A method for model training, a method for information recommendation, and a device.

Abstract

This specification discloses a method for model training, an information recommendation method, and an apparatus. The method involves acquiring historical operation information of a specified user across various business scenarios. Next, based on the click objects involved in each historical operation and other users associated with the specified user in each business scenario, the common and unique representations of the specified user across each business scenario are determined. Then, the common representations, unique representations, and historical recommendation information are input into a recommendation model to be trained to predict the first click-through rate (CTR) in the given business scenario and the second CTR in each business scenario. Finally, the recommendation model is trained with the optimization objective of minimizing the deviation between the first and second CTRs and the actual click situation. This method can improve the recommendation effect for information to target users by leveraging the common preferences of users across different business scenarios and the differences in user representations between different business scenarios.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a method for model training, a method for information recommendation, and an apparatus. Background Technology

[0002] Currently, in most recommendation scenarios, there is a problem of data sparsity. That is, for low-frequency users or new users, due to insufficient interaction data, it is impossible to obtain a relatively accurate user representation.

[0003] Therefore, in practical applications, a comprehensive user feature vector across multiple scenarios is typically obtained by fusing historical user behavior data from different scenarios to improve the recommendation performance for that user in the current scenario. However, this approach leads to an overall improvement in the recommendation model's performance across various scenarios, but a decrease in its performance within a single scenario.

[0004] Therefore, how to recommend information to users more effectively is an urgent problem to be solved. Summary of the Invention

[0005] This specification provides a method, apparatus, storage medium, and electronic device for model training, in order to partially solve the aforementioned problems existing in the prior art.

[0006] The following technical solution is adopted in this specification:

[0007] This manual provides a method for model training, including:

[0008] Retrieve historical operation information of a specified user in various business scenarios;

[0009] Based on the click objects involved in each historical operation information and other users associated with the specified user in each business scenario, the common representation of the specified user in each business scenario is determined. For each business scenario, based on the click objects involved in the historical operation information corresponding to that business scenario and other users associated with the specified user in that business scenario, the unique representation of the specified user in that business scenario is determined.

[0010] The common representation, the unique representation, and the historical recommendation information are input into the recommendation model to be trained. The click-through rate for the historical recommendation information in the business scenario is predicted as the first click-through rate, and the click-through rate for the historical recommendation information in each business scenario is predicted as the second click-through rate.

[0011] The recommendation model is trained with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks.

[0012] Optionally, based on the clicked objects involved in each historical operation information and other users associated with the specified user in each business scenario, a common representation of the specified user in each business scenario is determined, specifically including:

[0013] Based on the historical operation information of each user in each business scenario, each user is treated as a user node and each clicked object in each business scenario is treated as an object node. A multi-scenario relationship graph corresponding to each business scenario is constructed. If a user node has historically performed a click operation on a clicked object, then the user node and the object node corresponding to the clicked object are connected by an edge.

[0014] Based on the multi-scenario relationship diagram, the common representation of the specified user in each business scenario is determined.

[0015] Optionally, based on the multi-scenario relationship diagram, the common representation of the specified user in each business scenario is determined, specifically including:

[0016] From the multi-scenario relationship graph, identify the user nodes that satisfy the set adjacency relationship with the specified user and use them as associated nodes;

[0017] Based on the representations corresponding to the associated nodes, the common representations of the specified user in each business scenario are determined.

[0018] Optionally, based on the representations corresponding to the associated nodes, a common representation of the specified user across the various business scenarios is determined, specifically including:

[0019] For each associated node, the weight between the associated node and the user node corresponding to the specified user is determined based on the degree of association between the object corresponding to the associated node and the specified user. This weight is used as the weight of the associated node. The object corresponding to the associated node includes the clicked object or the associated user.

[0020] Based on the representation corresponding to each associated node and the weight corresponding to each associated node, the common representation of the specified user in each business scenario is determined.

[0021] Optionally, for each business scenario, based on the clicked objects involved in the historical operation information corresponding to that business scenario and other users associated with the specified user in that business scenario, a unique representation of the specified user in that business scenario is determined, specifically including:

[0022] For each business scenario, based on the historical operation information of each user in the business scenario, each user is treated as a user node and each clicked object in the business scenario is treated as an object node. A single scenario relationship graph is constructed for the business scenario. If the user corresponding to a user node has performed a click operation on a clicked object in the past, then the user node and the object node corresponding to the clicked object are connected by an edge.

[0023] Based on the single-scenario relationship diagram, the unique representation of the specified user in this business scenario is determined.

[0024] Optionally, based on the single-scenario relationship graph, the unique representation of the specified user in this business scenario is determined, specifically including:

[0025] From the single-scene relationship graph, identify the user nodes that satisfy the set adjacency relationship with the specified user and use them as associated nodes;

[0026] Based on the representations corresponding to the associated nodes, the unique representation of the specified user in this business scenario is determined.

[0027] Optionally, based on the representations corresponding to the associated nodes, a unique representation of the specified user in this business scenario is determined, specifically including:

[0028] For each associated node, the weight between the associated node and the user node corresponding to the specified user is determined based on the degree of association between the object corresponding to the associated node and the specified user. This weight is used as the weight of the associated node. The object corresponding to the associated node includes the clicked object or the associated user.

[0029] Based on the representation corresponding to each associated node and the weight corresponding to each associated node, the unique representation of the specified user in this business scenario is determined.

[0030] Optionally, the recommendation model includes multiple recommendation sub-models;

[0031] The common representation, the unique representation, and historical recommendation information are input into the recommendation model to be trained. The click-through rate (CTR) for the historical recommendation information in the given business scenario is predicted as the first CTR, and the CTR for the historical recommendation information in each of the business scenarios is predicted as the second CTR. Specifically, this includes:

[0032] The unique representation and historical recommendation information are input into the sub-recommendation model corresponding to the business scenario to be trained, and the click rate for the historical recommendation information in the business scenario is predicted as the first click rate.

[0033] The shared representation and the historical recommendation information are input into the sub-recommendation model corresponding to each business scenario to be trained, and the click-through rate for the historical recommendation information in each business scenario is predicted as the second click-through rate.

[0034] Optionally, the recommendation model is trained with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks. Specifically, this includes:

[0035] Based on the common representation of the specified user in each business scenario and the specific representation of the specified user in that business scenario, a correlation value between the common representation and the specific representation is determined, wherein the higher the correlation value, the higher the correlation between the common representation and the specific representation.

[0036] The optimization objective is to minimize the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks, and to maximize the relevant representation values, thereby training the recommendation model.

[0037] Optionally, the recommendation model is trained with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks. Specifically, this includes:

[0038] Based on the unique characteristics of the specified user in each business scenario, determine the difference value between the unique characteristics of the specified user in each business scenario, wherein the lower the difference value, the greater the difference between the unique characteristics of the specified user in each business scenario.

[0039] The optimization objective is to minimize the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks. The recommendation model is trained to minimize the difference representation value.

[0040] This specification provides an information recommendation method, including:

[0041] In response to a target user's request for information in a specified business scenario, obtain candidate recommendation information in the specified business scenario and the target user's historical operation information in the specified business scenario;

[0042] For each candidate recommendation information, determine the representation corresponding to the candidate recommendation information in the specified business scenario based on the candidate recommendation information in the specified business scenario;

[0043] Based on the target user's historical operation information in the specified business scenario, determine the unique characteristics of the target user in the specified business scenario;

[0044] The representation of the candidate recommendation information in the specified business scenario and the unique representation of the target user in the specified business scenario are input into the pre-trained recommendation model to predict the click-through rate of the candidate recommendation information in the specified business scenario. The recommendation model is trained by the above model training method.

[0045] Based on the click-through rate of each candidate recommendation information in a specified business scenario, the recommendation information to be recommended to the target user is determined as the target recommendation information, and the target recommendation information is recommended to the target user.

[0046] This specification provides a model training apparatus, comprising:

[0047] The acquisition module is used to acquire historical operation information of a specified user in various business scenarios;

[0048] The determination module is used to determine the common representation of the specified user in each business scenario based on the click objects involved in each historical operation information and other users associated with the specified user in each business scenario, and to determine the unique representation of the specified user in each business scenario based on the click objects involved in the historical operation information corresponding to that business scenario and other users associated with the specified user in that business scenario.

[0049] The prediction module is used to input the common representation, the unique representation, and the historical recommendation information into the recommendation model to be trained, and predict the click rate for the historical recommendation information in the business scenario as the first click rate, and the click rate for the historical recommendation information in each business scenario as the second click rate;

[0050] The training module is used to train the recommendation model with the optimization objective of minimizing the deviation between the first click-through rate and the actual click situation corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual click situation.

[0051] This specification provides an information recommendation device, comprising:

[0052] The acquisition module is used to respond to the information acquisition request of the target user in the specified business scenario, and to acquire the candidate recommendation information in the specified business scenario and the historical operation information of the target user in the specified business scenario.

[0053] The first determining module is used to determine the representation corresponding to each candidate recommendation information in the specified business scenario based on the candidate recommendation information in the specified business scenario.

[0054] The second determining module is used to determine the unique characteristics of the target user in the specified business scenario based on the target user's historical operation information in the specified business scenario.

[0055] The prediction module is used to input the representation of the candidate recommendation information in the specified business scenario and the unique representation of the target user in the specified business scenario into the pre-trained recommendation model to predict the click-through rate of the candidate recommendation information in the specified business scenario. The recommendation model is trained by the above model training method.

[0056] The recommendation module is used to determine the recommended information to be recommended to the target user based on the click-through rate of each candidate recommendation information in a specified business scenario, and recommend the target recommendation information to the target user.

[0057] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described methods for model training and information recommendation.

[0058] This specification provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method for model training and information recommendation.

[0059] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects:

[0060] The model training method provided in this specification involves obtaining historical operation information of a specified user in various business scenarios. Next, based on the click objects involved in each historical operation and other users associated with the specified user in each business scenario, a common representation of the specified user in each business scenario is determined. Then, for each business scenario, based on the click objects involved in the corresponding historical operation information and other users associated with the specified user in that business scenario, a unique representation of the specified user in that business scenario is determined. Subsequently, the common representation, unique representation, and historical recommendation information are input into the recommendation model to be trained to predict the click-through rate (CTR) for historical recommendation information in that business scenario, which is used as the first CTR. The CTR for historical recommendation information in each business scenario is used as the second CTR. Finally, the recommendation model is trained with the optimization objective of minimizing the deviation between the first CTR and the actual clicks corresponding to the historical recommendation information, and the deviation between the second CTR and the actual clicks.

[0061] As can be seen from the above method, this method can determine the correlation between user representations in various business scenarios by specifying common representations of users in each business scenario, thereby obtaining the common preferences of users in each business scenario. Furthermore, by specifying unique representations for users in each business scenario, differences in user representations exist between different business scenarios. Finally, the recommendation model is trained with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to historical recommendation information, as well as the deviation between the second click-through rate and the actual clicks. This approach ensures both the correlation and the diversity of user representations across different business scenarios. Attached Figure Description

[0062] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings:

[0063] Figure 1 A schematic flowchart illustrating the model training method provided in the embodiments of this specification;

[0064] Figure 2 This is a schematic diagram of a multi-scenario relationship diagram provided in the embodiments of this specification;

[0065] Figure 3 A schematic diagram of a single-scene relationship diagram provided in the embodiments of this specification;

[0066] Figure 4 A flowchart illustrating the method for recommending information provided in the embodiments of this specification;

[0067] Figure 5 A schematic diagram of the structure of the model training apparatus provided in the embodiments of this specification;

[0068] Figure 6 A schematic diagram of the device recommended for the information provided in the embodiments of this specification;

[0069] Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this specification. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this specification clearer, the technical solutions of this specification will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of them. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this specification.

[0071] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0072] In the embodiments of this specification, when recommending information in response to a target user's information retrieval request in a specified business scenario, a pre-trained recommendation model is required. Therefore, the process of training the recommendation model will be described below, such as... Figure 1 As shown.

[0073] Figure 1 The flowchart of the model training method provided in the embodiments of this specification is shown in the figure, which specifically includes the following steps:

[0074] S100: Obtain historical operation information of a specified user in various business scenarios.

[0075] In the embodiments of this specification, the execution entity for training the recommendation model can be a server or an electronic device such as a desktop computer. For ease of description, the following description will only use a server as the execution entity to illustrate the training method of the recommendation model provided in this specification.

[0076] In the embodiments described in this specification, the server can obtain historical operation information of a specified user in various business scenarios. The specified user mentioned here can be any user who has historical online operation information in the past. The business scenarios mentioned here can be various, for example, search scenarios to meet users' active search needs, which can include search scenarios in various fields such as food search scenarios, travel search scenarios, and product search scenarios. Another example is recommendation scenarios to meet users' passive personalized recommendation needs, which can include recommendation scenarios in various fields such as food recommendation scenarios, travel recommendation scenarios, and product recommendation scenarios.

[0077] The historical operation information mentioned here can refer to a user's historical click information and historical order information in various business scenarios. For example, the food a user clicks on in a food search scenario. Another example is the products a user orders in a product recommendation scenario.

[0078] In different business scenarios, the business scenario information obtained by the server varies. The server can pre-set the business scenario information to be obtained based on the actual needs of the business scenario. For example, in a product recommendation scenario, the information for each product may include at least one of the following: the actual name of each product, the attribute information of each product (product type, material, etc.), and the business information of each product (product price, sales volume, number of times the product has been recommended, etc.). As another example, in a food search scenario, the information for each food item may include at least one of the following: the actual name of each food item, the attribute information of each food item (food flavor, ingredients, etc.), and the business information of each food item (food price, sales volume, number of times the food has been recommended, etc.).

[0079] It should be noted that before obtaining a user's historical operation information, the server needs to send an authorization request to the user. If the user does not agree to the authorization, the server cannot obtain the user's historical operation information. If the user agrees to the authorization, the server can delete the obtained user's historical operation information when the user cancels the service using this method.

[0080] S102: Based on the click objects involved in each historical operation information and other users associated with the specified user in each business scenario, determine the common representation of the specified user in each business scenario; and for each business scenario, based on the click objects involved in the historical operation information corresponding to that business scenario and other users associated with the specified user in that business scenario, determine the unique representation of the specified user in that business scenario.

[0081] In the embodiments of this specification, the click objects involved in the historical operation information of a specified user under various business scenarios obtained by the server can reflect the specified user's personal preferences to a certain extent. Furthermore, if the click objects involved in the historical operation information of two users are the same, it can be considered that these two users have similar personal preferences. Therefore, the personal preferences of other users who click on the same historical operation information as the specified user can also reflect the specified user's personal preferences to a certain extent.

[0082] In the embodiments of this specification, the server can determine the common representation of a specified user in various business scenarios based on the click objects involved in each historical operation information and other users associated with the specified user in each business scenario. The "other users associated with the specified user in each business scenario" mentioned here refers to, for each user, if the specified user has performed a click operation on a click object in each business scenario, and the user has also performed a click operation on the same click object, then that user is an "other user associated with the specified user." In other words, if two users have historically performed click operations on the same click object, then these two users are mutually associated users.

[0083] Specifically, the server can construct a multi-scenario relationship graph based on the historical operation information of each user in various business scenarios, treating each user as a user node and each clicked object in each business scenario as an object node. If a user node's corresponding user has historically clicked on a clicked object, then that user node is connected to the object node corresponding to that clicked object by an edge. Then, based on the multi-scenario relationship graph, the common representation of a specified user across various business scenarios is determined. For example... Figure 2 As shown.

[0084] Figure 2 This is a schematic diagram of a multi-scenario relationship diagram provided in the embodiments of this specification.

[0085] exist Figure 2 In this diagram, circles with letters represent user nodes. A circle with the letter 'A' represents user node A corresponding to a specific user. Circles with numbers represent object nodes corresponding to clicked objects. A square represents a business scenario. If a specified user has historically clicked on object 1 in business scenario 1, object 2 in business scenario 2, and object 3 in business scenario 3, the server can connect user node A corresponding to the specified user to object nodes 1, 2, and 3 with edges. Based on this, the server can determine the common representation of the specified user across various business scenarios according to the connection relationships in the multi-scenario relationship graph.

[0086] In the embodiments described in this specification, the server can determine, from the multi-scenario relationship graph, all nodes that satisfy a set adjacency relationship with the user nodes corresponding to a specified user, and use them as associated nodes. Then, based on the representations corresponding to the associated nodes, the server determines the common representations of the specified user in various business scenarios.

[0087] Specifically, if a user node corresponding to a specified user is connected to the same object node as user nodes corresponding to other users, then the user node corresponding to the specified user has a first-order adjacency relationship with the other user nodes. If the user node corresponding to the specified user is connected to the same object node as user nodes corresponding to other users, and the user nodes corresponding to other users are also connected to other object nodes in various business scenarios, then the user node corresponding to the specified user has a second-order adjacency relationship with the other object nodes in various business scenarios. And so on, multi-order adjacency relationships for user nodes corresponding to a specified user can be obtained. The adjacency relationship mentioned here can refer to an adjacency relationship of a specified order.

[0088] For example, in Figure 2 In the given scenario, user node A corresponding to a specified user is connected to object node 1, and user node B corresponding to other users is connected to object node 2. Therefore, user node A corresponding to the specified user and user node B corresponding to other users have a first-order adjacency relationship. Furthermore, user node B corresponding to other users is connected to object node 4, so user node A corresponding to the specified user and object node 4 have a second-order adjacency relationship.

[0089] Furthermore, for each associated node, the server can determine the weight between the associated node and the user node corresponding to the specified user based on the degree of association between the object corresponding to that associated node and the specified user. This weight is then used as the weight for that associated node. The object corresponding to the associated node mentioned here includes either the clicked object or the associated user. The specific formula is as follows:

[0090]

[0091] In the above formula, e t This can be used to represent the user representation corresponding to a specified user t, which is a randomly generated standard vector. The user representation corresponding to a specified user will be continuously optimized during the model training process to determine a more accurate user representation. s It can be used to represent the object node s corresponding to each business scenario. || It can be used for vector concatenation operations. a lThe transformation parameters of the l-th layer can be used to characterize the transformation parameters, which are continuously optimized during the model training process to determine the accurate correlation between the user representation corresponding to a specified user t and the representation corresponding to the object node s in the l-th layer. ReLU() can be used to characterize the activation function. Softmax() can be used to characterize the normalization function, ensuring that the formula result is controlled between 0 and 1. This can be used to characterize the weights corresponding to the association between the user representation of a specified user t in layer l and the representation of the object node s. As can be seen from the above formula, the server can determine the weights corresponding to each associated node.

[0092] Then, the server can determine the common representation of a specified user across various business scenarios based on the representation corresponding to each associated node and the weight corresponding to each associated node. The specific formula is as follows:

[0093]

[0094] In the above formula, It can be used to represent the set of associated nodes of a user node t corresponding to a specified user. It can be used to characterize the comprehensive representation of the l-th layer determined based on the representations corresponding to each associated node in the l-th layer and the weights corresponding to each associated node.

[0095] Next, the server can determine the common representation of a specified user across various business scenarios based on the comprehensive representations of each layer. The specific formula is as follows:

[0096]

[0097] In the above formula, f() can be used to represent a deep neural network, which can take many forms. Examples include a fully connected feedforward network (FFN) or a multi-head self-attention (MSA) network. This specification does not limit the specific method used for deep neural networks. t It can be used to represent the common characteristics of a specified user in various business scenarios.

[0098] Similarly, for each business scenario, the server can determine the unique characteristics of a specified user in that business scenario based on the click objects involved in the historical operation information corresponding to that business scenario and other users associated with the specified user in that business scenario.

[0099] Specifically, for each business scenario, the server can construct a single-scenario relationship graph based on the historical operation information of each user within that scenario, treating each user as a user node and each clicked object within that scenario as an object node. If a user node's corresponding user has historically clicked on an object, then that user node is connected to the object node corresponding to that clicked object by an edge. Then, based on the single-scenario relationship graph, the unique representation of a specific user within that business scenario is determined. For example... Figure 3 As shown.

[0100] Figure 3 This is a schematic diagram of a single-scene relationship diagram provided in the embodiments of this specification.

[0101] exist Figure 3 In this diagram, circles with letters represent user nodes. A circle with the letter 'A' represents user node A corresponding to a specific user. Circles with numbers represent object nodes corresponding to clicked objects. Squares represent business scenarios; a single-scenario relationship graph contains only one business scenario. If a specified user has historically clicked on object 1, object 2, and object 3, the server can connect user node A corresponding to that user to object nodes 1, 2, and 3 with edges. Based on this, the server can determine the unique representation of a specified user within that business scenario from the connection relationships in the single-scenario relationship graph.

[0102] In the embodiments described in this specification, the server can determine, from a single-scenario relationship graph, all nodes that satisfy a set adjacency relationship with the user nodes corresponding to a specified user, and use them as associated nodes. Then, based on the representations corresponding to the associated nodes, the server determines the unique representation of the specified user in that business scenario.

[0103] Furthermore, for each associated node, the server can determine the weight between the associated node and the user node corresponding to the specified user based on the degree of association between the object corresponding to the associated node and the specified user. This weight is then used as the weight for the associated node. The object corresponding to the associated node includes either the clicked object or the associated user. In this single-scenario relationship graph, all object nodes and user nodes belong to the same business scenario. The specific formula is as follows:

[0104]

[0105] In the above formula, e s-k It can be used to represent the object node s corresponding to business scenario k. This can be used to characterize the weights corresponding to the association between the user representation of a specified user t at layer l and the representation of the object node s in business scenario k. As can be seen from the formula above, the server can determine the weights corresponding to each associated node in this business scenario.

[0106] Then, the server can determine the unique representation of a specific user in this business scenario based on the representation corresponding to each associated node and the weight corresponding to each associated node. The specific formula is as follows:

[0107]

[0108] In the above formula, It can be used to represent the set of associated nodes of a user node t corresponding to a specified user in a business scenario k. It can be used to characterize the comprehensive representation of the l-th layer in business scenario k, based on the representations of each associated node in the l-th layer and the weights of each associated node.

[0109] Next, the server can determine the unique representation of a specific user within a given business scenario k based on the comprehensive representation of each layer. The specific formula is as follows:

[0110]

[0111] In the above formula, f() can be used to characterize deep neural networks. t-k It can be used to represent the unique characteristics of a specified user in business scenario k.

[0112] As can be seen from the above methods, determining the common representation of a specified user in various business scenarios is basically the same as determining the unique representation of a specified user in that business scenario. The difference is that the object nodes and user nodes in the common representation can be from multiple business scenarios, while the object nodes and user nodes in the unique representation can only be from a single business scenario.

[0113] S104: Input the common representation, the unique representation, and the historical recommendation information into the recommendation model to be trained, predict the click-through rate of the historical recommendation information in the business scenario as the first click-through rate, and the click-through rate of the historical recommendation information in each business scenario as the second click-through rate.

[0114] In the embodiments of this specification, the server can input common representations, unique representations, and historical recommendation information into the recommendation model to be trained, predict the click-through rate (CTR) for historical recommendation information in the current business scenario, which is used as the first CTR, and the CTR for the historical recommendation information in each business scenario, which is used as the second CTR. The historical recommendation information mentioned here can refer to historical recommendation information in each business scenario.

[0115] In the embodiments of this specification, the recommendation model includes multiple recommendation sub-models, and different recommendation sub-models correspond to single-scenario relationship graphs for different business scenarios. Each business scenario's multi-scenario relationship graph corresponds to one recommendation sub-model. The server can input unique representations and historical recommendation information into the sub-recommendation model corresponding to the business scenario to be trained, predicting the click-through rate for historical recommendation information within that business scenario, which is then used as the first click-through rate.

[0116] Meanwhile, the server can input shared representations and historical recommendation information into the sub-recommendation models corresponding to each business scenario to be trained, and predict the click-through rate for historical recommendation information in each business scenario, which will be used as the second click-through rate.

[0117] S106: The recommendation model is trained with the optimization objective of minimizing the deviation between the first click-through rate and the actual click situation corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual click situation.

[0118] In the embodiments of this specification, the server may train the recommendation model with the optimization objective of minimizing the deviation between the first click-through rate and the actual click situation corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual click situation.

[0119] In the embodiments described in this specification, in order to effectively migrate the user representation of a specified user in different business scenarios, the server can utilize the historical operation information of the specified user in other business scenarios to optimize the user representation of the specified user in the current business scenario. Based on this, the server can reduce the distance between the common representation and the unique representation of the specified user in each business scenario, thereby optimizing the correlation between the common representation and the unique representation.

[0120] Specifically, the server can determine the correlation value between the common representation and the unique representation of a specified user in various business scenarios, based on the user's common representation across different business scenarios and the user's unique representation within that specific business scenario. The higher the correlation value, the stronger the correlation between the common and unique representations. The specific formula is as follows:

[0121]

[0122] In the above formula, U can be used to represent the number of users. t, It can be used to represent the common characteristics of a specified user across various business scenarios. t-k It can be used to represent the unique representation of a specified user in business scenario k. c can be used to represent the common representation space, s k It can be used to represent the unique representation space under business scenario k. cos (c,s k The cosine vector distance can be used to represent the common and specific representations, with values ​​ranging from 0 to 1. A value of 1 indicates that the common and specific representations are completely similar; a value of 0 indicates that they are completely dissimilar. This determines the correlation value between the common and specific representations in business scenario k. As can be seen from the formula above, the higher the correlation value, the higher the correlation between the common and specific representations.

[0123] Of course, the correlation value between common and unique representations can also be expressed by other formulas. The specific formula is as follows:

[0124]

[0125] As can be seen from the above formula, the lower the correlation value, the higher the correlation between the common and unique characteristics. This specification does not limit the specific formula for determining the correlation value between the common and unique characteristics.

[0126] Then, the server can train the recommendation model with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks, and with the objective of maximizing the relevant representation values.

[0127] Of course, since the formulas for determining the correlation values ​​between common and specific representations vary across applications, the server can also optimize by minimizing the deviations between the first click-through rate (CTR) and the actual clicks corresponding to historical recommendation information, as well as the deviations between the second CTR and the actual clicks, and train the recommendation model by minimizing the correlation values. The server can choose the appropriate formula based on the specific formula used in the application.

[0128] In practical applications, because the above methods limit the distance between the common representation and the specific representations of each business scenario, it may lead to a lack of differentiation among the specific representations of each business scenario. In order to distinguish the specific representations of each business scenario, the server can improve the differentiation of the specific representations among each business scenario through orthogonal constraints.

[0129] In the embodiments described in this specification, the server can determine the difference value between the unique representations of a specified user in various business scenarios based on the user's unique representations in each business scenario. The lower the difference value, the greater the difference between the unique representations of the specified user in each business scenario. The specific formula is as follows:

[0130]

[0131] In the above formula, z t-k It can be used to represent the transpose of a specific user's unique representation in business scenario k. t_j It can be used to represent the unique characteristics of a specified user in business scenario j. k It can be used to represent a unique representation space under business scenario k. j It can be used to represent the unique representation space under business scenario j. orth (s k ,s j This value can be used to represent orthogonal constraints, ranging from 0 to 1. A value of 1 indicates complete similarity between unique representations, while a value of 0 indicates complete dissimilarity. The formula above shows that the lower the difference value, the greater the difference between the unique representations of a specified user across various business scenarios.

[0132] Then, the server can train the recommendation model with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks, and minimize the difference representation value.

[0133] Of course, the server can simultaneously determine the correlation between the common representation and the unique representation of a specified user in each business scenario, as well as the differences between the unique representations of a specified user in each business scenario. This ensures that the common representation of a specified user in different scenarios reflects the common preferences of the specified user in different scenarios, and that the unique representation of each business scenario reflects the preferences of the specified user in different aspects.

[0134] In other words, the server can optimize by minimizing the deviation between the first click-through rate (CTR) and the actual clicks corresponding to historical recommendation information, as well as the deviation between the second CTR and the actual clicks. It can then train the recommendation model by minimizing the relevance and difference representation values. Through multiple rounds of iterative training, the deviations can be continuously reduced and converged within a certain numerical range, thus completing the training process of the recommendation model.

[0135] It should be noted that the framework used in the recommendation model mentioned above can be an Enc-Dec network structure such as a fully connected forward propagation network (FFN), a fully connected neural network (FCNN), or a multi-head self-attention (MSA) network, or other forms of neural network structure. This specification does not limit the specific form of the framework used in the recommendation model.

[0136] As can be seen from the above process, this method can determine the correlation between user representations in various business scenarios by specifying the common representations of users in each business scenario, thereby obtaining the common preferences of users in each business scenario. Furthermore, by specifying the unique representations of users in each business scenario, differences in user representations exist between different business scenarios. Simultaneously, it determines the correlation between the common and unique representations of specified users in various business scenarios, as well as the differences between the unique representations of specified users in different business scenarios. This ensures that the common representations of specified users in different scenarios reflect their common preferences in different scenarios, and that the unique representations of each business scenario reflect their preferences in different aspects. Finally, with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to historical recommendation information, and the deviation between the second click-through rate and the actual clicks, and by minimizing the correlation representation values ​​and the difference representation values, the recommendation model is trained. This ensures both the correlation and the difference in user representations between different business scenarios.

[0137] After the recommendation model is trained, the embodiments in this specification can use the recommendation model to recommend information to users. The specific process is as follows: Figure 4 As shown.

[0138] Figure 4 A flowchart illustrating the method for recommending information provided in the embodiments of this specification.

[0139] S400: In response to a target user's information retrieval request in a specified business scenario, obtain candidate recommendation information in the specified business scenario and the target user's historical operation information in the specified business scenario.

[0140] In the embodiments of this specification, the server can respond to a target user's information retrieval request in a specified business scenario, and obtain candidate recommendation information in the specified business scenario as well as the target user's historical operation information in the specified business scenario. Here, the target user refers to the user currently performing a specified business in the specified business scenario. The candidate recommendation information may refer to filtered information in the specified business scenario, or it may refer to all information in the specified business scenario.

[0141] S402: For each candidate recommendation information, determine the representation corresponding to the candidate recommendation information in the specified business scenario based on the candidate recommendation information in the specified business scenario.

[0142] S404: Based on the target user's historical operation information in the specified business scenario, determine the unique characteristics of the target user in the specified business scenario.

[0143] After obtaining the candidate recommendation information in the specified business scenario and the target user's historical operation information in the specified business scenario, the representation corresponding to each candidate recommendation information in the specified business scenario and the target user's unique representation in the specified business scenario can be determined respectively. The method for determining the representation corresponding to each candidate recommendation information in the specified business scenario and the target user's unique representation in the specified business scenario is basically the same as the method mentioned in the model training process above, and will not be elaborated upon here.

[0144] S406: The representation of the candidate recommendation information in the specified business scenario and the unique representation of the target user in the specified business scenario are input into the pre-trained recommendation model to predict the click-through rate of the candidate recommendation information in the specified business scenario. The recommendation model is trained by the above model training method.

[0145] S408: Based on the click-through rate of each candidate recommendation information in the specified business scenario, determine the recommendation information to be recommended to the target user, and recommend the target recommendation information to the target user.

[0146] In the embodiments of this specification, the server can input the representation of the candidate recommendation information in a specified business scenario and the unique representation of the target user in the specified business scenario into a pre-trained recommendation model to predict the click-through rate of the candidate recommendation information in the specified business scenario.

[0147] Then, based on the click-through rate of each candidate recommendation information in the specified business scenario, the recommendation information to be recommended to the target user is determined as the target recommendation information, and the target recommendation information is recommended to the target user.

[0148] Specifically, the server can determine the click-through rate of each candidate recommendation in a specified business scenario, sort the candidate recommendations in descending order of click-through rate, and use the candidate recommendations ranked before the set sorting position as the target recommendation for the current time.

[0149] As can be seen from the above, the server can improve the recommendation effect of recommending information to target users by using the target recommendation information determined by the recommendation model.

[0150] The above describes a model training method provided by one or more embodiments of this specification. Based on the same idea, this specification also provides a corresponding model training apparatus, such as... Figure 5 As shown.

[0151] Figure 5 A schematic diagram of the structure of the model training apparatus provided in the embodiments of this specification, specifically including:

[0152] The acquisition module 500 is used to acquire historical operation information of a specified user in various business scenarios.

[0153] The determining module 502 is used to determine the common representation of the specified user in each business scenario based on the click objects involved in each historical operation information and other users associated with the specified user in each business scenario, and to determine the unique representation of the specified user in each business scenario based on the click objects involved in the historical operation information corresponding to that business scenario and other users associated with the specified user in that business scenario.

[0154] The prediction module 504 is used to input the common representation, the unique representation and the historical recommendation information into the recommendation model to be trained, and predict the click rate of the historical recommendation information in the business scenario as the first click rate, and the click rate of the historical recommendation information in each business scenario as the second click rate.

[0155] The training module 506 is used to train the recommendation model with the optimization objective of minimizing the deviation between the first click-through rate and the actual click situation corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual click situation.

[0156] Optionally, the determining module 502 is specifically used to construct a multi-scenario relationship graph corresponding to each business scenario based on the historical operation information of each user in each business scenario, taking each user as a user node and each clicked object in each business scenario as an object node. If a user node has historically performed a click operation on a clicked object, then the user node is connected to the object node corresponding to the clicked object by an edge. Based on the multi-scenario relationship graph, the common representation of the specified user in each business scenario is determined.

[0157] Optionally, the determining module 502 is specifically used to determine, from the multi-scenario relationship graph, each node that satisfies the set adjacency relationship with the user node corresponding to the specified user as an associated node, and determine the common representation of the specified user in each business scenario based on the representation corresponding to the associated node.

[0158] Optionally, the determining module 502 is specifically used to, for each associated node, determine the weight between the associated node and the user node corresponding to the specified user based on the degree of association between the object corresponding to the associated node and the specified user, and use this weight as the weight corresponding to the associated node. The object corresponding to the associated node includes clicked objects or associated users. Based on the representation corresponding to each associated node and the weight corresponding to each associated node, determine the common representation of the specified user in each business scenario.

[0159] Optionally, the determining module 502 is specifically used to construct a single-scenario relationship graph for each business scenario based on the historical operation information of each user in the business scenario, taking each user as a user node and each clicked object in the business scenario as an object node. If a user node has historically performed a click operation on a clicked object, then the user node and the object node corresponding to the clicked object are connected by an edge. Based on the single-scenario relationship graph, the unique representation of the specified user in the business scenario is determined.

[0160] Optionally, the determining module 502 is specifically used to determine, from the single-scene relationship graph, each node that satisfies the set adjacency relationship with the user node corresponding to the specified user as an associated node, and determine the unique representation of the specified user in the business scenario based on the representation corresponding to the associated node.

[0161] Optionally, the determining module 502 is specifically used to, for each associated node, determine the weight between the associated node and the user node corresponding to the specified user based on the degree of association between the object corresponding to the associated node and the specified user, and use this weight as the weight corresponding to the associated node. The object corresponding to the associated node includes clicked objects or associated users. Based on the representation corresponding to each associated node and the weight corresponding to each associated node, determine the unique representation of the specified user in this business scenario.

[0162] Optionally, the prediction module 504 is specifically used to include multiple recommendation sub-models in the recommendation model;

[0163] The unique representation and historical recommendation information are input into the sub-recommendation model corresponding to the business scenario to be trained, and the click-through rate for the historical recommendation information in the business scenario is predicted as the first click-through rate. The common representation and the historical recommendation information are input into the sub-recommendation model corresponding to each business scenario to be trained, and the click-through rate for the historical recommendation information in each business scenario is predicted as the second click-through rate.

[0164] Optionally, the training module 506 is specifically used to determine the correlation value between the common representation and the specific representation based on the common representation of the specified user in each business scenario and the specific representation of the specified user in that business scenario. The higher the correlation value, the higher the correlation between the common representation and the specific representation. The optimization objective is to minimize the deviation between the first click-through rate and the actual click situation corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual click situation. The recommendation model is trained to maximize the correlation value.

[0165] Optionally, the training module 506 is specifically used to determine the difference representation value between the unique representations of the specified user in each business scenario based on the unique representation of the specified user in each business scenario, wherein the lower the difference representation value, the greater the difference between the unique representations of the specified user in each business scenario, with the optimization objective of minimizing the deviation between the first click rate and the actual click situation corresponding to the historical recommendation information, and the deviation between the second click rate and the actual click situation, and to train the recommendation model by minimizing the difference representation value.

[0166] Figure 6 The schematic diagram of the device recommended for the information provided in the embodiments of this specification specifically includes:

[0167] The acquisition module 600 is used to respond to the information acquisition request of the target user in the specified business scenario, and acquire the candidate recommendation information in the specified business scenario and the historical operation information of the target user in the specified business scenario;

[0168] The first determining module 602 is used to determine the representation corresponding to each candidate recommendation information in a specified business scenario based on the candidate recommendation information in a specified business scenario.

[0169] The second determining module 604 is used to determine the unique characteristics of the target user in the specified business scenario based on the target user's historical operation information in the specified business scenario.

[0170] The prediction module 606 is used to input the representation of the candidate recommendation information in the specified business scenario and the unique representation of the target user in the specified business scenario into a pre-trained recommendation model to predict the click-through rate of the candidate recommendation information in the specified business scenario. The recommendation model is trained by the above model training method.

[0171] The recommendation module 608 is used to determine the recommendation information to be recommended to the target user based on the click-through rate of each candidate recommendation information in a specified business scenario, and recommend the target recommendation information to the target user.

[0172] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The provided model training method and the above Figure 4 The information provided is recommended by the method.

[0173] This instruction manual also provides Figure 7 The diagram shows the structure of the electronic device. Figure 7 At the hardware level, the electronic device includes a processor, internal bus, network interface, memory, and non-volatile memory, and may also include other hardware required for the business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to achieve the above-mentioned functions. Figure 1 The model training method and the above Figure 4 The information provided recommends certain methods. Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. In other words, the execution subject of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0174] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should understand that by simply performing some logic programming on the method flow using one of these hardware description languages ​​and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.

[0175] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0176] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0177] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0178] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0179] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0180] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0181] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0182] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0183] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0184] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0185] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0186] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this specification may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0187] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0188] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0189] The above description is merely an embodiment of this specification and is not intended to limit this specification. Various modifications and variations can be made to this specification by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this specification should be included within the scope of the claims of this specification.

Claims

1. A method for training a model, characterized in that, include: Obtain historical operation information of a specified user in each business scenario; based on the obtained historical operation information of each user in each business scenario, construct a multi-scenario relationship graph corresponding to each business scenario, with each user as a user node and each clicked object in each business scenario as an object node. If a user node has historically performed a click operation on a clicked object, then the user node and the object node corresponding to the clicked object are connected by an edge. From the multi-scenario relationship graph, nodes that satisfy a set adjacency relationship with the specified user are identified as associated nodes. For each associated node, the weight between the associated node and the specified user is determined based on the correlation between the object corresponding to the associated node and the specified user. This weight is used as the weight corresponding to the associated node. The object corresponding to the associated node includes clicked objects or associated users. Based on the representation of each associated node and the weight of each associated node, the common representation of the specified user in each business scenario is determined. Furthermore, for each business scenario, based on the historical operation information of each user in that business scenario, each user is treated as a user node, and each clicked object in that business scenario is treated as an object node, constructing the corresponding representation for that business scenario. A single-scenario relationship graph is constructed, wherein if a user node has historically clicked on a clicked object, the user node is connected to the object node corresponding to that clicked object by an edge. Based on the single-scenario relationship graph, a unique representation of the specified user in the business scenario is determined. The shared representation, the unique representation, and historical recommendation information are input into the recommendation model to be trained to predict the click-through rate (CTR) for the historical recommendation information in the business scenario, which is taken as the first CTR, and the CTR for the historical recommendation information in each business scenario is taken as the second CTR. The recommendation model is trained with the optimization objective of minimizing the deviation between the first CTR and the actual click situation corresponding to the historical recommendation information, and the deviation between the second CTR and the actual click situation.

2. The method as described in claim 1, characterized in that, Based on the single-scenario relationship graph, the unique representation of the specified user in the business scenario is determined, specifically including: identifying from the single-scenario relationship graph each node that satisfies a set adjacency relationship with the user node corresponding to the specified user as an associated node; and determining the unique representation of the specified user in the business scenario based on the representation corresponding to the associated node.

3. The method as described in claim 2, characterized in that, Based on the representations corresponding to the associated nodes, the unique representation of the specified user in this business scenario is determined, specifically including: for each associated node, determining the weight between the associated node and the user node corresponding to the specified user based on the degree of association between the object corresponding to the associated node and the specified user, which serves as the weight corresponding to the associated node, wherein the object corresponding to the associated node includes clicked objects or associated users; and determining the unique representation of the specified user in this business scenario based on the representations corresponding to each associated node and the weights corresponding to each associated node.

4. The method as described in claim 1, characterized in that, The recommendation model comprises multiple recommendation sub-models. The common representation, the unique representation, and historical recommendation information are input into the recommendation model to be trained to predict the click-through rate (CTR) for the historical recommendation information in the given business scenario, which is used as the first CTR. The click-through rate (CTR) for the historical recommendation information in each of the business scenarios is used as the second CTR. Specifically, this includes: inputting the unique representation and historical recommendation information into the sub-recommendation model corresponding to the business scenario to be trained, predicting the CTR for the historical recommendation information in the given business scenario, which is used as the first CTR; and inputting the common representation and historical recommendation information into the sub-recommendation models corresponding to each of the business scenarios to be trained, predicting the CTR for the historical recommendation information in each of the business scenarios, which is used as the second CTR.

5. The method as described in claim 1, characterized in that, The recommendation model is trained with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks. Specifically, this includes: determining a correlation value between the common representation and the specific representation based on the common representation of the specified user in each business scenario and the specific representation of the specified user in that business scenario; wherein a higher correlation value indicates a higher correlation between the common representation and the specific representation; and training the recommendation model with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks, and with the objective of maximizing the correlation value.

6. The method as described in claim 1 or 5, characterized in that, The recommendation model is trained with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks. Specifically, this includes: determining the difference representation value between the unique representations of the specified user in each business scenario based on the user's unique representations in each business scenario, wherein the lower the difference representation value, the greater the difference between the user's unique representations in each business scenario; and training the recommendation model with the optimization objective of minimizing the deviation between the first click-through rate and the actual clicks corresponding to the historical recommendation information, and the deviation between the second click-through rate and the actual clicks, and minimizing the difference representation value.

7. A method for information recommendation, characterized in that, include: In response to a target user's request for information in a specified business scenario, obtain candidate recommendation information in the specified business scenario and the target user's historical operation information in the specified business scenario; For each candidate recommendation information, determine the representation corresponding to the candidate recommendation information in the specified business scenario based on the candidate recommendation information in the specified business scenario; Based on the target user's historical operation information in the specified business scenario, determine the unique characteristics of the target user in the specified business scenario; The representation of the candidate recommendation information in the specified business scenario and the unique representation of the target user in the specified business scenario are input into a pre-trained recommendation model to predict the click-through rate of the candidate recommendation information in the specified business scenario. The recommendation model is trained by the method described in any one of claims 1 to 6. Based on the click-through rate of each candidate recommendation information in the specified business scenario, the recommendation information to be recommended to the target user is determined as the target recommendation information, and the target recommendation information is recommended to the target user.

8. A device for model training, characterized in that, include: The acquisition module is used to acquire historical operation information of a specified user in various business scenarios; The determination module is used to construct a multi-scenario relationship graph corresponding to each business scenario based on the historical operation information of each user in each business scenario, taking each user as a user node and each clicked object in each business scenario as an object node. If a user node has historically performed a click operation on a clicked object, then the user node and the object node corresponding to the clicked object are connected by an edge. From the multi-scenario relationship graph, nodes that satisfy the set adjacency relationship with the user nodes corresponding to the specified user are identified as associated nodes. For each associated node, the weight between the associated node and the user nodes corresponding to the specified user is determined based on the degree of association between the associated node and the object corresponding to the associated node. The object corresponding to the associated node includes clicked objects or associated users. Based on the representation of each associated node and the weight of each associated node, the common representation of the specified user in each business scenario is determined. And, for each business scenario, based on the historical operation information of each user in the business scenario, a single-scenario relationship graph corresponding to the business scenario is constructed, with each user as a user node and each clicked object in the business scenario as an object node. Wherein, if a user node corresponds to a user who has historically performed a click operation on a clicked object, then the user node and the object node corresponding to the clicked object are connected by an edge; according to the single-scenario relationship graph, the unique representation of the specified user in the business scenario is determined; the prediction module is used to input the common representation, the unique representation, and historical recommendation information into the recommendation model to be trained, and predict the click rate for the historical recommendation information in the business scenario as the first click rate, and the click rate for the historical recommendation information in each business scenario as the second click rate; the training module is used to train the recommendation model with the optimization objective of minimizing the deviation between the first click rate and the actual click situation corresponding to the historical recommendation information, and the deviation between the second click rate and the actual click situation.

9. An information recommendation device, characterized in that, include: The acquisition module is used to respond to the information acquisition request of the target user in the specified business scenario, and to acquire the candidate recommendation information in the specified business scenario and the historical operation information of the target user in the specified business scenario. The first determining module is used to determine the representation corresponding to each candidate recommendation information in the specified business scenario based on the candidate recommendation information in the specified business scenario. The second determining module is used to determine the unique characteristics of the target user in the specified business scenario based on the target user's historical operation information in the specified business scenario. The prediction module is used to input the representation of the candidate recommendation information in the specified business scenario and the unique representation of the target user in the specified business scenario into a pre-trained recommendation model to predict the click-through rate of the candidate recommendation information in the specified business scenario. The recommendation model is trained by the method described in any one of claims 1 to 6. The recommendation module is used to determine the recommended information to be recommended to the target user based on the click-through rate of each candidate recommendation information in a specified business scenario, and recommend the target recommendation information to the target user.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 6 or 7.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method described in any one of claims 1 to 6 or 7.

Images