A feature extraction method and apparatus

Abstract

This application discloses a feature extraction method that can acquire multimedia content to be processed, input the information of the multimedia content to be processed into a target feature extraction model, and obtain the target features of the multimedia content to be processed. In scenarios where the probability of a user interacting with the multimedia content in a specific category is predicted, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. Using the scheme of this application's embodiments, target features that are complementary to the historical features of the multimedia content to be processed in scenarios where the probability of a user interacting with the multimedia content in a specific category is predicted can be obtained based on the target feature extraction model. Based on the fused features obtained by fusing the target features and historical features, a more accurate model for predicting the probability of a user interacting with the multimedia content in a specific category can be obtained, thereby improving the prediction accuracy of the probability of a user interacting with the multimedia content in a specific category.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a feature extraction method and apparatus. Background Technology

[0002] With the development of science and technology, machine learning models can be used to predict certain indicators. For example, machine learning models can be used to predict the click-through rate (CTR), forwarding rate, and collection rate of multimedia content, etc., which will not be listed here. The machine learning model can be pre-trained using "historical" features related to multimedia content, and as time goes on, more and more "new" features different from those "historical" features will appear.

[0003] How to use these "new" features to make them contribute more to the model's prediction accuracy is a problem that remains to be solved. Summary of the Invention

[0004] To solve, or at least partially solve, the above-mentioned technical problems, this application provides a feature extraction method and apparatus.

[0005] In a first aspect, embodiments of this application provide a feature extraction method, the method comprising:

[0006] Acquire the multimedia content to be processed;

[0007] The information of the multimedia content to be processed is input into the target feature extraction model to obtain the target features of the multimedia content to be processed. In the scenario of predicting the probability of a user interacting with the multimedia content in a specific category, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. The target features and the historical features are fused together to obtain the fused features used to predict the probability of a user interacting with the multimedia content in a specific category. The historical features are features obtained by processing the information of the multimedia content to be processed using a preset feature extraction method.

[0008] Optionally, the target feature extraction model is trained in the following manner:

[0009] Obtain target training multimedia content and its tags, wherein the tags of the target training multimedia content are used to indicate the likelihood of a user interacting with the target training multimedia content in the specific category;

[0010] The target feature extraction model is trained using the information of the target training multimedia content, the first prediction result, and the label of the target training multimedia content. The first prediction result is the prediction result output by the initial model after the target training multimedia content is input into the initial model. The initial model is trained using historical features of historical training multimedia content. The initial model is used to predict the probability of a user interacting with multimedia content in a specific category.

[0011] Optionally, training the target feature extraction model using the information of the target training multimedia content, the first prediction result, and the labels of the target training multimedia content includes:

[0012] A first feature is obtained, which is obtained by an intermediate feature extraction model processing information of the target training multimedia content, wherein the target feature extraction model is the intermediate feature extraction model that has completed training;

[0013] Based on the first feature, the first prediction result, and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated.

[0014] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0015] A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type.

[0016] Based on the second prediction result and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated, wherein the target feature extraction model is the intermediate feature extraction model that has been trained.

[0017] Optionally, the fused feature is obtained by fusing the first feature, the first prediction result, and at least one auxiliary feature by the auxiliary model, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

[0018] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0019] Obtain the third prediction result obtained by the auxiliary model after processing the first feature;

[0020] Based on the third prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

[0021] Optionally, updating the parameters of the intermediate feature extraction model based on the third prediction result, the labels of the target training multimedia content, and the first prediction result includes:

[0022] Based on the third prediction result, the tags of the target training multimedia content, and the first prediction result, a loss function is calculated;

[0023] The parameters of the intermediate feature extraction model are updated based on the loss function.

[0024] Optionally, calculating the loss function based on the third prediction result, the labels of the target training multimedia content, and the first prediction result includes:

[0025] Based on the third prediction result, the label of the target training multimedia content, the first prediction result, and at least one auxiliary feature, the loss function is calculated, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

[0026] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0027] A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type.

[0028] Based on the second prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

[0029] Optionally, updating the parameters of the intermediate feature extraction model includes:

[0030] Update the parameters of the intermediate feature extraction model and the auxiliary model.

[0031] Optionally, the method further includes:

[0032] Based on the historical features and target features of the multimedia content to be processed, a target model is trained. The target model is used to predict the probability of a user interacting with the multimedia content in a specific category.

[0033] Secondly, embodiments of this application provide a feature extraction apparatus, the apparatus comprising:

[0034] The acquisition unit is used to acquire the multimedia content to be processed.

[0035] A determining unit is used to input the information of the multimedia content to be processed into a target feature extraction model to obtain the target features of the multimedia content to be processed. In a scenario where the probability of a user interacting with the multimedia content in a specific category is predicted, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. The fused features obtained after the target features and the historical features are fused are used to predict the probability of a user interacting with the multimedia content in a specific category. The historical features are features obtained by processing the information of the multimedia content to be processed using a preset feature extraction method.

[0036] Optionally, the target feature extraction model is trained in the following manner:

[0037] Obtain target training multimedia content and its tags, wherein the tags of the target training multimedia content are used to indicate the likelihood of a user interacting with the target training multimedia content in the specific category;

[0038] The target feature extraction model is trained using the information of the target training multimedia content, the first prediction result, and the label of the target training multimedia content. The first prediction result is the prediction result output by the initial model after the target training multimedia content is input into the initial model. The initial model is trained using historical features of historical training multimedia content. The initial model is used to predict the probability of a user interacting with multimedia content in a specific category.

[0039] Optionally, training the target feature extraction model using the information of the target training multimedia content, the first prediction result, and the labels of the target training multimedia content includes:

[0040] A first feature is obtained, which is obtained by an intermediate feature extraction model processing information of the target training multimedia content, wherein the target feature extraction model is the intermediate feature extraction model that has completed training;

[0041] Based on the first feature, the first prediction result, and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated.

[0042] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0043] A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type.

[0044] Based on the second prediction result and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated, wherein the target feature extraction model is the intermediate feature extraction model that has been trained.

[0045] Optionally, the fused feature is obtained by fusing the first feature, the first prediction result, and at least one auxiliary feature by the auxiliary model, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

[0046] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0047] Obtain the third prediction result obtained by the auxiliary model after processing the first feature;

[0048] Based on the third prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

[0049] Optionally, updating the parameters of the intermediate feature extraction model based on the third prediction result, the labels of the target training multimedia content, and the first prediction result includes:

[0050] Based on the third prediction result, the tags of the target training multimedia content, and the first prediction result, a loss function is calculated;

[0051] The parameters of the intermediate feature extraction model are updated based on the loss function.

[0052] Optionally, calculating the loss function based on the third prediction result, the labels of the target training multimedia content, and the first prediction result includes:

[0053] Based on the third prediction result, the label of the target training multimedia content, the first prediction result, and at least one auxiliary feature, the loss function is calculated, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

[0054] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0055] A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type.

[0056] Based on the second prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

[0057] Optionally, updating the parameters of the intermediate feature extraction model includes:

[0058] Update the parameters of the intermediate feature extraction model and the auxiliary model.

[0059] Optionally, the device further includes:

[0060] The training unit is used to train a target model based on the historical features and target features of the multimedia content to be processed. The target model is used to predict the probability of a user interacting with the multimedia content in a specific category.

[0061] Thirdly, embodiments of this application provide an apparatus, the apparatus including a processor and a memory;

[0062] The processor is configured to execute instructions stored in the memory to cause the device to perform the method as described in any one of the first aspects above.

[0063] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that instruct a device to perform the method as described in any one of the first aspects above.

[0064] Fifthly, embodiments of this application provide a computer program product that, when run on a computer, causes the computer to perform the method described in any of the first aspects above.

[0065] Compared with the prior art, the embodiments of this application have the following advantages:

[0066] This application provides a feature extraction method. In one example, multimedia content to be processed can be acquired, and then the information of the multimedia content to be processed can be input into a target feature extraction model to obtain the target features of the multimedia content to be processed. In a scenario predicting the probability of a user interacting with the multimedia content in a specific category, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. Therefore, using the solution of this application embodiment, a target feature complementary to the historical features of the multimedia content to be processed can be obtained based on the target feature extraction model in a scenario predicting the probability of a user interacting with the multimedia content in a specific category. Since the target features are complementary to the historical features, the target features can contribute more to training the model predicting the probability of a user interacting with the multimedia content in a specific category. Accordingly, in one example, based on the target features and the historical features, a more accurate model predicting the probability of a user interacting with the multimedia content in a specific category can be obtained, thereby improving the prediction accuracy of the probability of a user interacting with the multimedia content in a specific category. Attached Figure Description

[0067] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0068] Figure 1 This application provides a flowchart illustrating a feature extraction method.

[0069] Figure 2 This application provides a flowchart illustrating a method for training a target feature extraction model.

[0070] Figure 3 This application provides a schematic diagram of the structure of a feature extraction device. Detailed Implementation

[0071] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0072] The inventors of this application discovered through research that machine learning models can be pre-trained using "historical" features related to multimedia content. However, as time goes on, more and more "new" features, different from these "historical" features, emerge. Currently, the "new" features can be fused with the "historical" features, and the corresponding model can be retrained using the fused features. However, this method cannot fully realize the potential of the aforementioned "new" features; that is, it cannot maximize the contribution of the aforementioned "new" features to the model's prediction accuracy.

[0073] To address the aforementioned issues, this application provides a feature extraction method and apparatus.

[0074] The various non-limiting embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0075] Exemplary methods

[0076] See Figure 1 The figure is a flowchart illustrating a feature extraction method provided in an embodiment of this application.

[0077] Figure 1 The method shown can be executed by the client or the server, and this application embodiment does not make specific limitations.

[0078] In one example Figure 1 The method shown may include, for example, the following steps: S101-S102.

[0079] S101: Obtain the multimedia content to be processed.

[0080] The multimedia content to be processed mentioned here includes, but is not limited to: advertisements, short videos, articles, videos, etc.

[0081] The multimedia content to be processed mentioned here may be multimedia content that the user has interacted with, or multimedia content that is recommended to the user but which the user has not interacted with. This application does not make any specific limitations.

[0082] The user interaction with multimedia content mentioned in this application embodiment includes, but is not limited to: user clicking, browsing, liking, collecting, commenting on, or purchasing the multimedia content.

[0083] S102: Input the information of the multimedia content to be processed into the target feature extraction model to obtain the target features of the multimedia content to be processed. In the scenario of predicting the probability of a user interacting with the multimedia content in a specific category, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. The fused features obtained by fusing the target features and the historical features are used to predict the probability of a user interacting with the multimedia content in a specific category. The historical features are features obtained by processing the information of the multimedia content to be processed using a preset feature extraction method.

[0084] To maximize the effectiveness of the "new" features of the multimedia content to be processed in predicting the probability of a user interacting with the multimedia content in a specific category, this embodiment utilizes a target feature extraction model to process the information of the multimedia content to be processed, obtaining target features. In predicting the probability of a user interacting with the multimedia content in a specific category, the target features are complementary to the historical features of the multimedia content to be processed. This complementarity between the historical features and the target features can be understood as mutual reinforcement, resulting in a more complete set of features. In other words:

[0085] The fused features, obtained by fusing the historical features of the multimedia content to be processed with the target features, can be used to predict the probability of a user interacting with the multimedia content in a specific category. Furthermore, the accuracy of predicting the probability of a user interacting with the multimedia content in a specific category based on the fused features is higher. Specifically, the prediction accuracy corresponding to the fused features is higher than that corresponding to the historical features and also higher than that corresponding to the target features. The prediction accuracy corresponding to the fused features can be the accuracy of predicting the probability of a user interacting with the multimedia content in a specific category using the fused features; the prediction accuracy corresponding to the historical features can be the accuracy of predicting the probability of a user interacting with the multimedia content in a specific category using the historical features of the multimedia content to be processed; and the prediction accuracy corresponding to the target features can be the accuracy of predicting the probability of a user interacting with the multimedia content in a specific category using the target features.

[0086] The specific categories of interactions mentioned here include, but are not limited to, user clicks, views, likes, favorites, comments, or purchases mentioned above.

[0087] In this embodiment, the historical features of multimedia content (e.g., multimedia content to be processed, historical training multimedia content) can be obtained by processing the information of the multimedia content using a preset feature extraction method. The preset feature extraction method mentioned herein can be a traditional feature extraction method (e.g., a traditional feature extraction model). Therefore, in one example, the historical features of the multimedia content to be processed can be obtained by processing the information of the multimedia content to be processed using a preset feature extraction method. The preset feature extraction method mentioned herein can be a traditional feature extraction method (e.g., a traditional feature extraction model).

[0088] The information about the multimedia content to be processed mentioned in the embodiments of this application can be information related to the multimedia content to be processed. The information about the multimedia content to be processed includes, but is not limited to, the specific content of the multimedia content to be processed. For example, if the multimedia content to be processed is a video, then the information about the multimedia content to be processed can include: video content and video title.

[0089] As can be seen from the above description, since the target features are complementary to the historical features of the multimedia content to be processed, the target features can contribute more to training a model that predicts the probability of users interacting with multimedia content in a specific category.

[0090] In one example, after obtaining the target features of the multimedia content to be processed, a target model can be trained based on the historical features of the multimedia content and the target features. This target model is used to predict the probability of a user engaging in a specific type of interaction with the multimedia content. In another example, when training the target model, it can be trained based on the historical features of the multimedia content to be processed, the target features, and the labels of the multimedia content to be processed, wherein the labels of the multimedia content to be processed are used to indicate the probability of a user engaging in a specific category of interaction with the multimedia content to be processed.

[0091] As mentioned earlier, the multimedia content to be processed can be multimedia content that the user has previously interacted with. In this case, if the user has had a specific type of interaction with the multimedia content to be processed, then the tag of the multimedia content to be processed is used to indicate that the user has had a specific type of interaction with the multimedia content to be processed. In other words, the tag of the multimedia content to be processed is used to indicate that the probability of the user having a specific type of interaction with the multimedia content to be processed is very high. For example, the tag of the multimedia content to be processed is used to indicate that the probability value of the user having a specific type of interaction with the multimedia content to be processed is 1.

[0092] As mentioned earlier, the multimedia content to be processed can be multimedia content recommended to the user, but which the user has not interacted with. In this case, the tag of the multimedia content to be processed is used to indicate that the user has no specific type of interaction with the multimedia content to be processed, or in other words, the tag of the multimedia content to be processed is used to indicate that the probability of the user interacting with the multimedia content to be processed in a specific type is very low; for example, the tag of the multimedia content to be processed is used to indicate that the probability of the user interacting with the multimedia content to be processed in a specific type is 0.

[0093] Because the target features and the historical features of the multimedia content to be processed are complementary in scenarios where the likelihood of a user interacting with multimedia content in a specific category is predicted, the trained target model can more accurately predict the likelihood of a user interacting with multimedia content in a specific category.

[0094] In this embodiment of the application, the target feature extraction model mentioned in S102 can be pre-trained. Next, combined with... Figure 2 The method for training the target feature extraction model is introduced.

[0095] See Figure 2 The figure is a flowchart illustrating a method for training a target feature extraction model according to an embodiment of this application. Figure 2 The method shown may include, for example, the following S201-S202.

[0096] S201: Obtain the target training multimedia content and the tags of the target training multimedia content, wherein the tags of the target training multimedia content are used to indicate the likelihood of a user interacting with the target training multimedia content in the specific category.

[0097] The target training multimedia content is similar to the multimedia content to be processed, and includes, but is not limited to: advertisements, short videos, articles, videos, etc.

[0098] In one example, the target training multimedia content could be multimedia content that the user has previously interacted with. In this case, if the user has interacted with the target training multimedia content in a specific category, then the tag of the target training multimedia content indicates that the user has interacted with the target training multimedia content in a specific category. In other words, the tag of the target training multimedia content indicates that the probability of the user interacting with the target training multimedia content in a specific category is very high. For example, the tag of the target training multimedia content indicates that the probability value of the user interacting with the target training multimedia content in a specific category is 1.

[0099] In another example, the target training multimedia content could be multimedia content recommended to the user, but which the user has not interacted with. In this case, the tag of the target training multimedia content indicates that the user has no specific type of interaction with the target training multimedia content, or in other words, the tag of the target training multimedia content indicates that the probability of the user interacting with the target training multimedia content in a specific type is very low; for example, the tag of the target training multimedia content indicates that the probability of the user interacting with the target training multimedia content in a specific type is 0.

[0100] S202: Using the information of the target training multimedia content, the first prediction result, and the label of the target training multimedia content, train the target feature extraction model, wherein the first prediction result is the prediction result output by the initial model after the target training multimedia content is input into the initial model, the initial model is trained using historical features of historical training multimedia content, and the initial model is used to predict the probability of a user interacting with multimedia content in a specific category.

[0101] In this embodiment, to enable the trained target feature extraction model to extract target features complementary to the historical features of the multimedia content being processed, the target feature extraction model can be trained using the first prediction result obtained after processing the target training multimedia content with the initial model. The initial model is trained using the historical features of the historical training multimedia content and is used to predict the probability of a user interacting with the multimedia content in a specific category.

[0102] Regarding the historical training multimedia content, please refer to the description of the target training multimedia content; it will not be detailed here. As for the training method of the initial model, this application does not impose specific limitations on the embodiments. In one example, the initial model may be trained based on the historical features and labels of the historical training multimedia content. Regarding the labels of the historical training multimedia content, please refer to the description of the labels of the target training multimedia content above; it will not be detailed here.

[0103] In one example, S202 may include the following steps S2021-S2022 in a specific implementation.

[0104] S2021: Obtain a first feature, which is obtained by the intermediate feature extraction model processing the information of the target training multimedia content, wherein the target feature extraction model is the intermediate feature extraction model that has completed training.

[0105] In this embodiment of the application, the information of the target training multimedia content can be input into an intermediate feature extraction model, and the intermediate feature extraction model can process the target training multimedia content to obtain the first feature of the target training multimedia content.

[0106] S2022: Based on the first feature, the first prediction result, and the tags of the target training multimedia content, update the parameters of the intermediate feature extraction model.

[0107] In this embodiment of the application, after obtaining the first feature, the parameters of the intermediate feature extraction model can be updated by combining the first feature, the first prediction result, and the tags of the target training multimedia content.

[0108] In its implementation, S2022 can be combined with an auxiliary model. This auxiliary model is used to predict the probability of a user engaging in a specific type of interaction with multimedia content. In one example, the auxiliary model can be a Multilayer Perceptron (MLP).

[0109] In one example, in the specific implementation of S2022, in addition to updating the parameters of the intermediate feature extraction model, the parameters of the auxiliary model can also be updated.

[0110] S2022 can be implemented in multiple ways. The following introduces two possible implementation methods.

[0111] The first implementation method can be achieved through the following steps A1-A2.

[0112] A1: Obtain the second prediction result obtained by the auxiliary model processing the fused features, wherein the fused features are obtained by the auxiliary model fusing the first features and the first prediction result.

[0113] In this embodiment, the auxiliary model can first fuse the first feature and the first prediction result to obtain a fused feature. Further, the fused feature is processed to obtain a second prediction result. The second prediction result indicates the likelihood of a user engaging in a specific type of interaction with the target training multimedia content. It is easy to understand that the second prediction result combines the first feature and the first prediction result; therefore, the accuracy of the second prediction result can, to a certain extent, reflect the complementarity between the first feature and the historical features of the target training multimedia content.

[0114] In one example, to more comprehensively utilize the correlation between the first feature and the first prediction result, the auxiliary model can fuse the first feature, the first prediction result, and at least one auxiliary feature during feature fusion to obtain the fused feature. The at least one auxiliary feature can be obtained based on the first feature and the first prediction result.

[0115] This application does not specifically limit the at least one auxiliary feature. In one example, the at least one auxiliary feature may include the product of the first feature and the first prediction result. Wherein, assuming the feature dimension of the first feature is d and the dimension of the first prediction result is 1-dimensional, the product of the first feature and the first prediction result can be: multiplying each dimension of the first feature by the first prediction result to obtain a d-dimensional auxiliary feature.

[0116] The at least one auxiliary feature may further include: the difference feature between the first feature and the first prediction result. Wherein, assuming the feature dimension of the first feature is d and the dimension of the first prediction result is 1-dimensional, the difference between the first feature and the first prediction result can be obtained by subtracting the first prediction result from each dimension of the first feature to obtain a d-dimensional auxiliary feature.

[0117] The at least one auxiliary feature may further include: the product of the first feature and the result output after the first prediction result is processed through an MLP layer.

[0118] In this embodiment of the application, fusing the first feature and the first prediction result can be achieved by concatenating the first feature and the first prediction result in terms of dimensions. Similarly, fusing the first feature, the first prediction result, and at least one auxiliary feature can be achieved by concatenating the first feature, the first prediction result, and at least one auxiliary feature in terms of dimensions.

[0119] A2: Based on the second prediction result and the tags of the target training multimedia content, update the parameters of the intermediate feature extraction model, wherein the target feature extraction model is the intermediate feature extraction model that has completed training.

[0120] In specific implementation, step A2 can be implemented by, for example, calculating a first loss function based on the second prediction result and the target training multimedia content labels, and then updating the parameters of the intermediate feature extraction model and the auxiliary model based on the first loss function.

[0121] The second implementation method can be achieved through the following steps B1-B2.

[0122] B1: Obtain the third prediction result obtained by the auxiliary model processing the first feature.

[0123] In this embodiment, the auxiliary model can process the first feature to obtain a third prediction result. The third prediction result is used to indicate the likelihood of a user engaging in a specific type of interaction with the target training multimedia content.

[0124] B2: Based on the third prediction result, the labels of the target training multimedia content, and the first prediction result, update the parameters of the intermediate feature extraction model.

[0125] In one example, step B2 can be implemented by calculating a loss function based on the third prediction result, the labels of the target training multimedia content, and the first prediction result. Then, based on the loss function, the parameters of the intermediate feature extraction model and the auxiliary model are updated. In one example, the loss function may be, for example, the aforementioned first loss function multiplied by a corresponding weight parameter, where the weight parameter is determined based on the first prediction result.

[0126] In one example, to more comprehensively utilize the correlation between the first feature and the first prediction result, the loss function can be calculated based on the third prediction result, the label of the target training multimedia content, the first prediction result, and at least one auxiliary feature. The at least one auxiliary feature is obtained based on the first feature and the first prediction result. In one example, the loss function can be, for example, the aforementioned first loss function multiplied by a corresponding weight parameter, where the weight parameter is determined based on the first prediction result and the at least one auxiliary feature. Regarding the at least one auxiliary feature, please refer to the relevant description section above; it will not be described in detail here.

[0127] When performing feature fusion, the auxiliary model can fuse the first feature, the first prediction result, and at least one auxiliary feature to obtain the fused feature. The at least one auxiliary feature can be obtained based on the first feature and the first prediction result.

[0128] In one example, the first and second methods described above can be combined, provided there is no conflict. For example, S2022 can be achieved through steps A1 and B3.

[0129] A1: Obtain the second prediction result obtained by the auxiliary model processing the fused features, wherein the fused features are obtained by the auxiliary model fusing the first features and the first prediction result.

[0130] For step A1, please refer to the relevant description above, which will not be repeated here.

[0131] B3: Based on the second prediction result, the labels of the target training multimedia content, and the first prediction result, update the parameters of the intermediate feature extraction model.

[0132] Regarding step B3, its principle is the same as that of step B2. The difference is that step B3 updates the parameters of the intermediate feature extraction model based on "the second prediction result, the label of the target training multimedia content, and the first prediction result," while step B2 updates the parameters of the intermediate feature extraction model based on "the third prediction result, the label of the target training multimedia content, and the first prediction result." For the specific implementation of step B3, please refer to the description of step B2 above; it will not be repeated here.

[0133] Exemplary device

[0134] Based on the methods provided in the above embodiments, this application also provides an apparatus, which will be described below with reference to the accompanying drawings.

[0135] See Figure 3 The figure is a schematic diagram of a feature extraction device provided in an embodiment of this application. The device 300 may specifically include, for example, an acquisition unit 301 and a determination unit 302.

[0136] Acquisition unit 301 is used to acquire multimedia content to be processed;

[0137] The determining unit 302 is used to input the information of the multimedia content to be processed into the target feature extraction model to obtain the target features of the multimedia content to be processed. In the scenario of predicting the probability of a user interacting with the multimedia content in a specific category, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. The fused features obtained by fusing the target features and the historical features are used to predict the probability of a user interacting with the multimedia content in a specific category. The historical features are features obtained by processing the information of the multimedia content to be processed using a preset feature extraction method.

[0138] Optionally, the target feature extraction model is trained in the following manner:

[0139] Obtain target training multimedia content and its tags, wherein the tags of the target training multimedia content are used to indicate the likelihood of a user interacting with the target training multimedia content in the specific category;

[0140] The target feature extraction model is trained using the information of the target training multimedia content, the first prediction result, and the label of the target training multimedia content. The first prediction result is the prediction result output by the initial model after the target training multimedia content is input into the initial model. The initial model is trained using historical features of historical training multimedia content. The initial model is used to predict the probability of a user interacting with multimedia content in a specific category.

[0141] Optionally, training the target feature extraction model using the information of the target training multimedia content, the first prediction result, and the labels of the target training multimedia content includes:

[0142] A first feature is obtained, which is obtained by an intermediate feature extraction model processing information of the target training multimedia content, wherein the target feature extraction model is the intermediate feature extraction model that has completed training;

[0143] Based on the first feature, the first prediction result, and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated.

[0144] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0145] A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type.

[0146] Based on the second prediction result and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated, wherein the target feature extraction model is the intermediate feature extraction model that has been trained.

[0147] Optionally, the fused feature is obtained by fusing the first feature, the first prediction result, and at least one auxiliary feature by the auxiliary model, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

[0148] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0149] Obtain the third prediction result obtained by the auxiliary model after processing the first feature;

[0150] Based on the third prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

[0151] Optionally, updating the parameters of the intermediate feature extraction model based on the third prediction result, the labels of the target training multimedia content, and the first prediction result includes:

[0152] Based on the third prediction result, the tags of the target training multimedia content, and the first prediction result, a loss function is calculated;

[0153] The parameters of the intermediate feature extraction model are updated based on the loss function.

[0154] Optionally, calculating the loss function based on the third prediction result, the labels of the target training multimedia content, and the first prediction result includes:

[0155] Based on the third prediction result, the label of the target training multimedia content, the first prediction result, and at least one auxiliary feature, the loss function is calculated, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

[0156] Optionally, updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes:

[0157] A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type.

[0158] Based on the second prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

[0159] Optionally, updating the parameters of the intermediate feature extraction model includes:

[0160] Update the parameters of the intermediate feature extraction model and the auxiliary model.

[0161] Optionally, the device further includes:

[0162] The training unit is used to train a target model based on the historical features and target features of the multimedia content to be processed. The target model is used to predict the probability of a user interacting with the multimedia content in a specific category.

[0163] Since the device 300 is a device corresponding to the method provided in the above method embodiments, the specific implementation of each unit of the device 300 is based on the same concept as the above method embodiments. Therefore, the specific implementation of each unit of the device 300 can be referred to the description section of the above method embodiments, and will not be repeated here.

[0164] This application also provides an apparatus, which includes a processor and a memory;

[0165] The processor is used to execute instructions stored in the memory so that the device performs the feature extraction method provided in the above method embodiments.

[0166] This application provides a computer-readable storage medium including instructions that instruct a device to perform the feature extraction method provided in the above method embodiments.

[0167] This application also provides a computer program product that, when run on a computer, causes the computer to execute the feature extraction method provided in the above method embodiments.

[0168] It is understood that before using the technical solutions of the various embodiments in this disclosure, users will be informed of the type, scope of use, and usage scenarios of the personal information involved in an appropriate manner, and user authorization will be obtained.

[0169] For example, upon receiving a user's active request, a prompt message is sent to the user to explicitly inform them that the requested operation will require the acquisition and use of the user's personal information. This allows the user to independently choose, based on the prompt message, whether to provide personal information to the software or hardware such as electronic devices, applications, servers, or storage media performing the operations of this disclosed technical solution.

[0170] As an optional but not limited implementation, in response to a user's active request, sending a prompt message to the user can be done via a pop-up window, where the prompt message can be presented in text format. Furthermore, the pop-up window can also include a selection control allowing the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0171] It is understood that the above notification and user authorization process are merely illustrative and do not constitute a limitation on the implementation of this disclosure. Other methods that comply with relevant laws and regulations may also be applied to the implementation of this disclosure.

[0172] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0173] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

[0174] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A feature extraction method, characterized in that, The method includes: Acquire the multimedia content to be processed; The information of the multimedia content to be processed is input into the target feature extraction model to obtain the target features of the multimedia content to be processed. In scenarios predicting the probability of a user interacting with multimedia content of a specific category, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. After the target feature and the historical feature are fused, the resulting fused feature is used to predict the probability of a user interacting with multimedia content in a specific category. The historical feature is a feature obtained by processing the information of the multimedia content to be processed using a preset feature extraction method. The target feature extraction model is trained in the following manner: Obtain target training multimedia content and its tags, wherein the tags of the target training multimedia content are used to indicate the likelihood of a user interacting with the target training multimedia content in the specific category; The target feature extraction model is trained using the information of the target training multimedia content, the first prediction result, and the label of the target training multimedia content. The first prediction result is the prediction result output by the initial model after the target training multimedia content is input into the initial model. The initial model is trained using historical features of historical training multimedia content. The initial model is used to predict the probability of a user interacting with multimedia content in a specific category. The step of training the target feature extraction model using information from the target training multimedia content, a first prediction result, and the tags of the target training multimedia content includes: A first feature is obtained, which is obtained by an intermediate feature extraction model processing information of the target training multimedia content, wherein the target feature extraction model is the intermediate feature extraction model that has completed training; Based on the first feature, the first prediction result, and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated.

2. The feature extraction method according to claim 1, characterized in that, The step of updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes: A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type. Based on the second prediction result and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated, wherein the target feature extraction model is the intermediate feature extraction model that has been trained.

3. The feature extraction method according to claim 2, characterized in that, The fusion feature is obtained by fusing the first feature, the first prediction result, and at least one auxiliary feature by the auxiliary model, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

4. The feature extraction method according to claim 1, characterized in that, The step of updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes: Obtain the third prediction result obtained by the auxiliary model processing the first feature; Based on the third prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

5. The feature extraction method according to claim 4, characterized in that, The step of updating the parameters of the intermediate feature extraction model based on the third prediction result, the tags of the target training multimedia content, and the first prediction result includes: Based on the third prediction result, the tags of the target training multimedia content, and the first prediction result, a loss function is calculated; The parameters of the intermediate feature extraction model are updated based on the loss function.

6. The feature extraction method according to claim 5, characterized in that, The step of calculating the loss function based on the third prediction result, the tags of the target training multimedia content, and the first prediction result includes: Based on the third prediction result, the label of the target training multimedia content, the first prediction result, and at least one auxiliary feature, the loss function is calculated, wherein the at least one auxiliary feature is obtained based on the first feature and the first prediction result.

7. The feature extraction method according to claim 1, characterized in that, The step of updating the parameters of the intermediate feature extraction model based on the first feature, the first prediction result, and the tags of the target training multimedia content includes: A second prediction result is obtained by processing the fused features by the auxiliary model. The fused features are obtained by the auxiliary model by fusing the first features and the first prediction result. The auxiliary model is used to predict the probability of a user interacting with multimedia content in a specific type. Based on the second prediction result, the labels of the target training multimedia content, and the first prediction result, the parameters of the intermediate feature extraction model are updated.

8. The feature extraction method according to claim 2, 4, or 7, characterized in that, The updating of the parameters of the intermediate feature extraction model includes: Update the parameters of the intermediate feature extraction model and the auxiliary model.

9. The feature extraction method according to claim 1, characterized in that, The method further includes: Based on the historical features and target features of the multimedia content to be processed, a target model is trained. The target model is used to predict the probability of a user interacting with the multimedia content in a specific category.

10. A feature extraction device, characterized in that, The device includes: The acquisition unit is used to acquire the multimedia content to be processed. A determining unit is used to input the information of the multimedia content to be processed into a target feature extraction model to obtain the target features of the multimedia content to be processed. In a scenario where the probability of a user interacting with the multimedia content in a specific category is predicted, the target features of the multimedia content to be processed are complementary to the historical features of the multimedia content to be processed. The fused features obtained by fusing the target features and the historical features are used to predict the probability of a user interacting with the multimedia content in a specific category. The historical features are features obtained by processing the information of the multimedia content to be processed using a preset feature extraction method. The target feature extraction model is trained in the following manner: Obtain target training multimedia content and its tags, wherein the tags of the target training multimedia content are used to indicate the likelihood of a user interacting with the target training multimedia content in the specific category; The target feature extraction model is trained using the information of the target training multimedia content, the first prediction result, and the label of the target training multimedia content. The first prediction result is the prediction result output by the initial model after the target training multimedia content is input into the initial model. The initial model is trained using historical features of historical training multimedia content. The initial model is used to predict the probability of a user interacting with multimedia content in a specific category. The step of training the target feature extraction model using information from the target training multimedia content, a first prediction result, and the tags of the target training multimedia content includes: A first feature is obtained, which is obtained by an intermediate feature extraction model processing information of the target training multimedia content, wherein the target feature extraction model is the intermediate feature extraction model that has completed training; Based on the first feature, the first prediction result, and the tags of the target training multimedia content, the parameters of the intermediate feature extraction model are updated.

11. A device, characterized in that, The device includes a processor and a memory; The processor is configured to execute instructions stored in the memory to cause the device to perform the feature extraction method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, Includes instructions that instruct the device to perform the feature extraction method as described in any one of claims 1 to 9.

Images