Recommendation method, device and equipment for input method and medium

Abstract

The application discloses a recommendation method, device and equipment for an input method and a medium. The recommendation method comprises the following steps: obtaining multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to the candidate recommendation functions; obtaining the click probability of each candidate recommendation function and the click probability of each candidate recommendation content by using a prediction model; determining the joint click probability between each candidate recommendation function and the corresponding candidate recommendation content based on the click probability of each candidate recommendation function and the click probability of each candidate recommendation content; and determining the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function based on the joint click probability. In this way, the function and the content corresponding to the function that the user is interested in can be accurately recommended, so that the click rate of the user can be improved.

Description

Technical Field

[0001] This invention relates to the field of input method technology, and in particular to a recommended method, apparatus, device, and medium for input methods. Background Technology

[0002] Input methods are an essential application for terminal devices. With the continuous development of input method technology, in addition to inputting text or symbols, input methods are also equipped with other different functions to meet users' needs for shopping, social interaction, and entertainment.

[0003] However, when an input method has a large number of configured functions, users may find it difficult to locate the functions they are interested in and their corresponding content within the input method interface. This can negatively impact click-through rates. Therefore, accurately recommending relevant functions and their corresponding content within the input method is a pressing issue that needs to be addressed. Summary of the Invention

[0004] The main technical problem solved by this invention is to provide a recommendation method, apparatus, device, and medium for input methods, which can accurately recommend functions that users are interested in and the corresponding content, thereby improving the user's click-through rate.

[0005] To address the aforementioned technical problems, this application provides a recommendation method for input methods, comprising: acquiring multiple candidate recommendation functions and multiple candidate recommendation contents corresponding to each candidate recommendation function in the input method; predicting the click probability of each candidate recommendation function and the click probability of each candidate recommendation content using a prediction model, wherein the click probability of a candidate recommendation content represents the likelihood of that the candidate recommendation content being clicked among the multiple candidate recommendation contents of its respective candidate recommendation function; determining the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content based on the click probabilities of each candidate recommendation function and the click probabilities of each candidate recommendation content; and determining the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function based on the joint click probability.

[0006] The process of using a prediction model to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content includes: using the function prediction network of the prediction model to predict the click probability of each candidate recommendation function based on the function reference parameters of each candidate recommendation function; and using the content prediction network of the prediction model to predict the click probability of each candidate recommendation content based on the content reference parameters of each candidate recommendation content.

[0007] Among them, the functional reference parameters of each candidate recommendation function include at least one historical functional feature in the user's historical click records, the candidate functional features corresponding to each candidate recommendation function, and functional auxiliary features. The functional auxiliary features include at least one of user features and context features. The context features include at least one current state information of the device running the input method.

[0008] The content reference parameters for each candidate recommended content include at least one historical content feature from the user's historical click records, the candidate content feature corresponding to each candidate recommended content, and content auxiliary features. The content auxiliary features include at least one of user features and context features.

[0009] The functional reference parameters for each candidate recommended function include at least one historical functional feature from the user's historical click records and the candidate functional features corresponding to each candidate recommended function. Based on the functional reference parameters of each candidate recommended function, the click probability of each candidate recommended function is predicted, including: for each candidate recommended function, obtaining the parameter feature vector corresponding to each functional reference parameter of the candidate recommended function; using the functional interest extraction layer of the functional prediction network, extracting the user's functional interest vector for each candidate recommended function based on the parameter feature vector corresponding to each historical functional feature and the parameter feature vector corresponding to each candidate functional feature; and using the first prediction layer of the functional prediction network, predicting the click probability of each candidate recommended function based on the functional interest vector corresponding to each candidate recommended function.

[0010] Specifically, based on the parameter feature vectors corresponding to each historical functional feature and the parameter feature vectors corresponding to each candidate functional feature, the user's functional interest vector for each candidate recommended function is extracted. This includes: for each candidate functional feature, attention processing is performed based on the parameter feature vectors corresponding to each historical functional feature and the parameter feature vectors corresponding to the candidate functional feature to obtain the attention weights of each historical functional feature with respect to the candidate functional feature. The attention weights of each historical functional feature represent the degree of correlation between each historical functional feature and the candidate functional feature. Using the attention weights of each historical functional feature with respect to the candidate functional feature, the parameter feature vectors corresponding to each historical functional feature are weighted to obtain the functional interest vectors corresponding to the candidate functional features. And / or, the functional reference parameters of the candidate recommended function also include functional auxiliary features. Based on the functional interest vectors corresponding to each candidate recommended function, the click probability of each candidate recommended function is predicted. This includes: for each candidate recommended function, the click probability of the candidate recommended function is predicted based on the functional interest vectors corresponding to the candidate recommended function, the parameter feature vectors corresponding to each candidate functional feature, and the parameter feature vectors corresponding to the functional auxiliary features.

[0011] The content reference parameters for candidate recommended content include at least one historical content feature from the user's historical click records. Based on the content reference parameters of each candidate recommended content, the click probability of each candidate recommended content is predicted, including: for each candidate recommended content, obtaining the parameter feature vector corresponding to each content reference parameter of the candidate recommended content; using the content interest extraction layer of the content prediction network to extract the user's content interest vector based on the parameter feature vector corresponding to each historical content feature; and using the second prediction layer of the content prediction network to predict the click probability of each candidate recommended content based on the content interest vector.

[0012] Specifically, based on the parameter feature vectors corresponding to each historical content feature, the user's content interest vector is extracted, including: performing self-attention processing on the parameter feature vectors corresponding to each historical content feature to obtain the association vectors corresponding to each historical content feature; concatenating the association vectors corresponding to each historical content feature to obtain the content interest vector; and / or, the content reference parameters of the candidate recommended content also include content auxiliary features. Based on the content interest vector, the click probability of each candidate recommended content is predicted, including: for each candidate recommended content, the click probability of the candidate recommended content is predicted based on the content interest vector, the parameter feature vectors corresponding to each candidate recommended content feature, and the parameter feature vectors corresponding to the content auxiliary features.

[0013] The feature vectors of the functional reference parameters and the feature vectors of the content reference parameters are either extracted by the same feature extraction layer or by the feature extraction layers in their respective prediction networks.

[0014] Wherein, the joint click probability between the candidate recommendation function and the corresponding candidate recommendation content is the product of the click probability of the candidate recommendation function and the click probability of the corresponding candidate recommendation content; and / or, based on the click probability of each candidate recommendation function and the click probability of each candidate recommendation content, the joint click probability between each candidate recommendation function and the corresponding candidate recommendation content is determined, including: for each candidate recommendation content, using the output layer of the prediction model based on the click probability of the candidate recommendation content and the click probability of the candidate recommendation function to which it belongs, to obtain the joint click probability between the candidate recommendation content and the candidate recommendation function to which it belongs.

[0015] The recommendation method further includes: obtaining multiple candidate recommendation functions and multiple candidate recommendation contents corresponding to each candidate recommendation function, wherein at least one labeled recommendation function exists among the multiple candidate recommendation functions, and at least one labeled recommendation content exists among the multiple candidate recommendation contents corresponding to each candidate recommendation function; using a prediction model to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content; based on the click probabilities of each candidate recommendation function and each candidate recommendation content, determining the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content; based on the joint click probability, determining the target recommendation function and its corresponding target recommendation content; and adjusting the network parameters of the prediction model based on the differences between the target recommendation function and the labeled recommendation function, and the differences between the target recommendation content and the labeled recommendation content.

[0016] The steps of obtaining multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function are performed in response to the pop-up of the input keyboard of the input method or the user's selection operation of the preset icon of the input method; and / or, based on the joint click probability, determining the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function, including: determining the candidate recommendation functions and candidate recommendation contents whose joint click probabilities meet the recommendation requirements as the target recommendation function and target recommendation content.

[0017] To address the aforementioned technical problems, another technical solution adopted in this application is: providing a recommendation device for an input method, the device comprising: an acquisition module, configured to acquire multiple candidate recommendation functions and multiple candidate recommendation contents corresponding to each candidate recommendation function in the input method; a prediction module, configured to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content using a prediction model, wherein the click probability of a candidate recommendation content represents the likelihood of that the candidate recommendation content being clicked among the multiple candidate recommendation contents of its respective candidate recommendation function; a first determination module, configured to determine the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content based on the click probabilities of each candidate recommendation function and the click probabilities of each candidate recommendation content; and a second determination module, configured to determine the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function based on the joint click probability.

[0018] To solve the above-mentioned technical problems, another technical solution adopted in this application is: to provide a processing device, including a memory and a processor coupled to each other, wherein the memory stores program instructions; and the processor is used to execute the program instructions stored in the memory to implement the above-mentioned recommended method for input methods.

[0019] To solve the above-mentioned technical problems, another technical solution adopted in this application is to provide a computer-readable storage medium for storing program instructions that can be executed to implement the above-mentioned recommended method for input methods.

[0020] The above scheme utilizes a prediction model to obtain the click probability of each candidate recommended function and the click probability of each candidate recommended content. Furthermore, based on the joint click probability between the click probabilities of each candidate recommended function and its corresponding candidate recommended content, the target recommended function and its corresponding target recommended content for the input method are determined. Since the same prediction model is used to simultaneously predict the click probabilities of both candidate recommended functions and candidate recommended content, the resulting joint click probability can simultaneously express the user's functional and content interests. Therefore, the joint click probability can accurately determine the target recommended function and its corresponding target recommended content. This allows for accurate recommendations of functions and corresponding content that the user is interested in within the input method, thereby improving the click-through rate. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating an embodiment of the recommendation method for input methods provided in this application;

[0022] Figure 2 This is a schematic diagram of the framework of an embodiment of the prediction model provided in this application;

[0023] Figure 3 This is a flowchart illustrating another embodiment of the recommendation method for input methods provided in this application;

[0024] Figure 4 This is a schematic diagram of the input method display interface provided in this application;

[0025] Figure 5 This is a flowchart illustrating an embodiment of the training method for the prediction model provided in this application;

[0026] Figure 6 This is a schematic diagram of a framework of an embodiment of the recommendation device for input methods provided in this application;

[0027] Figure 7 This is a schematic diagram of the framework of an embodiment of the processing device provided in this application;

[0028] Figure 8 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium provided in this application. Detailed Implementation

[0029] To make the purpose, technical solution and effects of this application clearer and more explicit, the following describes this application in further detail with reference to the accompanying drawings and embodiments.

[0030] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this article means two or more. Moreover, the term "at least one" in this article means any combination of at least two of any one or more of a plurality of elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0031] Furthermore, if the embodiments of this application involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, features defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed in this application.

[0032] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the recommended method for an input method provided in this application. It should be noted that if substantially the same result is achieved, the method of this invention is not necessarily identical. Figure 1 The illustrated process sequence is limited. For example... Figure 1 As shown, the method includes the following steps:

[0033] S101: Obtain multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function.

[0034] In this embodiment, "multiple candidate recommendation functions" refers to two or more candidate recommendation functions, and "multiple candidate recommendation contents" refers to two or more candidate recommendation contents. For example, candidate recommendation functions may include meme battle functions, chat functions, fun chat functions, fancy text functions, and product functions. Candidate recommendation contents may be specific content corresponding to the aforementioned candidate recommendation functions, such as specific emoticons in the meme battle function or specific chat text in the chat function.

[0035] Specifically, all or some of the functions configured in the input method can be used as candidate recommended functions, and all or some of the content corresponding to each function can be used as candidate recommended content.

[0036] S102: Use the prediction model to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content.

[0037] The click probability of a candidate recommendation feature represents the likelihood that it will be clicked among multiple candidate recommendation features. The click probability of candidate recommended content represents the likelihood that it will be clicked among multiple candidate recommended content within its respective candidate recommendation feature.

[0038] In this embodiment, the function prediction network of the prediction model can be used to predict the click probability of each candidate recommendation function based on the function reference parameters of each candidate recommendation function. Similarly, the content prediction network of the prediction model can be used to predict the click probability of each candidate recommendation content based on the content reference parameters of each candidate recommendation content.

[0039] S103: Based on the click probability of each candidate recommendation function and the click probability of each candidate recommendation content, determine the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content.

[0040] Specifically, the product of the click probability of the candidate recommendation function and the click probability of the corresponding candidate recommendation content is taken as the joint click probability between the candidate recommendation function and the corresponding candidate recommendation content.

[0041] S104: Based on the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content, determine the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function.

[0042] Specifically, candidate recommended functions and candidate recommended content whose joint click probabilities meet the recommendation requirements are identified as target recommended functions and target recommended content. There can be one or more candidate recommended functions that meet the recommendation requirements, and correspondingly, there can be one or more candidate recommended content items for each candidate recommended function that meets the recommendation requirements. For example, the candidate recommended function and its corresponding candidate recommended content corresponding to the highest joint click probability can be used as the target recommended function and target recommended content. Another example is using multiple candidate recommended functions and their corresponding candidate recommended content corresponding to joint click probabilities greater than a probability threshold as the target recommended function and target recommended content.

[0043] In this embodiment, a prediction model is used to obtain the click probability of each candidate recommended function and the click probability of each candidate recommended content. Furthermore, based on the joint click probability between the click probabilities of each candidate recommended function and the corresponding click probabilities of each candidate recommended content, the target recommended function of the input method and the target recommended content corresponding to the target recommended function are determined. Since the same prediction model is used to simultaneously predict the click probabilities of both candidate recommended functions and candidate recommended content, the resulting joint click probability can simultaneously express the user's functional and content interests. Therefore, based on the joint click probability between the click probabilities of each candidate recommended function and the corresponding click probabilities of each candidate recommended content, the target recommended function and the target recommended content corresponding to the target recommended function can be accurately determined. This allows for accurate recommendation of functions that the user is interested in within the input method, as well as the corresponding content, thereby improving the user's click-through rate.

[0044] Please see Figure 2 , Figure 2 This is a schematic diagram of the framework of an embodiment of the prediction model provided in this application. Figure 2 As shown, the prediction model includes a function prediction network 10, a content prediction network 20, and an output layer 30. The function prediction network 10 includes a first input layer 11, a first feature extraction layer 12, a function interest extraction layer 13, and a first prediction layer 14. The content prediction network 20 includes a second input layer 21, a second feature extraction layer 22, a content interest extraction layer 23, and a second prediction layer 24.

[0045] It should be noted that, in this embodiment, the first input layer 11 and the second input layer 21 can be the same layer, that is, the function prediction network 10 and the content prediction network 20 share the same input layer; or, the first input layer 11 and the second input layer 21 can be two different layers corresponding to the function prediction network 10 and the content prediction network 20, respectively. The first feature extraction layer 12 and the second feature extraction layer 22 can be the same layer, that is, the function prediction network 10 and the content prediction network 20 can share the same feature extraction layer; or, the first feature extraction layer 12 and the second feature extraction layer 22 can be two different layers corresponding to the function prediction network 10 and the content prediction network 20, respectively. The first prediction layer 14 and the second prediction layer 24 can be the same layer, that is, the function prediction network 10 and the content prediction network 20 can share the same prediction layer; or, the first prediction layer 14 and the second prediction layer 24 can be two different layers corresponding to the function prediction network 10 and the content prediction network 20, respectively.

[0046] In this embodiment, the first input layer 11 is used to input the functional reference parameters of each candidate recommendation function. For each candidate recommendation function, the first feature extraction layer 12 performs feature extraction (embedding) on ​​the functional reference parameters of the candidate recommendation function to obtain the parameter feature vectors corresponding to each functional reference parameter of the candidate recommendation function. The functional interest extraction layer 13 extracts the user's functional interest vector for each candidate recommendation function based on the parameter feature vectors corresponding to each historical function feature and the parameter feature vectors corresponding to each candidate function feature. And, the first prediction layer 14 predicts the click probability of each candidate recommendation function based on the functional interest vectors corresponding to each candidate recommendation function. Simultaneously, the second input layer 21 is used to input the content reference parameters of each candidate recommendation content. For each candidate recommendation content, the second feature extraction layer 22 performs feature extraction on the content reference parameters of the candidate recommendation content to obtain the parameter feature vectors corresponding to each content reference parameter of the candidate recommendation content. The content interest extraction layer 23 extracts the user's content interest vector based on the parameter feature vectors corresponding to each historical content feature. And, the second prediction layer 24 predicts the click probability of each candidate recommendation content based on the content interest vector. Finally, by using the click probability of the candidate recommended content and the click probability of the candidate recommended function based on the click probability of the candidate recommended content in the output layer 30, the joint click probability between the candidate recommended content and the candidate recommended function is obtained.

[0047] The first input layer 11 and the second input layer 21, the first feature extraction layer 12 and the second feature extraction layer 22, the functional interest extraction layer 13 and the content interest extraction layer 23, the first prediction layer 14 and the second prediction layer 24 are obtained through synchronous training during model training.

[0048] In related technologies, a function prediction model is first used to predict the click probability of each candidate recommended function, and then the target recommended function is determined based on the click probabilities of each candidate recommended function. On this basis, a content prediction model is then used to predict the click probability of each candidate recommended content corresponding to the target recommended function, and finally the target recommended content corresponding to the target recommended function is determined based on the click probabilities of each candidate recommended content. This approach has two drawbacks. First, it requires building separate function and content prediction models. The data processing for these two models is asynchronous and data cannot be shared during model training and click probability prediction, resulting in low accuracy in predicting the target recommended function and content. Second, as the number of functions configured in the input method increases, the number of function and content prediction models also increases, increasing the maintenance cost for engineers.

[0049] In this embodiment, on the one hand, the data processing is synchronized and data can be shared during model training and click probability prediction, which can improve the accuracy of the predicted target recommendation functions and target recommendation content. On the other hand, only one prediction model is needed to simultaneously predict the click probability of each candidate recommendation function and the click probability of the corresponding candidate recommendation content, reducing the complexity of the prediction model and making it easier for engineers to optimize the prediction model, thereby reducing the maintenance cost for engineers.

[0050] Please see Figure 3 , Figure 3 This is a flowchart illustrating an embodiment of the recommendation method for input methods provided in this application. This method can be based on... Figure 2 The prediction model shown is implemented. Figure 3 As shown, the method includes the following steps:

[0051] S301: Obtain multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function.

[0052] The candidate recommendation function, candidate recommendation content, and related information on obtaining the candidate recommendation function and candidate recommendation content are described in step S101 above and will not be repeated here.

[0053] In one embodiment, step S301 is executed in response to the pop-up of the input keyboard of the input method. For example, when an input command is received, it is determined that the input keyboard of the input method will pop up, and then multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function are obtained.

[0054] In another embodiment, step S301 is executed in response to a user's selection of a preset icon in the input method. That is, it is determined whether recommendations are needed based on the user's individual needs. For example, when a user's selection instruction for a preset icon is received, it is determined that the user needs recommendations. In this case, multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function are acquired. When no user selection instruction for a preset icon is received, or a user's cancellation instruction for a preset icon is received, it is determined that the user does not need recommendations. In this case, multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function are not acquired.

[0055] S302: Utilize the predictive model's functional prediction network to predict the click probability of each candidate recommendation function based on the functional reference parameters of each candidate recommendation function.

[0056] In one embodiment, considering that a user's historical click habits have a significant impact on their current click function—for example, if a user frequently clicks the meme-sharing function while chatting using the input method, it indicates that the user is quite interested in this function, and they are highly likely to click it again when using the input method. Therefore, to improve the accuracy of predicting the click probability of each candidate recommended function, the functional reference parameters for each candidate recommended function may include at least one historical functional feature from the user's historical click records and the candidate functional features corresponding to each candidate recommended function.

[0057] The historical functional features include at least one of the following: the functional identifier corresponding to the historical function and the functional type corresponding to the historical function. For example, the functional identifier can be an ID (Identity Document), a number, etc., and the functional type can be image-based, language-based, or product-based, etc. At least one historical functional feature in the user's historical click records can be all historical functional features recorded in the user's historical click records; or, it can be one or more historical functional features recorded within a first set time period in the user's historical click records. For example, the first set time period closest to the current time, or any first set time period in the historical click records. The duration of the first set time period can be set according to actual needs, and this embodiment does not specifically limit it. For example, the first set time period is one week. Candidate functional features include at least one of the following: the functional identifier corresponding to the candidate function and the functional type.

[0058] In another implementation, considering that in addition to historical click habits, other auxiliary information may also affect the user's current click function, for example, younger users may be more interested in the meme-sharing function or the chat function; or, when a user uses their device at night and opens the input method, they may be more interested in the product function. Therefore, to further improve the accuracy of predicting the click-through rate of each candidate recommended function, the functional reference parameters of each candidate recommended function may include at least one historical functional feature from the user's historical click records, the candidate functional features corresponding to each candidate recommended function, and functional auxiliary features.

[0059] The functional auxiliary features include at least one of user features and contextual features. That is, functional auxiliary features may include only user features; or only contextual features; or both user features and contextual features. User features include basic user features and user preference features. For example, basic user features include user ID, user age, user gender, user height, and user weight; user preference features include the user's preference for various functions and the user's preference for various function types. Contextual features include at least one piece of current state information of the device running the input method. Current state information includes device model, the current time the device is running the input method, and the device's location when running the input method.

[0060] In this embodiment, the click probability of each candidate recommendation function is predicted based on the functional reference parameters of each candidate recommendation function, including the following sub-steps:

[0061] Sub-step one: For each candidate recommendation function, obtain the parameter feature vector corresponding to each function reference parameter of the candidate recommendation function.

[0062] In this embodiment, for each candidate recommendation function, the first feature extraction layer of the function prediction network model is used to extract features from the corresponding function reference parameters to obtain parameter feature vectors for each function reference parameter. For example, the function reference parameters of the candidate recommendation function include at least one historical function feature from the user's historical click records, candidate function features, user features, and context features. The first feature extraction layer extracts features from each historical function feature, each candidate function feature, and each function auxiliary feature to obtain parameter feature vectors corresponding to each historical function feature, each candidate function feature, and each function auxiliary feature. For example, the parameter feature vector is an M*S*N dense feature vector, where M represents the number of features, S represents the number of feature types, and N represents the feature dimension. For instance, the first feature extraction layer converts user IDs into an m*1*n parameter feature vector, where m represents the number of user IDs, S represents the number of user ID types (1), and N represents the feature dimension of user IDs (64 dimensions). By extracting features from the corresponding function reference parameters of the candidate recommendation function, sparse features can be converted into dense features, thereby improving the fitting ability of the prediction model.

[0063] Optionally, in this embodiment, features that are already dense features in the functional reference parameters are not extracted. Furthermore, when the functional reference parameters include multiple dense features, these dense features are normalized to ensure that their dimensions are consistent. For example, if the user features of the functional reference parameters include dense features such as user height and user weight, the values ​​of user height and user weight are mapped to a range of 0 to 1 to ensure that the dimensions of user height and user weight are consistent.

[0064] Sub-step two involves using the functional interest extraction layer of the functional prediction network to extract the user's functional interest vector for each candidate recommended function based on the parameter feature vectors corresponding to each historical functional feature and the parameter feature vectors corresponding to each candidate functional feature.

[0065] In this embodiment, for each candidate feature of the candidate recommendation function, firstly, attention processing is performed based on the parameter feature vectors corresponding to each historical feature and the parameter feature vectors corresponding to the candidate features to obtain the attention weights of each historical feature with respect to the candidate features. The attention weights corresponding to each historical feature represent the degree of correlation between each historical feature and the candidate feature. That is, historical features with a high degree of correlation with the candidate feature have higher attention weights, while historical features with a low degree of correlation with the candidate feature have lower attention weights. For example, historical features in the user's historical click records include the "Chat" feature ID and the "Meme" feature ID, and candidate features include the "Fancy Text" feature ID. Since both the "Chat" and "Fancy Text" features belong to the language category, while the "Meme" feature belongs to the image category, it can be seen that the correlation between the "Chat" feature ID and the "Fancy Text" feature ID is higher than that between the "Meme" feature ID and the "Fancy Text" feature ID. Specifically, attention processing is implemented using activation units in the functional interest extraction layer. This attention processing includes the following steps: First, the inner product of the parameter feature vectors corresponding to each historical functional feature and the parameter feature vectors corresponding to the candidate functional features is taken to obtain the inner product vector for each historical functional feature. Then, the parameter feature vectors corresponding to each historical functional feature, the inner product vectors corresponding to each historical functional feature, and the parameter feature vectors corresponding to the candidate functional features are concatenated to obtain a first concatenated vector. Finally, this first concatenated vector is input into the fully connected layer in the activation unit, and the attention weights of each historical functional feature with respect to the candidate functional features are obtained using this fully connected layer.

[0066] Secondly, the attention weights of each historical functional feature with respect to the candidate functional feature are used to weight the parameter feature vectors corresponding to each historical functional feature, thus obtaining the functional interest vectors corresponding to the candidate functional features. Specifically, the parameter feature vectors corresponding to each historical functional feature are multiplied by the attention weights corresponding to the candidate functional features, and then the parameter feature vectors multiplied by the attention weights are summed to obtain the functional interest vectors corresponding to the candidate functional features.

[0067] It should be noted that for each candidate recommendation function, when its corresponding function reference parameters include multiple candidate function features, the function interest vector corresponding to each candidate function feature is obtained separately according to the method in sub-step two. For example, if the function reference parameters of a candidate recommendation function include the function identifier and the function type of the candidate recommendation function, then the function interest vector corresponding to the function identifier and the function interest vector corresponding to the function type of the candidate recommendation function are obtained separately.

[0068] In the above approach, since not all historical functional features are related to the candidate functional features of the current candidate recommendation function, attention processing is performed to obtain the attention weights of each historical functional feature with respect to the candidate functional features, thereby obtaining the functional interest vector corresponding to the candidate recommendation feature. In the subsequent prediction of the click probability of the candidate recommendation function, the interference of historical functional features with a low degree of correlation with the candidate functional features on the click probability prediction can be eliminated, making the prediction model pay more attention to historical functional features with a high degree of correlation with the candidate functional features, thereby improving the accuracy of predicting the click probability of the candidate recommendation function.

[0069] Sub-step three involves using the first prediction layer of the functional prediction network to predict the click probability of each candidate recommendation function based on the functional interest vector corresponding to each candidate recommendation function.

[0070] In one embodiment, the functional reference parameters for each candidate recommendation function include at least one historical functional feature from the user's historical click records, candidate functional features corresponding to each candidate recommendation function, and functional auxiliary features. For each candidate recommendation function, the click probability of the candidate recommendation function is predicted based on the functional interest vector corresponding to the candidate recommendation function, the parameter feature vector corresponding to each candidate functional feature, and the parameter feature vector corresponding to the functional auxiliary features. Specifically, for each candidate recommendation function, the functional interest vector corresponding to the candidate recommendation function, the parameter feature vector corresponding to each candidate functional feature, and the parameter feature vector corresponding to the functional auxiliary features are concatenated to obtain a second concatenated vector. This second concatenated vector is input into the first prediction layer of the functional prediction network, and the click probability of the candidate recommendation function is predicted using the first prediction layer.

[0071] In another embodiment, the functional reference parameters for each candidate recommendation function include at least one historical functional feature from the user's historical click records and the candidate functional features corresponding to each candidate recommendation function. For each candidate recommendation function, the click probability of the candidate recommendation function is predicted based on the functional interest vector corresponding to the candidate recommendation function and the parameter feature vector corresponding to each candidate functional feature. Specifically, for each candidate recommendation function, the functional interest vector corresponding to the candidate recommendation function and the parameter feature vector corresponding to each candidate functional feature are concatenated to obtain a third concatenated vector. This third concatenated vector is input into the first prediction layer of the functional prediction network, and the click probability of the candidate recommendation function is predicted using the first prediction layer.

[0072] Furthermore, in both of the above embodiments, the functional interest intensity value of the candidate recommendation function is first predicted using the first prediction layer. This functional interest intensity value is a value greater than 1. Then, the functional interest intensity value of the candidate recommendation function is normalized to obtain the click probability of the candidate recommendation function, which is a value between 0 and 1. For example, the sigmoid function (also known as the S-shaped function or S-shaped growth curve) or other normalization functions can be used to map the functional interest intensity value of the candidate recommendation function to a value between 0 and 1. The first prediction layer includes a first set number of fully connected layers. The first set number of layers can be set according to actual conditions, and this embodiment does not specifically limit it. For example, the first set number of layers is 3.

[0073] In the above sub-steps one to three, since the function interest vector corresponding to each candidate recommended function is obtained based on at least one historical function feature in the user's historical click record, and the user's current click function largely depends on the user's historical click habits, predicting the click probability of each candidate recommended function based on the function interest vector corresponding to each candidate recommended function can make the predicted click probability of each candidate recommended function more accurate.

[0074] S303: The content prediction network of the prediction model predicts the click probability of each candidate recommended content based on the content reference parameters of each candidate recommended content.

[0075] In one implementation, considering that a user's historical click habits have a significant impact on their current click content—for example, if a user frequently clicks on funny emoticons in the meme-sharing feature while chatting using an input method—it indicates that the user is quite interested in funny emoticons, and they are highly likely to click on funny emoticons in the meme-sharing feature again when using the input method. Therefore, to improve the accuracy of predicting the click probability of each candidate recommended content, the content reference parameters for each candidate recommended content can include at least one historical content feature from the user's historical click records and the candidate function features corresponding to each candidate recommended function.

[0076] The historical content features include at least one of the following: a content identifier corresponding to the historical content, a content type corresponding to the historical content, and multiple keywords corresponding to the historical content. For example, the content identifier can be an ID, a number, etc., and the content type can be humorous, artistic, refreshing, cute, etc. The multiple keywords corresponding to the historical content can be multiple words obtained by splitting a chat sentence. At least one historical content feature in the user's historical click records can be all historical content features recorded in the user's historical click records; or, it can be one or more historical content features within a second set time period recorded in the user's historical click records. For example, the second set time period closest to the current time, or any second set time period in the historical click records. The duration of the second set time period can be set according to actual needs, and this embodiment does not specifically limit it. For example, the second set time period is one week. The candidate content features include at least one of the following: a content identifier corresponding to the candidate content and a content type.

[0077] In another implementation, considering that in addition to historical click habits, other auxiliary information may also affect the user's current click content, in order to further improve the accuracy of predicting the click probability of each candidate recommended content, the content reference parameters of each candidate recommended content include at least one historical content feature from the user's historical click records, the candidate content feature corresponding to each candidate recommended content, and content auxiliary features. Content auxiliary features include at least one of user features and context features. The relevant content of user features and context features is described in step S302 above and will not be repeated here.

[0078] In this embodiment, the click probability of each candidate recommended content is predicted based on the content reference parameters of each candidate recommended content, including the following sub-steps:

[0079] Sub-step one: For each candidate recommended content, obtain the parameter feature vector corresponding to each content reference parameter of the candidate recommended content.

[0080] In this embodiment, for each candidate recommendation content, the second feature extraction layer of the functional prediction network model is used to extract features from each content reference parameter of the candidate recommendation content, thereby obtaining the parameter feature vector corresponding to each content reference parameter of the candidate recommendation content. The relevant details can be found in sub-step one of step S302 above, and will not be repeated here.

[0081] It should be noted that in this embodiment, the parameter feature vectors of the functional reference parameters obtained in sub-step one of step S302 and the parameter feature vectors of the content reference parameters obtained in sub-step one of step S304 are extracted by the same feature extraction layer. In this case, the first feature extraction layer of the aforementioned functional prediction network and the second feature extraction layer of the content prediction network are the same layer. That is, the functional prediction network and the content prediction network can share the parameter feature vectors obtained by the same feature extraction layer to reduce the computational load of the prediction model. Alternatively, the parameter feature vectors of the functional reference parameters obtained in sub-step one of step S302 and the parameter feature vectors of the content reference parameters obtained in sub-step one of step S304 are extracted by the feature extraction layers in their respective prediction networks. In this case, the first feature extraction layer of the aforementioned functional prediction network and the second feature extraction layer of the content prediction network are different layers.

[0082] Sub-step two involves using the content interest extraction layer of the content prediction network to extract the user's content interest vector based on the parameter feature vectors corresponding to each historical content feature.

[0083] In this embodiment, firstly, self-attention processing is performed on the parameter feature vectors corresponding to each historical content feature to obtain the association vectors corresponding to each historical content feature. The association vectors corresponding to each historical content feature represent the degree of association between each historical content feature and other historical content features. Specifically, self-attention processing is implemented using the self-attention mechanism in the content interest extraction layer. For example, the self-attention mechanism can be a single-head self-attention mechanism or a multi-head self-attention mechanism. The number of heads in a multi-head self-attention mechanism can be set according to actual needs, and this embodiment does not impose a specific limitation on this. The self-attention processing process may include: inputting the parameter feature vectors corresponding to each historical functional feature into the self-attention mechanism, and using the self-attention mechanism to obtain the association vectors corresponding to each historical functional feature.

[0084] Secondly, the association vectors corresponding to each historical feature are concatenated to obtain the content interest vector. Specifically, the association vectors corresponding to each historical functional feature are concatenated, and the resulting vector is used as the content interest vector.

[0085] In the above method, by performing self-attention processing on the parameter feature vectors corresponding to each historical content feature, the prediction model can pay more attention to the correlation between each historical content feature in the subsequent process of predicting the click probability of candidate recommended content. For example, the correlation between keywords in a chat sentence can improve the accuracy of predicting the click probability of candidate recommended content.

[0086] Sub-step three involves using the second prediction layer of the content prediction network to predict the click probability of each candidate recommended content based on the content interest vector.

[0087] In one embodiment, the content reference parameters for each candidate recommended content include: at least one historical content feature from the user's historical click records, candidate content features corresponding to each candidate recommended content, and content auxiliary features. For each candidate recommended content, the click probability of the candidate recommended content is predicted based on the content interest vector, the parameter feature vector corresponding to each candidate recommended content feature, and the parameter feature vector corresponding to the content auxiliary features. Specifically, for each candidate recommended content, the content interest vector corresponding to the candidate recommended content, the parameter feature vector corresponding to each candidate recommended function, and the parameter feature vector corresponding to the content auxiliary features are concatenated to obtain a fourth concatenated vector. This fourth concatenated vector is input into the second prediction layer of the content prediction network, and the click probability of the candidate recommended content is predicted using this second prediction layer.

[0088] In another embodiment, the content reference parameters for each candidate recommended content include: at least one historical content feature from the user's historical click records and candidate content features corresponding to each candidate recommended content. For each candidate recommended content, the click probability of the candidate recommended content is predicted based on the content interest vector and the parameter feature vector corresponding to each candidate recommended content feature. Specifically, for each candidate recommended content, the content interest vector corresponding to the candidate recommended content and the parameter feature vector corresponding to each candidate recommended function are concatenated to obtain a fifth concatenated vector. This fifth concatenated vector is input into the second prediction layer of the content prediction network, and the click probability of the candidate recommended content is predicted using this second prediction layer.

[0089] Furthermore, in both of the above embodiments, the content interest intensity value of the candidate recommendation function is first predicted using the second prediction layer. This content interest intensity value is a value greater than 1. Then, the content interest intensity value of the candidate recommendation content is normalized to obtain the click probability of the candidate recommendation content. The click probability of the candidate recommendation content is a value between 0 and 1. For example, the sigmoid function or other normalization functions can be used to map the content interest intensity value of the candidate recommendation content to a value between 0 and 1. The second prediction layer includes a fully connected layer of a second set number of layers. The second set number of layers can be set according to actual conditions, and this embodiment does not specifically limit it. For example, the second set number of layers is 3 layers.

[0090] In the above sub-steps one to three, since the content interest vector is obtained based on at least one historical content feature in the user's historical click record, and the user's current click content largely depends on the user's historical click habits, predicting the click probability of each candidate recommended content based on the content interest vector can make the predicted click probability of each candidate recommended content more accurate.

[0091] S304: Based on the click probability of each candidate recommendation function and the click probability of each candidate recommendation content, determine the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content.

[0092] In this embodiment, for each candidate recommendation content, the output layer of the prediction model uses the click probability of the candidate recommendation content and the click probability of its corresponding candidate recommendation function to obtain the joint click probability between the candidate recommendation content and its corresponding candidate recommendation function. Specifically, the joint click probability between the candidate recommendation function and its corresponding candidate recommendation content is the product of the click probability of the candidate recommendation function and the click probability of the corresponding candidate recommendation content.

[0093] S305: Based on the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content, determine the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function.

[0094] In this embodiment, candidate recommended functions and candidate recommended content whose combined click probabilities meet the recommendation requirements are identified as target recommended functions and target recommended content.

[0095] In one embodiment, the candidate recommendation function and the corresponding candidate recommendation content corresponding to the largest joint click probability among the multiple joint click probabilities are taken as the target recommendation function and the target recommendation content.

[0096] In another embodiment, multiple candidate recommendation functions and their corresponding candidate recommendation content that have a probability greater than a probability threshold from the obtained multiple joint click probabilities are used as the target recommendation function and target recommendation content. The probability threshold is set according to actual needs, and this embodiment does not impose a specific limitation on it.

[0097] Furthermore, once the target recommendation function and target recommendation content are determined, they are displayed in the corresponding display interface of the input method to make recommendations to the user.

[0098] Please see Figure 4 , Figure 4 This is a schematic diagram of the input method display interface provided in this application. For example... Figure 4 As shown, in the current input method display interface, the corresponding target recommended content is the "Chat" function, and the corresponding target recommended content is "Strengthen your body, persevere to the end, never back down, taro boba."

[0099] In this embodiment, on the one hand, the click probability of each candidate recommended function is predicted using the function prediction network of the prediction model, and the click probability of each candidate recommended content is predicted using the content prediction network of the prediction model. Since the same prediction model is used to simultaneously predict the click probabilities of both candidate recommended functions and candidate recommended content, the joint click probability between the click probabilities of each candidate recommended function and the corresponding click probabilities of each candidate recommended content can simultaneously express the user's functional and content interests. Therefore, based on the joint click probability between the click probabilities of each candidate recommended function and the corresponding click probabilities of each candidate recommended content, the target recommended function and the target recommended content corresponding to the target recommended function can be accurately determined, thereby improving the user's click-through rate.

[0100] On the other hand, since the user's current click behavior largely depends on the user's historical click habits, and the click probability of each candidate recommended function is predicted based on the corresponding function interest vector, and the click probability of each candidate recommended content is predicted based on the content interest vector, which is obtained based on at least one historical content feature in the user's historical click records, and the function interest vector is obtained based on at least one historical function feature in the user's historical click records, the accuracy of the predicted click probability of each candidate recommended function and the accuracy of the predicted click probability of each candidate recommended content can be improved, thereby further improving the user's click-through rate.

[0101] Please see Figure 5 , Figure 5 This is a flowchart illustrating an embodiment of the training method for the prediction model provided in this application. Figure 5 As shown, the method includes the following steps:

[0102] S501: Obtain multiple sample candidate recommendation functions and the multiple sample candidate recommendation content corresponding to each sample candidate recommendation function.

[0103] Among the multiple candidate recommendation functions, at least one is labeled, and among the multiple candidate recommendation contents corresponding to each candidate recommendation function, at least one is labeled. The labeled recommendation function is used to determine the correctness of the predicted target recommendation function, and the labeled recommendation content is used to determine the correctness of the predicted target recommendation content.

[0104] S502: Use the prediction model to predict the click probability of the candidate recommendation function and the click probability of the candidate recommendation content for each sample.

[0105] S503: Based on the click probability of each sample candidate recommendation function and the click probability of each sample candidate recommendation content, determine the sample joint click probability between each sample candidate recommendation function and the corresponding sample candidate recommendation content.

[0106] S504: Based on the joint click probability of the samples, determine the sample target recommendation function and the sample target recommendation content corresponding to the sample target recommendation function.

[0107] Steps S502, S503 and S504 can be referred to the aforementioned steps S302 to S305, and will not be repeated here.

[0108] S505: Adjust the network parameters of the prediction model based on the differences between the sample target recommendation function and the labeled recommendation function, and the differences between the sample target recommendation content and the labeled recommendation content.

[0109] Specifically, the training of the prediction model is determined to be completed when the difference between the sample target recommendation function and the labeled recommendation function is less than a first difference threshold, and the difference between the sample target recommendation content and the labeled recommendation content is less than a second difference threshold.

[0110] In this embodiment, when training the prediction model, the process of predicting the click probability of each sample candidate recommendation function and the click probability of each sample candidate recommendation content is carried out simultaneously. This improves the accuracy of the prediction model in predicting the click probability of candidate recommendation functions and candidate recommendation content, thereby increasing the user's click-through rate.

[0111] Please see Figure 6 , Figure 6This is a schematic diagram of a framework of an embodiment of a recommendation device for an input method provided in this application. In this embodiment, the recommendation device 60 for an input method includes: an acquisition module 61, a prediction module 62, a first determination module 63, and a second determination module 64. The acquisition module 61 is used to acquire multiple candidate recommendation functions and multiple candidate recommendation contents corresponding to each candidate recommendation function in the input method. The prediction module 62 is used to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content using a prediction model. The click probability of a candidate recommendation content represents the likelihood that the candidate recommendation content will be clicked among the multiple candidate recommendation contents of the corresponding candidate recommendation function. The first determination module 63 is used to determine the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content based on the click probabilities of each candidate recommendation function and its corresponding click probability. The second determination module 64 is used to determine the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function based on the joint click probability.

[0112] Optionally, the prediction module 62 is used to predict the click probability of each candidate recommendation function based on the functional reference parameters of each candidate recommendation function using the functional prediction network of the prediction model, and to predict the click probability of each candidate recommendation content based on the content reference parameters of each candidate recommendation content using the content prediction network of the prediction model.

[0113] Optionally, the functional reference parameters of each candidate recommendation function include at least one historical functional feature in the user's historical click records, candidate functional features corresponding to each candidate recommendation function, and functional auxiliary features. The functional auxiliary features include at least one of user features and context features, and the context features include at least one current state information of the device running the input method. The content reference parameters of each candidate recommendation content include at least one historical content feature in the user's historical click records, candidate content features corresponding to each candidate recommendation content, and content auxiliary features. The content auxiliary features include at least one of user features and context features.

[0114] Optionally, the functional reference parameters of each candidate recommendation function include at least one historical functional feature from the user's historical click records and candidate functional features corresponding to each candidate recommendation function; for each candidate recommendation function, the prediction module 62 is used to obtain the parameter feature vector corresponding to each of the functional reference parameters of the candidate recommendation function; the functional interest extraction layer of the functional prediction network extracts the user's functional interest vector for each candidate recommendation function based on the parameter feature vector corresponding to each of the historical functional features and the parameter feature vector corresponding to each of the candidate functional features; the first prediction layer of the functional prediction network predicts the click probability of each candidate recommendation function based on the functional interest vector corresponding to each candidate recommendation function.

[0115] Optionally, for each of the candidate functional features, the prediction module 62 is used to perform attention processing based on the parameter feature vectors corresponding to each of the historical functional features and the parameter feature vectors corresponding to the candidate functional features to obtain the attention weights of each of the historical functional features with respect to the candidate functional features. The attention weights of each of the historical functional features represent the degree of correlation between each of the historical functional features and the candidate functional features. Using the attention weights of each of the historical functional features with respect to the candidate functional features, the parameter feature vectors corresponding to each of the historical functional features are weighted to obtain the functional interest vectors corresponding to the candidate functional features. And / or, the functional reference parameters of the candidate recommendation function further include functional auxiliary features. For each of the candidate recommendation functions, the prediction module 62 is used to predict the click probability of the candidate recommendation function based on the functional interest vectors corresponding to the candidate recommendation function, the parameter feature vectors corresponding to each of the candidate functional features, and the parameter feature vectors corresponding to the functional auxiliary features.

[0116] Optionally, the content reference parameters of the candidate recommended content include at least one historical content feature from the user's historical click records; the prediction module 62 is used to obtain, for each candidate recommended content, a parameter feature vector corresponding to each of the content reference parameters of the candidate recommended content; extract the user's content interest vector based on the parameter feature vector corresponding to each of the historical content features using the content interest extraction layer of the content prediction network; and predict the click probability of each candidate recommended content based on the content interest vector using the second prediction layer of the content prediction network.

[0117] Optionally, the prediction module 62 is used to perform self-attention processing on the parameter feature vectors corresponding to each of the historical content features to obtain the association vectors corresponding to each of the historical content features; concatenate the association vectors corresponding to each of the historical content features to obtain the content interest vector; and / or, the content reference parameters of the candidate recommended content further include content auxiliary features. For each candidate recommended content, the prediction module 62 is used to predict the click probability of the candidate recommended content based on the content interest vector, the parameter feature vectors corresponding to each of the candidate recommended content features, and the parameter feature vectors corresponding to the content auxiliary features.

[0118] Optionally, the parameter feature vectors of the functional reference parameters and the parameter feature vectors of the content reference parameters are extracted by the same feature extraction layer, or they are extracted by the feature extraction layers in the corresponding prediction networks, respectively.

[0119] Optionally, the joint click probability between the candidate recommendation function and the corresponding candidate recommendation content is the product of the click probability of the candidate recommendation function and the click probability of the corresponding candidate recommendation content; and / or, for each candidate recommendation content, the first determining module 63 is used to obtain the joint click probability between the candidate recommendation content and the candidate recommendation function based on the click probability of the candidate recommendation content and the click probability of the candidate recommendation function to which it belongs by using the output layer of the prediction model.

[0120] Optionally, the recommendation device further includes a model training module 65, which is used to acquire multiple candidate recommendation functions and multiple candidate recommendation contents corresponding to each candidate recommendation function, wherein at least one labeled recommendation function is among the multiple candidate recommendation functions, and at least one labeled recommendation content is among the multiple candidate recommendation contents corresponding to each candidate recommendation function; the model uses a prediction model to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content; based on the click probabilities of each candidate recommendation function and the click probabilities of each candidate recommendation content, the model determines the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content; based on the joint click probability, the model determines the target recommendation function and the target recommendation content corresponding to the target recommendation function; and based on the differences between the target recommendation function and the labeled recommendation function, and the differences between the target recommendation content and the labeled recommendation content, the model adjusts the network parameters of the prediction model.

[0121] Optionally, the acquisition module 61 acquires multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function in response to the pop-up of the input keyboard of the input method or the user's selection operation of the preset icon of the input method; and / or, the second determination module 64 is used to determine the candidate recommendation functions and candidate recommendation contents whose joint click probability meets the recommendation requirements as the target recommendation function and the target recommendation contents.

[0122] It should be noted that the apparatus of this embodiment can perform the steps in the above method. For detailed descriptions of the relevant content, please refer to the method section above, which will not be repeated here.

[0123] Please see Figure 7 , Figure 7 This is a schematic diagram of a framework of an embodiment of the voice interaction device provided in this application. In this embodiment, the processing device 70 includes a memory 71 and a processor 72.

[0124] Processor 72 can also be referred to as CPU (Central Processing Unit). Processor 72 may be an integrated circuit chip with signal processing capabilities. Processor 72 can also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component. A general-purpose processor can be a microprocessor, or processor 72 can be any conventional processor 72, etc.

[0125] The memory 71 in the processing device 70 is used to store the program instructions required for the processor 72 to run.

[0126] The processor 72 is used to execute program instructions to implement the recommended method for input methods in this application.

[0127] Please see Figure 8 , Figure 8This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium provided in this application. The computer-readable storage medium 80 of this application embodiment stores program instructions 81, which, when executed, implement the recommended method for input methods provided in this application. The program instructions 81 can be formed into a program file and stored in the aforementioned computer-readable storage medium 80 in the form of a software product, so that a computer device (which may be a personal computer, server, or network device, etc.) can execute all or part of the steps of the methods of various embodiments of this application. The aforementioned computer-readable storage medium 80 includes various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or terminal devices such as computers, servers, mobile phones, and tablets.

[0128] The above scheme utilizes a prediction model to obtain the click probability of each candidate recommended function and the click probability of each candidate recommended content. Furthermore, based on the joint click probability between the click probabilities of each candidate recommended function and its corresponding candidate recommended content, the target recommended function and its corresponding target recommended content for the input method are determined. Since the same prediction model is used to simultaneously predict the click probabilities of both candidate recommended functions and candidate recommended content, the resulting joint click probability can simultaneously express the user's functional and content interests. Therefore, the joint click probability can accurately determine the target recommended function and its corresponding target recommended content. This allows for accurate recommendations of functions and corresponding content that the user is interested in within the input method, thereby improving the click-through rate.

[0129] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0130] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0131] In the several embodiments provided in this application, it should be understood that the disclosed methods, apparatuses, and systems can be implemented in other ways. For example, the apparatus implementations described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection of apparatuses or units may be electrical, mechanical, or other forms.

[0132] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0133] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0134] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0135] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A recommendation method for input methods, characterized in that, The method includes: Obtain multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function; The click probability of each candidate recommendation function and the click probability of each candidate recommendation content are predicted using a prediction model. The click probability of the candidate recommendation content represents the probability that the candidate recommendation content is clicked among the multiple candidate recommendation contents of the candidate recommendation function to which it belongs. Based on the click probability of each candidate recommendation function and the click probability of each candidate recommendation content, the joint click probability between each candidate recommendation function and the corresponding candidate recommendation content is determined respectively. Based on the joint click probability, the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function are determined; Training the prediction model includes: acquiring multiple candidate recommendation functions and multiple candidate recommendation contents corresponding to each candidate recommendation function, wherein at least one labeled recommendation function exists among the multiple candidate recommendation functions, and at least one labeled recommendation content exists among the multiple candidate recommendation contents corresponding to each candidate recommendation function; using the prediction model to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content; based on the click probabilities of each candidate recommendation function and each candidate recommendation content, determining the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content; based on the joint click probability, determining the target recommendation function and its corresponding target recommendation content; and adjusting the network parameters of the prediction model based on the differences between the target recommendation function and the labeled recommendation function, and the differences between the target recommendation content and the labeled recommendation content.

2. The method according to claim 1, characterized in that, The step of using a prediction model to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content includes: The function prediction network of the prediction model predicts the click probability of each candidate recommendation function based on the function reference parameters of each candidate recommendation function, and... The content prediction network of the prediction model predicts the click probability of each candidate recommended content based on the content reference parameters of each candidate recommended content.

3. The method according to claim 2, characterized in that, The functional reference parameters for each of the candidate recommendation functions include at least one historical functional feature in the user's historical click records, the candidate functional features corresponding to each of the candidate recommendation functions, and functional auxiliary features. The functional auxiliary features include at least one of user features and context features. The context features include at least one current state information of the device running the input method. The content reference parameters for each candidate recommended content include at least one historical content feature from the user's historical click records, candidate content features corresponding to each candidate recommended content, and content auxiliary features, wherein the content auxiliary features include at least one of the user features and the context features.

4. The method according to claim 2, characterized in that, The functional reference parameters for each of the candidate recommendation functions include at least one historical functional feature from the user's historical click records and the candidate functional features corresponding to each of the candidate recommendation functions; The prediction of the click probability of each candidate recommendation function based on the function reference parameters of each candidate recommendation function includes: For each of the candidate recommendation functions, obtain the parameter feature vector corresponding to each of the function reference parameters of the candidate recommendation function; The functional interest extraction layer of the functional prediction network extracts the user's functional interest vector for each of the candidate recommended functions based on the parameter feature vectors corresponding to each of the historical functional features and the parameter feature vectors corresponding to each of the candidate functional features. The first prediction layer of the function prediction network predicts the click probability of each candidate recommendation function based on the function interest vector corresponding to each candidate recommendation function.

5. The method according to claim 4, characterized in that, The step of extracting the user's functional interest vector for each candidate recommended function based on the parameter feature vector corresponding to each of the historical functional features and the parameter feature vector corresponding to each of the candidate functional features includes: For each of the candidate functional features, attention processing is performed based on the parameter feature vectors corresponding to each of the historical functional features and the parameter feature vectors corresponding to the candidate functional features to obtain the attention weights of each of the historical functional features with respect to the candidate functional features. The attention weights corresponding to each of the historical functional features represent the degree of correlation between each of the historical functional features and the candidate functional features. By utilizing the attention weights of each historical functional feature with respect to the candidate functional feature, the parameter feature vectors corresponding to each historical functional feature are weighted to obtain the functional interest vectors corresponding to the candidate functional features. And / or, the functional reference parameters of the candidate recommendation functions further include functional auxiliary features, and the prediction of the click probability of each candidate recommendation function based on the functional interest vector corresponding to each candidate recommendation function includes: For each of the candidate recommendation functions, the click probability of the candidate recommendation function is predicted based on the function interest vector corresponding to the candidate recommendation function, the parameter feature vector corresponding to each candidate function feature, and the parameter feature vector corresponding to the function auxiliary feature.

6. The method according to claim 2, characterized in that, The content reference parameters for the candidate recommended content include at least one historical content feature from the user's historical click records; The prediction of the click probability of each candidate recommended content based on the content reference parameters of each candidate recommended content includes: For each of the candidate recommendation contents, obtain the parameter feature vector corresponding to each of the content reference parameters of the candidate recommendation contents; The content interest extraction layer of the content prediction network extracts the user's content interest vector based on the parameter feature vectors corresponding to each of the historical content features. The second prediction layer of the content prediction network predicts the click probability of each candidate recommended content based on the content interest vector.

7. The method according to claim 6, characterized in that, The step of extracting the user's content interest vector based on the parameter feature vectors corresponding to each of the historical content features includes: Self-attention processing is performed on the parameter feature vectors corresponding to each of the historical content features to obtain the association vectors corresponding to each of the historical content features. The association vectors corresponding to each of the aforementioned historical content features are concatenated to obtain the content interest vector; And / or, the content reference parameters of the candidate recommended content further include content auxiliary features, and the prediction of the click probability of each candidate recommended content based on the content interest vector includes: For each of the candidate recommended content, the click probability of the candidate recommended content is predicted based on the content interest vector, the parameter feature vector corresponding to the features of each candidate recommended content, and the parameter feature vector corresponding to the content auxiliary features.

8. The method according to any one of claims 4 to 7, characterized in that, The feature vectors of the functional reference parameters and the content reference parameters are extracted by the same feature extraction layer, or by the feature extraction layers in their respective prediction networks.

9. The method according to claim 1, characterized in that, The joint click probability between the candidate recommendation function and the corresponding candidate recommendation content is the product of the click probability of the candidate recommendation function and the click probability of the corresponding candidate recommendation content. And / or, determining the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content based on the click probability of each candidate recommendation function and the click probability of each candidate recommendation content includes: For each of the candidate recommended content, the output layer of the prediction model is used to obtain the joint click probability between the candidate recommended content and the candidate recommended function to which it belongs, based on the click probability of the candidate recommended content and the click probability of the candidate recommended function to which it belongs.

10. The method according to claim 1, characterized in that, The step of obtaining multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function is performed in response to the pop-up of the input keyboard of the input method or the user's selection operation of the preset icon of the input method; And / or, determining the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function based on the joint click probability includes: The candidate recommendation functions and candidate recommendation content whose joint click probabilities meet the recommendation requirements are determined as the target recommendation function and the target recommendation content.

11. A recommendation device for an input method, characterized in that, The device includes: The acquisition module is used to acquire multiple candidate recommendation functions in the input method and multiple candidate recommendation contents corresponding to each candidate recommendation function. The prediction module is used to predict the click probability of each of the candidate recommendation functions and the click probability of each of the candidate recommendation contents using a prediction model. The click probability of the candidate recommendation contents represents the probability that the candidate recommendation contents will be clicked among the multiple candidate recommendation contents of the candidate recommendation function. The first determining module is used to determine the joint click probability between each candidate recommendation function and the corresponding candidate recommendation content based on the click probability of each candidate recommendation function and the click probability of each candidate recommendation content. The second determining module is used to determine the target recommendation function of the input method and the target recommendation content corresponding to the target recommendation function based on the joint click probability. The model training module is used to acquire multiple candidate recommendation functions and corresponding candidate recommendation content for each candidate function. Among the multiple candidate recommendation functions, at least one is annotated, and among the corresponding candidate recommendation content, at least one is annotated. The module uses a prediction model to predict the click probability of each candidate recommendation function and the click probability of each candidate recommendation content. Based on these probabilities, the module determines the joint click probability between each candidate recommendation function and its corresponding candidate recommendation content. Based on the joint click probability, the module determines the target recommendation function and its corresponding target recommendation content. Finally, based on the differences between the target recommendation function and the annotated recommendation function, and between the target recommendation content and the annotated recommendation content, the module adjusts the network parameters of the prediction model.

12. A processing apparatus, characterized in that, Including interconnected memory and processor, The memory stores program instructions; The processor is used to execute program instructions stored in the memory to implement the method according to any one of claims 1-10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium is used to store program instructions that can be executed to implement the method of any one of claims 1-10.

Images