detailed introduction to the information recommendation training method, system, storage medium and server provided in the embodiments of the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A training method for information recommendation, characterized in that: include: Acquire original training samples, where the original training samples include associated sample information corresponding to a plurality of positive sample users respectively; An initial information recommendation model is obtained by training according to the original training samples; Determining the initial recommended information for the positive sample user according to the initial information recommendation model; Fusion the associated sample information and the initial recommendation information of the positive sample user to obtain a fused original training sample; The final information recommendation model is obtained by training the fused original training samples, and the information recommendation model is used to obtain corresponding recommended information according to user-related information.

2. The method according to claim 1, characterized in that The obtaining of the original training samples specifically includes: A plurality of multimedia information pieces in which the positive sample user performs a specific type of operation are obtained from the target business application system as associated sample information corresponding to the positive sample user.

3. The method according to claim 1, characterized in that The determining the initial recommended information of the positive sample user according to the initial information recommendation model specifically includes: The initial information recommendation model determines a plurality of recommended information according to the relevant information of the positive sample user; At least one piece of recommendation information is selected from the plurality of recommendation information as the initial recommendation information.

4. The method according to claim 3, characterized in that The selecting at least one piece of recommendation information from the plurality of recommendation information as the initial recommendation information specifically includes: Determine a first similarity between the relevant information of the positive sample user and the associated sample information, and determine a second similarity between the relevant information of the positive sample user and the recommended information; At least one piece of recommendation information whose second similarity is greater than the first similarity is selected from the plurality of recommendation information.

5. The method according to claim 1, characterized in that The step of training the final information recommendation model according to the fused original training samples specifically includes: Acquire, according to the initial information recommendation model, a first similarity between a positive sample user and associated sample information in the fused original training sample, and acquire a second similarity between the positive sample user and the initial recommendation information; Calculating a loss function associated with the initial information recommendation model according to the first similarity and the second similarity; The initial information recommendation model is adjusted according to the loss function to obtain a final information recommendation model.

6. The method according to claim 5, characterized in that The calculating a loss function related to the initial information recommendation model according to the first similarity and the second similarity specifically includes: Setting a corresponding first weight value and a second weight value for the first similarity and the second similarity corresponding to the positive sample user, respectively; the second weight value is less than the preset weight value; adding the first weight value and the second weight value to the first similarity and the second similarity respectively to obtain a weighted first similarity and a weighted second similarity; The loss function is calculated according to the weighted first similarity and the weighted second similarity.

7. The method according to any one of claims 1 to 6, characterized in that: Before determining the initial recommendation information of the positive sample user according to the initial information recommendation model, the method further includes: When it is determined that the total number of users in the target business application system is less than a preset value, or the online duration of the target business application system is less than a preset duration, the step of determining the initial recommendation information is performed.

8. A training system for information recommendation, characterized in that: include: An original sample unit, used to obtain an original training sample, wherein the original training sample includes associated sample information corresponding to a plurality of positive sample users; An initial training unit, used for training an initial information recommendation model according to the original training samples; An initial recommendation unit, configured to determine the initial recommendation information of the positive sample user according to the initial information recommendation model; A fusion unit, used to fuse the associated sample information and the initial recommendation information of the positive sample user to obtain a fused original training sample; The final training unit is used to train the final information recommendation model according to the fused original training samples, and the information recommendation model is used to obtain corresponding recommendation information according to the user-related information.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of computer programs, and the computer programs are suitable for being loaded by a processor and executing the information recommendation training method according to any one of claims 1 to 7.

10. A server, characterized in that: including a processor and a memory; The memory is used to store multiple computer programs, and the computer programs are used to be loaded by the processor and execute the information recommendation training method as described in any one of claims 1 to 7; the processor is used to implement each computer program in the multiple computer programs.

Images