Information recommendation method, object recognition method and device

Abstract

The present disclosure relates to an information recommendation method, an object recognition method and a device, the information recommendation method comprising: obtaining object attribute information of a target object; inputting the object attribute information into an object recognition model for label recognition to obtain an identification label of the target object; the object recognition model is trained based on target samples corresponding to multiple sample labels; the target sample is obtained by sampling the original sample in a target sampling manner based on sample balance data of the original sample corresponding to the multiple sample labels; the sample balance data is used to characterize the statistical feature information of the number of samples corresponding to each of the multiple sample labels; and recommending the information to be recommended corresponding to the identification label to the target object. The present disclosure can improve the generalization ability of the object recognition model and the accuracy of label recognition, thereby improving the accuracy of information recommendation.

Description

Technical Field

[0001] The present disclosure relates to the technical field of machine learning, and in particular to an information recommendation method, an object recognition method and a device. Background Art

[0002] As machine learning technology is increasingly used in the e-commerce field, using models to accurately identify users' first-time purchase preferences on the platform is also one of the more important business scenarios. However, due to the large number of preferred categories, sample imbalance is a common problem in supervised learning, which also poses a great challenge to model training.

[0003] In the related art, the model is generally trained directly based on the acquired multi-label samples. The recognition model trained by this method has the problem of low recognition accuracy. Summary of the invention

[0004] The present disclosure provides an information recommendation method, an object recognition method and a device to at least solve the problem of low recognition accuracy of the recognition model in the related art. The technical solution of the present disclosure is as follows:

[0005] According to a first aspect of an embodiment of the present disclosure, there is provided an information recommendation method, comprising:

[0006] Get the object attribute information of the target object;

[0007] The object attribute information is input into the object recognition model for label recognition to obtain the recognition label of the target object; the object recognition model is trained based on the target samples corresponding to the multiple sample labels; the target samples are obtained by sampling the original samples in a target sampling manner based on the sample balance data of the original samples corresponding to the multiple sample labels; the sample balance data is used to characterize the statistical feature information of the number of samples corresponding to each of the multiple sample labels; the target sampling manner is obtained based on the total number of samples of the original samples and the sample balance data;

[0008] Recommending the target object with the to-be-recommended information corresponding to the identification tag.

[0009] In an exemplary embodiment, the step of training the object recognition model includes:

[0010] Acquire original samples corresponding to the target task; the original samples include original samples corresponding to the multiple sample labels;

[0011] Determining the sample balance data based on the original samples corresponding to the multiple sample labels;

[0012] Sampling the original sample in the target sampling manner to obtain a target sample corresponding to the target task;

[0013] The object recognition model is obtained based on the target sample training.

[0014] According to a second aspect of an embodiment of the present disclosure, there is provided an object recognition method, including:

[0015] Acquire original samples corresponding to the target task; the original samples include original samples corresponding to multiple sample labels;

[0016] Determine sample balance data based on the original samples corresponding to the multiple sample labels; the sample balance data is used to characterize statistical feature information of the number of samples corresponding to each of the multiple sample labels;

[0017] Sampling the original samples in a target sampling manner to obtain target samples corresponding to the target task; the target sampling manner is obtained based on the total number of samples of the original samples and the sample balance data;

[0018] The target object is identified based on an object recognition model corresponding to the target task to obtain an identification label of the target object; the object recognition model is trained based on the target sample.

[0019] In an exemplary embodiment, sampling the original sample in a target sampling manner to obtain a target sample corresponding to the target task includes:

[0020] Based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels, a first target sample label is determined from the multiple sample labels; the number of samples corresponding to the first target sample label is greater than the sample balance data;

[0021] Performing sample deletion processing on the original sample corresponding to the first target sample label to obtain a first sample;

[0022] Based on the first sample, the target sample is obtained.

[0023] In an exemplary embodiment, the sample balance data includes original sample means corresponding to the plurality of sample labels;

[0024] The determining a first target sample label from the multiple sample labels based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels comprises:

[0025] Determine the difference between the original sample quantities corresponding to the multiple sample labels and the original sample means based on the original sample quantities corresponding to the multiple sample labels;

[0026] Among the multiple sample labels, the sample label whose original sample quantity difference is greater than or equal to a first preset threshold is determined as the first target sample label.

[0027] In an exemplary embodiment, performing sample deletion processing on the original sample corresponding to the first target sample label to obtain the first sample includes:

[0028] Performing sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes;

[0029] Performing deletion processing on samples under each sample class to obtain processed samples corresponding to the sample class;

[0030] The first sample is obtained based on processed samples corresponding to the multiple sample classes.

[0031] In an exemplary embodiment, the pruning of samples under each sample class to obtain processed samples corresponding to the sample class includes:

[0032] Determine information about the ratio of the number of samples under the sample class to the number of original samples corresponding to the first target sample label;

[0033] Determine difference information between the sample balance data and the sample quantity of the original sample corresponding to the first target sample label;

[0034] Determining a first quantity corresponding to the sample class based on the ratio information and the difference information;

[0035] The first number of samples are deleted from the samples under the sample class to obtain processed samples corresponding to the sample class.

[0036] In an exemplary embodiment, sampling the original sample in a target sampling manner to obtain a target sample corresponding to the target task includes:

[0037] Based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels, a second target sample label is determined from the multiple sample labels; the number of samples corresponding to the second target sample label is smaller than the sample balance data;

[0038] Performing sample augmentation processing on the original sample corresponding to the second target sample label to obtain a second sample;

[0039] Based on the second sample, the target sample is obtained.

[0040] In an exemplary embodiment, the sample balance data includes original sample means corresponding to the plurality of sample labels;

[0041] The determining a second target sample label from the multiple sample labels based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels comprises:

[0042] Determine the difference between the original sample quantities corresponding to the multiple sample labels and the original sample means based on the original sample quantities corresponding to the multiple sample labels;

[0043] Among the multiple sample labels, the sample label whose original sample quantity difference is greater than or equal to a second preset threshold is determined as the second target sample label.

[0044] In an exemplary embodiment, performing sample augmentation processing on the original sample corresponding to the second target sample label to obtain the second sample includes:

[0045] Determine a second quantity corresponding to the second target sample label based on the original sample quantity corresponding to the second target sample label and the sample balance data;

[0046] Based on the original samples corresponding to the second target sample labels, generating the second number of newly added samples;

[0047] The second number of newly added samples is added to the second target sample label to obtain the second sample.

[0048] In an exemplary embodiment, the generating the second number of new samples based on the original samples corresponding to the second target sample label includes:

[0049] Traversing each original sample corresponding to the second target sample label;

[0050] A reference sample was identified from the original sample that was not visited;

[0051] Determining the similarity between the reference sample and each original sample corresponding to the plurality of sample labels;

[0052] Based on the similarity between the reference sample and the original samples, sequentially determining a third number of original samples similar to the reference sample from the original samples;

[0053] Determining, from the third number of original samples, matching samples that have the same sample label as the reference sample;

[0054] Generating new samples corresponding to the reference samples based on the matching samples; the number of the new samples is obtained based on the number of matching samples corresponding to each reference sample and the second number;

[0055] Based on the new samples corresponding to the reference samples, the second number of new samples is obtained.

[0056] In an exemplary embodiment, the method further comprises:

[0057] Normalizing the number of matching samples corresponding to each reference sample to obtain a target coefficient corresponding to each reference sample;

[0058] Based on the second number and the target coefficient, the number of newly added samples corresponding to each reference sample is obtained.

[0059] In an exemplary embodiment, the generating a new sample corresponding to the reference sample based on the matching sample includes:

[0060] Determine samples to be operated in the number of the newly added samples from the third number of original samples;

[0061] Determine an actual difference between the reference sample and the sample to be operated; the actual difference is a multidimensional vector;

[0062] Randomly weighting each dimension of the actual difference to obtain a weighted difference corresponding to the actual difference;

[0063] Based on the reference sample and the weighted vector, a new sample corresponding to the reference sample is obtained.

[0064] According to a third aspect of an embodiment of the present disclosure, there is provided an information recommendation device, including:

[0065] An object attribute information acquisition unit, configured to acquire object attribute information of a target object;

[0066] The object label recognition unit is configured to input the object attribute information into an object recognition model for label recognition to obtain an identification label of the target object; the object recognition model is trained based on target samples corresponding to multiple sample labels; the target sample is obtained by sampling the original sample in a target sampling manner based on sample balance data of the original samples corresponding to the multiple sample labels; the sample balance data is used to characterize statistical feature information of the number of samples corresponding to each of the multiple sample labels; the target sampling manner is obtained based on the total number of samples of the original sample and the sample balance data;

[0067] The information recommendation unit is configured to recommend the to-be-recommended information corresponding to the identification tag to the target object.

[0068] In an exemplary embodiment, the apparatus further comprises:

[0069] A first acquisition unit is configured to acquire original samples corresponding to the target task; the original samples include original samples corresponding to the multiple sample labels;

[0070] A first determining unit is configured to determine the sample balance data based on the original samples corresponding to the multiple sample labels;

[0071] A first sampling unit is configured to perform sampling on the original sample in the target sampling manner to obtain a target sample corresponding to the target task;

[0072] The model training unit is configured to perform training based on the target sample to obtain the object recognition model.

[0073] According to a fourth aspect of an embodiment of the present disclosure, there is provided an object recognition device, including:

[0074] A second acquisition unit is configured to acquire original samples corresponding to the target task; the original samples include original samples corresponding to multiple sample labels;

[0075] A second determining unit is configured to determine sample balance data based on the original samples corresponding to the multiple sample labels; the sample balance data is used to characterize statistical feature information of the number of samples corresponding to each of the multiple sample labels;

[0076] A second sampling unit is configured to perform sampling of the original sample in a target sampling manner to obtain a target sample corresponding to the target task; the target sampling manner is obtained based on the total number of samples of the original sample and the sample balance data;

[0077] The information identification unit is configured to perform identification of the target object based on an object identification model corresponding to the target task to obtain an identification label of the target object; the object identification model is trained based on the target sample.

[0078] In an exemplary embodiment, the second sampling unit includes:

[0079] A first target sample label determination unit is configured to determine a first target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the first target sample label is greater than the sample balance data;

[0080] A first sample determination unit is configured to perform sample deletion processing on an original sample corresponding to the first target sample label to obtain a first sample;

[0081] The second determining unit is configured to obtain the target sample based on the first sample.

[0082] In an exemplary embodiment, the sample balance data includes original sample means corresponding to the plurality of sample labels;

[0083] The first target sample label determination unit includes:

[0084] A first difference determination unit is configured to determine the difference of the original sample quantity corresponding to the multiple sample labels based on the original sample quantity corresponding to the multiple sample labels and the original sample mean;

[0085] The third determining unit is configured to determine, among the multiple sample labels, a sample label whose original sample quantity difference is greater than or equal to a first preset threshold as the first target sample label.

[0086] In an exemplary embodiment, the first sample determination unit includes:

[0087] a sample clustering unit, configured to perform sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes;

[0088] A sample deletion unit is configured to perform deletion processing on samples under each sample class to obtain processed samples corresponding to the sample class;

[0089] The fourth determining unit is configured to obtain the first sample based on the processed samples corresponding to the multiple sample classes.

[0090] In an exemplary embodiment, the sample deletion unit includes:

[0091] a ratio information determining unit, configured to determine ratio information of the number of samples under the sample class and the number of samples of the original samples corresponding to the first target sample label;

[0092] A second difference information determining unit is configured to determine difference information of the sample quantity between the sample balance data and the original sample corresponding to the first target sample label;

[0093] A first quantity determining unit is configured to determine a first quantity corresponding to the sample class based on the ratio information and the difference information;

[0094] The processed sample determining unit is configured to delete the first number of samples from the samples under the sample class to obtain processed samples corresponding to the sample class.

[0095] In an exemplary embodiment, the second sampling unit includes:

[0096] A second target sample label determination unit is configured to determine a second target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the second target sample label is less than the sample balance data;

[0097] a sample adding unit, configured to perform sample adding processing on the original sample corresponding to the second target sample label to obtain a second sample;

[0098] The fifth determining unit is configured to obtain the target sample based on the second sample.

[0099] In an exemplary embodiment, the sample balance data includes original sample means corresponding to the plurality of sample labels;

[0100] The second target sample label determination unit includes:

[0101] A third difference information determining unit is configured to determine the difference of the original sample quantity corresponding to the multiple sample labels based on the original sample quantity corresponding to the multiple sample labels and the original sample mean;

[0102] The sixth determining unit is configured to determine, among the multiple sample labels, a sample label whose original sample quantity difference is greater than or equal to a second preset threshold as the second target sample label.

[0103] In an exemplary embodiment, the sample adding unit includes:

[0104] A second quantity determining unit is configured to determine a second quantity corresponding to the second target sample label based on the original sample quantity corresponding to the second target sample label and the sample balancing data;

[0105] A new sample generating unit is configured to generate the second number of new samples based on the original samples corresponding to the second target sample label;

[0106] The second sample determination unit is configured to add the second number of newly added samples to the second target sample label to obtain the second sample.

[0107] In an exemplary embodiment, the new sample generation unit includes:

[0108] A traversal unit, configured to traverse each original sample corresponding to the second target sample label;

[0109] A reference sample determination unit, configured to determine a reference sample from an original sample that has not been accessed;

[0110] a similarity unit, configured to determine the similarity between the reference sample and each original sample corresponding to the plurality of sample labels;

[0111] A seventh determining unit is configured to sequentially determine a third number of original samples similar to the reference sample from the original samples based on the similarity between the reference sample and the original samples;

[0112] a matching sample determining unit, configured to determine, from the third number of original samples, matching samples that have the same sample label as the reference sample;

[0113] A sample generating unit is configured to generate new samples corresponding to the reference samples based on the matching samples; the number of the new samples is obtained based on the number of matching samples corresponding to each reference sample and the second number;

[0114] The new sample determination unit is configured to execute the new samples corresponding to the reference samples to obtain the second number of new samples.

[0115] In an exemplary embodiment, the apparatus further comprises:

[0116] a normalization processing unit, configured to perform normalization processing on the number of matching samples corresponding to each reference sample to obtain a target coefficient corresponding to each reference sample;

[0117] The newly added sample quantity determination unit is configured to obtain the quantity of newly added samples corresponding to each reference sample based on the second quantity and the target coefficient.

[0118] In an exemplary embodiment, the sample generation unit includes:

[0119] a sample-to-be-operated determining unit, configured to determine the sample-to-be-operated of the number of the newly added samples from the third number of original samples;

[0120] an actual difference determination unit, configured to determine an actual difference between the reference sample and the sample to be operated; the actual difference is a multi-dimensional vector;

[0121] A random weighting unit, configured to perform random weighting on each dimension of the actual difference to obtain a weighted difference corresponding to the actual difference;

[0122] The eighth determining unit is configured to obtain a new sample corresponding to the reference sample based on the reference sample and the weighted vector.

[0123] According to a fifth aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: a processor; and a memory for storing instructions executable by the processor; wherein the processor is configured to execute the instructions to implement the information recommendation method or object recognition method as described above.

[0124] According to a sixth aspect of an embodiment of the present disclosure, a computer-readable storage medium is provided. When instructions in the computer-readable storage medium are executed by a processor of a server, the server is enabled to execute the information recommendation method or object recognition method as described above.

[0125] According to the seventh aspect of an embodiment of the present disclosure, a computer program product is provided, which includes a computer program, and the computer program is stored in a readable storage medium. At least one processor of a computer device reads and executes the computer program from the readable storage medium, so that the device performs the above-mentioned information recommendation method or object recognition method.

[0126] The technical solution provided by the embodiments of the present disclosure brings at least the following beneficial effects:

[0127] The present disclosure uses an object recognition model to perform label recognition on the object attribute information of the target object to obtain the recognition label of the target object; wherein the object recognition model is trained based on the target sample corresponding to the multiple sample labels; the target sample is obtained by sampling the original sample in a target sampling manner based on the sample balance data of the original sample corresponding to the multiple sample labels; the sample balance data is used to characterize the statistical feature information of the number of samples corresponding to each of the multiple sample labels; the target sampling manner is obtained based on the total number of samples of the original sample and the sample balance data; and then recommends the information to be recommended corresponding to the recognition label to the target object. The present disclosure determines the sample balance data of the original sample, and then samples the original sample in a target sampling manner based on the sample balance data to obtain the corresponding target sample, thereby improving the accuracy of the balanced sample data and the sample balance efficiency; further generates the corresponding object recognition model based on the balanced target sample training, improves the generalization ability of the object recognition model, and improves the accuracy of model recognition, thereby recommending information based on the recognition label recognized by the object recognition model, and improving the accuracy of information recommendation.

[0128] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0129] The drawings herein are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure, and do not constitute improper limitations on the present disclosure.

[0130] Figure 1 It is a schematic diagram of an implementation environment according to an exemplary embodiment.

[0131] Figure 2 The present invention is a flow chart of an information recommendation method according to an exemplary embodiment.

[0132] Figure 3 The figure is a flow chart of an object recognition method according to an exemplary embodiment.

[0133] Figure 4 The figure is a flow chart of a method for generating a target sample according to an exemplary embodiment.

[0134] Figure 5 The figure is a flow chart of a method for determining a target sample label according to an exemplary embodiment.

[0135] Figure 6 is a flow chart of a first sample generating method according to an exemplary embodiment.

[0136] Figure 7 The figure is a flow chart of a method for deleting samples under a sample class according to an exemplary embodiment.

[0137] Figure 8 is a flow chart of a second sample generating method according to an exemplary embodiment.

[0138] Fig. 9 The figure is a flow chart of a method for generating a new sample according to an exemplary embodiment.

[0139] Fig.10 The present invention is a flow chart of a method for generating new samples based on normalization processing according to an exemplary embodiment.

[0140] Fig.11 The figure is a flow chart of a method for generating new samples based on a weighted vector according to an exemplary embodiment.

[0141] Fig.12 The figure is a flow chart of a new user's first purchase preference model according to an exemplary embodiment.

[0142] Fig.13The present invention is a block diagram of an information recommendation device according to an exemplary embodiment.

[0143] Fig.14 The figure is a block diagram of an object recognition device according to an exemplary embodiment.

[0144] Fig.15 The figure is a structural diagram of an electronic device according to an exemplary embodiment. DETAILED DESCRIPTION

[0145] In order to enable ordinary persons in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings.

[0146] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure 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 disclosure described herein can be implemented in an order other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0147] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0148] See also Figure 1 , which shows a schematic diagram of an implementation environment provided by an embodiment of the present disclosure, the implementation environment may include: at least one first computer device 110 and a second computer device 120, the first computer device 110 and the second computer device 120 can communicate data through a network.

[0149] Specifically, the second computer device 120 may collect object attribute information of the object corresponding to the first computer device 110, and the sample object label to generate the corresponding original sample; the second computer device 120 may perform a sample balancing operation on the original sample to obtain the target sample; the second computer device 120 may train a preset machine learning model based on the target sample to obtain an object recognition model. Further, the second computer device 120 may recognize the target object based on the object attribute information of the target object and the object recognition model, obtain the recognition label corresponding to the target object, and recommend the to-be-recommended information corresponding to the recognition label to the target object.

[0150] The second computer device 120 may specifically include: an original sample generation terminal 1201, a sample data balancing terminal 1203, an object recognition model training terminal 1205, an information recommendation terminal 1207, etc., and data communication can be performed between any two ends of the original sample generation terminal 1201, the sample data balancing terminal 1203, the object recognition model training terminal 1205, and the information recommendation terminal 1207.

[0151] The first computer device 110 can communicate with the second computer device 120 based on a browser / server mode (B / S) or a client / server mode (C / S). The first computer device 110 may include: physical devices such as smart phones, tablet computers, laptops, digital assistants, smart wearable devices, vehicle terminals, servers, etc., and may also include software running in physical devices, such as applications, etc. The operating system running on the first computer device 110 in the embodiment of the present disclosure may include but is not limited to Android system, IOS system, Linux, Windows, etc.

[0152] The second computer device 120 can establish a communication connection with the first computer device 110 via wired or wireless communication. The second computer device 120 can include an independently operated server, or a distributed server, or a server cluster composed of multiple servers, wherein the server can be a cloud server.

[0153] In order to solve the problem of low recognition accuracy of the recognition model in the related art, the present disclosure provides an information recommendation method, which can be performed by the above-mentioned Figure 1 The second computer device in Figure 2 , the method may include:

[0154] S210. Obtain object attribute information of the target object.

[0155] In the embodiment of the present disclosure, the target object may be the user to be recommended, and the corresponding object attribute information may be attribute information related to the user to be recommended, and may specifically include static feature information and historical behavior feature information of the user to be recommended. Specifically, the static feature information may include relevant feature information such as the user's age, gender, geographic location, educational background, and income information, and the user's historical behavior may include user clicks, likes, comments, and other behaviors. By analyzing the user's clicks, likes, comments, and other behaviors on certain information, the user's historical behavior feature information may be obtained.

[0156] In a specific embodiment, in a scenario where the first conversion preference of a target object is predicted, the object attribute information of the target object can be periodically updated before the target object performs the first conversion; because as the static information filled in by the target object becomes richer and richer, and the historical behavior of the target object becomes more and more, more feature information of the target object can be obtained, and the attribute information of the target object is updated based on the updated information, so that the object attribute information can better represent the feature information of the target object, that is, improve the feature representation ability of the object attribute information for the target object, and accordingly, the subsequent information recommendation for the target object based on the continuously updated object attribute information can make the recommended information more in line with the conversion preference of the target object.

[0157] S220. Input the object attribute information into the object recognition model for label recognition to obtain the identification label of the target object; the object recognition model is trained based on the target samples corresponding to the multiple sample labels; the target samples are obtained by sampling the original samples in a target sampling manner based on the sample balance data of the original samples corresponding to the multiple sample labels; the sample balance data is used to characterize the statistical feature information of the number of samples corresponding to each of the multiple sample labels; the target sampling method is obtained based on the total number of samples of the original samples and the sample balance data.

[0158] Specifically, the sample balance data can be used to characterize the statistical characteristic information of the number of samples corresponding to each of the multiple sample labels, that is, the sample balance data can reflect the overall sample data distribution information of the original sample, rather than being limited to the individual analysis of a single original sample; further, the sample balance data can indicate the degree of concentration or concentration trend of each sample in the original sample.

[0159] In a specific embodiment, the original samples corresponding to the multiple sample labels may be generated by the object attribute information of each object and the label category of the first conversion of each object. The original samples may specifically include the object attribute information of each object; each label category corresponds to the object attribute information of at least one object. In the original samples, there may be some label categories corresponding to the object attribute information of more objects, and some label categories corresponding to the object attribute information of fewer objects, which leads to an imbalance in the number of samples corresponding to each label category, so it is necessary to perform a sample balancing operation on the original samples to obtain the target samples.

[0160] In one example, the target sampling method may include oversampling, undersampling and mixed sampling. Mixed sampling may be a sampling method that uses both oversampling and undersampling. Specifically, the corresponding sampling method may be determined based on the comparison result between the number of original samples under each sample label and the sample balance data. When the number of original samples is too small, the oversampling sampling method is used, which can increase the number of samples. When the number of original samples is too large, the undersampling sampling method is used, which can reduce the number of samples. Thus, different sampling methods may be used for different sample labels. For example, for sample label 1, the number of original samples corresponding to it is small, so it is oversampled. For sample label 2, the number of original samples corresponding to it is large, so it is downsampled. Thus, the original samples may be mixed and sampled to obtain the target samples.

[0161] In an optional embodiment, by determining the sample balance data of the original sample, and then performing mixed sampling on the original sample based on the sample balance data, the mixed sampling can specifically be a sampling method that simultaneously adopts oversampling and undersampling, that is, oversampling samples corresponding to some sample labels, and undersampling samples corresponding to some sample labels, so as to obtain the target sample. Processing the original sample by a mixed sampling method rather than a single sampling method can improve the flexibility of sampling the sample, and thus improve the efficiency of sample balance. In addition, the corresponding sampling method is determined based on the sample balance data, so that the sampling method is adapted to the actual distribution of the sample, avoiding the introduction of too much noise data, thereby improving the accuracy of the balanced sample.

[0162] S230. Recommend the to-be-recommended information corresponding to the identification tag to the target object.

[0163] The information to be recommended may specifically be multimedia information for recommending a recommendation target corresponding to the identification tag, and the recommendation target may be an article, music, video, commodity, etc. corresponding to the identification tag.

[0164] The present disclosure determines sample balance data of original samples, and then samples the original samples in a target sampling manner based on the sample balance data to obtain corresponding target samples, thereby improving the accuracy of the balanced sample data and improving the sample balancing efficiency; further, the corresponding object recognition model is generated based on the balanced target sample training, thereby improving the generalization ability of the object recognition model and the accuracy of model recognition, thereby performing information recommendation based on the recognition tags recognized by the object recognition model, thereby improving the accuracy of information recommendation.

[0165] In one example, the step of training the object recognition model includes:

[0166] Acquire original samples corresponding to the target task; the original samples include original samples corresponding to the multiple sample labels; determine the sample balance data based on the original samples corresponding to the multiple sample labels; sample the original samples in the target sampling manner to obtain target samples corresponding to the target task; and obtain the object recognition model based on the target sample training.

[0167] In a specific embodiment, see Figure 3 , which shows an object recognition method, the execution subject of the method may be the second computer device mentioned above, and the method may specifically include:

[0168] S310. Obtain original samples corresponding to the target task; the original samples include original samples corresponding to multiple sample labels.

[0169] S320. Determine sample balance data based on the original samples corresponding to the multiple sample labels; the sample balance data is used to characterize statistical feature information of the number of samples corresponding to each of the multiple sample labels.

[0170] S330. Sampling the original samples in a target sampling manner to obtain target samples corresponding to the target task; the target sampling manner is obtained based on the total number of samples of the original samples and the sample balance data.

[0171] S340. Identify the target object based on the object recognition model corresponding to the target task to obtain an identification label of the target object; the object recognition model is trained based on the target sample.

[0172] In an example, an object may refer to a user or multimedia information. When the object is a user, the user's identification tag can be obtained through the object recognition model. The identification tag may specifically be the user's category tag or interest tag. When the user is multimedia information, the multimedia information's identification tag can be obtained through the object recognition model. The identification tag may specifically be the multimedia information's domain label, type label, etc.

[0173] The target task may be to identify user tags, to identify multimedia information tags, etc. In a specific embodiment, the target task is to identify user tags. By screening the first order data in the user's historical transactions and mining user characteristics from the basic portrait data and behavior data before the user's first order day, object attribute information is obtained; then, according to the category mapping relationship of the goods, the goods of each transaction are mapped to a unique product category C, and each category Ci is uniquely id-encoded to obtain a sample data label, thereby obtaining the original sample corresponding to the target task.

[0174] Since there are many categories of goods, if a multi-classification model is used for training directly, the model will prefer to predict goods in categories with more orders, resulting in a low accuracy rate of overall model recognition. Therefore, the original samples can be sampled based on the target sampling method to obtain target samples. The target sampling method can be determined based on the sample balance data and the total number of samples of the original samples, so that the sample balance data can reflect the overall sample data distribution of the original samples, rather than being limited to the individual analysis of a single original sample. Furthermore, the sample balance data can indicate the degree of concentration or concentration trend of each sample in the original sample, so that the target sampling method determined based on the sample balance data is consistent with the actual distribution of the samples, which can improve the accuracy of the balanced samples and avoid introducing too much noise data, thereby improving the recognition accuracy of the object recognition model trained based on the target samples.

[0175] In an alternative embodiment, see Figure 4 , which shows a method for generating a target sample, which may include:

[0176] S410. Based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels, determine a first target sample label from the multiple sample labels; the number of samples corresponding to the first target sample label is greater than the sample balance data.

[0177] S420. Perform sample deletion processing on the original sample corresponding to the first target sample label to obtain a first sample.

[0178] S430. Based on the first sample, obtain the target sample.

[0179] Specifically, the first target sample label needs to be determined based on the sample balance data, that is, the sample balance data can be used as a benchmark for dividing the sample label type, and the sample label type includes the first target sample label; thus, different sample balance operations are performed on different sample label types, specifically, the samples corresponding to the first target sample label are pruned; the first target sample label may include at least one sample label, so when the original sample corresponding to the first target sample label is pruned, the original sample corresponding to each sample label included in the first target sample label needs to be pruned separately. Thus, by using the sample balance data as a benchmark for distinguishing the sample label type, different sample balance operations are performed on different types of sample labels to achieve mixed sampling of the original samples, avoiding the problems of inaccurate sample balance and low sample balance efficiency caused by using only one sampling method, thereby improving the accuracy and efficiency of sample balance.

[0180] Optionally, the sample balance data corresponds to the original sample, and different original samples correspond to different sample balance data. The sample balance data may specifically be the average or median of the number of samples corresponding to each of the multiple sample labels.

[0181] In an optional embodiment, the sample balance data includes the original sample means corresponding to the multiple sample labels; accordingly, the sample balance data calculation method may include: determining the total number of samples of the original sample; and obtaining the original sample mean based on the total number of samples and the number of sample labels.

[0182] According to the above content, specifically, the number of original samples under each sample label is counted, and based on the number of original samples under each sample label, the total number of samples corresponding to multiple sample labels is obtained; then, the original sample mean is calculated based on the total number of samples and the number of sample labels. Since the average value can characterize the central position where the observations are relatively concentrated, it can reflect the general level of the phenomenon or the central trend of the distribution, so that the original sample mean is used as the sample balance data, and the sample balance data is used as the benchmark for distinguishing the sample label type. Different types of sample labels correspond to different sample balance operations, which can improve the rationality and accuracy of the sample label type classification, and further improve the reliability and accuracy of the samples after the balance operation.

[0183] In a specific embodiment, see Figure 5 , which shows a method for determining a target sample label, which may include:

[0184] S510. Based on the original sample quantities corresponding to the multiple sample labels and the original sample means, determine the difference in the original sample quantities corresponding to the multiple sample labels.

[0185] S520. Determine, among the multiple sample labels, a sample label whose original sample quantity difference is greater than or equal to a first preset threshold as a first target sample label.

[0186] Specifically, when determining the first target sample label, the difference between the original sample quantity corresponding to each sample label and the original sample mean can be calculated first, and sample balancing operations can be performed on sample labels whose original sample quantity difference is greater than or equal to a first preset threshold, while sample labels whose original sample quantity difference is less than the preset threshold do not need to be processed.

[0187] For example, the labels of the training samples are counted to obtain the original number of samples N (C i ), and use the mean of the number of samples under each label as the balance value of the number of sample labels That is, sample balanced data, the number of original samples corresponding to the first target sample label N (Ci ) are greater than the equilibrium value Right now Calculate the original number of samples N(C i )and The difference And set the first preset threshold θ, when Δ i <θ, the number of samples for this label is not adjusted; when Δ i ≥θ, then the sample label is determined to be the first target sample label.

[0188] Because the original sample mean is greater than the preset threshold, it means that the deviation from the data center is far and a balancing operation is needed. However, the original sample mean is less than the preset threshold, which means that the deviation from the data center is within an acceptable range and can be ignored, thereby reducing the amount of calculation and improving data balancing efficiency.

[0189] In an alternative embodiment, see Figure 6 , which shows a first sample generation method, which may include:

[0190] S610. Perform sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes.

[0191] S620. Perform deletion processing on samples under each sample class to obtain processed samples corresponding to the sample class.

[0192] S630. Obtain the first sample based on the processed samples corresponding to the multiple sample classes.

[0193] Specifically, the first target sample label may include multiple sample labels, and the original samples under each sample label may be clustered to obtain multiple sample classes; then the samples under each sample class are randomly deleted, and the remaining original samples in the sample class are determined as processed samples corresponding to the sample class; thereby, the first sample is obtained based on the processed samples corresponding to the multiple sample classes. The methods for implementing sample clustering in the related art can be applied to the embodiments of the present disclosure, and will not be repeated here.

[0194] In an optional embodiment, when a sample class includes samples corresponding to multiple sample labels, the corresponding number of samples to be deleted can be determined according to the number of samples corresponding to each sample label; specifically, the more samples the sample label corresponds to, the more samples are deleted, and the fewer samples the sample label corresponds to, the fewer samples are deleted; thereby, sample labels with a larger number of samples have more available samples and can delete more samples, and sample labels with a smaller number of samples have fewer available samples and can delete fewer samples, and as many samples as possible are retained under the sample label. Furthermore, for sample labels with a smaller number of samples, the samples under the sample label may not be deleted, so that the sample balancing operation under each sample class can be adapted to the number of samples corresponding to each label.

[0195] Through sample clustering, samples with similar characteristics can be brought together, and then samples can be randomly deleted from the concentrated samples, which can ensure that sample deletion operations are performed on each sample class, avoiding sample deletion being concentrated on samples of a small number of sample classes, so that the remaining original samples can still cover all sample classes, thereby improving the uniformity of sample deletion and the reliability of the first sample generation.

[0196] In a specific embodiment, see Figure 7 , which shows a method for deleting samples under a sample class, the method may include:

[0197] S710. Determine information on the ratio of the number of samples under the sample class to the number of original samples corresponding to the first target sample label.

[0198] S720. Determine the difference information between the sample balance data and the sample quantity of the original sample corresponding to the first target sample label.

[0199] S730. Determine a first quantity corresponding to the sample class based on the ratio information and the difference information.

[0200] S740. Delete the first number of samples from the samples under the sample class to obtain processed samples corresponding to the sample class.

[0201] Specifically, when deleting the original samples under each sample class, the number of original samples to be deleted can be determined first, and the number of original samples to be deleted can be determined based on sample balance data, the number of samples under the sample class, and the number of original samples corresponding to the first target sample label. Determining the first number may include: determining the ratio information of the number of samples under the sample class and the number of original samples corresponding to the first target sample label. Determine the difference information Δ between the sample balance data and the sample quantity of the original sample corresponding to the first target sample label i Based on the ratio information and the difference information, determine the first quantity δ corresponding to the sample class j Specifically, we calculate each class τ j The number of samples that should be randomly deleted N(C i ) is the number of original samples under each label, N(τ j ) represents the number of samples under the jth cluster.

[0202] Therefore, the number of deleted original samples is consistent with the relevant data, rather than being deleted arbitrarily, thereby improving the rationality of the number of deleted samples.

[0203] In one example, the method for generating a target sample may also include: determining a second target sample label from the multiple sample labels based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the second target sample label is less than the sample balance data; performing sample addition processing on the original sample corresponding to the second target sample label to obtain a second sample; and obtaining the target sample based on the second sample.

[0204] Specifically, the second target sample label needs to be determined based on the sample balance data, that is, the sample balance data can be used as a basis for dividing the sample label types, and the sample label types include the second target sample label; thereby different sample balance operations are performed on different sample label types, specifically, the samples corresponding to the second target sample label are added; the second target sample label may include at least one sample label, so when the original sample corresponding to the second target sample label is added, it is necessary to add samples to the original samples corresponding to each sample label included in the second target sample label.

[0205] Therefore, by using the sample balance data as a basis for distinguishing the sample label type, the corresponding first target sample label and the second target sample label can be determined. Further, different sample balancing operations can be performed on different types of sample labels to achieve mixed sampling of the original samples, avoiding the problems of inaccurate sample balance and low sample balancing efficiency caused by using only one sampling method, thereby improving the accuracy and efficiency of sample balancing.

[0206] In one example, a method for determining a second target sample label may include: determining a difference in the original sample quantity corresponding to the multiple sample labels based on the original sample quantity corresponding to the multiple sample labels and the original sample mean; and determining a sample label among the multiple sample labels whose original sample quantity difference is greater than or equal to a second preset threshold as the second target sample label.

[0207] Specifically, when determining the second target sample label, the difference between the original sample quantity corresponding to each sample label and the original sample mean can be calculated first, and sample balancing operations can be performed on sample labels whose original sample quantity difference is greater than or equal to the second preset threshold, while sample labels whose original sample quantity difference is less than the preset threshold do not need to be processed.

[0208] For example, the labels of the training samples are counted to obtain the original number of samples N (C i ), and use the mean of the number of samples under each label as the balance value of the number of sample labels That is, sample balanced data, the original number of samples corresponding to the second target sample label N (C i ) are both less than the equilibrium value Right now Calculate the original number of samples N(C i )and The difference And set the second preset threshold θ, when Δ i <θ, the number of samples for this label is not adjusted; when Δ i ≥θ, then the sample label is determined to be the second target sample label.

[0209] In the embodiment of the present disclosure, the first preset threshold and the second preset threshold can be the same or different. The first preset threshold and the second preset threshold can be determined based on the overall number of original samples, the number of samples required for model training, the resources consumed by model training, the accuracy of the model, and other factors. For example, when the overall number of original samples is large and the number of samples required for model training is small, more samples need to be adaptively deleted, and the first preset threshold can be set smaller; when the resources allocated to model training are limited, more samples need to be adaptively deleted, and the first preset threshold can be set smaller; for another example, when the accuracy of the model to be generated is high, it is necessary to retain as many samples as possible based on the preset training resources, so that the first preset threshold can be set larger and the second preset threshold can be set smaller; thereby achieving flexibility in setting the preset thresholds.

[0210] Because the original sample mean is greater than the preset threshold, it means that the deviation from the data center is far and a balancing operation is needed. However, the original sample mean is less than the preset threshold, which means that the deviation from the data center is within an acceptable range and can be ignored, thereby reducing the amount of calculation and improving data balancing efficiency.

[0211] In an alternative embodiment, see Figure 8 , which shows a second sample generation method, which may include:

[0212] S810. Determine a second quantity corresponding to the second target sample label based on the original sample quantity corresponding to the second target sample label and the sample balance data.

[0213] S820. Generate the second number of new samples based on the original samples corresponding to the second target sample label.

[0214] S830. Add the second number of newly added samples to the second target sample label to obtain the second sample.

[0215] Specifically, the second number can be determined based on s=Δ i *λ implementation, where λ can be pre-specified based on the number of samples corresponding to each label, Δ i The difference information of the sample quantity between the sample balance data and the original sample corresponding to the second target sample label may be used to determine the corresponding second quantity.

[0216] Similarly, the number of new samples required to generate the second target sample label also needs to be determined based on the original number of samples corresponding to the second target sample label and the sample balance data, so that the number of new samples is consistent with the relevant data, rather than being determined arbitrarily, thereby improving the rationality of the number of new samples.

[0217] In a specific embodiment, see Fig. 9 , which shows a method for generating a new sample, which may include:

[0218] S910. Traverse each original sample corresponding to the second target sample label.

[0219] S920. Determine a reference sample from the original sample that has not been visited.

[0220] S930. Determine the similarity between the reference sample and each original sample corresponding to the multiple sample labels.

[0221] S940. Based on the similarity between the reference sample and the original samples, determine in sequence from the original samples a third number of original samples that are similar to the reference sample.

[0222] S950. Determine, from the third number of original samples, matching samples that have the same sample label as the reference sample.

[0223] S960. Generate new samples corresponding to the reference samples based on the matching samples; the number of the new samples is obtained based on the number of matching samples corresponding to each reference sample and the second number.

[0224] S970. Obtain the second number of new samples based on the new samples corresponding to the reference samples.

[0225] Specifically, for each sample label in the second target sample label, each original sample under the sample label is traversed, that is, a corresponding new sample is generated for each original sample.

[0226] In a specific embodiment, the first target sample label includes sample labels a and b, sample label a includes samples a1 and a2, and sample label b includes sample b1; the second target sample label includes sample labels c and d, sample label c includes samples c1, c2, c3 and c4, and sample label d includes samples d1, d2 and d3; if the samples in sample label c are currently traversed, that is, samples c1, c2, c3 and c4 are traversed, sample c1 is determined as reference sample t, and the similarity between sample c1 and samples a1, a2, b1, c2, c3, c4, d1, d2 and d3 is calculated respectively, and then according to the similarity calculation result, the original samples with the similarity to sample c1 as the first third number K are determined, that is, the calculated similarities are sorted in order from large to small, and the original samples with the first third number K are determined. From the original samples of the third number K, the matching samples belonging to the same sample label c are determined, and the number of newly added samples corresponding to each second target sample label is determined based on the number of matching samples. The more matching samples corresponding to the reference sample, the more samples with higher similarity to the reference sample belong to the same sample label as the reference sample, and the more confident the samples under the sample label are. For samples with high confidence, more new samples can be generated. Therefore, when generating new samples, the more reference samples with more matching samples can generate more new samples, thereby improving the rationality of the generation of new samples and further improving the accuracy of sample balance.

[0227] In an alternative embodiment, see Fig.10 , which shows a method for generating new samples based on normalization processing, which may include:

[0228] S1010. Normalize the number of matching samples corresponding to each reference sample to obtain a target coefficient corresponding to each reference sample.

[0229] S1020. Based on the second number and the target coefficient, obtain the number of new samples corresponding to each reference sample.

[0230] The specific methods for determining the number of additional samples may include:

[0231] 1. For the second target sample label, traverse each sample t under the label, calculate the K samples closest to it, and from these K samples, determine the number of samples μ that belong to the same label as sample t t .

[0232] 2. After traversing the samples, μ corresponding to each sample t t Normalization is performed, that is, Get the number of samples G that should be generated for each sample t t =r t *s.

[0233] Since the number of additional samples required to correspond to the second target sample label has been determined, specifically the second number, the second number can be distributed to each original sample of the second target sample label, that is, the number of additional samples required to be generated corresponding to each original sample is determined. According to the above content, the number of matching samples corresponding to the reference sample can be used to characterize the confidence information of the reference sample, so as to generate the corresponding target coefficient according to the number of matching samples of each original sample, and based on the target coefficient and the second number, the number of additional samples specifically corresponding to each reference sample is obtained, which further improves the rationality of the determination of the additional samples and improves the accuracy of the sample balance.

[0234] In a specific embodiment, see Fig.11 , which shows a method for generating additional samples based on a weighted vector, which may include:

[0235] S1110. Determine samples to be operated in the number of the newly added samples from the third number of original samples.

[0236] S1120. Determine the actual difference between the reference sample and the sample to be operated; the actual difference is a multi-dimensional vector.

[0237] S1130. Randomly weight each dimension of the actual difference to obtain a weighted difference corresponding to the actual difference.

[0238] S1140. Based on the reference sample and the weighted vector, obtain a new sample corresponding to the reference sample.

[0239] Specifically, for each reference sample, the number n of newly added samples corresponding to it is determined, and then n samples to be operated are selected from the third number of original samples. The selection of n samples to be operated can be randomly selected from the third number of original samples, or can be the first n original samples with the highest similarity to the reference sample, which is not specifically limited in this embodiment.

[0240] In a specific embodiment, a random vector may be used to randomly weight each dimension of the actual difference, wherein the number of dimensions of the random vector is consistent with the number of dimensions of the actual difference, each dimension of the random vector may be a random number between 0 and 1, and a random vector needs to be randomly generated each time the random vector is used. Furthermore, a random zeroing function may be used, which may randomly set the value of each dimension of the random vector to zero.

[0241] Specifically, for each t in the second target sample label, the method of generating a corresponding newly added sample may include:

[0242] 1. Select G from the K samples closest to sample t t samples;

[0243] 2. Based on Generate new samples. Among them, d i Indicates the newly added sample, x i represents sample t, x zi Represents G t One sample out of the samples, is a random vector whose value of each dimension is a random number between 0 and 1. Its number of dimensions is consistent with the number of dimensions of the sample. drop is a function used to The numbers in each dimension are randomly set to zero.

[0244] Therefore, by using random vectors to randomly weight the actual differences instead of using a fixed scalar to calculate the actual differences, the differences between the features can be reflected, different new samples can be generated, and the diversity of the new samples can be improved. In addition, by randomly setting the random vectors to zero through a random zeroing function, on the one hand, calculations based on the random vectors after being set to zero can reduce the amount of calculations; on the other hand, the new samples calculated based on the random vectors after being set to zero can generate noise data caused by dead samples, thereby improving robustness.

[0245] The following uses the user's first purchase preference prediction scenario as an example to illustrate the implementation details of the present disclosure. Fig.12 , which shows a new user first purchase preference model process structure diagram; which may specifically include:

[0246] 1. Sample data construction

[0247] Filter the first order data in the user's historical transactions, and mine user characteristics from the basic portrait data and behavior data before the user's first order day to obtain object attribute information; then map each transaction product to a unique product category C based on the product category mapping relationship, and perform a i Perform unique ID encoding and obtain sample data labels to obtain training sample data.

[0248] 2. Label data analysis

[0249] Since there are many categories of goods, if we directly use a multi-classification model for training, the model will prefer to predict goods in categories with more orders, resulting in a low accuracy rate of overall model prediction. Therefore, we first count the labels of the training samples and get the number of samples N (C i ), and use the mean of the number of samples under each label as the balance value of the number of sample labels That is, the sample is balanced data.

[0250] 3. Mixed Sampling

[0251] Calculate the number of samples N(C i )and The difference And set a preset threshold θ, when Δ i <θ, the number of samples for this label is not adjusted.

[0252] When Δ i ≥θ, if Then the samples under this label are downsampled and deleted. The specific deletion process is as follows:

[0253] a) Get all samples under the label, and then use a clustering algorithm, such as K-means, to cluster them into τ classes.

[0254] b) Calculate τ for each class j The number of samples that should be randomly deleted N(τ j ) represents the number of samples under the jth cluster.

[0255] like It is necessary to oversample the samples under this label. Here, an adaptive sample uniform generation algorithm is used and optimized. The detailed process is as follows:

[0256] a) Calculate the number of new samples that should be generated under this label s = Δ i *λ, where λ can be pre-specified based on the number of samples corresponding to each label.

[0257] b) Traverse each sample t under the label, calculate the K samples closest to it, and from these K samples, determine the number of samples μ that have the same label as sample t t .

[0258] c) After traversing the samples, μ corresponding to each sample t t Normalization is performed, that is, Get the number of samples G that should be generated for each sample t t =r t *s.

[0259] d) For each sample t, generate the corresponding new sample. The process is as follows:

[0260] 1) Select G from the K samples closest to sample t t samples.

[0261] 2) Logic for generating new samples:

[0262]

[0263] Among them, d i Indicates the newly added sample, x i represents sample t, x zi Represents G t One sample out of the samples, is a random vector whose value of each dimension is a random number between 0 and 1. Its number of dimensions is consistent with the number of dimensions of the sample. drop is a function used to The numbers in each dimension are randomly set to zero.

[0264] Each sample corresponds to the object attribute information of a user, that is, the user feature information, and each item of user feature information can be mapped to a feature vector; thus, a new feature vector, that is, a new sample, can be obtained by calculating the feature vector and the random vector.

[0265] When generating new samples, the present invention adds random vectors instead of scalars to take into account the differences between features. On the other hand, it adds a drop function, which can reduce the noise caused by sample generation to some extent and improve robustness.

[0266] 4. Model training and prediction

[0267] Use the reconstructed samples and the labels corresponding to the samples to train the machine learning model and obtain the preference label prediction results; when recommending products to new users, use the model's prediction results to obtain the user's product category preferences and recommend products in the preferred categories to the user.

[0268] The multi-category sample balancing method based on mixed sampling in the present disclosure, on the one hand, improves the time efficiency in the sampling process and can complete the construction of balanced samples more quickly; on the other hand, it increases the accuracy and generalization of the model in prediction, and has a certain improvement in the accuracy of predicting the purchase preferences of new users.

[0269] Fig.13 is a block diagram of an information recommendation device according to an exemplary embodiment. Fig.13 , the device comprises:

[0270] The object attribute information acquisition unit 1310 is configured to acquire the object attribute information of the target object;

[0271] The object label identification unit 1320 is configured to perform label identification by inputting the object attribute information into an object identification model to obtain an identification label of the target object; the object identification model is obtained by training a target sample corresponding to a plurality of sample labels; the target sample is obtained by sampling the original sample in a target sampling manner based on sample balance data of the original samples corresponding to the plurality of sample labels; the sample balance data is used to characterize statistical feature information of the number of samples corresponding to each of the plurality of sample labels; the target sampling manner is obtained based on the total number of samples of the original sample and the sample balance data;

[0272] The information recommendation unit 1330 is configured to recommend the to-be-recommended information corresponding to the identification tag to the target object.

[0273] In an exemplary embodiment, the apparatus further comprises:

[0274] A first acquisition unit is configured to acquire original samples corresponding to the target task; the original samples include original samples corresponding to the multiple sample labels;

[0275] A first determining unit is configured to determine the sample balance data based on the original samples corresponding to the multiple sample labels;

[0276] A first sampling unit is configured to perform sampling on the original sample in the target sampling manner to obtain a target sample corresponding to the target task;

[0277] The model training unit is configured to perform training based on the target sample to obtain the object recognition model.

[0278] Fig.14 is a block diagram of an object recognition device according to an exemplary embodiment. Fig.14 , the device comprises:

[0279] The second acquisition unit 1410 is configured to acquire original samples corresponding to the target task; the original samples include original samples corresponding to multiple sample labels;

[0280] The second determining unit 1420 is configured to determine sample balance data based on the original samples corresponding to the multiple sample labels; the sample balance data is used to characterize statistical feature information of the number of samples corresponding to each of the multiple sample labels;

[0281] The second sampling unit 1430 is configured to perform sampling on the original sample in a target sampling manner to obtain a target sample corresponding to the target task; the target sampling manner is obtained based on the total number of samples of the original sample and the sample balance data;

[0282] The information identification unit 1440 is configured to perform identification of the target object based on an object identification model corresponding to the target task to obtain an identification label of the target object; the object identification model is trained based on the target sample.

[0283] In an exemplary embodiment, the second sampling unit 1430 includes:

[0284] A first target sample label determination unit is configured to determine a first target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the first target sample label is greater than the sample balance data;

[0285] A first sample determination unit is configured to perform sample deletion processing on an original sample corresponding to the first target sample label to obtain a first sample;

[0286] The second determining unit is configured to obtain the target sample based on the first sample.

[0287] In an exemplary embodiment, the sample balance data includes original sample means corresponding to the plurality of sample labels;

[0288] The first target sample label determination unit includes:

[0289] A first difference determination unit is configured to determine the difference of the original sample quantity corresponding to the multiple sample labels based on the original sample quantity corresponding to the multiple sample labels and the original sample mean;

[0290] The third determining unit is configured to determine, among the multiple sample labels, a sample label whose original sample quantity difference is greater than or equal to a first preset threshold as the first target sample label.

[0291] In an exemplary embodiment, the first sample determination unit includes:

[0292] a sample clustering unit, configured to perform sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes;

[0293] A sample deletion unit is configured to perform deletion processing on samples under each sample class to obtain processed samples corresponding to the sample class;

[0294] The fourth determining unit is configured to obtain the first sample based on the processed samples corresponding to the multiple sample classes.

[0295] In an exemplary embodiment, the sample deletion unit includes:

[0296] a ratio information determining unit, configured to determine ratio information of the number of samples under the sample class and the number of samples of the original samples corresponding to the first target sample label;

[0297] A second difference information determining unit is configured to determine difference information of the sample quantity between the sample balance data and the original sample corresponding to the first target sample label;

[0298] A first quantity determining unit is configured to determine a first quantity corresponding to the sample class based on the ratio information and the difference information;

[0299] The processed sample determining unit is configured to delete the first number of samples from the samples under the sample class to obtain processed samples corresponding to the sample class.

[0300] In an exemplary embodiment, the second sampling unit 1430 includes:

[0301] A second target sample label determination unit is configured to determine a second target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the second target sample label is less than the sample balance data;

[0302] a sample adding unit, configured to perform sample adding processing on the original sample corresponding to the second target sample label to obtain a second sample;

[0303] The fifth determining unit is configured to obtain the target sample based on the second sample.

[0304] In an exemplary embodiment, the sample balance data includes original sample means corresponding to the plurality of sample labels;

[0305] The second target sample label determination unit includes:

[0306] A third difference information determining unit is configured to determine the difference of the original sample quantity corresponding to the multiple sample labels based on the original sample quantity corresponding to the multiple sample labels and the original sample mean;

[0307] The sixth determining unit is configured to determine, among the multiple sample labels, a sample label whose original sample quantity difference is greater than or equal to a second preset threshold as the second target sample label.

[0308] In an exemplary embodiment, the sample adding unit includes:

[0309] A second quantity determining unit is configured to determine a second quantity corresponding to the second target sample label based on the original sample quantity corresponding to the second target sample label and the sample balancing data;

[0310] A new sample generating unit is configured to generate the second number of new samples based on the original samples corresponding to the second target sample label;

[0311] The second sample determination unit is configured to add the second number of newly added samples to the second target sample label to obtain the second sample.

[0312] In an exemplary embodiment, the new sample generation unit includes:

[0313] A traversal unit, configured to traverse each original sample corresponding to the second target sample label;

[0314] A reference sample determination unit, configured to determine a reference sample from an original sample that has not been accessed;

[0315] a similarity unit, configured to determine the similarity between the reference sample and each original sample corresponding to the plurality of sample labels;

[0316] A seventh determining unit is configured to sequentially determine a third number of original samples similar to the reference sample from the original samples based on the similarity between the reference sample and the original samples;

[0317] a matching sample determining unit, configured to determine, from the third number of original samples, matching samples that have the same sample label as the reference sample;

[0318] A sample generating unit is configured to generate new samples corresponding to the reference samples based on the matching samples; the number of the new samples is obtained based on the number of matching samples corresponding to each reference sample and the second number;

[0319] The new sample determination unit is configured to execute the new samples corresponding to the reference samples to obtain the second number of new samples.

[0320] In an exemplary embodiment, the apparatus further comprises:

[0321] a normalization processing unit, configured to perform normalization processing on the number of matching samples corresponding to each reference sample to obtain a target coefficient corresponding to each reference sample;

[0322] The newly added sample quantity determination unit is configured to obtain the quantity of newly added samples corresponding to each reference sample based on the second quantity and the target coefficient.

[0323] In an exemplary embodiment, the sample generation unit includes:

[0324] a sample-to-be-operated determining unit, configured to determine the sample-to-be-operated of the number of the newly added samples from the third number of original samples;

[0325] an actual difference determination unit, configured to determine an actual difference between the reference sample and the sample to be operated; the actual difference is a multi-dimensional vector;

[0326] A random weighting unit, configured to perform random weighting on each dimension of the actual difference to obtain a weighted difference corresponding to the actual difference;

[0327] The eighth determining unit is configured to obtain a new sample corresponding to the reference sample based on the reference sample and the weighted vector.

[0328] Regarding the device in the above embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be elaborated here.

[0329] In an exemplary embodiment, a computer-readable storage medium including instructions is also provided. Optionally, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a tape, a floppy disk, an optical data storage device, etc.; when the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute any of the methods described above.

[0330] In an exemplary embodiment, a computer program product is also provided, which includes a computer program stored in a readable storage medium, and at least one processor of a computer device reads and executes the computer program from the readable storage medium, so that the device performs any of the above methods.

[0331] This embodiment also provides an electronic device, whose structure diagram is shown in FIG. Fig.15 The device 1500 may vary greatly due to different configurations or performances, and may include one or more central processing units (CPU) 1522 (e.g., one or more processors) and memory 1532, one or more storage media 1530 (e.g., one or more mass storage devices) storing application programs 1542 or data 1544. Among them, the memory 1532 and the storage medium 1530 may be short-term storage or persistent storage. The program stored in the storage medium 1530 may include one or more modules (not shown in the figure), each module may include a series of instruction operations on the device. Furthermore, the central processing unit 1522 may be configured to communicate with the storage medium 1530 and execute a series of instruction operations in the storage medium 1530 on the device 1500. The device 1500 may also include one or more power supplies 1526, one or more wired or wireless network interfaces 1550, one or more input and output interfaces 1558, and / or, one or more operating systems 1541, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM Etc. Any of the above methods in this embodiment can be based on Fig.15 The device shown is implemented.

[0332] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0333] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. An information recommendation method, characterized in that: include: Get the object attribute information of the target object; Inputting the object attribute information into an object recognition model for label recognition to obtain an identification label of the target object; The object recognition model is obtained by training based on target samples corresponding to multiple sample labels; The target sample is obtained by sampling the original sample in a target sampling manner based on sample balance data of the original sample corresponding to the multiple sample labels; The sample balance data is used to characterize statistical characteristic information of the number of samples corresponding to each of the multiple sample labels; The target sampling method is obtained based on the total number of samples of the original sample and the sample balance data; The target object is a user or multimedia information. When the target object is the user, an identification tag of the user is obtained through the object recognition model, and the identification tag is a category tag or an interest tag of the user; when the target object is the multimedia information, an identification tag of the multimedia information is obtained through the object recognition model, and the identification tag is a domain tag or a type tag of the multimedia information; The steps of training the object recognition model include: Acquire original samples corresponding to the target task; the original samples include original samples corresponding to the multiple sample labels; determine the sample balance data based on the original samples corresponding to the multiple sample labels; determine a first target sample label from the multiple sample labels based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the first target sample label is greater than the sample balance data; perform sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes; determine the ratio information of the number of samples under the sample class and the number of samples of the original samples corresponding to the first target sample label; determine the difference information between the sample balance data and the number of samples of the original samples corresponding to the first target sample label; determine the first number corresponding to the sample class based on the ratio information and the difference information; delete the first number of samples from the samples under the sample class to obtain processed samples corresponding to the sample class; obtain a first sample based on the processed samples corresponding to the multiple sample classes; obtain the target sample based on the first sample; obtain the object recognition model based on the target sample training; and recommend the information to be recommended corresponding to the recognition label to the target object.

2. An object recognition method, characterized in that: include: Acquire original samples corresponding to the target task; the original samples include original samples corresponding to multiple sample labels; Determining sample balance data based on original samples corresponding to the multiple sample labels; The sample balance data is used to characterize statistical characteristic information of the number of samples corresponding to each of the multiple sample labels; Sampling the original sample in a target sampling manner to obtain a target sample corresponding to the target task; The target sampling method is obtained based on the total number of samples of the original sample and the sample balance data; The sampling of the original sample in a target sampling manner to obtain a target sample corresponding to the target task includes: determining a first target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the sample quantities corresponding to each of the multiple sample labels; the sample quantity corresponding to the first target sample label is greater than the sample balance data; Performing sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes; determining the ratio information of the sample quantity of the samples under the sample class to the sample quantity of the original samples corresponding to the first target sample label; determining the difference information of the sample balance data and the sample quantity of the original samples corresponding to the first target sample label; Based on the ratio information and the difference information, determine a first quantity corresponding to the sample class; delete the first quantity of samples from the samples under the sample class to obtain processed samples corresponding to the sample class; obtain a first sample based on the processed samples corresponding to the multiple sample classes; obtain the target sample based on the first sample; The target object is identified based on an object recognition model corresponding to the target task to obtain an identification label of the target object; the object recognition model is trained based on the target sample; the target object is a user or multimedia information, and when the target object is the user, the identification label of the user is obtained through the object recognition model, and the identification label is a category label or an interest label of the user; when the target object is the multimedia information, the identification label of the multimedia information is obtained through the object recognition model, and the identification label is a domain label or a type label of the multimedia information.

3. The method according to claim 2, characterized in that The sample balance data includes original sample means corresponding to the multiple sample labels; The determining a first target sample label from the multiple sample labels based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels comprises: Determine the difference between the original sample quantities corresponding to the multiple sample labels and the original sample means based on the original sample quantities corresponding to the multiple sample labels; Among the multiple sample labels, the sample label whose original sample quantity difference is greater than or equal to a first preset threshold is determined as the first target sample label.

4. The method according to claim 2, characterized in that: The sampling of the original sample in a target sampling manner to obtain a target sample corresponding to the target task includes: Based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels, a second target sample label is determined from the multiple sample labels; the number of samples corresponding to the second target sample label is less than the sample balance data; Performing sample augmentation processing on the original sample corresponding to the second target sample label to obtain a second sample; Based on the second sample, the target sample is obtained.

5. The method according to claim 4, characterized in that The sample balance data includes original sample means corresponding to the multiple sample labels; The determining a second target sample label from the multiple sample labels based on the relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels comprises: Determine the difference between the original sample quantities corresponding to the multiple sample labels and the original sample means based on the original sample quantities corresponding to the multiple sample labels; Among the multiple sample labels, the sample label whose original sample quantity difference is greater than or equal to a second preset threshold is determined as the second target sample label.

6. The method according to claim 4, characterized in that The performing sample augmentation processing on the original sample corresponding to the second target sample label to obtain the second sample includes: Determine a second quantity corresponding to the second target sample label based on the original sample quantity corresponding to the second target sample label and the sample balance data; Based on the original samples corresponding to the second target sample labels, generating the second number of newly added samples; The second number of newly added samples is added to the second target sample label to obtain the second sample.

7. The method according to claim 6, characterized in that The generating the second number of newly added samples based on the original samples corresponding to the second target sample labels includes: Traversing each original sample corresponding to the second target sample label; A reference sample was identified from the original sample that was not visited; Determining the similarity between the reference sample and each original sample corresponding to the plurality of sample labels; Based on the similarity between the reference sample and the original samples, sequentially determining a third number of original samples similar to the reference sample from the original samples; Determining, from the third number of original samples, matching samples that have the same sample label as the reference sample; Generating new samples corresponding to the reference samples based on the matching samples; the number of the new samples is obtained based on the number of matching samples corresponding to each reference sample and the second number; Based on the new samples corresponding to the reference samples, the second number of new samples is obtained.

8. The method according to claim 7, characterized in that The method further comprises: Normalizing the number of matching samples corresponding to each reference sample to obtain a target coefficient corresponding to each reference sample; Based on the second number and the target coefficient, the number of newly added samples corresponding to each reference sample is obtained.

9. The method according to claim 7, characterized in that: The generating a new sample corresponding to the reference sample based on the matching sample includes: Determine samples to be operated in the number of the newly added samples from the third number of original samples; Determine an actual difference between the reference sample and the sample to be operated; the actual difference is a multidimensional vector; Randomly weighting each dimension of the actual difference to obtain a weighted difference corresponding to the actual difference; Based on the reference sample and the weighted difference, a new sample corresponding to the reference sample is obtained.

10. An information recommendation device, characterized in that: include: An object attribute information acquisition unit, configured to acquire object attribute information of a target object; An object label recognition unit is configured to input the object attribute information into an object recognition model for label recognition to obtain an identification label of the target object; The object recognition model is obtained by training based on target samples corresponding to multiple sample labels; The target sample is obtained by sampling the original sample in a target sampling manner based on sample balance data of the original sample corresponding to the multiple sample labels; The sample balance data is used to characterize statistical characteristic information of the number of samples corresponding to each of the multiple sample labels; The target sampling method is obtained based on the total number of samples of the original sample and the sample balance data; The target object is a user or multimedia information. When the target object is the user, an identification tag of the user is obtained through the object recognition model, and the identification tag is a category tag or an interest tag of the user; when the target object is the multimedia information, an identification tag of the multimedia information is obtained through the object recognition model, and the identification tag is a domain tag or a type tag of the multimedia information; The steps of training the object recognition model include: Acquire original samples corresponding to the target task; the original samples include original samples corresponding to the multiple sample labels; determine the sample balance data based on the original samples corresponding to the multiple sample labels; determine a first target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the first target sample label is greater than the sample balance data; Performing sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes; determining the ratio information of the sample quantity of the samples under the sample class to the sample quantity of the original samples corresponding to the first target sample label; determining the difference information of the sample balance data and the sample quantity of the original samples corresponding to the first target sample label; Based on the ratio information and the difference information, determine a first quantity corresponding to the sample class; delete the first quantity of samples from the samples under the sample class to obtain processed samples corresponding to the sample class; obtain a first sample based on the processed samples corresponding to the multiple sample classes; obtain the target sample based on the first sample; and obtain the object recognition model based on the training of the target sample; The information recommendation unit is configured to recommend the to-be-recommended information corresponding to the identification tag to the target object.

11. An object recognition device, characterized in that: include: A second acquisition unit is configured to acquire original samples corresponding to the target task; the original samples include original samples corresponding to multiple sample labels; A second determining unit is configured to determine sample balance data based on the original samples corresponding to the multiple sample labels; The sample balance data is used to characterize statistical characteristic information of the number of samples corresponding to each of the multiple sample labels; A second sampling unit is configured to perform sampling on the original sample in a target sampling manner to obtain a target sample corresponding to the target task; The target sampling method is obtained based on the total number of samples of the original sample and the sample balance data; The sampling of the original sample in a target sampling manner to obtain a target sample corresponding to the target task includes: determining a first target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the sample quantities corresponding to each of the multiple sample labels; the sample quantity corresponding to the first target sample label is greater than the sample balance data; Performing sample clustering processing on the original samples corresponding to the first target sample label to obtain multiple sample classes; determining the ratio information of the sample quantity of the samples under the sample class to the sample quantity of the original samples corresponding to the first target sample label; determining the difference information of the sample balance data and the sample quantity of the original samples corresponding to the first target sample label; Based on the ratio information and the difference information, determine a first quantity corresponding to the sample class; delete the first quantity of samples from the samples under the sample class to obtain processed samples corresponding to the sample class; obtain a first sample based on the processed samples corresponding to the multiple sample classes; obtain the target sample based on the first sample; The information identification unit is configured to perform identification of the target object based on the object recognition model corresponding to the target task to obtain an identification label of the target object; the object recognition model is trained based on the target sample; the target object is a user or multimedia information, when the target object is the user, the identification label of the user is obtained through the object recognition model, and the identification label is the category label or interest label of the user; when the target object is the multimedia information, the identification label of the multimedia information is obtained through the object recognition model, and the identification label is the domain label or type label of the multimedia information.

12. The device according to claim 11, characterized in that The sample balance data includes original sample means corresponding to the multiple sample labels; The first target sample label determination unit includes: A first difference determination unit is configured to determine the difference of the original sample quantity corresponding to the multiple sample labels based on the original sample quantity corresponding to the multiple sample labels and the original sample mean; The third determining unit is configured to determine, among the multiple sample labels, a sample label whose original sample quantity difference is greater than or equal to a first preset threshold as the first target sample label.

13. The device according to claim 11, characterized in that The second sampling unit comprises: A second target sample label determination unit is configured to determine a second target sample label from the multiple sample labels based on a relative relationship between the sample balance data and the number of samples corresponding to each of the multiple sample labels; the number of samples corresponding to the second target sample label is less than the sample balance data; a sample adding unit, configured to perform sample adding processing on the original sample corresponding to the second target sample label to obtain a second sample; The fifth determining unit is configured to obtain the target sample based on the second sample.

14. The device according to claim 13, characterized in that The sample balance data includes original sample means corresponding to the multiple sample labels; The second target sample label determination unit includes: A third difference information determining unit is configured to determine the difference of the original sample quantity corresponding to the multiple sample labels based on the original sample quantity corresponding to the multiple sample labels and the original sample mean; The sixth determining unit is configured to determine, among the multiple sample labels, a sample label whose original sample quantity difference is greater than or equal to a second preset threshold as the second target sample label.

15. The device according to claim 13, characterized in that The sample adding unit comprises: A second quantity determining unit is configured to determine a second quantity corresponding to the second target sample label based on the original sample quantity corresponding to the second target sample label and the sample balancing data; A new sample generating unit is configured to generate the second number of new samples based on the original samples corresponding to the second target sample label; The second sample determination unit is configured to add the second number of newly added samples to the second target sample label to obtain the second sample.

16. The device according to claim 15, characterized in that The newly added sample generation unit includes: A traversal unit, configured to traverse each original sample corresponding to the second target sample label; A reference sample determination unit, configured to determine a reference sample from an original sample that has not been accessed; a similarity unit, configured to determine the similarity between the reference sample and each original sample corresponding to the plurality of sample labels; A seventh determining unit is configured to sequentially determine a third number of original samples similar to the reference sample from the original samples based on the similarity between the reference sample and the original samples; a matching sample determining unit, configured to determine, from the third number of original samples, matching samples that have the same sample label as the reference sample; A sample generating unit is configured to generate new samples corresponding to the reference samples based on the matching samples; the number of the new samples is obtained based on the number of matching samples corresponding to each reference sample and the second number; The newly added sample determining unit is configured to execute the newly added samples corresponding to the reference samples to obtain the second number of newly added samples.

17. The device according to claim 16, characterized in that The device also includes: a normalization processing unit, configured to perform normalization processing on the number of matching samples corresponding to each reference sample to obtain a target coefficient corresponding to each reference sample; The newly added sample quantity determination unit is configured to obtain the quantity of newly added samples corresponding to each reference sample based on the second quantity and the target coefficient.

18. The device according to claim 16, characterized in that The sample generation unit comprises: a sample-to-be-operated determining unit, configured to determine the sample-to-be-operated of the number of the newly added samples from the third number of original samples; an actual difference determination unit, configured to determine an actual difference between the reference sample and the sample to be operated; the actual difference is a multi-dimensional vector; A random weighting unit, configured to perform random weighting on each dimension of the actual difference to obtain a weighted difference corresponding to the actual difference; An eighth determining unit is configured to obtain a new sample corresponding to the reference sample based on the reference sample and the weighted difference.

19. An electronic device, characterized in that: include: processor; a memory for storing instructions executable by the processor; The processor is configured to execute the instructions to implement the information recommendation method according to claim 1 or the object recognition method according to any one of claims 2-9.

20. A computer-readable storage medium, characterized in that: When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the electronic device is enabled to execute the information recommendation method as described in claim 1, or the object recognition method as described in any one of claims 2-9.

21. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the information recommendation method according to claim 1 or the object recognition method according to any one of claims 2 to 9 is implemented.

Images