A model training method, a model application method, and related devices

Abstract

The model training method, the model application method and the related device disclosed by the embodiments of the present application can be used for training a model. Since the feature values of the fixed-value type features are relatively fixed, even if the content information amount corresponding to a certain sample content is small, each feature value of the fixed-value type features has relatively more corresponding content information for model training. Therefore, in the model training process, the values of the non-fixed-value type features are replaced by default values to generate a comparison vector for comparison training. Through a first loss function, the model learns how to generate a vector expression that can reflect the content specificity of the sample content based on the fixed-value type features, and the expression of the fixed-value type information on the content specificity in content sorting is strengthened, so that the model can also have a relatively accurate content expression for the sample content with a small content information amount. The model trained by the present application can be applied to the fields of advertisement pushing, multimedia content pushing and the like.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a model training method, a model application method and related devices. Background Art

[0002] The rationality and accuracy of content push are important factors affecting users' content browsing experience. In order to improve users' content browsing experience, content pushers usually first analyze the user's intention level for multiple contents, and then push content with higher intention level to users for browsing.

[0003] In related technologies, when analyzing content, a content ranking model is mainly used to rank the intention levels of multiple content to be pushed based on the content features corresponding to the content to be pushed, and push the content to the user based on the ranking results.

[0004] However, in the related art, when sorting, analysis based on content features is difficult to accurately reflect the differences between different contents, resulting in poor accuracy of content analysis and thus inability to reasonably push content. Summary of the Invention

[0005] In order to solve the above technical problems, the present application provides a model training method, and the model trained by this method can effectively improve the accuracy of content analysis.

[0006] The embodiments of this application disclose the following technical solutions:

[0007] In a first aspect, an embodiment of the present application discloses a model training method, the method comprising:

[0008] Acquire a content information set, the content information set including content information corresponding to a plurality of sample contents, the plurality of sample contents having corresponding sample content orderings, the content information including fixed-value features and non-fixed-value features;

[0009] Taking each of the plurality of content information as target content information, generating a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information through the first vector generation part of the initial content ranking model, and generating a second content representation vector corresponding to the target content information based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set;

[0010] generating a first loss function based on a first similarity between a first content representation vector corresponding to each piece of content information and a second content representation vector corresponding to each piece of content information and a second content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the first loss function decreases as the first similarity increases and decreases as the second similarity decreases;

[0011] Determining, by means of the initial content ranking model, pending content rankings corresponding to the plurality of sample contents according to first content representation vectors respectively corresponding to the plurality of content information in the content information set;

[0012] generating a second loss function according to a difference between the pending content ranking and the sample content ranking;

[0013] The parameters of the initial content ranking model are adjusted according to the first loss function and the second loss function to obtain a content ranking model, and the content ranking model is used to perform content ranking on a plurality of contents to be ranked.

[0014] In a second aspect, an embodiment of the present application discloses a model application method, which is applied to a content ranking model trained by any one of the first aspects, and the method includes:

[0015] Obtaining sorting content information corresponding to a plurality of to-be-sorted contents, wherein the sorting content information has the same feature type as the content information in the content information set, and the content information set is used to train the content sorting model;

[0016] Determining, by means of the content sorting model, content sorting corresponding to the plurality of to-be-sorted contents according to the sorting content information;

[0017] Content is pushed to the object to be pushed according to the content sorting.

[0018] In a third aspect, an embodiment of the present application discloses a model training device, comprising an acquisition unit, a first generation unit, a second generation unit, a first determination unit, a third generation unit, and a parameter adjustment unit:

[0019] The acquiring unit is configured to acquire a content information set, wherein the content information set includes content information corresponding to a plurality of sample contents, the plurality of sample contents have corresponding sample content orderings, and the content information includes fixed-value features and non-fixed-value features;

[0020] The first generation unit is configured to generate, as target content information, a plurality of content information respectively, generate, by a first vector generation part in an initial content ranking model, a first content representation vector corresponding to the target content information based on fixed-value type features and non-fixed-value type features in the target content information, and generate a second content representation vector corresponding to the target content information based on the fixed-value type features in the target content information and default values of the non-fixed-value type features corresponding to the content information set;

[0021] The second generation unit is configured to generate a first loss function based on a first similarity between the first content representation vector and the second content representation vector corresponding to each content information, and a second similarity between the first content representation vector corresponding to each content information and the second content representation vector corresponding to the content information different from the sample content in the plurality of content information, the first loss function decreasing with an increase of the first similarity and decreasing with a decrease of the second similarity;

[0022] The first determination unit is configured to determine, by the initial content ranking model, a pending content ranking corresponding to the plurality of sample contents according to the first content representation vectors corresponding to the plurality of content information in the content information set.

[0023] The third generation unit is configured to generate a second loss function according to a difference between the pending content ranking and the sample content ranking.

[0024] The parameter adjustment unit is configured to adjust the initial content ranking model according to the first loss function and the second loss function to obtain a content ranking model, the content ranking model being used for content ranking of a plurality of content to be ranked.

[0025] In a possible implementation, the first generation unit is specifically configured to:

[0026] The second vector generation part in the initial content ranking model is used to generate a first feature vector corresponding to the fixed-value type features in the target content information, a second feature vector corresponding to the non-fixed-value type features, and a third feature vector corresponding to the default values of the non-fixed-value type features.

[0027] The first vector generation part in the initial content ranking model is used to generate the first content representation vector according to the first feature vector and the second feature vector, and generate the second content representation vector according to the first feature vector and the third feature vector.

[0028] In a possible implementation, the first determination unit is specifically configured to:

[0029] Determining, by means of a vector ranking portion in the initial content ranking model, pending content rankings corresponding to the plurality of sample contents based on first content representation vectors respectively corresponding to the plurality of content information in the content information set;

[0030] The parameter adjustment unit is specifically used for:

[0031] The first vector generation part, the second vector generation part, and the vector sorting part in the initial content sorting model are adjusted according to the first loss function and the second loss function to obtain the content sorting model.

[0032] In one possible implementation, the second vector generating portion includes a first feature input layer and a second feature input layer, the first feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a second feature vector corresponding to a non-fixed-value feature, and the second feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a third feature vector corresponding to a default value of the non-fixed-value feature;

[0033] The first vector generating part includes a first fully connected layer and a second fully connected layer, the first fully connected layer is used to generate the first content representation vector according to the first feature vector and the second feature vector generated by the first feature input layer, and the second fully connected layer is used to generate the second content representation vector according to the first feature vector and the third feature vector generated by the second feature input layer;

[0034] The device further comprises a synchronization unit:

[0035] The synchronization unit is used to synchronously adjust the parameters of the first feature input layer and the second feature input layer, and synchronously adjust the parameters of the first fully connected layer and the second fully connected layer during the process of adjusting the parameters of the initial content sorting model.

[0036] In one possible implementation, the second fully connected layer includes multiple vector fusion layers, each vector fusion layer includes multiple fusion nodes, and the fusion nodes are used to fuse and output vectors output by fusion nodes in a previous vector fusion layer. The apparatus further includes a second determining unit:

[0037] The second determining unit is used to determine a target fusion node in the multi-layer vector fusion layer;

[0038] The first generating unit is specifically configured to:

[0039] The second content representation vector is generated according to the first feature vector and the third feature vector by removing the second fully connected layer of the target fusion node.

[0040] In a possible implementation, the apparatus further includes a third determining unit:

[0041] The third determining unit is configured to determine object information corresponding to each of the plurality of content information in the content information set, the object information being used to identify an object corresponding to the content information.

[0042] The first determining unit is specifically configured to:

[0043] The initial content ranking model is used to determine the pending content ranking of the plurality of sample contents according to the first content representation vector and the object information corresponding to each of the plurality of content information in the content information set.

[0044] In a possible implementation, the content information set includes content information belonging to the first information part and content information belonging to the second information part, and the first generating unit is specifically configured to:

[0045] In response to the target content information being content information belonging to the first information part, the second content representation vector corresponding to the target content information is generated based on the fixed-value class feature in the target content information and the default value of the non-fixed-value class feature corresponding to the content information set.

[0046] The apparatus further includes a fourth generating unit:

[0047] The fourth generating unit is configured to, in response to the target content information being content information belonging to the second information part, generate the second content representation vector corresponding to the target content information based on the fixed-value class feature and the non-fixed-value class feature in the target content information.

[0048] In a possible implementation, the content information set includes content information corresponding to a first grouping manner and content information corresponding to a second grouping manner, and the plurality of content information in the content information set correspond to a same plurality of feature categories, and the first generating unit is specifically configured to:

[0049] The plurality of content information corresponding to the first grouping manner are respectively taken as first target content information, and a first vector generation part in the initial content ranking model is used to generate a first content representation vector corresponding to the first target content information based on fixed-value type features and non-fixed-value type features in the first target content information, and generate a second content representation vector corresponding to the first target content information based on the fixed-value type features in the first target content information and default values of the non-fixed-value type features corresponding to the content information set;

[0050] The apparatus further includes a first grouping unit and a fifth generation unit:

[0051] The first grouping unit is configured to divide the plurality of feature categories into a first feature category group and a second feature category group based on similarity between feature categories.

[0052] The fifth generation unit is configured to take the plurality of content information corresponding to the second grouping manner as second target content information, and use the first vector generation part to generate a first content representation vector corresponding to the second target content information based on a plurality of features corresponding to the first feature category group in the second target content information, and generate a second content representation vector corresponding to the second target content information based on a plurality of features corresponding to the second feature category group in the second target content information.

[0053] In a possible implementation, the plurality of content information in the content information set correspond to the same plurality of feature categories, and the apparatus further includes a second grouping unit, a sixth generation unit, and a seventh generation unit:

[0054] The second grouping unit is configured to divide the plurality of feature categories into a first feature category group and a second feature category group based on similarity between feature categories.

[0055] The sixth generation unit is configured to use the first vector generation part in the initial content ranking model to generate a third content representation vector corresponding to the target content information based on a plurality of features corresponding to the first feature category group in the target content information, and generate a fourth content representation vector corresponding to the target content information based on a plurality of features corresponding to the second feature category group in the target content information.

[0056] the seventh generating unit is configured to generate a third loss function based on a third similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector, and a fourth similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the third loss function decreases as the third similarity increases and decreases as the fourth similarity decreases;

[0057] The parameter adjustment unit is specifically used for:

[0058] The initial content ranking model is adjusted according to the first loss function, the second loss function, and the third loss function.

[0059] In a possible implementation, the parameter adjustment unit is specifically configured to:

[0060] Determine a first weight parameter corresponding to the first loss function and a second weight parameter corresponding to the third loss function, where the first weight parameter is used to identify the degree of influence of the first loss function on the parameter adjustment, and the second weight parameter is used to identify the degree of influence of the third loss function on the parameter adjustment;

[0061] Generate a comprehensive loss function according to the first loss function, the first weight parameter, the second loss function, the third loss function and the second weight parameter;

[0062] The parameters of the initial content ranking model are adjusted according to the comprehensive loss function.

[0063] In a fourth aspect, an embodiment of the present application discloses a model application device, which is applied to a content ranking model trained by any one of the first aspects, and includes a first acquisition unit, a ranking unit, and a push unit:

[0064] The first acquisition unit is configured to acquire sorting content information corresponding to a plurality of to-be-sorted contents, wherein the sorting content information has the same feature type as the content information in the content information set, and the content information set is used to train the content sorting model;

[0065] The sorting unit is configured to determine the content sorting corresponding to the plurality of to-be-sorted contents according to the sorting content information using the content sorting model;

[0066] The pushing unit is used to push content to the object to be pushed according to the content ranking.

[0067] In a possible implementation, the apparatus further includes a second acquiring unit:

[0068] The second acquiring unit is configured to acquire object information corresponding to the object to be pushed;

[0069] The sorting unit is specifically used for:

[0070] The content ranking corresponding to the plurality of to-be-ranked contents is determined by using the content ranking model according to the ranking content information and the object information.

[0071] In a fifth aspect, an embodiment of the present application discloses a computer device, comprising a processor and a memory:

[0072] The memory is used to store program code and transmit the program code to the processor;

[0073] The processor is used to execute the model training method described in any one of the first aspects, or the model application method described in any one of the second aspects according to the instructions in the program code.

[0074] In a sixth aspect, an embodiment of the present application discloses a computer-readable storage medium, which is used to store a computer program, and the computer program is used to execute the model training method described in any one of the first aspects, or the model application method described in any one of the second aspects.

[0075] In the seventh aspect, an embodiment of the present application discloses a computer program product comprising instructions, which, when run on a computer, enables the computer to execute the model training method described in any one of the first aspects, or the model application method described in any one of the second aspects.

[0076] It can be seen from the above technical solution that this application will distinguish between fixed-value features and non-fixed-value features in content information when conducting model training. Since the feature values ​​of fixed-value features are relatively fixed, even if the amount of content information corresponding to a certain sample content is small, each feature value of the fixed-value feature has more corresponding content information that can be used for model training; since there are more feature values ​​of non-fixed-value features, when the amount of content information corresponding to a certain sample content is small, each feature value of the non-fixed-value feature has less content information corresponding to it, making it difficult for the model to fully learn the non-fixed-value features of the sample content. Based on this, this application will replace the values ​​of non-fixed value features with default values ​​during the model training process to generate comparative vectors for comparative training. Through the first loss function, the model learns how to generate vector expressions that can reflect the specificity of sample content based on fixed value features. Through the second loss function, the model learns how to accurately sort multiple sample contents. Therefore, combined with the first loss function and the second loss function, the expression of content specificity of fixed value information during content sorting can be enhanced, so that the content representation vector determined by the model can represent the content more accurately, and can also have a more accurate content expression for sample content with less content information, thereby improving the accuracy of content sorting. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0078] Figure 1 A schematic diagram of a model training method in an actual application scenario provided in an embodiment of the present application;

[0079] Figure 2 A flowchart of a model training method provided in an embodiment of the present application;

[0080] Figure 3 A schematic diagram of a model structure provided in an embodiment of the present application;

[0081] Figure 4 A schematic diagram of a model structure provided in an embodiment of the present application;

[0082] Figure 5 A schematic diagram of a model structure provided in an embodiment of the present application;

[0083] Figure 6A flowchart of a model application method provided in an embodiment of the present application;

[0084] Figure 7 A schematic diagram of a model structure provided in an embodiment of the present application;

[0085] Figure 8 A parameter diagram provided for an embodiment of the present application;

[0086] Figure 9 A structural block diagram of a device provided in an embodiment of the present application;

[0087] Figure 10 A structural block diagram of a device provided in an embodiment of the present application;

[0088] Figure 11 A structural diagram of a terminal provided in an embodiment of the present application;

[0089] Figure 12 A structural diagram of a server provided in an embodiment of the present application. DETAILED DESCRIPTION

[0090] The embodiments of the present application are described below with reference to the accompanying drawings.

[0091] Content ranking is the process of analyzing the intended audience's interest in multiple pieces of content and ranking them based on that interest. To achieve accurate content ranking, the content ranking model must be able to learn from the features contained in the content information how to accurately represent the content to be pushed.

[0092] In related art, content information typically includes fixed-value features and non-fixed-value features. Fixed-value features are features whose value can only be one of multiple fixed values ​​corresponding to the feature. For example, the feature "product type corresponding to the content" can only take a value from a fixed set of product types. Non-fixed-value features are features that do not have corresponding fixed values, such as content conversion rate and recall rate. It is understandable that for a model, the more samples corresponding to a feature value, the better it can learn how to analyze content based on that feature. However, not all content has sufficient content information. For content with a large amount of corresponding content information, each value of a non-fixed-value feature has a large number of samples corresponding to it, allowing the model to fully learn these features. However, for content with a small amount of corresponding content information, each value of a non-fixed-value feature has a small number of samples corresponding to it, making it difficult for the model to fully learn these features, making it difficult to accurately rank the content.

[0093] In order to solve the above technical problems, an embodiment of the present application provides a model training method, in which the processing device can strengthen the model's learning of the content-specific performance of fixed-value class features, thereby making up for the deficiencies in learning non-fixed-value class features, improving the accuracy of the model's overall content analysis, and thereby improving the accuracy of the content ranking determined by the model.

[0094] It is understandable that the method can be applied to a processing device, which is a processing device capable of performing model training, for example, a terminal device or a server with a model training function. The method can be executed independently by a terminal device or a server, or it can be applied to a network scenario in which a terminal device and a server communicate, and is executed in cooperation with the terminal device and the server. Among them, the terminal device can be a computer, a mobile phone or other device. The server can be understood as an application server or a web server. In actual deployment, the server can be an independent server or a cluster server. The terminal device may include an object terminal, and the object terminal includes but is not limited to a mobile phone, a computer, an intelligent voice interaction device, a smart home appliance, a vehicle-mounted terminal, an aircraft, etc. The content sorting determined by the model obtained by training based on this application can push the content to the relevant object through the object terminal.

[0095] In order to facilitate understanding of the technical solution provided by the embodiment of the present application, the following will introduce a model training method provided by the embodiment of the present application in combination with an actual application scenario.

[0096] See also Figure 1 , Figure 1 This is a schematic diagram of a model training method in an actual application scenario provided in an embodiment of the present application. In this actual application scenario, the processing device is a server 101 with a model training function.

[0097] Firstly, the server 101 obtains a content information set corresponding to a plurality of sample contents, the content information set including n content information (i.e., content information 1-content information n) corresponding to four sample contents, i.e., sample content 1-sample content 4, and taking each of the n content information as a target content information, the target content information including a non-fixed value type feature and a fixed value type feature. The server 101 replaces the non-fixed value type feature in the target content information with a default value of the non-fixed value type feature, and can obtain enhanced content information corresponding to the target content information. Based on the first vector generation part in the initial content ranking model, according to the fixed value type feature and the non-fixed value type feature in the target content information, a first content representation vector corresponding to the target content information can be generated, and according to the default value corresponding to the fixed value type feature and the non-fixed value type feature in the target content information, a second content representation vector corresponding to the target content information can be generated. In order to strengthen the expression of the specificity of the sample content by the fixed value type feature, the server 101 can obtain a first loss function according to a first similarity between the first content representation vector and the second content representation vector corresponding to each content information, and a second similarity between the first content representation vector corresponding to each content information and the second content representation vector corresponding to other content information different from the sample content. The more accurate the expression of the specificity of the sample content by the fixed value type feature is, the greater the first similarity is, the smaller the second similarity is, and the smaller the first loss function is.

[0098] At the same time, the initial content ranking model can determine a pending content ranking corresponding to a plurality of sample contents based on the first content representation vector corresponding to each of the n content information. The second loss function can be generated by the difference between the pending content ranking and the accurate sample content ranking corresponding to the content information set, and the learning of the ranking ability of the model can be realized through the second loss function. Therefore, by adjusting the initial content ranking model in combination with the first loss function and the second loss function, the model can learn how to strengthen the expression of the content specificity by the fixed value type feature in the ranking process. Because even if the content information corresponding to the sample content is less, there are still a large number of sample quantities for each value of the fixed value type feature, the content ranking model can also accurately rank some long-tail content in the sample, and the accuracy of the content ranking is improved.

[0099] Next, a model training method provided by the embodiments of the present application will be introduced in combination with the drawings.

[0100] Referring to Figure 2 , Figure 2 The flowchart of the model training method provided by the embodiments of the present application, the method comprises:

[0101] S201: Obtain a content information set.

[0102] The content information set includes content information corresponding to multiple sample contents. The sample contents can be of any type, such as advertising. This content information is generated based on the interaction between an object and the content and is used to analyze the object's level of interest in each sample content. An object refers to an object that can interact with the content, such as a user. For example, if the sample content is advertising, the content information may include features such as the ad's exposure rate, conversion rate, click-through rate, conversion rate, and click-through rate.

[0103] The plurality of sample contents have corresponding sample content rankings, which are relatively accurate content rankings corresponding to the plurality of sample contents. The content rankings are determined based on the degree of object's intention towards the plurality of sample contents analyzed based on the model.

[0104] Content information includes fixed-value features and non-fixed-value features. Fixed-value features are features whose value can only be one of the multiple fixed values ​​corresponding to the feature. For example, for advertising content, for a piece of content information corresponding to the advertising content, assuming that the product type corresponding to the advertising content has only four optional product types, with values ​​of "1, 2, 3, and 4" respectively, then the feature "product type corresponding to the advertising content" can only have four values: "1, 2, 3, and 4." Non-fixed-value features are features that do not have corresponding fixed values. For example, for advertising content, these non-fixed-value features can be the unique code corresponding to the advertising content, the conversion rate of the content, the recall rate, etc.

[0105] S202: Using the plurality of content information as target content information respectively, generating a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information through the first vector generation part of the initial content ranking model, and generating a second content representation vector corresponding to the target content information based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set.

[0106] It can be understood that in the content information set used for sorting, the amount of content information corresponding to each sample content is usually different. For some popular content, the amount of content information is large because of more opportunities for interaction with the object. For some unpopular content or new content, the amount of content information is usually small because of fewer opportunities for interaction with the object. For non-fixed value type features, when the amount of content information is small, the feature values of these features are not fixed, so there are more possible feature values. When the amount of content information is small, the amount of information corresponding to each feature value is small. Therefore, for content with less content information, the model is difficult to accurately learn the specific expression of these contents from non-fixed value type features. Because the feature values of the fixed value type features are relatively fixed, there are fewer possible feature values, so even if the amount of content information is small, each feature value has sufficient amount of information for the model to learn.

[0107] Based on this, in order to improve the accuracy of the model for content analysis, thereby improving the rationality of content sorting, the processing device can strengthen the learning of the model for the content-specific expression of the fixed value type features. In an embodiment of the present application, the processing device can set a corresponding default value for the non-fixed value type features in the content information in the content information set, and take a plurality of content information as target content information. For the target content information, through the first vector generation part in the initial content sorting model, on the one hand, the first content representation vector corresponding to the target content information can be generated based on the fixed value type features and the non-fixed value type features in the target content information. On the other hand, the second content representation vector corresponding to the target content information can be generated based on the default values corresponding to the fixed value type features and the non-fixed value type features. The content representation vector is used to represent the sample content corresponding to the target content information. Because the original non-fixed value type features are replaced by default values, the expression of the non-fixed value type features on the content specificity is reduced in the second content representation vector, and the expression of the fixed value type features on the content specificity is retained.

[0108] S203: generating a first loss function based on the first similarity between the first content representation vector and the second content representation vector corresponding to each content information, and the second similarity between the second content representation vector corresponding to each content information and the second content representation vector corresponding to the content information different from the sample content in the plurality of content information.

[0109] It can be understood that, since the expression of content specificity of the non-fixed value class is reduced in the second content representation vector, the influence degree of the fixed value class feature on the content expression can be reflected based on the first similarity between the first content representation vector corresponding to each content information and the second content representation vector. The greater the first similarity, the greater the influence degree of the fixed value class feature on the content expression, and vice versa. The second similarity between the first content representation vector corresponding to each content information and the second content representation vector corresponding to the content information different from the sample content in the plurality of content information can reflect the difference between the sample contents based on the fixed value class feature. The smaller the second similarity, the more obvious the difference between the sample contents based on the fixed value class feature.

[0110] As can be seen, when the model learns better for the fixed value class feature in the expression of content specificity, the first similarity should be greater and the second similarity should be smaller, therefore, the processing device can generate a first loss function according to the first similarity and the second similarity, the first loss function decreases with the increase of the first similarity, and decreases with the decrease of the second similarity.

[0111] S204: determining, by the initial content ranking model, a to-be-determined content ranking corresponding to the plurality of sample contents according to the first content representation vectors respectively corresponding to the plurality of content information in the content information set.

[0112] Since the final purpose of the model training is to rank the plurality of contents, when the model learns for the fixed value class feature, the learned fixed value class feature should be more suitable for content ranking; in addition, since the non-fixed value class feature can usually reflect the actual delivery performance of the content when delivered to the object, retaining the non-fixed value class feature also helps to improve the overall accuracy of the content ranking, therefore, the processing device can determine, by the initial content ranking model, a to-be-determined content ranking corresponding to the plurality of sample contents according to the first content representation vectors respectively corresponding to the plurality of content information in the content information set.

[0113] S205: generating a second loss function according to the difference between the to-be-determined content ranking and the sample content ranking.

[0114] Since the sample content ranking is a relatively accurate content ranking corresponding to the plurality of sample contents, and the to-be-determined content ranking is the content ranking output by the model, the difference between the to-be-determined content ranking and the sample content ranking can reflect the accuracy of the initial content ranking model in determining the content ranking based on the content representation vector.

[0115] S206: parameterizing the initial content ranking model according to the first loss function and the second loss function to obtain a content ranking model.

[0116] As mentioned above, the first loss function can reflect the learning accuracy of the initial content ranking model in expressing content-specific features for fixed-value features, and the second loss function can reflect the accuracy of the initial content ranking model in content ranking. Therefore, combining the first and second loss functions can enhance the role of fixed-value features in expressing content-specific features during the model's ranking of multiple contents, thereby enabling the trained content ranking model to perform relatively accurate content ranking even for content with less content information. During application, the content ranking model is used to sort multiple contents to be sorted, where the contents to be sorted are the contents that need to be sorted.

[0117] It can be seen from the above technical solution that this application will distinguish between fixed-value features and non-fixed-value features in content information when conducting model training. Since the feature values ​​of fixed-value features are relatively fixed, even if the amount of content information corresponding to a certain sample content is small, each feature value of the fixed-value feature has more corresponding content information that can be used for model training; since there are more feature values ​​of non-fixed-value features, when the amount of content information corresponding to a certain sample content is small, each feature value of the non-fixed-value feature has less content information corresponding to it, making it difficult for the model to fully learn the non-fixed-value features of the sample content. Based on this, this application will replace the values ​​of non-fixed value features with default values ​​during the model training process to generate comparative vectors for comparative training. Through the first loss function, the model learns how to generate vector expressions that can reflect the specificity of sample content based on fixed value features. Through the second loss function, the model learns how to accurately sort multiple sample contents. Therefore, combined with the first loss function and the second loss function, the expression of content specificity of fixed value information during content sorting can be enhanced, so that the content representation vector determined by the model can represent the content more accurately, and can also have a more accurate content expression for sample content with less content information, thereby improving the accuracy of content sorting.

[0118] To further strengthen the model's learning of fixed-value features, the processing device can also train the model's feature input portion, which converts features in content information into corresponding vectors for input into the model for processing. To improve the expression of content-specific fixed-value features, the processing device can extract more reasonable feature vectors corresponding to fixed-value features from this feature input portion.

[0119] In a possible implementation, the feature input part in the initial content ranking model may be a second vector generating part, and the second vector generating part is used to generate a feature vector corresponding to each feature in the content information.

[0120] The processing device can first generate a first feature vector corresponding to a fixed-value feature in the target content information, a second feature vector corresponding to a non-fixed-value feature, and a third feature vector corresponding to the default value of the non-fixed-value feature through the second vector generating part in the initial content sorting model. Then, through the first vector generating part in the initial content sorting model, a first content representation vector is generated based on the first feature vector and the second feature vector, and a second content representation vector is generated based on the first feature vector and the third feature vector. Thus, in the process of training the model, the processing device can not only adjust the parameters of the second vector generating part in the model used to fuse multiple feature vectors to obtain a content representation vector, but also adjust the parameters of the first vector generating part used to generate feature vectors, thereby refining the training process of the model and obtaining more accurate model training results.

[0121] Specifically, in one possible implementation, when performing content sorting, the processing device can determine the pending content sorting corresponding to multiple sample contents based on the first content representation vectors corresponding to the multiple content information in the content information set through the vector sorting part in the initial content sorting model. When adjusting the model parameters, the processing device can adjust the first vector generation part, the second vector generation part, and the vector sorting part in the initial content sorting model according to the first loss function and the second loss function to obtain the content sorting model. That is, during the model training process, the processing device can achieve synchronous adjustment of these three parts of the model, thereby strengthening the expression of content-specificity of fixed-value features in the overall content sorting process.

[0122] In one possible implementation, to effectively implement the aforementioned training process, the processing device may introduce a dual-tower model during the training process. The second vector generation portion includes a first feature input layer and a second feature input layer. The first feature input layer is configured to generate a first feature vector corresponding to a fixed-value feature in the target content information and a second feature vector corresponding to a non-fixed-value feature. The second feature input layer is configured to generate a first feature vector corresponding to a fixed-value feature in the target content information and a third feature vector corresponding to a default value of the non-fixed-value feature.

[0123] The first vector generation portion includes a first fully connected layer and a second fully connected layer. The first fully connected layer is used to generate a first content representation vector based on the first feature vector and the second feature vector generated by the first feature input layer. The second fully connected layer is used to generate a second content representation vector based on the first feature vector and the third feature vector generated by the second feature input layer. When generating the content representation vector, the fully connected layer can fuse and connect multiple input feature vectors to generate the desired content representation vector.

[0124] During parameter adjustment of the initial content ranking model, the processing device can synchronously adjust the parameters of the first feature input layer and the second feature input layer, as well as the parameters of the first fully connected layer and the second fully connected layer. This ensures that during training, the first content vector and the second content vector can accurately reflect the content-specific expression of fixed-value features, eliminating interference caused by different parameters on the vectors. In addition, since the first content vector will participate in the content ranking process, and the second content vector is mainly used for the model's learning of the content-specificity of fixed-value features, parameter synchronization can be used to adjust the model parameters in combination with the two learning dimensions of content ranking and content-specific expression, thereby improving the accuracy of the model.

[0125] See also Figure 3 , Figure 3 A schematic diagram of a model structure provided in an embodiment of the present application, in which the same content information can be input into two branches of the model. In the first branch, a first content representation vector can be directly generated through the first feature input layer and the first fully connected layer. In the second branch, a second content representation vector can be obtained by replacing non-fixed value features through the second feature input layer and the second fully connected layer. Based on the first content representation vector and the second content representation vector, a first loss function can be determined. The first loss function is used to strengthen the model's learning of content-specific expression of fixed value features.

[0126] In addition to obtaining enhanced vectors by replacing non-fixed value features in content information during feature input, the processing device can also perform further vector enhancement at the fully connected layer to further improve the richness of the samples.

[0127] In a possible implementation, the second fully connected layer may include multiple layers of vector fusion layers, each of which includes multiple fusion nodes, and the fusion nodes are used to fuse and output vectors output by the fusion nodes in the previous vector fusion layer. Figure 4 As shown, Figure 4This is a schematic diagram of a model structure provided by this application. The second fully connected layer includes three layers of fusion nodes. Each layer of fusion nodes includes multiple fusion nodes. Each fusion node can receive feature vectors output by multiple fusion nodes in the previous layer, fuse them, and output them to the fusion nodes in the next layer. The multi-layer vector fusion layer can include one or more target fusion nodes, which are fusion nodes used for data enhancement.

[0128] In order to achieve data enhancement in the fully connected layer, the processing device can first determine the target fusion node in the multi-layer vector fusion layer. When generating the second content representation vector, the processing device can remove the second fully connected layer of the target fusion node and generate the second content representation vector based on the first feature vector and the third feature vector. That is, in the process of generating the second content representation vector, the processing device can hide the influence of this part of the fusion node on the second content representation vector, thereby providing a certain amount of noise for the second content representation vector, which plays a role in data enhancement. For example, Figure 4 The fusion node 1 in the figure is the target fusion node. When the fusion node in the next layer receives the features output by the fusion node in the previous layer, it may not receive the feature output of the fusion node 1.

[0129] It can be understood that since the content ranking determined by this application is a ranking based on the dimension of the object's intention towards the content, in order to improve the accuracy of the content ranking, in one possible implementation, the processing device can also analyze the content ranking in combination with the object information.

[0130] The processing device may determine object information corresponding to each of the plurality of content information in the content information set. The object information is used to identify the object corresponding to the content information. That is, the content identification information is information generated by the interaction between the object and a certain sample content.

[0131] The processing device can use the initial content sorting model to analyze the association relationship between objects and content based on the first content representation vectors and object information corresponding to multiple content information in the content information set, thereby determining the pending content sorting corresponding to the multiple sample contents, so that the pending content sorting can more accurately reflect the object's intention for the content.

[0132] like Figure 5 As shown, Figure 5A schematic diagram of a model structure provided for an embodiment of the present application, wherein a branch for object information processing is added to the left side of the model, and a feature vector corresponding to the object information can be obtained through the third feature input layer in the model, and the feature vector can be fused and expressed through the third fully connected layer to obtain an object representation vector corresponding to the object information, and the object representation vector and the first feature vector corresponding to the content information can be used to determine the pending content ranking corresponding to multiple sample contents, and then the second loss function can be determined through the difference between the pending content ranking and the sample content ranking.

[0133] In one possible implementation, in addition to strengthening the model's learning of content-specific expressions for fixed-value features, the processing device may also strengthen the model's learning of content-specific expressions for non-fixed-value features.

[0134] The processing device can divide the content information set into a first information portion and a second information portion according to a certain ratio (e.g., 50%). That is, the content information set includes content information belonging to the first information portion and content information belonging to the second information portion. The content information of the first information portion is used to strengthen the learning of fixed-value features, and the content information of the second information portion is used to strengthen the learning of non-fixed-value features.

[0135] When processing target content information, the model can first determine the part to which the target content information belongs. In response to the target content information being content information belonging to the first information part, the processing device can generate a second content representation vector corresponding to the target content information based on the fixed-value class features in the target content information and the default values ​​of the non-fixed-value class features corresponding to the content information set, thereby enabling learning of the content-specific expression of the fixed-value class features based on the second content representation vector.

[0136] In response to the target content information being content information belonging to the second information portion, the processing device can directly generate a second content representation vector corresponding to the target content information based on the fixed-value features and non-fixed-value features in the target content information. Since the present application generates a first loss function based on the second similarity between the first content representation vector corresponding to each piece of content information and the second content representation vector corresponding to content information with different corresponding sample content in multiple pieces of content information, the more accurately the non-fixed-value features express content specificity, the smaller the second similarity will theoretically be. Therefore, this part of the content information can enable the model to learn the content-specific expression of non-fixed-value features.

[0137] The first information part and the second information part may be randomly selected.

[0138] In addition to dividing multiple features in the content information based on non-fixed value class features and fixed value class features, in a possible implementation, the processing device may also combine other feature grouping methods to form a comparison for feature learning.

[0139] For example, the processing device can divide the features into two groups based on the feature similarity between the feature types of multiple features included in the content information. If the two groups of features determined for the same sample content are relatively consistent in the expression of the sample content, and the two groups of features determined for different sample contents are relatively different in the expression of the sample content, then it means that the model is more effective in learning the overall features of multiple feature types, otherwise the learning accuracy is low.

[0140] Based on this, the processing device can combine the grouping method based on feature type similarity and the grouping method based on feature value type to perform model training. Among them, the training method combining the two grouping methods can include the following two methods:

[0141] First, in one possible implementation, the processing device may divide the content information in the content information set into two parts: one part is trained using a grouping method based on feature type similarity, and the other part is trained using a grouping method based on feature value types. The content information set may include content information corresponding to the first grouping method and content information corresponding to the second grouping method, and multiple content information in the content information set may correspond to the same multiple feature types.

[0142] The content information corresponding to the first grouping method adopts a grouping method based on the type of feature values. The processing device can use multiple content information corresponding to the first grouping method as first target content information respectively, and generate a first content representation vector corresponding to the first target content information based on the fixed-value features and non-fixed-value features in the first target content information through the first vector generation part in the initial content sorting model. The second content representation vector corresponding to the first target content information is generated based on the fixed-value features in the first target content information and the default values ​​of the non-fixed-value features corresponding to the content information set, thereby strengthening the model's learning of the content-specific expression of the fixed-value features.

[0143] The content information corresponding to the second grouping method adopts a grouping method based on feature type similarity. The processing device can divide multiple feature types into a first feature type group and a second feature type group based on the similarity between the feature types. The processing device can use the multiple content information corresponding to the second grouping method as the second target content information, and generate a first content representation vector corresponding to the second target content information based on the multiple features in the second target content information corresponding to the first feature type group through the first vector generation part, and generate a second content representation vector corresponding to the second target content information based on the multiple features in the second target content information corresponding to the second feature type group. Therefore, when the first loss function is determined based on the first content representation vector and the second content representation vector for model training, the model can learn how to strengthen the expression of content specificity based on the features of each feature type, thereby improving the overall expression accuracy of the multiple features for the content. In this training method, the two grouping methods can be calculated as the same loss function, namely the first loss function, to train the model.

[0144] Second, in order to increase the sample size analyzed during the training process, the processing device may process the entire content information in the content information set in two grouping ways.

[0145] The multiple content information in the content information set corresponds to the same multiple feature types. Based on the above content, on the basis of the function of obtaining the first loss by processing the multiple content information in the content information set based on the feature value type, the processing device can also divide the multiple feature types into a first feature type group and a second feature type group based on the similarity between the feature types.

[0146] Similarly, the processing device can generate, through the first vector generation part in the initial content ranking model, a third content representation vector corresponding to the target content information based on the plurality of features corresponding to the first feature category group in the target content information, and generate a fourth content representation vector corresponding to the target content information based on the plurality of features corresponding to the second feature category group in the target content information. Based on a third similarity between the third content representation vector and the fourth content representation vector corresponding to each content information, and a fourth similarity between the third content representation vector corresponding to each content information and the fourth content representation vector corresponding to the content information different from the sample content in the plurality of content information, a third loss function can be generated, which decreases with the increase of the third similarity and decreases with the decrease of the fourth similarity. Since the third similarity can reflect the consistency of the plurality of features of the same content information in expressing the content characteristics of the same content, and the fourth similarity can reflect the difference of the plurality of features corresponding to different content information in content expression, the model training through the third loss function can enable the model to learn how to perform more accurate content expression based on the features of the plurality of feature categories contained in the content information during the parameter adjustment.

[0147] Therefore, the processing device can adjust the initial content ranking model according to the first loss function, the second loss function and the third loss function to obtain a content ranking model with higher accuracy.

[0148] In a possible implementation, in order to obtain more accurate training effect when training the model in combination with the three loss functions, the processing device can assign corresponding weight parameters to the three loss functions. Since the main function of the content ranking model is still to rank a plurality of contents, the second loss function is a loss function for ranking accuracy, and the first loss function and the third loss function are both used to strengthen the learning of the model on feature content expression, the processing device can increase the influence degree of the second loss function in the training process, and appropriately reduce the influence degree of the other two loss functions.

[0149] The processing device can determine a first weight parameter corresponding to the first loss function, and a second weight parameter corresponding to the third loss function, the first weight parameter being used to identify the influence degree of the first loss function on the parameter adjustment, and the second weight parameter being used to identify the influence degree of the third loss function on the parameter adjustment.

[0150] The processing device can generate a comprehensive loss function based on the first loss function, the first weight parameter, the second loss function, the third loss function, and the second weight parameter, and then adjust the parameters of the initial content ranking model based on the comprehensive loss function. Of course, when adjusting the parameters based only on the first loss function and the second loss function, the processing device can also set corresponding weight parameters for the first loss function. The above-mentioned weight parameters can be the optimal values ​​determined by the processing device based on the model training results corresponding to the various weight parameter values.

[0151] In actual application, the processing device can use the content sorting model trained by the above process to sort the content. Figure 6 , Figure 6 This is a flowchart of a model application method provided in an embodiment of the present application. The method can be applied to a content ranking model trained by any of the above embodiments. The method includes:

[0152] S601: Obtain sorting content information corresponding to a plurality of contents to be sorted.

[0153] In order to ensure the validity of the content ranking results, the ranked content information has the same feature types as the content information in the content information set, that is, it also includes fixed value features and non-fixed value features. The content information set is used to train the content ranking model.

[0154] S602: Determine content rankings corresponding to a plurality of to-be-ranked contents according to ranking content information using a content ranking model.

[0155] Through the features in the content information, the processing device can analyze the content characteristics of the multiple contents to be sorted, and then analyze the intention levels corresponding to the multiple contents to be pushed for the objects to be pushed.

[0156] S603: Push the content to the object to be pushed according to the content ranking.

[0157] The processing device may push content that is of high interest to the object to be pushed based on the degree of interest reflected by the content ranking, thereby improving the content browsing experience of the object to be pushed.

[0158] In one possible implementation, to more accurately push content to the target object, the processing device may also obtain object information corresponding to the target object. This object information is used to identify the target object's characteristics, such as the target object's preferences for various types of content and its own object characteristics. The processing device may use a content ranking model to determine the content ranking corresponding to the multiple target content based on the ranking content information and the object information, making the content ranking more targeted to the target content.

[0159] In order to facilitate understanding of the technical solution provided by the embodiment of the present application, the following will introduce a model training method provided by the embodiment of the present application in combination with an actual application scenario.

[0160] In this actual application scenario, the sample content can be sample advertising content, and the content information corresponding to the advertising content is information generated based on the interaction between the object and the advertising content. This interaction can include clicks, exposures, conversions, etc. The content information includes multiple advertising features, of which non-fixed value features mainly include advertising identifiers (Identity Document, abbreviated as ID) and statistical features. For example, they can be the unique code corresponding to the advertising content, the click-through rate, conversion rate, and exposure rate corresponding to the advertising content in a historical period, etc. Fixed value features can include whether the advertising content corresponding to the content information has been clicked or converted, etc.

[0161] See also Figure 7 , Figure 7 This is a schematic diagram of a model structure provided in an embodiment of the present application. After the content information corresponding to the advertising content is input into the model, a randomly selected portion of the content information is replaced with default values. The first feature input layer in the model generates corresponding first and second feature vectors based on the fixed-value features and non-fixed-value features in the content information. The second feature input layer can replace the non-fixed-value features with corresponding default values, such as the default recall rate and default exposure rate, and determine the corresponding third feature vector based on the default values, thereby achieving data enhancement for the feature input layer.

[0162] The first fully connected layer can generate a corresponding first content representation vector based on the first eigenvector and the second eigenvector. The second fully connected layer includes a random dropout structure. Each fusion node in the fully connected layer can be regarded as a neuron. The target fusion node is the neuron discarded by the dropout structure and is expressed as 0 on the output. The role of this part of the fusion node in the fully connected layer is hidden to achieve data enhancement in the fully connected layer and obtain the second content representation vector. The first content representation vector and the second content representation vector can be used to obtain a first loss function for learning the expression of fixed-value features in the advertising content information that is specific to the advertising content.

[0163] For example, the processing device can perform data enhancement by randomly discarding advertisement IDs and statistical features with a certain probability. where f k is the kth feature among the m features included in the advertising content information, let f j Represents the non-fixed value class feature in the advertising content information. Assume that the i-th advertising content ad among multiple advertising contentsi With feature set After data enhancement of the feature input layer, the feature set obtained is in The default value corresponding to the non-fixed value class feature. constituted Positive samples.

[0164] In the second fully connected layer of the sorting model, there is a Dropout structure, which can provide a certain amount of noise and play a role in data enhancement. Through this Dropout structure, multiple feature vectors can be further enhanced to obtain the second content representation vector In another approach, the first content representation vector can be obtained by using fixed-value features and non-fixed-value features in the advertisement content information.

[0165] The same advertising content information is obtained by using data enhancement and As a positive example, the advertisement ad i The first content representation vector and other ads j The second content representation vector obtained by data augmentation As a negative example, self-supervised contrastive learning loss function (Info Noise Contrastive Estimation Loss, referred to as InfoNCE Loss) is used for self-supervised contrastive learning to obtain the first loss function. The formula of the first loss function is as follows:

[0166]

[0167] in, To calculate the similarity function of two content representation vectors, this solution uses cosine similarity. As can be seen from the formula, the value of the loss function increases with As the similarity between The other ads here ad j It can be all other advertisement contents in the plurality of sample advertisement contents, or part of the sample advertisement contents can be randomly selected as other advertisement contents to enhance the randomness of the training.

[0168] The loss function of self-supervised contrastive learning can be accumulated with the main task loss function MainLoss (i.e., the second loss function) of the ranking learning model for joint training, as shown in the following formula, where w is the weight of the loss function of self-supervised contrastive learning. The main task loss function of the ranking learning model can be trained using cross-entropy loss, Bayesian Personalized Ranking (BPR) loss, etc.

[0169] Loss=MainLoss+w*InfoNCE Loss

[0170] The content ranking model trained using this application can effectively improve the model's learning of advertising content, especially advertising content with less information content. This can include, for example, newly emerged advertising content and less popular long-tail advertising content. Long-tail advertising content refers to advertising content that appears fewer than N times in the training sample advertising content information set, while new advertising refers to advertising that has been exposed to the target audience fewer than M times in the advertising push system.

[0171] The actual test results of the model are shown in the following table:

[0172]

[0173] Among them, the coarse sorting algorithm is the algorithm adopted by the content sorting model of this application. Compared with the fine sorting algorithm, the coarse sorting algorithm is faster and more efficient, and the fine sorting algorithm is relatively more accurate. Therefore, this application can use the content sorting determined by the fine sorting algorithm as the sample content sorting.

[0174] The Group AUC (GAUC) is calculated by aggregating the top-ranked ad content within a request as positive examples and the non-top-ranked ad content as negative examples. This is used to evaluate the overall ranking capability of the coarse-ranked scores. The GAUC calculation formula is as follows:

[0175]

[0176] Here, #pv is the number of pieces of information in the advertisement content information.

[0177] Recall@N_K is the recall ratio of the top K ads in the fine ranking to the top N ads in the coarse ranking. This ratio is aggregated by request and evaluates the recall of the top N ads in the coarse ranking relative to the top K ads in the fine ranking, with K being less than N. A higher recall indicates a more accurate comparison between the fine ranking and the coarse ranking, indicating a better learning effect of the coarse ranking algorithm. The recall formula is as follows:

[0178]

[0179] As can be seen from the table, the model trained by this application has improved the ranking accuracy of overall advertising content, new advertising content and long-tail advertising content.

[0180] At the same time, the model trained by this application has a better convergence degree for features with non-fixed values. For example, see Figure 8 , Figure 8 A parameter diagram provided for an embodiment of the present application shows that, for advertising content, for the fixed-value feature "secondary industry to which the advertiser belongs," Curve 1 shows the convergence of the model in the related art to the emb mean of this feature vector, and Curve 2 shows the convergence of the model in the present application to the emb mean of this feature vector. As can be seen from the figure, the emb mean in the related art is constantly fluctuating, while it converges quickly in the present application. At the same time, after visualizing the content representation vectors corresponding to the various advertising contents determined by the model of the present application, it can be seen that the present application can simultaneously achieve long-tail advertising aggregation within the same advertiser industry, while also achieving a high degree of differentiation between different advertiser industries.

[0181] Based on the model training method provided in the above embodiment, the present application embodiment also provides a model training device, see Figure 9 , Figure 9 This is a structural block diagram of a model training device 900 provided in an embodiment of the present application, which includes an acquisition unit 901, a first generation unit 902, a second generation unit 903, a first determination unit 904, a third generation unit 905, and a parameter adjustment unit 906:

[0182] The acquiring unit 901 is configured to acquire a content information set, wherein the content information set includes content information corresponding to a plurality of sample contents, the plurality of sample contents having corresponding sample content orderings, and the content information includes fixed-value features and non-fixed-value features;

[0183] The first generating unit 902 is configured to use the plurality of content information as target content information, generate a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information, and generate a second content representation vector corresponding to the target content information based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set, using the first vector generating portion of the initial content ranking model.

[0184] The second generating unit 903 is configured to generate a first loss function based on a first similarity between a first content representation vector corresponding to each piece of content information and a second content representation vector corresponding to each piece of content information and a second content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the first loss function decreases as the first similarity increases and decreases as the second similarity decreases;

[0185] The first determining unit 904 is configured to determine, by using the initial content ranking model, the pending content rankings corresponding to the plurality of sample contents according to first content representation vectors respectively corresponding to the plurality of content information in the content information set;

[0186] The third generating unit 905 is configured to generate a second loss function according to the difference between the pending content ranking and the sample content ranking;

[0187] The parameter adjustment unit 906 is used to adjust the parameters of the initial content ranking model according to the first loss function and the second loss function to obtain a content ranking model, and the content ranking model is used to perform content ranking on multiple content to be ranked.

[0188] In a possible implementation, the first generating unit 902 is specifically configured to:

[0189] Generate, by means of a second vector generating portion in the initial content ranking model, a first feature vector corresponding to the fixed-value feature in the target content information, a second feature vector corresponding to the non-fixed-value feature, and a third feature vector corresponding to the default value of the non-fixed-value feature;

[0190] The first vector generating part in the initial content ranking model generates the first content representation vector according to the first feature vector and the second feature vector, and generates the second content representation vector according to the first feature vector and the third feature vector.

[0191] In a possible implementation, the first determining unit 904 is specifically configured to:

[0192] Determining, by means of a vector ranking portion in the initial content ranking model, pending content rankings corresponding to the plurality of sample contents based on first content representation vectors respectively corresponding to the plurality of content information in the content information set;

[0193] The parameter adjustment unit 906 is specifically used to:

[0194] The first vector generation part, the second vector generation part, and the vector sorting part in the initial content sorting model are adjusted according to the first loss function and the second loss function to obtain the content sorting model.

[0195] In one possible implementation, the second vector generating portion includes a first feature input layer and a second feature input layer, the first feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a second feature vector corresponding to a non-fixed-value feature, and the second feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a third feature vector corresponding to a default value of the non-fixed-value feature;

[0196] The first vector generating part includes a first fully connected layer and a second fully connected layer, the first fully connected layer is used to generate the first content representation vector according to the first feature vector and the second feature vector generated by the first feature input layer, and the second fully connected layer is used to generate the second content representation vector according to the first feature vector and the third feature vector generated by the second feature input layer;

[0197] The device further comprises a synchronization unit:

[0198] The synchronization unit is used to synchronously adjust the parameters of the first feature input layer and the second feature input layer, and synchronously adjust the parameters of the first fully connected layer and the second fully connected layer during the process of adjusting the parameters of the initial content sorting model.

[0199] In one possible implementation, the second fully connected layer includes multiple vector fusion layers, each vector fusion layer includes multiple fusion nodes, and the fusion nodes are used to fuse and output vectors output by fusion nodes in a previous vector fusion layer. The apparatus further includes a second determining unit:

[0200] The second determining unit is used to determine a target fusion node in the multi-layer vector fusion layer;

[0201] The first generating unit 902 is specifically configured to:

[0202] The second content representation vector is generated according to the first feature vector and the third feature vector by removing the second fully connected layer of the target fusion node.

[0203] In a possible implementation, the apparatus further includes a third determining unit:

[0204] The third determining unit is configured to determine object information corresponding to each of the plurality of content information in the content information set, wherein the object information is used to identify the object corresponding to the content information;

[0205] The first determining unit 904 is specifically configured to:

[0206] The undetermined content rankings corresponding to the plurality of sample contents are determined by using the initial content ranking model and according to the first content representation vectors and object information respectively corresponding to the plurality of content information in the content information set.

[0207] In a possible implementation, the content information set includes content information belonging to the first information part and content information belonging to the second information part, and the first generating unit 902 is specifically configured to:

[0208] In response to the target content information being content information belonging to the first information portion, generating a second content representation vector corresponding to the target content information based on fixed-value features in the target content information and default values ​​of non-fixed-value features corresponding to the content information set;

[0209] The apparatus further includes a fourth generating unit:

[0210] The fourth generating unit is configured to generate, in response to the target content information being content information belonging to the second information part, a second content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information.

[0211] In one possible implementation, the content information set includes content information corresponding to the first grouping method and content information corresponding to the second grouping method. Multiple pieces of content information in the content information set correspond to the same multiple feature types. The first generating unit 902 is specifically configured to:

[0212] Using the plurality of content information corresponding to the first grouping method as first target content information, generating a first content representation vector corresponding to the first target content information based on fixed-value features and non-fixed-value features in the first target content information through the first vector generation part of the initial content ranking model, and generating a second content representation vector corresponding to the first target content information based on the fixed-value features in the first target content information and default values ​​of the non-fixed-value features corresponding to the content information set;

[0213] The apparatus further includes a first grouping unit and a fifth generating unit:

[0214] The first grouping unit is configured to divide the plurality of feature categories into a first feature category group and a second feature category group based on similarities between the feature categories;

[0215] The fifth generating unit is configured to use the plurality of content information corresponding to the second grouping method as second target content information, and generate, through the first vector generating part, a first content representation vector corresponding to the second target content information based on the plurality of features in the second target content information corresponding to the first feature category grouping, and to generate a second content representation vector corresponding to the second target content information based on the plurality of features in the second target content information corresponding to the second feature category grouping.

[0216] In a possible implementation, the multiple content information in the content information set correspond to the same multiple feature types, and the apparatus further includes a second grouping unit, a sixth generating unit, and a seventh generating unit:

[0217] The second grouping unit is configured to divide the plurality of feature categories into a first feature category group and a second feature category group based on similarities between the feature categories;

[0218] the sixth generating unit is configured to generate, by means of the first vector generating portion in the initial content ranking model, a third content representation vector corresponding to the target content information based on a plurality of features in the target content information that are grouped according to the first feature category, and to generate a fourth content representation vector corresponding to the target content information based on a plurality of features in the target content information that are grouped according to the second feature category;

[0219] the seventh generating unit is configured to generate a third loss function based on a third similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector, and a fourth similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the third loss function decreases as the third similarity increases and decreases as the fourth similarity decreases;

[0220] The parameter adjustment unit 906 is specifically used to:

[0221] The initial content ranking model is adjusted according to the first loss function, the second loss function, and the third loss function.

[0222] In a possible implementation, the parameter adjustment unit 906 is specifically configured to:

[0223] Determine a first weight parameter corresponding to the first loss function and a second weight parameter corresponding to the third loss function, where the first weight parameter is used to identify the degree of influence of the first loss function on the parameter adjustment, and the second weight parameter is used to identify the degree of influence of the third loss function on the parameter adjustment;

[0224] Generate a comprehensive loss function according to the first loss function, the first weight parameter, the second loss function, the third loss function and the second weight parameter;

[0225] The parameters of the initial content ranking model are adjusted according to the comprehensive loss function.

[0226] Based on the model application method provided in the above embodiment, the present application embodiment also embodies a model application device, see Figure 10 , Figure 10 This is a structural block diagram of a model application device 1000 provided in an embodiment of the present application. The device 1000 is applied to a content ranking model trained by any of the above embodiments. The device 1000 includes a first acquisition unit 1001, a ranking unit 1002, and a push unit 1003:

[0227] The first acquisition unit 1001 is configured to acquire sorting content information corresponding to a plurality of content to be sorted, wherein the sorting content information has the same feature type as the content information in the content information set, and the content information set is used to train the content sorting model;

[0228] The sorting unit 1002 is configured to determine the content sorting corresponding to the plurality of to-be-sorted contents according to the sorting content information using the content sorting model;

[0229] The pushing unit 1003 is configured to push content to the object to be pushed according to the content ranking.

[0230] In a possible implementation, the apparatus 1000 further includes a second acquiring unit:

[0231] The second acquiring unit is configured to acquire object information corresponding to the object to be pushed;

[0232] The sorting unit 1002 is specifically configured to:

[0233] The content ranking corresponding to the plurality of to-be-ranked contents is determined by using the content ranking model according to the ranking content information and the object information.

[0234] The present application also provides a computer device, which is described below with reference to the accompanying drawings. Figure 11As shown, the embodiment of the present application provides a device, which may also be a terminal device. The terminal device may be any intelligent terminal including a mobile phone, a tablet computer, a personal digital assistant (PDA), a point of sales (POS), an in-vehicle computer, etc., taking a mobile phone as an example:

[0235] Figure 11 The block diagram shows a partial structure of a mobile phone related to the terminal device provided in the embodiment of the present application. Figure 11 The mobile phone includes components such as a radio frequency (RF) circuit 710, a memory 720, an input unit 730, a display unit 740, a sensor 750, an audio circuit 760, a wireless fidelity (WiFi) module 770, a processor 780, and a power supply 790. Those skilled in the art will appreciate that Figure 11 The mobile phone structure shown in the figure does not constitute a limitation to the mobile phone, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0236] The following combination Figure 11 A detailed introduction to the various components of a mobile phone:

[0237] The RF circuit 710 can be used to receive and send signals during information transmission or calls. In particular, after receiving the downlink information from the base station, it is sent to the processor 780 for processing; in addition, the designed uplink data is sent to the base station. Generally, the RF circuit 710 includes but is not limited to an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier (LNA), a duplexer, etc. In addition, the RF circuit 710 can also communicate with the network and other devices through wireless communication. The above-mentioned wireless communication can use any communication standard or protocol, including but not limited to Global System of Mobile communication (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, Short Messaging Service (SMS), etc.

[0238] The memory 720 can be used to store software programs and modules. The processor 780 executes the various functional applications and data processing of the mobile phone by running the software programs and modules stored in the memory 720. The memory 720 can mainly include a program storage area and a data storage area. The program storage area can store an operating system and at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created based on the use of the mobile phone (such as audio data, a phone book, etc.). In addition, the memory 720 can include high-speed random access memory and non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0239] The input unit 730 can be used to receive input digital or character information, and to generate key signal input related to the user settings and function control of the mobile phone. Specifically, the input unit 730 may include a touch panel 731 and other input devices 732. The touch panel 731, also known as a touch screen, can collect user touch operations on or near it (such as operations performed by the user using any suitable object or accessory such as a finger, stylus, etc. on or near the touch panel 731) and drive the corresponding connection device according to a pre-set program. Optionally, the touch panel 731 may include two parts: a touch detection device and a touch controller. Among them, the touch detection device detects the user's touch direction and detects the signal caused by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection device and converts it into touch point coordinates, which are then sent to the processor 780. It can also receive commands sent by the processor 780 and execute them. In addition, the touch panel 731 can be implemented using various types such as resistive, capacitive, infrared and surface acoustic wave. In addition to the touch panel 731, the input unit 730 may further include other input devices 732. Specifically, the other input devices 732 may include, but are not limited to, one or more of a physical keyboard, function keys (such as volume control keys, switch keys, etc.), a trackball, a mouse, and a joystick.

[0240] The display unit 740 can be used to display information input by the user or information provided to the user and various menus of the mobile phone. The display unit 740 may include a display panel 741. Optionally, the display panel 741 may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), etc. Further, the touch panel 731 may cover the display panel 741. When the touch panel 731 detects a touch operation on or near it, it is transmitted to the processor 780 to determine the type of touch event. Subsequently, the processor 780 provides corresponding visual output on the display panel 741 according to the type of touch event. Although in Figure 11 In the embodiment, the touch panel 731 and the display panel 741 are used as two independent components to realize the input and output functions of the mobile phone, but in some embodiments, the touch panel 731 and the display panel 741 can be integrated to realize the input and output functions of the mobile phone.

[0241] The mobile phone may also include at least one sensor 750, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor may include an ambient light sensor and a proximity sensor, wherein the ambient light sensor may adjust the brightness of the display panel 741 according to the brightness of the ambient light, and the proximity sensor may turn off the display panel 741 and / or the backlight when the mobile phone is moved to the ear. As a type of motion sensor, the accelerometer sensor can detect the magnitude of acceleration in all directions (generally three axes), and can detect the magnitude and direction of gravity when stationary. It can be used for applications that identify the posture of the mobile phone (such as horizontal and vertical screen switching, related games, magnetometer posture calibration), vibration recognition related functions (such as pedometer, tapping), etc.; as for other sensors that the mobile phone can also be configured with, such as gyroscopes, barometers, hygrometers, thermometers, infrared sensors, etc., they will not be described here.

[0242] Audio circuit 760, speaker 761, and microphone 762 provide an audio interface between the user and the phone. Audio circuit 760 converts received audio data into electrical signals and transmits them to speaker 761, which then converts them into sound signals for output. Microphone 762, on the other hand, converts collected sound signals into electrical signals, which are then received by audio circuit 760 and converted into audio data. The audio data is then processed by processor 780 and transmitted to, for example, another phone via RF circuit 710, or stored in memory 720 for further processing.

[0243] WiFi is a short-range wireless transmission technology. The mobile phone can help users send and receive emails, browse the web and access streaming media through the WiFi module 770. It provides users with wireless broadband Internet access. Figure 11A WiFi module 770 is shown, but it is understandable that it is not an essential component of the mobile phone and can be omitted as needed without changing the essence of the invention.

[0244] Processor 780 is the control center of the phone, connecting all parts of the phone using various interfaces and circuits. By running or executing software programs and / or modules stored in memory 720 and accessing data stored in memory 720, it performs various phone functions and processes data, thereby performing overall phone testing. Optionally, processor 780 may include one or more processing units; preferably, processor 780 may integrate an application processor and a modem processor, where the application processor primarily handles the operating system, user interface, and application programs, while the modem processor primarily handles wireless communications. It is understood that the modem processor may not be integrated into processor 780.

[0245] The mobile phone also includes a power supply 790 (such as a battery) for supplying power to various components. Preferably, the power supply can be logically connected to the processor 780 through a power management system, thereby managing charging, discharging, and power consumption management functions through the power management system.

[0246] Although not shown, the mobile phone may also include a camera, a Bluetooth module, etc., which will not be described in detail here.

[0247] In this embodiment, the processor 780 included in the terminal device further has the following functions:

[0248] Acquire a content information set, the content information set including content information corresponding to a plurality of sample contents, the plurality of sample contents having corresponding sample content orderings, the content information including fixed-value features and non-fixed-value features;

[0249] Taking each of the plurality of content information as target content information, generating a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information through the first vector generation part of the initial content ranking model, and generating a second content representation vector corresponding to the target content information based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set;

[0250] generating a first loss function based on a first similarity between a first content representation vector corresponding to each piece of content information and a second content representation vector corresponding to each piece of content information and a second content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the first loss function decreases as the first similarity increases and decreases as the second similarity decreases;

[0251] Determining, by means of the initial content ranking model, pending content rankings corresponding to the plurality of sample contents according to first content representation vectors respectively corresponding to the plurality of content information in the content information set;

[0252] generating a second loss function according to a difference between the pending content ranking and the sample content ranking;

[0253] The parameters of the initial content ranking model are adjusted according to the first loss function and the second loss function to obtain a content ranking model, and the content ranking model is used to perform content ranking on a plurality of contents to be ranked.

[0254] Alternatively, the processor 780 included in the terminal device further has the following functions:

[0255] Obtaining sorting content information corresponding to a plurality of to-be-sorted contents, wherein the sorting content information has the same feature type as the content information in the content information set, and the content information set is used to train the content sorting model;

[0256] Determining, by means of the content sorting model, content sorting corresponding to the plurality of to-be-sorted contents according to the sorting content information;

[0257] Content is pushed to the object to be pushed according to the content sorting.

[0258] This application embodiment also provides a server, see Figure 12 As shown, Figure 12The server 800 provided by the embodiments of the present application can have a large difference due to different configurations or performances, and can include one or more central processing units (CPUs) 822 (for example, one or more processors) and a memory 832, one or more storage media 830 (for example, one or more mass storage devices) storing application programs 842 or data 844. The memory 832 and the storage media 830 can be temporary storage or persistent storage. The programs stored in the storage media 830 can include one or more modules (not shown in the figure), and each module can include a series of instruction operations in the server. Further, the central processing unit 822 can be configured to communicate with the storage media 830 and execute the series of instruction operations in the storage media 830 on the server 800.

[0259] The server 800 can also include one or more power supplies 826, one or more wired or wireless network interfaces 850, one or more input / output interfaces 858, and / or one or more operating systems 841, such as Windows Server TM , Mac OS X TM , Unix TM , Linux TM , FreeBSD TM , and the like.

[0260] The steps performed by the server in the above embodiments can be based on the server structure shown in the above embodiments. Figure 12

[0261] The embodiments of the present application also provide a computer readable storage medium for storing a computer program, and the computer program is used to execute any one of the model training methods or model application methods described in the above embodiments.

[0262] Those skilled in the art can understand that all or part of the steps of the above-mentioned method embodiments can be completed by program instruction related hardware, and the above-mentioned program can be stored in a computer readable storage medium, and the program is executed to perform the steps of the above-mentioned method embodiments; and the above-mentioned storage medium can be at least one of the following media: read-only memory (English: read-only memory, abbreviation: ROM), RAM, magnetic disk or optical disk and various media that can store program codes.

[0263] ​It should be noted that the various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments. The device and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the scheme of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0264] The above is merely one specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A model training method, characterized in that: The method comprises: Acquire a content information set, the content information set including content information corresponding to a plurality of sample contents, the plurality of sample contents having corresponding sample content orderings, the content information including fixed-value features and non-fixed-value features; Taking each of the plurality of content information as target content information, generating a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information through the first vector generation part of the initial content ranking model, and generating a second content representation vector corresponding to the target content information based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set; generating a first loss function based on a first similarity between a first content representation vector corresponding to each piece of content information and a second content representation vector corresponding to each piece of content information and a second content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the first loss function decreases as the first similarity increases and decreases as the second similarity decreases; Determining, by means of the initial content ranking model, pending content rankings corresponding to the plurality of sample contents according to first content representation vectors respectively corresponding to the plurality of content information in the content information set; generating a second loss function according to a difference between the pending content ranking and the sample content ranking; The parameters of the initial content ranking model are adjusted according to the first loss function and the second loss function to obtain a content ranking model, and the content ranking model is used to perform content ranking on a plurality of contents to be ranked.

2. The method according to claim 1, characterized in that The step of generating a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information by a first vector generating portion in the initial content ranking model, and generating a second content representation vector corresponding to the target content information based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set, includes: Generate, by means of a second vector generating portion in the initial content ranking model, a first feature vector corresponding to the fixed-value feature in the target content information, a second feature vector corresponding to the non-fixed-value feature, and a third feature vector corresponding to the default value of the non-fixed-value feature; The first vector generating part in the initial content ranking model generates the first content representation vector according to the first feature vector and the second feature vector, and generates the second content representation vector according to the first feature vector and the third feature vector.

3. The method according to claim 2, characterized in that Determining the pending content rankings corresponding to the plurality of sample contents according to the content information set using the initial content ranking model includes: Determining, by means of a vector ranking portion in the initial content ranking model, pending content rankings corresponding to the plurality of sample contents based on first content representation vectors respectively corresponding to the plurality of content information in the content information set; The adjusting parameters of the initial content ranking model according to the first loss function and the second loss function to obtain a content ranking model includes: The first vector generation part, the second vector generation part, and the vector sorting part in the initial content sorting model are adjusted according to the first loss function and the second loss function to obtain the content sorting model.

4. The method according to claim 2, characterized in that The second vector generating portion includes a first feature input layer and a second feature input layer, the first feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a second feature vector corresponding to a non-fixed-value feature, and the second feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a third feature vector corresponding to a default value of the non-fixed-value feature; The first vector generating part includes a first fully connected layer and a second fully connected layer, the first fully connected layer is used to generate the first content representation vector according to the first feature vector and the second feature vector generated by the first feature input layer, and the second fully connected layer is used to generate the second content representation vector according to the first feature vector and the third feature vector generated by the second feature input layer; The method further comprises: In the process of adjusting the parameters of the initial content ranking model, the parameters of the first feature input layer and the second feature input layer are adjusted synchronously, and the parameters of the first fully connected layer and the second fully connected layer are adjusted synchronously.

5. The method according to claim 4, characterized in that The second fully connected layer includes multiple vector fusion layers, each vector fusion layer includes multiple fusion nodes, and the fusion nodes are used to fuse and output vectors output by fusion nodes in the previous vector fusion layer. The method further includes: Determining a target fusion node in the multi-layer vector fusion layer; Generating the second content representation vector according to the first feature vector and the third feature vector includes: The second content representation vector is generated according to the first feature vector and the third feature vector by removing the second fully connected layer of the target fusion node.

6. The method according to claim 1, characterized in that The method further comprises: Determining object information corresponding to each of the plurality of content information in the content information set, wherein the object information is used to identify the object corresponding to the content information; Determining the pending content rankings corresponding to the plurality of sample contents by using the initial content ranking model and according to first content representation vectors respectively corresponding to the plurality of content information in the content information set includes: The undetermined content rankings corresponding to the plurality of sample contents are determined by using the initial content ranking model and according to the first content representation vectors and object information respectively corresponding to the plurality of content information in the content information set.

7. The method according to claim 1, characterized in that The content information set includes content information belonging to a first information portion and content information belonging to a second information portion, and generating a second content representation vector corresponding to the target content information based on fixed-value features in the target content information and default values ​​of non-fixed-value features corresponding to the content information set includes: In response to the target content information being content information belonging to the first information portion, generating a second content representation vector corresponding to the target content information based on fixed-value features in the target content information and default values ​​of non-fixed-value features corresponding to the content information set; The method further comprises: In response to the target content information being content information belonging to the second information portion, a second content representation vector corresponding to the target content information is generated based on fixed-value features and non-fixed-value features in the target content information.

8. The method according to claim 1, characterized in that The content information set includes content information corresponding to a first grouping method and content information corresponding to a second grouping method, multiple content information in the content information set correspond to the same multiple feature types, and the multiple content information are respectively used as target content information, and a first vector generation part in an initial content ranking model is used to generate a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information, and a second content representation vector corresponding to the target content information is generated based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set, including: Using the plurality of content information corresponding to the first grouping method as first target content information, generating a first content representation vector corresponding to the first target content information based on fixed-value features and non-fixed-value features in the first target content information through the first vector generation part of the initial content ranking model, and generating a second content representation vector corresponding to the first target content information based on the fixed-value features in the first target content information and default values ​​of the non-fixed-value features corresponding to the content information set; The method further comprises: dividing the plurality of feature categories into a first feature category group and a second feature category group based on similarities between the feature categories; The multiple content information corresponding to the second grouping method are respectively used as second target content information. Through the first vector generation part, the first content representation vector corresponding to the second target content information is generated based on the multiple features in the second target content information corresponding to the first feature category grouping, and the second content representation vector corresponding to the second target content information is generated based on the multiple features in the second target content information corresponding to the second feature category grouping.

9. The method according to claim 1, characterized in that The plurality of content information in the content information set corresponds to the same plurality of feature types, and the method further includes: dividing the plurality of feature categories into a first feature category group and a second feature category group based on similarities between the feature categories; generating, by a first vector generating portion of an initial content ranking model, a third content representation vector corresponding to the target content information based on a plurality of features in the target content information that are grouped according to the first feature category, and generating a fourth content representation vector corresponding to the target content information based on a plurality of features in the target content information that are grouped according to the second feature category; generating a third loss function based on a third similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector, and a fourth similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the third loss function decreases as the third similarity increases and decreases as the fourth similarity decreases; The adjusting parameters of the initial content ranking model according to the first loss function and the second loss function includes: The initial content ranking model is adjusted according to the first loss function, the second loss function, and the third loss function.

10. The method according to claim 9, characterized in that The adjusting parameters of the initial content ranking model according to the first loss function, the second loss function, and the third loss function includes: Determine a first weight parameter corresponding to the first loss function and a second weight parameter corresponding to the third loss function, where the first weight parameter is used to identify the degree of influence of the first loss function on the parameter adjustment, and the second weight parameter is used to identify the degree of influence of the third loss function on the parameter adjustment; Generate a comprehensive loss function according to the first loss function, the first weight parameter, the second loss function, the third loss function and the second weight parameter; The parameters of the initial content ranking model are adjusted according to the comprehensive loss function.

11. A model application method, characterized in that: The method is applied to a content ranking model obtained by training according to any one of claims 1 to 10, and the method comprises: Obtaining sorting content information corresponding to a plurality of to-be-sorted contents, wherein the sorting content information has the same feature type as the content information in the content information set, and the content information set is used to train the content sorting model; Determining, by means of the content sorting model, content sorting corresponding to the plurality of to-be-sorted contents according to the sorting content information; Content is pushed to the object to be pushed according to the content sorting.

12. The method according to claim 11, characterized in that The method further comprises: Obtaining object information corresponding to the object to be pushed; Determining the content rankings corresponding to the plurality of to-be-ranked contents according to the ranking content information using the content ranking model includes: The content ranking corresponding to the plurality of to-be-ranked contents is determined by using the content ranking model according to the ranking content information and the object information.

13. A model training device, characterized in that: The device includes an acquisition unit, a first generation unit, a second generation unit, a first determination unit, a third generation unit, and a parameter adjustment unit: The acquiring unit is configured to acquire a content information set, wherein the content information set includes content information corresponding to a plurality of sample contents, the plurality of sample contents have corresponding sample content orderings, and the content information includes fixed-value features and non-fixed-value features; The first generating unit is configured to use the plurality of content information as target content information, generate a first content representation vector corresponding to the target content information based on fixed-value features and non-fixed-value features in the target content information, and generate a second content representation vector corresponding to the target content information based on the fixed-value features in the target content information and default values ​​of the non-fixed-value features corresponding to the content information set, using the first vector generating unit in the initial content ranking model. The second generating unit is configured to generate a first loss function based on a first similarity between a first content representation vector corresponding to each piece of content information and a second content representation vector corresponding to each piece of content information and a second content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the first loss function decreases as the first similarity increases and decreases as the second similarity decreases; The first determining unit is configured to determine, by using the initial content ranking model, the pending content rankings corresponding to the plurality of sample contents according to first content representation vectors respectively corresponding to the plurality of content information in the content information set; The third generating unit is configured to generate a second loss function according to a difference between the pending content ranking and the sample content ranking; The parameter adjustment unit is used to adjust the parameters of the initial content ranking model according to the first loss function and the second loss function to obtain a content ranking model, and the content ranking model is used to sort the multiple contents to be sorted.

14. The device according to claim 13, characterized in that The first generating unit is specifically configured to: Generate, by means of a second vector generating portion in the initial content ranking model, a first feature vector corresponding to the fixed-value feature in the target content information, a second feature vector corresponding to the non-fixed-value feature, and a third feature vector corresponding to the default value of the non-fixed-value feature; The first vector generating part in the initial content ranking model generates the first content representation vector according to the first feature vector and the second feature vector, and generates the second content representation vector according to the first feature vector and the third feature vector.

15. The device according to claim 14, characterized in that The first determining unit is specifically configured to: Determining, by means of a vector ranking portion in the initial content ranking model, pending content rankings corresponding to the plurality of sample contents based on first content representation vectors respectively corresponding to the plurality of content information in the content information set; The parameter adjustment unit is specifically used for: The first vector generation part, the second vector generation part, and the vector sorting part in the initial content sorting model are adjusted according to the first loss function and the second loss function to obtain the content sorting model.

16. The device according to claim 14, characterized in that The second vector generating portion includes a first feature input layer and a second feature input layer, the first feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a second feature vector corresponding to a non-fixed-value feature, and the second feature input layer being used to generate a first feature vector corresponding to a fixed-value feature in the target content information and a third feature vector corresponding to a default value of the non-fixed-value feature; The first vector generating part includes a first fully connected layer and a second fully connected layer, the first fully connected layer is used to generate the first content representation vector according to the first feature vector and the second feature vector generated by the first feature input layer, and the second fully connected layer is used to generate the second content representation vector according to the first feature vector and the third feature vector generated by the second feature input layer; The device further comprises a synchronization unit: The synchronization unit is used to synchronously adjust the parameters of the first feature input layer and the second feature input layer, and synchronously adjust the parameters of the first fully connected layer and the second fully connected layer during the process of adjusting the parameters of the initial content sorting model.

17. The device according to claim 16, characterized in that The second fully connected layer includes multiple vector fusion layers, each vector fusion layer includes multiple fusion nodes, and the fusion nodes are used to fuse and output vectors output by the fusion nodes in the previous vector fusion layer. The device also includes a second determining unit: The second determining unit is used to determine a target fusion node in the multi-layer vector fusion layer; The first generating unit is specifically configured to: The second content representation vector is generated according to the first feature vector and the third feature vector by removing the second fully connected layer of the target fusion node.

18. The device according to claim 13, characterized in that The apparatus further includes a third determining unit: The third determining unit is configured to determine object information corresponding to each of the plurality of content information in the content information set, wherein the object information is used to identify the object corresponding to the content information; The first determining unit is specifically configured to: The undetermined content rankings corresponding to the plurality of sample contents are determined by using the initial content ranking model and according to the first content representation vectors and object information respectively corresponding to the plurality of content information in the content information set.

19. The device according to claim 13, characterized in that The content information set includes content information belonging to a first information part and content information belonging to a second information part, and the first generating unit is specifically configured to: In response to the target content information being content information belonging to the first information portion, generating a second content representation vector corresponding to the target content information based on fixed-value features in the target content information and default values ​​of non-fixed-value features corresponding to the content information set; The apparatus further includes a fourth generating unit: In response to the target content information being content information belonging to the second information portion, a second content representation vector corresponding to the target content information is generated based on fixed-value features and non-fixed-value features in the target content information.

20. The device according to claim 13, wherein The content information set includes content information corresponding to the first grouping method and content information corresponding to the second grouping method. Multiple content information in the content information set correspond to the same multiple feature types. The first generating unit is specifically configured to: Using the plurality of content information corresponding to the first grouping method as first target content information, generating a first content representation vector corresponding to the first target content information based on fixed-value features and non-fixed-value features in the first target content information through the first vector generation part of the initial content ranking model, and generating a second content representation vector corresponding to the first target content information based on the fixed-value features in the first target content information and default values ​​of the non-fixed-value features corresponding to the content information set; The apparatus further includes a first grouping unit and a fifth generating unit: The first grouping unit is configured to divide the plurality of feature categories into a first feature category group and a second feature category group based on similarities between the feature categories; The fifth generating unit is configured to use the plurality of content information corresponding to the second grouping method as second target content information, and generate, through the first vector generating part, a first content representation vector corresponding to the second target content information based on the plurality of features in the second target content information corresponding to the first feature category grouping, and to generate a second content representation vector corresponding to the second target content information based on the plurality of features in the second target content information corresponding to the second feature category grouping.

21. The device according to claim 13, characterized in that The plurality of content information in the content information set corresponds to the same plurality of feature types, and the apparatus further includes a second grouping unit, a sixth generating unit, and a seventh generating unit: The second grouping unit is configured to divide the plurality of feature categories into a first feature category group and a second feature category group based on similarities between the feature categories; the sixth generating unit is configured to generate, by means of the first vector generating portion in the initial content ranking model, a third content representation vector corresponding to the target content information based on a plurality of features in the target content information that are grouped according to the first feature category, and to generate a fourth content representation vector corresponding to the target content information based on a plurality of features in the target content information that are grouped according to the second feature category; the seventh generating unit is configured to generate a third loss function based on a third similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector, and a fourth similarity between the third content representation vector corresponding to each piece of content information and the fourth content representation vector corresponding to content information having different sample content from the plurality of pieces of content information, wherein the third loss function decreases as the third similarity increases and decreases as the fourth similarity decreases; The parameter adjustment unit is specifically used for: The initial content ranking model is adjusted according to the first loss function, the second loss function, and the third loss function.

22. The device according to claim 21, characterized in that The parameter adjustment unit is specifically used for: Determine a first weight parameter corresponding to the first loss function and a second weight parameter corresponding to the third loss function, where the first weight parameter is used to identify the degree of influence of the first loss function on the parameter adjustment, and the second weight parameter is used to identify the degree of influence of the third loss function on the parameter adjustment; Generate a comprehensive loss function according to the first loss function, the first weight parameter, the second loss function, the third loss function and the second weight parameter; The parameters of the initial content ranking model are adjusted according to the comprehensive loss function.

23. A model application device, characterized in that: The device is applied to a content ranking model obtained by training according to any one of claims 1 to 10, and includes a first acquisition unit, a ranking unit, and a push unit: The first acquisition unit is configured to acquire sorting content information corresponding to a plurality of to-be-sorted contents, wherein the sorting content information has the same feature type as the content information in the content information set, and the content information set is used to train the content sorting model; The sorting unit is configured to determine the content sorting corresponding to the plurality of to-be-sorted contents according to the sorting content information using the content sorting model; The pushing unit is used to push content to the object to be pushed according to the content ranking.

24. The device according to claim 23, characterized in that The device further includes a second acquiring unit: The second acquiring unit is configured to acquire object information corresponding to the object to be pushed; The sorting unit is specifically used for: The content ranking corresponding to the plurality of to-be-ranked contents is determined by using the content ranking model according to the ranking content information and the object information.

25. A computer device, characterized in that: The computer device includes a processor and a memory: The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the model training method described in any one of claims 1-10, or the model application method described in any one of claims 11-12 according to the instructions in the program code.

26. A computer-readable storage medium, characterized in that The computer-readable storage medium is used to store a computer program, which is used to execute the model training method described in any one of claims 1-10, or the model application method described in any one of claims 11-12.

27. A computer program product comprising instructions, which, when run on a computer, enables the computer to execute the model training method described in any one of claims 1 to 10, or the model application method described in any one of claims 11 to 12.

Images