An information recommendation method, apparatus, electronic device, and readable storage medium

Abstract

This application provides an information recommendation method, apparatus, electronic device, and readable storage medium. The method includes: extracting features from multiple pieces of information to obtain general features and bias features for each piece of information, wherein the general feature of one piece of information is used to characterize the information content of the information, and the bias feature of one piece of information is used to characterize the display format of the information; inputting the general features and bias features of each piece of information into a pre-trained information recommendation model to determine the recommendation score for each piece of information; and obtaining target information to be recommended to the user based on the recommendation scores of each piece of information. This application fully considers the impact of the information display format on user preferences, more realistically reflects user preferences for information, reduces the deviation between target information and information that users are truly interested in, and can recommend information to users in a targeted manner, thereby enhancing the user experience.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to an information recommendation method, apparatus, electronic device, and readable storage medium. Background Technology

[0002] Currently, information providers design various display formats for their content to increase user engagement. For example, advertisers diversify the display format (e.g., size, color, brightness, title, placement) of their ads to enhance user attention. Similarly, a non-profit organization might use diverse display formats to attract users' attention to an event's information. However, each user has different preferences and needs, leading to varying degrees of favorability towards different display formats. Therefore, information display platforms need to select and display information in the most appealing formats based on user preferences.

[0003] Taking advertising as an example, existing multi-task prediction models for advertising primarily employ algorithms that mine user behavior characteristics and inherent attributes of advertisements to select and display the most attractive ads. For instance, based on the mined user behavior characteristics and inherent attributes of each ad, metrics such as Click-Through-Rate (CTR), Conversion Rate (CVR), and Gross Merchandise Volume (GMV) are determined. By evaluating these metrics, the most attractive ads are identified and displayed.

[0004] However, in related technologies, when the display format of advertisements is diversified, the display format factor is ignored, resulting in low accuracy of various predicted indicators such as click-through rate (CTR), conversion rate (CVR), and price. Therefore, it is impossible to overcome the influence of the display format of advertisements on user preferences. Summary of the Invention

[0005] This application provides an information recommendation method, apparatus, electronic device, and readable storage medium. By analyzing the display format of information, it can accurately obtain the user's preference for the display format, and select and display the information with the most attractive display format according to the user's preference, effectively enhancing the user experience.

[0006] A first aspect of this application provides an information recommendation method, the method comprising:

[0007] Feature extraction is performed on multiple pieces of information to obtain the general features and bias features of each piece of information. The general features of one piece of information are used to characterize the information content of the information, and the bias features of one piece of information are used to characterize the display format of the information.

[0008] The general features and bias features of each of the multiple pieces of information are input into a pre-trained information recommendation model to determine the recommendation score of each of the multiple pieces of information.

[0009] Based on the recommendation scores of the various pieces of information, target information to be recommended to the user is obtained;

[0010] The information recommendation model includes a preset task model, a bias model, and a fusion module. The preset task model is trained using the general features of multiple sample information and the labels corresponding to the preset task carried by each of the multiple sample information as input. The bias model is trained using the bias features of the multiple sample information and the labels corresponding to the preset task carried by each of the multiple sample information as input. The fusion module is used to fuse the general index value output by the preset task model and the bias coefficient output by the bias model, and output the recommendation score for each piece of information.

[0011] Optionally, the bias feature of an information includes multiple bias features of the information in different dimensions; the bias model includes multiple bias expert networks, and a bias expert network is obtained by training a first preset network model with the bias features of the same dimension of the multiple sample information and the bias factor labels of the same dimension that are pre-annotated to the multiple sample information respectively.

[0012] Optionally, the bias feature of an information includes multiple bias features of the information in different dimensions; the preset task model includes multiple sub-task networks; the bias model includes multiple gate switch networks, multiple bias expert networks and multiple fusion sub-modules, with one gate switch network and one fusion sub-module corresponding to one sub-task network;

[0013] A fusion submodule is used to fuse the bias factors output by multiple bias expert networks based on the weights of multiple different dimensions output by a gate switch network, and output the bias coefficients of each piece of information.

[0014] Optionally, the weights of the same dimension output by each of the multiple gate switch networks are different.

[0015] Optionally, the preset task model includes multiple sub-task networks, and the multiple sub-tasks represented by the multiple sub-task networks are interconnected; each sub-task network is trained by taking the general features of multiple sample information and the labels corresponding to the sub-tasks carried by the multiple sample information as input;

[0016] A gate switch network corresponding to a subtask network is obtained by training a second preset network model using the bias features of each dimension of the multiple sample information and the weight labels of each dimension of the subtask corresponding to the multiple sample information as training samples.

[0017] Optionally, the multiple subtasks represented by the multiple subtask network are any combination of the following: click-through rate (CTR), conversion rate (CVR), total sales volume (GMV), unit price (Price), and click-through rate (CTCVR).

[0018] Optionally, the bias characteristics of information include at least one of the following: the display location of the information, the display format of the information, the display channel of the information, whether the information contains an image, and the size of the image in the information.

[0019] A second aspect of this application provides an information recommendation device, the device comprising:

[0020] The extraction module is used to extract features from multiple pieces of information to obtain the general features and bias features of each piece of information. The general features of one piece of information are used to characterize the information content of the information, and the bias features of one piece of information are used to characterize the display format of the information.

[0021] The input module is used to input the general features and bias features of each of the multiple pieces of information into a pre-trained information recommendation model to determine the recommendation score of each of the multiple pieces of information;

[0022] The acquisition module is used to obtain target information to be recommended to the user based on the recommendation scores of the multiple pieces of information;

[0023] The information recommendation model includes a preset task model, a bias model, and a fusion module. The preset task model is trained using the general features of multiple sample information and the labels corresponding to the preset task carried by each of the multiple sample information as input. The bias model is trained using the bias features of the multiple sample information and the labels corresponding to the preset task carried by each of the multiple sample information as input. The fusion module is used to fuse the general index value output by the preset task model and the bias coefficient output by the bias model, and output the recommendation score for each piece of information.

[0024] Optionally, the bias feature of an information includes multiple bias features of the information in different dimensions; the bias model includes multiple bias expert networks, and a bias expert network is obtained by training a first preset network model with the bias features of the same dimension of the multiple sample information and the bias factor labels of the same dimension that are pre-annotated to the multiple sample information respectively.

[0025] Optionally, the bias feature of an information includes multiple bias features of the information in different dimensions; the preset task model includes multiple sub-task networks; the bias model includes multiple gate switch networks, multiple bias expert networks and multiple fusion sub-modules, with one gate switch network and one fusion sub-module corresponding to one sub-task network;

[0026] A fusion submodule is used to fuse the bias factors output by multiple bias expert networks based on the weights of multiple different dimensions output by a gate switch network, and output the bias coefficients of each piece of information.

[0027] Optionally, the weights of the same dimension output by each of the multiple gate switch networks are different.

[0028] Optionally, the preset task model includes multiple sub-task networks, and the multiple sub-tasks represented by the multiple sub-task networks are interconnected; each sub-task network is trained by taking the general features of multiple sample information and the labels corresponding to the sub-tasks carried by the multiple sample information as input;

[0029] A gate switch network corresponding to a subtask network is obtained by training a second preset network model using the bias features of each dimension of the multiple sample information and the weight labels of each dimension of the subtask corresponding to the multiple sample information as training samples.

[0030] Optionally, the multiple subtasks represented by the multiple subtask network are any combination of the following: click-through rate (CTR), conversion rate (CVR), total sales volume (GMV), unit price (Price), and click-through rate (CTCVR).

[0031] Optionally, the bias characteristics of information include at least one of the following: the display location of the information, the style of the information, the display channel of the information, whether the information contains an image, and the size of the image in the information.

[0032] A third aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the information recommendation method as described in the first aspect of this application.

[0033] A fourth aspect of this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the steps of the information recommendation method described in the first aspect of this application.

[0034] The information recommendation method of this embodiment first extracts features from multiple pieces of information to obtain general features and bias features for each piece of information. The general feature of one piece of information represents its content, and the bias feature of another piece of information represents its presentation format. Then, the general features and bias features of each piece of information are input into a pre-trained information recommendation model to determine the recommendation score for each piece of information. Finally, based on the recommendation scores of each piece of information, the target information to be recommended to the user is obtained. The information recommendation method of this application has the following several beneficial effects:

[0035] First, a bias model and a fusion module were set up on the basis of the traditional multi-task prediction model to obtain an information recommendation model. The bias coefficient output by the bias model was used to correct the general index values ​​output by the preset task model, thereby improving the accuracy of the various index values ​​output by the information recommendation model.

[0036] Second, by using the information presentation format as a bias feature input into the bias model, the influence of the presentation format (quantified as a bias coefficient) is used to correct the general index value output by the preset task model. This fully considers the impact of the information presentation format on user preferences, more accurately reflects user preferences for information, reduces the deviation between target information and information that users are truly interested in, and enhances the user experience. Attached Figure Description

[0037] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of a conventional advertising multi-task prediction model in related technologies;

[0039] Figure 2 This is a schematic diagram illustrating an implementation environment according to an embodiment of this application;

[0040] Figure 3 This is a flowchart illustrating an information recommendation method according to an embodiment of this application;

[0041] Figure 4This is a schematic diagram illustrating an information recommendation model according to an embodiment of this application;

[0042] Figure 5 This is a schematic diagram illustrating yet another information recommendation model according to an embodiment of this application;

[0043] Figure 6 This is a structural block diagram of an information recommendation device provided in an embodiment of this application;

[0044] Figure 7 This is a schematic diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0045] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0046] The information referred to in this application includes various types, such as advertisements placed by advertisers, activity information placed by social organizations, and policies issued by the government. This application does not specifically limit the type of information. It should be noted that, for the convenience of describing the following embodiments, the information referred to in this application below will all use advertisements as an example. When the type of information is other, the recommendation principle is the same as that of advertisements.

[0047] Taking advertising as an example, the process of selecting and displaying advertisements through a multi-task prediction model in related technologies is as follows: Figure 1 As shown. Figure 1 This is a schematic diagram of a conventional advertising multi-task prediction model in related technologies.

[0048] exist Figure 1 The model uses two different subtask networks, Subtask Network 1 and Subtask Network 2, to handle different metrics, such as Click-Through Rate (CTR), Conversion Rate (CVR), Gross Merchandise Volume (GMV), and Price. Of course, multiple other subtask networks can be set up in the multi-task prediction model. Figure 1 This example only illustrates the scenario with two sub-task networks. When an ad display platform selects and displays a target ad from multiple ads to be displayed, it inputs the common characteristics of each ad into each sub-task network to obtain the metric values ​​output by each sub-task network. Then, by analyzing each metric value, the target ad is obtained. For example, the target ad could be the ad with the highest conversion rate (CVR) or the ad with the highest total transaction value (GMV).

[0049] Among them, general features are features that do not consider the display format, and can include continuous features and categorical features. Continuous features are generally exponential value features, while categorical features are features that are enumerable or have a finite number of categories.

[0050] However, since the general features do not take into account the display format of the advertisement, the accuracy of the indicator values ​​output by the subtask network is low and cannot truly reflect the user's preference for the advertisement. This results in a large deviation between the identified target advertisement and the advertisement that the user is actually interested in, which greatly affects the user experience.

[0051] To overcome the problems existing in related technologies, this application proposes the following technical concept: using multiple bias networks to correct the traditional multi-task prediction model, and based on the corrected model, simultaneously considering the general characteristics and presentation characteristics of information, so that the model can comprehensively analyze the general characteristics and presentation characteristics, and more accurately predict the values ​​of various indicators, thereby obtaining target information that better reflects the user's true preferences.

[0052] Figure 2 This is a schematic diagram illustrating an implementation environment according to an embodiment of this application. Figure 2 In this process, after obtaining multiple pieces of information to be displayed, the information display platform inputs them into an information recommendation model to obtain target information, which is then input into the information display layer for display. The information recommendation model can be set up within the information display platform or on other third-party platforms. When set up on a third-party platform, the information display platform obtains the target information output by the information recommendation model through communication and interaction with that third-party platform.

[0053] The information recommendation method proposed in this application will be described in detail below. This method is applied to... Figure 2 The information display platform shown. Figure 3 This is a flowchart illustrating an information recommendation method according to an embodiment of this application. (Refer to...) Figure 3 The information recommendation method of this application may include the following steps:

[0054] Step S11: Perform feature extraction on multiple pieces of information to obtain the general features and bias features of each piece of information. The general features of one piece of information are used to characterize the information content of the information, and the bias features of one piece of information are used to characterize the display format of the information.

[0055] In this embodiment, a general feature represents the content of the information. For the text portion of the information, the content is the text; for the image portion, the content is the content displayed in the image; for the audio portion, the content is the content corresponding to the audio; for the video portion, the content is the content corresponding to the video, and so on. A general feature of information can be one or a combination of text, images, audio, video, etc.

[0056] For example, when the text portion of the information contains phrases like "End-of-Season Big Sale," the information display platform performs text recognition to obtain the content "End-of-Season Big Sale." When the image portion of the information contains pictures of various seafood, the platform performs image recognition to obtain the names of the seafood, such as "fish," "shrimp," and "shellfish," resulting in the content "fish, shrimp, shellfish." When the audio portion of the information contains voice content related to various furniture items, the platform converts the audio into text, such as "dining table," "sofa," and "cabinet," resulting in the content "dining table, sofa, cabinet." When the video portion of the information contains voice content related to various clothing items, the platform performs image recognition to obtain the names of the clothing items, such as "hat," "scarf," and "glasses," resulting in the content "hat, scarf, glasses."

[0057] In this embodiment, the bias feature represents the display format of information. The bias feature of information includes at least one of the following: the display position of the information, the style of the information, the display channel of the information, whether the information contains an image, and the size of the image in the information.

[0058] In this embodiment, "information" primarily refers to online information. The display location of the information mainly refers to its online display location, such as the center of the screen, the top of the screen, or the bottom right corner of the screen. The style of the information mainly refers to its design style, such as cartoon, retro, minimalist, or commercial. The display channel of the information mainly refers to online display channels, such as news media, social media, and OTV (Online TV). Of course, bias features can also include many other factors, such as the brightness of the information, whether it contains slogans, whether it contains titles, the color of the information, and the arrangement of the information; this embodiment does not impose specific limitations on these.

[0059] In this embodiment, the information display platform can also add image processing functions, such as rotation, cropping, mirroring, and brightness adjustment, to better identify the features of images in the information. The information display platform can also add text processing functions, such as using NLP (Neuro-Linguistic Programming) algorithms (including but not limited to LDA (Latent Dirichlet Allocation), word2vec (word to vector), and fasttext (fast text classification) methods to extract features from titles, slogans, etc. in the information. The functionality of the information display platform can be enriched according to actual needs in this application.

[0060] Step S12: Input the general features and bias features of each of the multiple pieces of information into the pre-trained information recommendation model to determine the recommendation score of each of the multiple pieces of information.

[0061] In one implementation, the general features and bias features of each of the plurality of information items are input into a pre-trained information recommendation model to determine the recommendation score of each of the plurality of information items, including:

[0062] The general features and bias features of each of the multiple pieces of information are input into a pre-trained information recommendation model to obtain preset index values ​​for each piece of information.

[0063] For each piece of information, a corresponding recommendation score is determined based on its preset indicator value.

[0064] The information recommendation model can set any number of indicators according to actual needs, so each piece of information can have one or more indicators.

[0065] In practice, the score corresponding to each indicator value can be determined based on the range in which each indicator value falls, and the sum of the scores corresponding to all indicator values ​​for each piece of information can be used as the recommendation score for that information.

[0066] For example, when the metrics set in the information recommendation model include Click-Through Rate (CTR), Conversion Rate (CVR), and Gross Merchandise Volume (GMV), and multiple pieces of information including information X and information Y are used as an example, to obtain the recommendation score for information X, we can first obtain the score G1 corresponding to the CTR, the score G2 corresponding to the CVR, and the score G3 corresponding to the GMV. The sum of G1, G2, and G3 (G1+G2+G3) is taken as the recommendation score for information X. Similarly, to obtain the recommendation score for information Y, we can first obtain the score G1' corresponding to the CTR, the score G2' corresponding to the CVR, and the score G3' corresponding to the GMV. The sum of G1', G2', and G3' (G1'+G2'+G3') is taken as the recommendation score for information Y.

[0067] Generally, indicator values ​​and scores are directly proportional; the higher the indicator value, the higher the corresponding score, and consequently, the higher the sum of all indicator values. Therefore, a higher recommendation score indicates greater effectiveness of the information, such as a higher conversion rate (CVR) or higher gross merchandise volume (GMV).

[0068] In this embodiment, the information recommendation model includes a preset task model, a bias model, and a fusion module. The preset task model is trained using the general features of multiple sample information and the labels corresponding to the preset tasks carried by each of the multiple sample information as input. The bias model is trained using the bias features of the multiple sample information and the labels corresponding to the preset tasks carried by each of the multiple sample information as input. The fusion module is used to fuse the general index value output by the preset task model and the bias coefficient output by the bias model, and output the recommendation score for each piece of information.

[0069] In this embodiment, each preset task corresponds to a metric. For example, when the preset task is click-through rate (CTR), the corresponding metric is CTR; when the preset task is conversion rate (CVR), the corresponding metric is conversion rate (CVR); and when the preset task is total sales (GMV), the corresponding metric is total sales volume (GMV). The tags corresponding to the preset tasks refer to the user's behavioral characteristics in response to the input information. For example, when the preset task is CTR, the corresponding tag is "clicked" or "not clicked," and when the preset task is conversion rate, the corresponding tag is "converted" or "not converted." The user's behavioral characteristics in response to the input information can be obtained by analyzing the user's historical behavioral data.

[0070] In this system, a pre-defined network model is trained using the general features of multiple sample information pieces and the labels corresponding to the pre-defined tasks carried by each sample information piece. This trains the model to obtain a pre-defined task model. When used, the general features are input into the pre-defined task model to obtain general index values. Similarly, a bias model is trained using the bias features of multiple sample information pieces and the labels corresponding to the pre-defined tasks carried by each sample information piece. This trains the model to obtain bias coefficients.

[0071] In this embodiment, the purpose of setting the bias model is to correct the output value of the preset task model. The bias model fully considers the bias characteristics of the information and corrects the general index value by the output bias coefficient, so as to obtain index values ​​with higher accuracy, namely the preset index values ​​mentioned later.

[0072] It is understandable that the process of correcting the general index value through the bias coefficient, that is, the process of the fusion module fusing the general index value and the bias coefficient, includes: multiplying the bias coefficient with the general index value.

[0073] Figure 4 This is a schematic diagram illustrating an information recommendation model according to an embodiment of this application. Combined with... Figure 4 In this embodiment, inputting the general features of information X into a preset task model yields a general index value, and inputting the bias features of information X into a bias model yields a bias coefficient. The fusion module uses both the general index value and the bias coefficient to obtain a preset index value, and the preset index value to obtain a recommendation score for information X. For example, the process of obtaining a recommendation score through the information recommendation model can be shown in the following table:

[0074]

[0075] Step S13: Based on the recommendation scores of the multiple pieces of information, obtain the target information to be recommended to the user.

[0076] In this embodiment, the information with the highest recommendation score among multiple pieces of information can be used as the target information. When multiple pieces of information have the same recommendation score, any one of them can be selected as the target information. Alternatively, the information with the same recommendation score can be re-evaluated according to other criteria, such as whether it conforms to current popular trends or whether the information delivery provider is a member, in order to obtain the target information.

[0077] Taking a food delivery platform as an example, this platform offers a user-defined food delivery app. The app's homepage features an ad box at the top, periodically displaying different ads. Therefore, the platform needs to periodically select target ads and display them in this ad box, for example, periodically selecting representative ads from different categories. Within a certain period, the platform extracts features from numerous merchants' ads within a specified category, obtaining general and bias features for each ad. Then, the general features of each ad are input into the pre-defined task model of the information recommendation model to obtain general index values. The bias features are input into the bias model of the information recommendation model to obtain bias coefficients. Based on the general index values ​​and bias coefficients, a high-precision pre-defined index value is obtained for each ad. Next, a recommendation score is obtained for each ad based on these pre-defined index values. Finally, the ad with the highest recommendation score is selected as the target ad to be displayed in the ad box of the food delivery app.

[0078] The information recommendation method of this embodiment first extracts features from multiple pieces of information to obtain general features and bias features for each piece of information. The general feature of one piece of information represents its content, and the bias feature of another piece of information represents its presentation format. Then, the general features and bias features of each piece of information are input into a pre-trained information recommendation model to determine the recommendation score for each piece of information. Finally, based on the recommendation scores of each piece of information, the target information to be recommended to the user is obtained. The information recommendation method of this application has the following several beneficial effects:

[0079] First, a bias model and a fusion module were set up on the basis of the traditional multi-task prediction model to obtain an information recommendation model. The bias coefficient output by the bias model was used to correct the general index values ​​output by the preset task model, thereby improving the accuracy of the various index values ​​output by the information recommendation model.

[0080] Second, by using the information presentation format as a bias feature input into the bias model, the influence of the presentation format (quantified as a bias coefficient) is used to correct the general index value output by the preset task model. This fully considers the impact of the information presentation format on user preferences, more accurately reflects user preferences for information, reduces the deviation between target information and information that users are truly interested in, and enhances the user experience.

[0081] In conjunction with the above embodiments, in one implementation, the bias feature of an information includes multiple bias features of the information in different dimensions; the bias model includes multiple bias expert networks, and a bias expert network is obtained by training a first preset network model using the bias features of the same dimension of the multiple sample information and the bias factor labels of the same dimension that are pre-annotated to the multiple sample information as training samples.

[0082] In this embodiment, for a piece of information, bias features can be obtained from multiple dimensions, such as the display location of the information, the style of the information, the display channel of the information, whether the information contains an image, and the size of the image in the information.

[0083] When training the bias expert network, the sample data used includes: bias features of the same dimension of multiple sample information, and bias factor labels of that dimension pre-annotated for each of the multiple sample information. The bias features of the same dimension of multiple sample information can be, for example, the display position of ad 1, the display position of ad 2, the display position of ad 3, ..., the display position of ad N. Its network structure can be, but is not limited to, an MLP (Multi-Layer Perceptron), and this embodiment does not impose any restrictions on this.

[0084] In this embodiment, the bias factor can be understood as a bias feature in vector form, and the bias factor label is used to represent the meaning of the bias feature in the corresponding dimension. For example, when the dimension of the bias feature is display location, the corresponding bias factor label is "display location"; as another example, when the dimension of the bias feature is display channel, the corresponding bias factor label is "display channel". The bias factor label can be represented by text information or by other symbols with preset meanings, such as Arabic numerals or letters, etc. This application does not impose specific limitations on this.

[0085] In conjunction with the above embodiments, in one implementation, the bias feature of an information includes multiple bias features of the information in different dimensions; the preset task model includes multiple sub-task networks; the bias model includes multiple gate switch networks, multiple bias expert networks and multiple fusion sub-modules, with one gate switch network and one fusion sub-module corresponding to one sub-task network.

[0086] A fusion submodule is used to fuse the bias factors output by multiple bias expert networks based on the weights of multiple different dimensions output by a gate switch network, and output the bias coefficients of each piece of information.

[0087] Figure 5 This is a schematic diagram illustrating yet another information recommendation model according to an embodiment of this application. Figure 5In this model, the bias model is a model composed of modules marked with shaded areas. The preset task model is a model composed of subtask network 1 and subtask network 2. The bias model includes multiple gate switch networks (e.g., gate switch network 1 and gate switch network 2), multiple bias expert networks (e.g., bias expert network 1-bias expert network 3), and multiple fusion sub-modules. The fusion sub-module includes an input layer and a bias network (e.g., input layer 1 and bias network 1 form a fusion sub-module 1, and input layer 2 and bias network 2 form a fusion sub-module 2). Figure 5 Only the input layer and bias network are shown in this example. In this embodiment, the information recommendation model can be configured with multiple subtasks. Figure 5 The example shown only illustrates the case with two subtasks.

[0088] The training process of the bias network is as follows: the cross-output results of the bias expert network and the gate switch network are used as training samples to train the preset network. The cross-output results can be obtained through operations such as product, subtraction, and weighted summation, and their dimensionality can be linearly related to the size of the vector output by the bias expert network. For example, the cross-output vectors of multiple bias expert networks can be concatenated to obtain the cross-output results, or the similarity of the vectors output by multiple bias expert networks can be calculated to obtain the cross-output results. The preset network can be, but is not limited to, an MLP (Multi-Layer Perceptron); this embodiment does not impose any restrictions on this.

[0089] In this embodiment, each subtask network has a corresponding bias network (for example, subtask network 1 corresponds to bias network 1, and subtask network 2 corresponds to bias network 2). A gate switch network corresponds to a fusion submodule and a subtask network (for example, gate switch network 1 corresponds to fusion submodule 1 and subtask network 1, and gate switch network 2 corresponds to fusion submodule 2 and subtask network 2).

[0090] In this embodiment, the gate switch network is used to control the weights of multiple different dimensions output by the bias expert network. For example, gate switch network 1 is used to control the weights of each bias expert network in the expert layer for each dimension of the bias features input by the fusion submodule 1. The fusion submodule is used to fuse the weights of multiple different dimensions output by the gate switch network and the bias factors output by each of the multiple bias expert networks to output the bias coefficients of each piece of information.

[0091] In this embodiment, the number of bias expert networks and the number of network layers can be adjusted as needed to meet business requirements. Weight learning in the gate switch network can employ multi-layer fully connected layers or an attenuation mechanism; this embodiment does not impose any restrictions on this approach.

[0092] In one implementation, the multiple subtasks represented by the multiple subtask network are any combination of the following: click-through rate (CTR), conversion rate (CVR), total sales volume (GMV), unit price (Price), and click-through rate (CTCVR).

[0093] In this embodiment, a subtask network is used to execute a subtask, and a subtask corresponds to a metric and a metric value. For example, subtask network 1 is used to obtain the click-through rate (CTR), and subtask network 2 is used to obtain the conversion rate (CVR).

[0094] In this embodiment, a bias expert network is used to transform the input bias features into a vector representation. The bias factor can be understood as a bias feature in vector form. One bias expert network is used to process the bias features of the same dimension.

[0095] In this embodiment, the gate switch network is used to set the weights of bias features in each dimension for each subtask. Specifically, the gate switch network is used to set weight values ​​for each bias expert network. For example, when... Figure 5 When the dimensions of the bias features corresponding to the bias expert networks 1-3 are the display position of the information, the style of the information, and whether the information contains an image, respectively, the door switch network 1 can set the weights of subtask 1 corresponding to subtask network 1 as {w1, w2, w3}, which means that the influence weight of the display position of the information on subtask 1 is w1, the influence weight of the style of the information on subtask 1 is w2, and the influence weight of whether the information contains an image on subtask 3 is w3.

[0096] The fusion submodule fuses the weights from multiple dimensions of the door switch network output with the bias factors output by each of the multiple bias expert networks. Specifically, it multiplies the weights output by the door switch network with the vector-like bias features output by each bias expert network to obtain a weighted vector-like bias feature. For example, in... Figure 5 In the input layer 1, the bias features in vector form input by bias expert networks 1-3 are {T1, T2, T3}, and the weights of different dimensions output by gate switch network 1 to input layer 1 are {w1, w2, w3}. Therefore, the result obtained after fusion by fusion submodule 1 is {w1*T1, w2*T2, w3*T3}, corresponding to... Figure 5 In the equation {t1, t2, t3}, t1, t2, and t3 are weighted vector-form bias features, respectively. Similarly, after fusing the vector-form bias features output to input layer 2 by bias expert networks 1-3, and the weights of different dimensions output to input layer 2 by gate switch network 2, fusion submodule 2 can obtain {t4, t5, t6}.

[0097] In one implementation, based on the above embodiments, the weights of the same dimension output by each of the multiple door switch networks are different.

[0098] In this embodiment, the door switch network can also set different weight values ​​for bias features of the same dimension. For example, in... Figure 5 In the diagram, the weights input from door switch network 1 to input layer 1 are {w1, w2, w3}, and the weights input from door switch network 2 to input layer 2 are {w1', w2', w3'}. The dimensions corresponding to bias expert networks 1-3 are display position, display style, and display channel, respectively. Therefore, door switch networks w1 and w1' can be different, w2 and w2' can be different, and w3 and w3' can be different.

[0099] In this embodiment, when there is only one subtask, a gate-switching network is not required. When there are multiple subtasks, a gate-switching network can be set for each subtask to improve the accuracy of its output metric values. When none of the subtasks in multiple subtasks have a gate-switching network set, the bias features of all dimensions have the same influence weight on all subtasks.

[0100] In one embodiment, based on the above embodiments, the preset task model includes multiple sub-task networks, and the multiple sub-tasks represented by the multiple sub-task networks are interconnected; each sub-task network is trained by taking the general features of multiple sample information and the labels corresponding to the sub-tasks carried by the multiple sample information as input.

[0101] A gate switch network corresponding to a subtask network is obtained by training a second preset network model using the bias features of each dimension of the multiple sample information and the weight labels of each dimension of the subtask corresponding to the multiple sample information as training samples.

[0102] When training the subtask network, the sample data used includes: the general features of each of the multiple sample pieces of information, and the labels corresponding to the subtasks carried by each of the multiple sample pieces of information. The labels corresponding to the subtasks have the same meaning as the labels corresponding to the preset tasks mentioned above, both referring to the user's behavioral characteristics in response to the input information. For example, when the subtask is click-through rate, the corresponding label is "click" or "not clicked," and when the subtask is conversion rate, the corresponding label is "converted" or "not converted." The network structure can use MLP, mainly adjusted according to the actual needs of each business, and the structure can differ between multiple subtasks.

[0103] When training the gate switch network, the sample data used includes: bias features of each dimension of multiple sample information, and pre-labeled weights for each dimension of the sub-tasks corresponding to the multiple sample information. The gate switch network can be an MLP structure, or it can be a fixed value or a simple rule design. It is recommended that the number of layers not be too deep and the vector output dimension not be too large.

[0104] In this embodiment, the door switch network can be trained end-to-end by the system, which can automatically learn the weights of different bias expert networks under different subtasks. It can also support users to manually add rules based on the learned weights, such as manually setting the weights for a certain bias expert network.

[0105] The following will combine Figure 5 The information recommendation method of this application will be described in detail with a specific embodiment.

[0106] exist Figure 5 In this model, assuming the bias features include display location, display style, and display channel, bias expert network 1 is used to convert the display location into a vector form, resulting in T1; bias expert network 2 is used to convert the display style into a vector form, resulting in T2; and bias expert network 3 is used to convert the display channel into a vector form, resulting in T3. Therefore, the values ​​input to input layer 1 by bias expert networks 1-3 are {T1, T2, T3}. Door switch network 1 sets the weights for bias expert network 1 to w1, bias expert network 2 to w2, and bias expert network 3 to w3. Therefore, the values ​​input to input layer 1 by door switch network 1 are {w1, w2, w3}. Input layer 1 multiplies w1 and T1, w2 and T2, and w3 and T3, resulting in {w1*T1, w2*T2, w3*T3}. Figure 5 The input layer 1 inputs {t1, t2, t3} into the bias network 1 to obtain a bias coefficient Z1. The fusion module multiplies the bias coefficient Z1 with the general index value output by the subtask network 1 to correct the general index value using the bias coefficient Z1, thus obtaining the preset index value 1.

[0107] Simultaneously, the values ​​input to input layer 2 by bias expert networks 1-3 are {T1, T2, T3}. Door switch network 2 sets the weights for bias expert network 1 to w4, bias expert network 2 to w5, and bias expert network 3 to w6. Therefore, the values ​​input to input layer 2 by door switch network 2 are {w4, w5, w6}. Input layer 2 multiplies w4 by T1, w5 by T2, and w6 by T3, obtaining {w4*T1, w5*T2, w6*T3}, corresponding to... Figure 5The input layer 2 inputs {t4, t5, t6} into the bias network 2, obtaining a bias coefficient Z2. The fusion module multiplies the bias coefficient Z2 with the general index value output by the subtask network 2 to correct the general index value using the bias coefficient Z2, thus obtaining the preset index value 2.

[0108] Finally, the fusion module obtains the recommended score of the information based on the preset index value 1 and preset index value 2, and further obtains the target information based on the recommended score of each piece of information (as described above).

[0109] In this embodiment, the multiple subtasks represented by the multiple subtask network can be interconnected. For example, the total transaction value (GMV) can be obtained through the click-through rate (CTR), conversion rate (CVR), and price of the subtasks. The correlation is: Preset indicator value (GMV) = Preset indicator value (CTR) * Preset indicator value (CVR) * Preset indicator value (Price). Specifically, Preset indicator value (GMV) = Bias coefficient (CTR) * General indicator value (CTR) * Bias coefficient (CVR) * General indicator value (CVR) * Bias coefficient (Price) * General indicator value (Price).

[0110] The information recommendation method in this application has the following beneficial effects:

[0111] In this application, for information presented in diverse formats, different bias networks are added to each sub-task network in the mainstream multi-task network. By learning the bias networks through bias expert networks and gate switch networks, the user's preferences and the impact of information presentation formats on the user can be accurately learned.

[0112] Second, in this application, bias networks and subtask networks are trained simultaneously, which can both obtain users' preferences for display formats and learn users' preferences for products.

[0113] Third, in this application, a gate switch network is introduced into the information recommendation model, which can better control the fitting accuracy of the bias network in each subtask.

[0114] Fourth, in this application, the biased expert network structure and number in the information recommendation model can be flexibly configured to better meet the user's business needs and enhance the user experience.

[0115] Based on the same inventive concept, one embodiment of this application provides an information recommendation device 600. (Reference) Figure 6 , Figure 6 This is a structural block diagram of an information recommendation device provided in an embodiment of this application. For example... Figure 6 As shown, the information recommendation device 600 includes:

[0116] The extraction module 601 is used to extract features from multiple pieces of information to obtain the general features and bias features of each piece of information. The general feature of one piece of information is used to characterize the information content of the information, and the bias feature of one piece of information is used to characterize the display format of the information.

[0117] Input module 602 is used to input the general features and bias features of each of the multiple pieces of information into a pre-trained information recommendation model to determine the recommendation score of each of the multiple pieces of information;

[0118] The module 603 is used to obtain target information to be recommended to the user based on the recommendation scores of the multiple pieces of information.

[0119] The information recommendation model includes a preset task model, a bias model, and a fusion module. The preset task model is trained using the general features of multiple sample information and the labels corresponding to the preset task carried by each of the multiple sample information as input. The bias model is trained using the bias features of the multiple sample information and the labels corresponding to the preset task carried by each of the multiple sample information as input. The fusion module is used to fuse the general index value output by the preset task model and the bias coefficient output by the bias model, and output the recommendation score for each piece of information.

[0120] Optionally, the bias feature of an information includes multiple bias features of the information in different dimensions; the bias model includes multiple bias expert networks, and a bias expert network is obtained by training a first preset network model with the bias features of the same dimension of the multiple sample information and the bias factor labels of the same dimension that are pre-annotated to the multiple sample information respectively.

[0121] Optionally, the bias feature of an information includes multiple bias features of the information in different dimensions; the preset task model includes multiple sub-task networks; the bias model includes multiple gate switch networks, multiple bias expert networks and multiple fusion sub-modules, with one gate switch network and one fusion sub-module corresponding to one sub-task network;

[0122] A fusion submodule is used to fuse the bias factors output by multiple bias expert networks based on the weights of multiple different dimensions output by a gate switch network, and output the bias coefficients of each piece of information.

[0123] Optionally, the weights of the same dimension output by each of the multiple gate switch networks are different.

[0124] Optionally, the preset task model includes multiple sub-task networks, and the multiple sub-tasks represented by the multiple sub-task networks are interconnected; each sub-task network is trained by taking the general features of multiple sample information and the labels corresponding to the sub-tasks carried by the multiple sample information as input;

[0125] A gate switch network corresponding to a subtask network is obtained by training a second preset network model using the bias features of each dimension of the multiple sample information and the weight labels of each dimension of the subtask corresponding to the multiple sample information as training samples.

[0126] Optionally, the multiple subtasks represented by the multiple subtask network are any combination of the following: click-through rate (CTR), conversion rate (CVR), total sales volume (GMV), unit price (Price), and click-through rate (CTCVR).

[0127] Optionally, the bias characteristics of information include at least one of the following: the display location of the information, the style of the information, the display channel of the information, whether the information contains an image, and the size of the image in the information.

[0128] Based on the same inventive concept, another embodiment of this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the information recommendation method as described in any of the above embodiments of this application.

[0129] Based on the same inventive concept, another embodiment of this application provides an electronic device 700, such as... Figure 7 As shown. Figure 7 This is a schematic diagram of an electronic device according to an embodiment of this application. The electronic device includes a memory 702, a processor 701, and a computer program stored in the memory and executable on the processor. When executed by the processor, the program implements the steps of the information recommendation method described in any of the above embodiments of this application.

[0130] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0131] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0132] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, embodiments of this application can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of this application can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0133] This application describes embodiments with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0134] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0135] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0136] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0137] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0138] The above provides a detailed description of the information recommendation method, apparatus, storage medium, and electronic device provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An information recommendation method, characterized in that, include: Feature extraction is performed on multiple pieces of information to obtain the general features and bias features of each piece of information. The general features of one piece of information are used to characterize the information content of the information, and the bias features of one piece of information are used to characterize the display format of the information. The general features and bias features of each of the multiple pieces of information are input into a pre-trained information recommendation model to determine the recommendation score of each of the multiple pieces of information. Based on the recommendation scores of the various pieces of information, target information to be recommended to the user is obtained; The information recommendation model includes a preset task model, a bias model, and a fusion module. The preset task model is trained using the general features of multiple sample information and the labels corresponding to the preset tasks carried by each of the multiple sample information as input. The bias model is trained using the bias features of multiple sample information and the labels corresponding to the preset tasks carried by each of the multiple sample information as input. The fusion module is used to fuse the general index value output by the preset task model and the bias coefficient output by the bias model, and output the recommendation scores of each of the multiple information. The bias feature of an information includes multiple bias features of the information in different dimensions; the bias model includes multiple bias expert networks, and a bias expert network is trained on a first preset network model using the bias features of the same dimension of the multiple sample information and the bias factor labels of the same dimension that are pre-annotated to the multiple sample information as training samples. The preset task model includes multiple sub-task networks; the bias model includes multiple door switch networks, multiple bias expert networks, and multiple fusion sub-modules, with one door switch network and one fusion sub-module corresponding to one sub-task network. A fusion submodule is used to fuse the bias factors output by multiple bias expert networks based on the weights of multiple different dimensions output by a gate switch network, and output the bias coefficients of each piece of information. The bias characteristics of information include: the display location of the information, the style of the information, the display channel of the information, whether the information contains images, and the size of the images in the information.

2. The method according to claim 1, characterized in that, The weights of the same dimension output by each of the multiple gate switch networks are different.

3. The method according to claim 2, characterized in that, The preset task model includes multiple sub-task networks, and the multiple sub-tasks represented by the multiple sub-task networks are interconnected; each sub-task network is trained by taking the general features of multiple sample information and the labels corresponding to the sub-tasks carried by the multiple sample information as input. A gate switch network corresponding to a subtask network is obtained by training a second preset network model using the bias features of each dimension of the multiple sample information and the weight labels of each dimension of the subtask corresponding to the multiple sample information as training samples.

4. The method according to claim 3, characterized in that, The multiple subtasks represented by the network are any combination of the following: click-through rate (CTR), conversion rate (CVR), total sales volume (GMV), unit price (Price), and click-through rate (CTCVR).

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the information recommendation method as described in any one of claims 1-4.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes, it implements the steps in the information recommendation method as described in any one of claims 1-4.

Images