A recommendation method and apparatus

Abstract

The specification discloses a recommendation method and device, determines each candidate recommendation sequence determined by combining a specified number of candidate recommendation objects according to different positions in a sequence, and a specified sequence in which a user has performed a specified operation in the past, obtains candidate recommendation features determined based on the content of each candidate recommendation object in the candidate recommendation sequence and the position of the candidate recommendation object, and each specified feature through a feature extraction layer of a click rate estimation model, and further determines the click rate of each candidate recommendation sequence through a prediction layer of the click rate estimation model, to determine a recommendation sequence recommended to the user based on each click rate. When predicting the click rate of each candidate recommendation sequence, the method not only determines the click rate based on the content of the candidate recommendation object contained in the candidate recommendation sequence, but also based on the position of each candidate recommendation object, so that the determined click rate is more accurate and the recommendation efficiency is higher.

Description

Technical Field

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

[0002] Currently, with the development of computer technology, how to recommend suitable objects to users (such as search results in search scenarios, and natural results displayed to users by service providers) has become one of the problems that service providers need to solve. Recommendation methods, because they can recommend suitable objects to users based on users' historical behavior and the characteristics of each object, are widely used in scenarios where service providers recommend objects to users.

[0003] In existing technologies, a common recommendation method is based on objects clicked historically by the user. Specifically, the service provider can obtain the user's historical behavior sequence to determine the objects the user has clicked in the past. Then, candidate recommendation objects can be obtained, and for each candidate object, the similarity between that candidate object and the objects clicked historically by the user can be determined. Finally, based on the determined similarity scores, the corresponding candidate recommendation objects are recommended to the user.

[0004] However, existing technologies do not take into account the impact of other candidate recommendations on the click-through rate of the recommended object when making recommendations, resulting in low recommendation efficiency when recommending to users based on the identified candidate objects. Summary of the Invention

[0005] This specification provides a recommended method and apparatus 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 recommended method, including:

[0008] Based on the recommendation request sent by the user, determine a specified number of candidate recommendation objects, as well as a specified sequence of operations that the user has historically performed;

[0009] The candidate recommendations are combined according to their different positions in the sequence to determine the candidate recommendation sequence;

[0010] For each candidate recommendation sequence, the candidate recommendation sequence and the specified sequences are used as inputs to the feature extraction layer of a pre-trained click-through rate prediction model. The result is the candidate recommendation features determined by the feature extraction layer based on the content of each candidate recommendation object in the candidate recommendation sequence and the position of each candidate recommendation object, as well as the specified features.

[0011] The candidate recommendation features and the specified features are used as inputs to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer.

[0012] Based on the click-through rate of each candidate recommendation sequence, a recommendation sequence is determined and recommended to the user.

[0013] Optionally, the candidate recommendation sequence is used as input to the feature extraction layer of a pre-trained click-through rate prediction model to obtain the candidate recommendation features output by the feature extraction layer, which are determined based on the content and position of each candidate recommendation object in the candidate recommendation sequence. Specifically, these features include:

[0014] For each candidate recommendation sequence, the candidate recommendation sequence is used as input and fed into the feature extraction layer of the pre-trained click-through rate prediction model;

[0015] For each candidate recommendation object in the candidate recommendation sequence, the object characteristics of the candidate recommendation object are determined based on the content of the candidate recommendation object and the position of the candidate recommendation object in the candidate recommendation sequence;

[0016] Determine the similarity between the candidate recommendation object and other candidate recommendation objects in the candidate recommendation sequence, and enhance the object features of the candidate recommendation object based on each similarity and the object features of the other candidate recommendation objects corresponding to each similarity.

[0017] Based on the enhancement results of each object's features, the candidate recommendation features of the candidate recommendation sequence are determined.

[0018] Optionally, the candidate recommendation features and the specified features are used as input to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer, specifically including:

[0019] The candidate recommendation features and the specified features are used as inputs to the fusion layer of the click-through rate prediction model. The enhanced result of the candidate recommendation features enhanced according to the specified features and the fusion result of the specified features are used as the fusion features.

[0020] The fused features are used as input to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer.

[0021] Optionally, the specified operation includes multiple operation types;

[0022] Based on the recommendation request sent by the user, determine the specified sequences of operations that the user has historically performed, specifically including:

[0023] Based on the recommendation request sent by the user, determine the user identifier carried in the recommendation request, and based on the user identifier, determine the sequences that the user has browsed in history, as each browsing sequence;

[0024] Determine the user's historical behavior, and for each browsing sequence, determine the user's operation on the browsing sequence based on the historical behavior, and determine the operation type to which the operation belongs, as a label for the browsing sequence;

[0025] For each operation type, each browsing sequence labeled with that operation type is determined from each browsing sequence and used as the designated sequence corresponding to that operation type.

[0026] Optionally, for each target sequence, the priorities of the objects contained in that target sequence are not entirely the same;

[0027] The candidate recommendation sequence is taken as input and fed into the feature extraction layer of a pre-trained click-through rate prediction model. The result is the output of the feature extraction layer, which determines the candidate recommendation features based on the content and position of each candidate recommendation object within the candidate recommendation sequence. Specifically, this includes:

[0028] Determine the priority of each candidate recommendation object included in the candidate recommendation sequence;

[0029] The candidate recommendation sequence is used as input and fed into the feature extraction layer of a pre-trained click-through rate prediction model to obtain the candidate recommendation features output by the feature extraction layer, which are determined based on the content of each candidate recommendation object contained in the candidate recommendation sequence, the position of each candidate recommendation object, and the priority of each candidate recommendation object.

[0030] Optionally, the click-through rate prediction model is trained in the following manner:

[0031] For each user, identify the sequences that the user has browsed in history as target sequences, and determine the labels of the target sequences based on the user's historical behavior. Also, identify the sequences from the target sequences in which the user has performed a specified operation in history as specified sequences.

[0032] The target sequences and the specified sequences are used as inputs to the feature extraction layer of the click-through rate prediction model to be trained. The output of the feature extraction layer is the target features determined based on the content of each object contained in the target sequence and the position of each object, as well as the specified features.

[0033] Each target feature and each specified feature are used as input to the prediction layer of the click-through rate prediction model to obtain the click-through rate corresponding to each target sequence output by the prediction layer.

[0034] The loss is determined based on the click-through rate and its label for each target sequence, and the model parameters of the click-through rate prediction model are adjusted based on the loss.

[0035] Optionally, the candidate recommendation features and the specified features are used as input to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer, specifically including:

[0036] The candidate recommendation features and the specified features are used as inputs to the prediction layer of the click-through rate prediction model;

[0037] For each operation type, the probability of a candidate recommendation feature corresponding to that operation type is determined based on the similarity between each specified sequence of that operation type and the candidate recommendation feature.

[0038] The click-through rate of the candidate recommendation sequence is determined based on the probability of each operation type corresponding to the candidate recommendation features and the preset weight of each operation type.

[0039] This specification provides a recommended device, comprising:

[0040] The receiving module is used to determine a specified number of candidate recommendation objects and a specified sequence of operations that the user has performed in the past, based on the recommendation request sent by the user.

[0041] The combination module is used to combine the candidate recommendation objects according to their different positions in the sequence to determine the candidate recommendation sequence;

[0042] The feature determination module is used to take each candidate recommendation sequence as input, the candidate recommendation sequence and the specified sequences as input, and input the feature extraction layer of the pre-trained click-through rate prediction model to obtain the candidate recommendation features determined by the feature extraction layer based on the content of each candidate recommendation object contained in the candidate recommendation sequence and the position of each candidate recommendation object, as well as the specified features.

[0043] The prediction module is used to take the candidate recommendation features and the specified features as inputs to the prediction layer of the click-through rate prediction model, and obtain the click-through rate of the candidate recommendation sequence output by the prediction layer.

[0044] The determination module is used to determine the recommendation sequence based on the click-through rate of each candidate recommendation sequence and to recommend it to the user.

[0045] This specification provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the recommended method described above.

[0046] 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 recommended method described above.

[0047] The above-mentioned technical solutions adopted in this specification can achieve the following beneficial effects:

[0048] In the recommendation method provided in this specification, candidate recommendation sequences are determined by combining a specified number of candidate recommendation objects according to different positions in the sequence, as well as a specified sequence of operations that the user has historically performed. Then, through the feature extraction layer of the click-through rate prediction model, candidate recommendation features are obtained based on the content and position of each candidate recommendation object in the candidate recommendation sequence, as well as each specified feature. Finally, through the prediction layer of the click-through rate prediction model, the click-through rate of each candidate recommendation sequence is determined, and based on each click-through rate, a recommendation sequence to be recommended to the user is determined.

[0049] As can be seen from the above method, when predicting the click-through rate of each candidate recommendation sequence, this method not only bases the prediction on the content of the candidate recommendation objects contained in the candidate recommendation sequence, but also on the position of each candidate recommendation object, making the determined click-through rate more accurate and the recommendation efficiency higher. Attached Figure Description

[0050] 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:

[0051] Figure 1 A flowchart illustrating the recommended method provided in this specification;

[0052] Figure 2 This is a flowchart illustrating the training method for the click-through rate prediction model provided in this manual.

[0053] Figure 3 This is a schematic diagram illustrating the training process of the click-through rate prediction model provided in this manual.

[0054] Figure 4 This is a schematic diagram of the recommended device provided in this specification;

[0055] Figure 5 The corresponding information provided in this specification Figure 1 A schematic diagram of an electronic device. Detailed Implementation

[0056] 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.

[0057] Generally, in the recommendation field, click-through rate (CTR) prediction models can be used to determine the click-through rate of users for different objects (such as content), and based on the CTR of each object, recommended objects to be shown to users can be determined from candidate objects. The determined recommended objects are then displayed on the page in sequence for recommendation to users.

[0058] However, when a user clicks on an object in a sequence, the reason for clicking may include not only the object's content itself, but also its position in the sequence and the influence of other objects in the sequence. This object can be a product, service, merchant, business object, etc., and the specific type of object can be set as needed; this manual does not impose any restrictions on it.

[0059] Taking a page containing three products, A, B, and C, as an example, where products A and C are more expensive, and product B is moderately priced, the reason a user clicks on product B might be because it is the cheapest on the page, rather than because the user prefers moderately priced products.

[0060] In summary, using only the objects clicked by users and candidate objects to determine similarity will result in an inaccurate click-through rate.

[0061] The technical solutions provided in the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0062] Figure 1 The flowchart of the recommended method provided in this specification includes the following steps:

[0063] S100: Based on the received recommendation request, determine a specified number of candidate recommendation objects and a specified sequence of operations that the user has historically performed.

[0064] Currently, in the recommendation field, recommendation methods can identify user data based on the user identifier in the received recommendation request. Then, based on this user data, a specified number of candidate recommendation objects matching the user's preferences are selected from a pool of candidate objects to form a recommendation sequence, which is then displayed to the user on the page. In other words, the service provider actually recommends a sequence containing a specified number of candidate recommendation objects to the user. Furthermore, determining the click-through rate (CTR) of a recommendation sequence containing recommended objects is more accurate than simply determining the CTR of a user on a candidate recommendation object.

[0065] Unlike current techniques that determine click-through rates (CTRs) based on the similarity between candidate recommendations and objects clicked historically by the user, this method fails to consider the influence of other candidate recommendations within the recommendation sequence, as well as the candidate recommendation's position within that sequence, leading to inaccurate CTRs. This specification provides a novel recommendation method that determines the CTR of a candidate recommendation sequence based on the content and position of each candidate recommendation, resulting in a more accurate CTR.

[0066] In one or more embodiments provided in this specification, the recommendation method is designed for scenarios in which candidate recommendation objects are presented to the user in units of candidate recommendation sequences, and the recommendation method can be specifically executed by a server.

[0067] Based on this, a specified number of candidate recommendation objects and user data can be determined according to the recommendation request sent by the user and the number of candidate recommendation objects contained in the candidate recommendation sequence.

[0068] Specifically, the server can receive recommendation requests sent by users and determine the specified number of items carried in the recommendation request and the user identifier corresponding to the user.

[0069] Then, the server can determine the candidate recommendation objects that specify the number of users from the pre-determined candidate recommendation objects.

[0070] Finally, the server can obtain the sequence of specified operations that the user has performed in history based on the user identifier, and use these as specified sequences, which are user data that represent the user's preferences.

[0071] Of course, each candidate recommendation object can be determined based on the keywords carried in the recommendation request, or it can be determined in advance by the server based on user data. The specific method of determining each candidate recommendation object can be set as needed, and this manual does not impose any restrictions on it.

[0072] S102: Combine the candidate recommendation objects according to their different positions in the sequence to determine the candidate recommendation queue.

[0073] In one or more embodiments provided in this specification, even if the candidate recommendation sequence consists of the same candidate recommendation objects, if the content of each candidate recommendation object in the sequence is different, the click rate of the user for the candidate recommendation content will also be different. Therefore, the server can combine each candidate recommendation object into different candidate recommendation queues according to their different positions in the sequence.

[0074] Specifically, the server can arrange and combine the recommended objects, sort the candidate recommended objects according to their different orders in the sequence, and use the sorting results as candidate recommendation sequences.

[0075] S104: For each candidate recommendation sequence, the candidate recommendation sequence and the specified sequences are used as inputs to the feature extraction layer of the pre-trained click-through rate prediction model to obtain the candidate recommendation features and the specified features output by the feature extraction layer, which are determined based on the content of each candidate recommendation object contained in the candidate recommendation sequence and the position of each candidate recommendation object.

[0076] In one or more embodiments provided in this specification, after determining a candidate recommendation sequence, the server can determine the characteristics of the candidate recommendation sequence to facilitate subsequent steps for determining the click-through rate (CTR) based on these characteristics. As previously described, unlike methods that determine CTR based on the similarity between user-clicked objects and candidate recommendation objects, the recommendation method provided in this specification determines the CTR not only based on the content of the candidate recommendation objects but also based on the position of the candidate recommendation objects within the candidate recommendation sequence. Therefore, the server can determine the candidate recommendation features of the candidate recommendation sequence based on the candidate recommendation objects included in the sequence and the position of each candidate recommendation object within the sequence.

[0077] Specifically, for each candidate recommendation sequence, the server can use that candidate recommendation sequence as input to the feature extraction layer of a pre-trained click-through rate prediction model.

[0078] The feature extraction layer of this click-through rate prediction model can determine the object features of each candidate recommendation object in the candidate recommendation sequence based on the content of the candidate recommendation object and the position of the candidate recommendation object in the candidate recommendation sequence.

[0079] Therefore, this feature extraction layer can fuse the identified candidate recommendation objects to determine the candidate recommendation features of the candidate recommendation sequence. This fusion method can include various techniques such as concatenation and addition. Since fusing multiple features into one feature is already a relatively mature technical solution, this solution will not elaborate on it further.

[0080] Furthermore, in this specification, the click-through rate (CTR) of each candidate recommendation sequence needs to be determined based on each specified sequence of operations performed by the user in the past. Based on this, while determining the features of each candidate recommendation, the server can also input each specified sequence as input into the feature extraction layer of the CTR prediction model, obtaining the specified features output by the feature extraction layer based on the content of the object contained in the specified sequence and the position of the object in the specified sequence.

[0081] Furthermore, each candidate recommendation object in the same candidate recommendation sequence will be affected by other candidate recommendation objects in the same sequence. Therefore, in order to accurately determine the object features and thus determine more accurate candidate recommendation features to ensure the accuracy of the estimated click-through rate, when determining the object features, the object features of each candidate recommendation object can be enhanced by using the object features of other candidate recommendation objects in the candidate recommendation sequence to which the candidate recommendation object belongs.

[0082] Specifically, the server can first input the candidate recommendation sequence into the feature extraction layer of the click-through rate prediction model. The feature extraction layer then determines the object features of each candidate recommendation object in the candidate recommendation sequence based on the content and location of the candidate recommendation object.

[0083] Then, the feature extraction layer can determine the similarity between each of the other candidate recommendation objects and the candidate recommendation object based on the object features of the candidate recommendation object and the object features of other candidate recommendation objects contained in the candidate recommendation sequence.

[0084] Finally, based on the determined similarities and the object features of other candidate recommendation objects corresponding to each similarity, the object features of the candidate recommendation object are enhanced.

[0085] Finally, the server can use the enhanced results as object features of the candidate recommendation object.

[0086] S106: The candidate recommendation features and the specified features are used as inputs to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer.

[0087] In one or more embodiments provided in this specification, the click-through rate (CTR) of each candidate recommendation sequence is predicted based on specified sequences of specified operations performed by the user in the past, and the similarity between candidate recommendation sequences composed of candidate recommendation objects. Therefore, after determining the candidate recommendation features and specified features, the server can use the determined candidate recommendation features and specified features as input to the prediction layer of the CTR prediction model to determine the CTR of each candidate recommendation sequence.

[0088] Specifically, the server can input the candidate recommendation features and the determined specified features of each candidate recommendation sequence into the prediction layer of the click-through rate prediction model.

[0089] This prediction layer can determine the similarity between the candidate recommendation feature and each specified feature based on the identified candidate recommendation feature and each specified feature, and determine the click-through rate of the candidate recommendation feature based on the type of the specified operation and the weight corresponding to each similarity.

[0090] Taking the specified operation as "click" as an example, if the similarity is 80%, the determined weight could be 0.8, etc. Taking the specified operation as "leave" as an example, if the similarity is 80%, the determined weight could be 0.2, etc. How to determine the weight based on similarity and the type of specified operation can be set as needed; this manual does not impose any restrictions on this.

[0091] S108: Determine the recommended sequence based on the click-through rate of each candidate recommended sequence, and recommend it to the user.

[0092] In one or more embodiments provided in this specification, after determining the click-through rate corresponding to each candidate recommendation sequence, the server can sort the multiple candidate recommendation sequences according to the determined click-through rates, and determine the recommendation sequence to be displayed to the user based on the sorting.

[0093] based on Figure 1 This recommendation method identifies candidate recommendation sequences by combining a specified number of candidate recommendation objects according to their different positions within the sequence, as well as a specified sequence of actions the user has historically performed. Then, through the feature extraction layer of a click-through rate (CTR) prediction model, candidate recommendation features are obtained based on the content and position of each candidate recommendation object in the candidate recommendation sequence, along with other specified features. Finally, the prediction layer of the CTR prediction model determines the CTR of each candidate recommendation sequence, and based on these CTRs, the recommended sequence to be presented to the user is determined. This method predicts the CTR of each candidate recommendation sequence not only based on the content of the candidate recommendation objects within the sequence but also on the position of each candidate object, resulting in more accurate CTRs and higher recommendation efficiency.

[0094] In addition, the click-through rate prediction model used in this recommendation method can typically be trained in the following way:

[0095] First, the server can identify each sequence that the user has browsed in history as a target sequence for each user, and determine the label of each target sequence based on the user's historical behavior. It can also identify each sequence from the target sequences in which the user has performed a specified operation in history as a specified sequence.

[0096] Secondly, for each target sequence, the server can take the target sequence and each specified sequence as input and input them into the feature extraction layer of the click-through rate prediction model to be trained, and obtain the target features and each specified feature output by the feature extraction layer based on the content and position of each object contained in the target sequence.

[0097] Then, the server takes the target feature and each specified feature as input and inputs them into the prediction layer of the click-through rate prediction model to obtain the click-through rate of the target sequence output by the prediction layer;

[0098] Finally, the server can determine the loss based on the click-through rate and its annotation for each target sequence, and adjust the model parameters of the click-through rate prediction model based on the loss.

[0099] It should be noted that the server used to train the model and the server used to execute the recommended method provided in this manual can be the same server or different servers.

[0100] Furthermore, the accuracy of results determined using only positive or negative feedback is obviously lower than that determined using both positive and negative feedback simultaneously. Therefore, in order to more accurately determine the click-through rate of each candidate recommendation sequence, the server can also set up multiple types of specified sequences.

[0101] Specifically, the server can determine the user identifier carried in the recommendation request sent by the user, and determine the series of browsing history that the user has viewed in the past based on the user identifier, as each browsing series.

[0102] Then, after determining each browsing sequence, the server can also determine the user's historical behavior, and for each browsing sequence, based on the determined historical behavior, determine the user's operation on that browsing sequence, and determine the operation type to which the operation belongs, as a label for that browsing sequence.

[0103] Finally, after determining the annotations for each browsing sequence, the server can identify the browsing sequences labeled with that operation type from the browsing sequences for each operation type, and use them as the designated sequences corresponding to that operation type.

[0104] Furthermore, when determining the click-through rate (CTR) of a candidate recommendation sequence, the server can simultaneously determine the CTR of the user for that candidate recommendation sequence based on multiple types of operations performed by the user.

[0105] Specifically, firstly, the server can use the candidate recommendation features and each specified feature as input to the prediction layer of the click-through rate prediction model.

[0106] Then, for each operation type, the prediction layer can determine the probability that the candidate recommendation feature corresponds to the operation type based on the similarity between each specified sequence of the operation type and the candidate recommendation feature.

[0107] Finally, based on the probability of each operation type corresponding to the candidate recommendation feature, and the preset weight of each operation type, the server can determine the click-through rate of the candidate recommendation sequence.

[0108] Furthermore, when making recommendations to users, the objects in the candidate recommendation sequence often correspond to different priorities. For example, the objects may include both advertisements and organic results, with organic results having higher priority and advertisements having lower priority, or vice versa. The service provider is more inclined to encourage users to click on advertisements. Therefore, when determining the candidate recommendation features of a candidate recommendation sequence, the server can also determine the priority of each candidate recommendation object within that sequence.

[0109] The server can then use the candidate recommendation sequence as input to the feature extraction layer of a pre-trained click-through rate prediction model, and obtain the candidate recommendation features output by the feature extraction layer, which are determined based on the content of each candidate recommendation object in the candidate recommendation sequence, the position of each candidate recommendation object, and the priority of each candidate recommendation object.

[0110] Following the same approach, this manual also provides a training method for a click-through rate prediction model, such as... Figure 2 As shown.

[0111] Figure 2 This is a flowchart illustrating the training method for the click-through rate prediction model provided in this manual, wherein:

[0112] S200: For each user, determine the sequences that the user has browsed in history as target sequences, and determine the labels of the target sequences based on the user's historical behavior, and determine the sequences from the target sequences that the user has performed specified operations in history as specified sequences.

[0113] Typically, click-through rate (CTR) prediction models are pre-trained using a server that trains the model, based on training samples. This specification provides a method for training a CTR prediction model, which can also be executed by the server used for training the model.

[0114] The training model can be divided into a sample generation stage and a model training stage. In the sample generation stage, samples for training the model can be determined according to the needs of the model and training. In this specification, the server can first determine the training samples for training the click-through rate prediction model. Since the click-through rate prediction model in this specification is used to predict the click-through rate of candidate recommendation sequences, the server can first determine the sequences that the user has viewed in history to determine the training samples.

[0115] Based on this, the server can determine the sequences that each user has browsed in history. These sequences can be pre-stored and retrieved by the server from the pre-stored historical browsing logs corresponding to each user based on the user's identifier, or they can be retrieved directly from the user's terminal. The historical browsing logs include the sequences browsed by the user in history, the objects contained in each sequence, and the operations performed by the user on each sequence. The specific content of the historical browsing logs and how they are retrieved can be configured as needed, and this specification does not impose any restrictions on this.

[0116] After determining the user's browsing history logs, the server can use these logs to identify the sequences the user has browsed in the past, which can then be used as the user's target sequences.

[0117] Similarly, the server can also determine the corresponding annotations for each target sequence based on the user's historical browsing logs, i.e., the user's historical behavior. These annotations can include two types: "clicked" and "not clicked".

[0118] Of course, for each target sequence of a user, the operations performed by the user in each target sequence will largely reflect the user's preferences for that target sequence. Therefore, based on the type of operations performed by the user in each target sequence, the target sequences in which the user has performed specified operations can be determined as the specified sequences. The specified operation can be a "click" operation, a "favorite" operation, a "leave" operation, etc., and the specific type of the specified operation can be set as needed; this manual does not impose any restrictions on this.

[0119] S202: For each target sequence, the target sequence and the specified sequences are used as inputs to the feature extraction layer of the click-through rate prediction model to be trained, and the target features and specified features are obtained from the output of the feature extraction layer based on the content of each object contained in the target sequence and the position of each object.

[0120] In one or more embodiments provided in this specification, after determining the target sequence, the server can determine the characteristics of each target sequence so that subsequent steps for determining the click-through rate can be performed based on the characteristics of the target sequence.

[0121] Unlike current methods that use user-clicked objects to determine the click-through rate (CTR) of candidate recommendations, the CTR prediction model provided in this specification not only determines the CTR of the candidate recommendation sequence based on the content of each candidate recommendation object, but also determines the CTR of the candidate recommendation sequence based on the position of each candidate recommendation content within the candidate recommendation sequence.

[0122] Based on this, the server can determine target features based on the objects contained in the target sequence and the position of each object in the target sequence.

[0123] Specifically, for each target sequence, the server can use that target sequence as input to the feature extraction layer of the click-through rate prediction model.

[0124] The feature extraction layer of this click-through rate prediction model can determine the object features corresponding to each object in the target sequence based on the object's own content and its position in the target sequence.

[0125] Therefore, this feature extraction layer can fuse the object features corresponding to each identified object to determine the target feature corresponding to the target sequence. This fusion method can include various techniques such as concatenation and addition. Since fusing multiple features into one feature is already a relatively mature technical solution, this solution will not elaborate on it further.

[0126] Meanwhile, the server can also input each specified sequence into the feature extraction layer to obtain the specified features output by the feature extraction layer for each specified sequence, based on each object contained in the specified sequence and the position of each object in the specified sequence.

[0127] Of course, when determining the target feature and the specified feature, the server can also determine the user feature corresponding to the user, and update the target feature and the specified feature based on the user feature. This update method can employ the technical means used in the aforementioned fusion process.

[0128] Furthermore, for each object, in order to more accurately reflect the influence of other objects in the target sequence on that object, the server can also use the object features of other objects to enhance the object features of that object when determining the target features.

[0129] Specifically, the server can first input the target sequence into the feature extraction layer, and the feature extraction layer can determine the object features of each object contained in the target sequence.

[0130] Then, the feature extraction layer can determine the similarity between the object and each of the other objects in the target sequence based on the object features and the object features of each other object contained in the target sequence. Based on the determined similarity and the object features of each other object corresponding to each similarity, the object features of the object are enhanced. The enhancement technique can be the same as the technique used in the update process described above.

[0131] Finally, the server can fuse the enhanced results of the object features of each object to determine the target features of the target sequence.

[0132] Of course, the above enhancements can also be used when determining the specified features to determine a more accurate click-through rate.

[0133] S204: The target feature and the specified features are used as inputs to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the target sequence output by the prediction layer.

[0134] In one or more embodiments provided in this specification, since it is necessary to predict the relationship between different features contained in the sequence and the different positions of different features and the click rate of the user on the sequence based on the target sequence that the user has browsed in history and the relationship between each specified sequence in which the user has performed a specified operation, the server can predict the click rate corresponding to the target sequence based on the target feature and each specified feature after determining the target feature and each specified feature.

[0135] Specifically, the server can use the target features and various specified features as inputs to the prediction layer of the click-through rate prediction model.

[0136] This prediction layer can determine the similarity between each specified feature and the target feature based on the acquired target feature and each specified feature, and determine the click-through rate of the target sequence based on the type of the specified operation and the weight of each similarity.

[0137] Taking the specified operation as "click" as an example, if the similarity is 80%, the determined weight could be 0.8, etc. Taking the specified operation as "leave" as an example, if the similarity is 80%, the determined weight could be 0.2, etc. How to determine the weight based on similarity and the type of specified operation can be set as needed; this manual does not impose any restrictions on this.

[0138] Furthermore, when determining the click-through rate (CTR) corresponding to the target sequence, the prediction layer can also fuse the target feature with various specified features, and input the fusion result into network structures such as fully connected layers or multilayer perceptrons to determine the CTR corresponding to the fused features. Of course, the specific network structure of this prediction layer can be set as needed, and this specification does not impose any restrictions on it.

[0139] Furthermore, in order to further enhance the features of the target sequence and determine a more accurate click-through rate, the server can also input the target features of the target sequence and each specified feature into the fusion layer of the click-through rate prediction model for fusion before inputting the target features of the target sequence into the prediction layer.

[0140] This fusion layer can determine the similarity between each target feature and the specified feature, determine the weights based on the similarity, and enhance the target feature according to the determined weights and the specified features. The fusion layer can then output the enhanced result as the fused feature.

[0141] The server can then input the determined fusion features into the prediction layer of the click-through rate prediction model, and the prediction layer will output the click-through rate corresponding to the target sequence.

[0142] S206: Determine the loss based on the click-through rate and its label for each target sequence, and adjust the model parameters of the click-through rate prediction model based on the loss.

[0143] In one or more embodiments provided in this specification, after determining the click-through rate of a target sequence, the server can train the click-through rate prediction model based on the annotation and click-through rate of each target sequence.

[0144] Specifically, the server can determine the loss based on the click-through rate and annotation of each target sequence, and adjust the model parameters of the click-through rate prediction model based on the loss to complete the training of the click-through rate prediction model.

[0145] In addition, based on Figure 2 The training method for the click-through rate (CTR) prediction model shown follows the same logic. This manual also provides a schematic diagram of the CTR prediction model training process, as follows: Figure 3 As shown.

[0146] Figure 3 This is a schematic diagram of the training process for the click-through rate (CTR) prediction model provided in this specification. The server first obtains the user's historical browsing logs and, based on these logs, determines the user's target sequences and the specified sequences in which the user has performed specified operations. Then, for each target sequence, the target sequence and the specified sequences are input to the feature extraction layer of the CTR prediction model to obtain the target features and the specified features of that target sequence. These target features and the specified features are then input to the fusion layer of the CTR prediction model to determine the fused features of the target sequence. The server can then input these fused features into the prediction layer of the CTR prediction model to obtain the CTR of the target sequence output by the prediction layer. Finally, based on the CTR and annotations of the target sequence, a loss is determined, and the CTR prediction model is trained according to this loss.

[0147] This historical browsing log is merely an example of how to obtain the user's target sequence. The specific method for obtaining the user's target sequence and specifying the sequence can be set as needed, and this manual does not impose any restrictions on this.

[0148] based on Figure 2 The recommended method, based on the user's target sequences and specified sequences of operations performed by the user, uses a click-through rate (CTR) prediction model's feature extraction layer to determine target features (based on the objects contained in the target sequence and their positions within the sequence) and specified features. Then, the CTR prediction model's prediction layer determines the CTR based on the target features and specified features. Finally, the loss is determined based on the CTR of each target sequence and the annotations, and the model is trained accordingly. This approach predicts page CTR not only based on the contained objects but also on their positions, resulting in a more accurate prediction model.

[0149] Furthermore, in order to train a more accurate click-through rate prediction model, various types of labels and specified sequences can be set when determining the labels of the target sequence and each specified sequence.

[0150] Specifically, for each target sequence, the server can determine the operation performed by the user in that target sequence based on the user's historical behavior, and determine the operation type to which the operation belongs, as a label for that target sequence.

[0151] Then, after determining the labels of each target sequence, the server can, for each operation type, determine each target sequence labeled as that operation type from the target sequences, as the designated sequences corresponding to that operation type.

[0152] Furthermore, when making recommendations to users, the objects in the candidate recommendation sequence often correspond to different priorities. For example, the objects may include both advertisements and organic results, with organic results having higher priority and advertisements having lower priority, or vice versa. The service provider is more inclined to encourage users to click on advertisements. Therefore, when determining the target features of a target sequence, the server can also determine the priority of each object contained in that target sequence.

[0153] The server can then use the target sequence as input to the feature extraction layer of the click-through rate prediction model to be trained, and obtain the target features of the target sequence based on the content of each object contained in the target sequence, the position of each object in the target sequence, and the priority of each object.

[0154] It should be noted that, in order to ensure the uniformity of the model input dimensions and facilitate better model training, when determining the user's specified sequence, a specified number of sequences must be determined as the user's specified sequence. If the number of sequences contained in the user's specified sequence is less than the specified number, the sequence in which the user performs the specified operation and the empty sequence can be combined into the specified sequence.

[0155] It should be noted that all actions involving the acquisition of signals, information, or data in this application are carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization granted by the owner of the relevant device.

[0156] The above are recommended methods provided by one or more embodiments of this specification. Based on the same idea, this specification also provides corresponding recommended devices, such as... Figure 4 As shown.

[0157] Figure 4 The recommended apparatus provided in this specification includes:

[0158] The receiving module 300 is used to determine a specified number of candidate recommendation objects and a specified sequence of specified operations that the user has performed in the past, based on the recommendation request sent by the user.

[0159] The combination module 302 is used to combine the candidate recommendation objects according to their different positions in the sequence to determine the candidate recommendation sequence.

[0160] The feature determination module 304 is used to take each candidate recommendation sequence as input, the candidate recommendation sequence and the specified sequences as input, and input the feature extraction layer of the pre-trained click-through rate prediction model to obtain the candidate recommendation features determined by the feature extraction layer based on the content of each candidate recommendation object contained in the candidate recommendation sequence and the position of each candidate recommendation object, as well as the specified features.

[0161] The prediction module 306 is used to input the candidate recommendation features and the specified features as inputs to the prediction layer of the click-through rate prediction model, and obtain the click-through rate of the candidate recommendation sequence output by the prediction layer.

[0162] The recommendation module 308 is used to determine the recommendation sequence based on the click rate corresponding to each candidate recommendation sequence and to recommend it to the user.

[0163] Optionally, the method further includes:

[0164] The training module 310 is used to determine, for each user, the sequences that the user has browsed in history as target sequences, and to determine the labels of the target sequences based on the user's historical behavior. It also determines the sequences from the target sequences that the user has historically performed specified operations as specified sequences. For each target sequence, the target sequence and the specified sequences are input into the feature extraction layer of the click-through rate (CTR) prediction model to be trained. The feature extraction layer outputs target features determined based on the content and position of objects contained in the target sequence, as well as specified features. The target features and specified features are then input into the prediction layer of the CTR prediction model to obtain the CTR of the target sequence output by the prediction layer. The loss is determined based on the CTR of each target sequence and its labels, and the model parameters of the CTR prediction model are adjusted based on the loss.

[0165] Optionally, the feature determination module 304 is configured to, for each candidate recommendation sequence, take the candidate recommendation sequence as input and input it into the feature extraction layer of a pre-trained click-through rate prediction model; for each candidate recommendation object in the candidate recommendation sequence, determine the object features of the candidate recommendation object based on the content of the candidate recommendation object and the position of the candidate recommendation object in the candidate recommendation sequence; determine the similarity between the candidate recommendation object and other candidate recommendation objects in the candidate recommendation sequence; enhance the object features of the candidate recommendation object based on each similarity and the object features of the other candidate recommendation objects corresponding to each similarity; and determine the candidate recommendation features of the candidate recommendation sequence based on the enhancement results of each object feature.

[0166] Optionally, the prediction module 306 is used to input the candidate recommendation features and the specified features as inputs into the fusion layer of the click-through rate prediction model to obtain the enhancement result of the candidate recommendation features enhanced according to the specified features and the fusion result of the specified features fused as the fusion feature. The fusion feature is then input into the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer.

[0167] Optionally, the specified operation includes multiple operation types. The receiving module 300 is used to determine the user identifier carried in the recommendation request sent by the user, and determine the sequences that the user has browsed in history based on the user identifier as each browsing sequence, determine the user's historical behavior, and for each browsing sequence, determine the user's operation on the browsing sequence based on the historical behavior, and determine the operation type to which the operation belongs as a label for the browsing sequence. For each operation type, determine each browsing sequence labeled as the operation type from each browsing sequence as each specified sequence corresponding to the operation type.

[0168] Optionally, for each target sequence, the priorities of the objects contained in the target sequence are not completely the same. The feature determination module 304 is used to determine the priority of each candidate recommendation object contained in the candidate recommendation sequence, and to input the candidate recommendation sequence as input to the feature extraction layer of a pre-trained click-through rate prediction model to obtain the candidate recommendation features output by the feature extraction layer, which are determined based on the content of each candidate recommendation object contained in the candidate recommendation sequence, the position of each candidate recommendation object, and the priority corresponding to each candidate recommendation direction.

[0169] Optionally, the recommendation module 308 is used to input the candidate recommendation features and the specified features as inputs into the prediction layer of the click-through rate prediction model. For each operation type, based on the similarity between the specified sequences of that operation type and the candidate recommendation features, the probability of the candidate recommendation feature corresponding to that operation type is determined. Based on the probability of the candidate recommendation feature corresponding to each operation type and the preset weights of each operation type, the click-through rate of the candidate recommendation sequence is determined.

[0170] This specification also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 Recommended methods provided.

[0171] This instruction manual also provides Figure 5 The diagram shows a schematic structural representation of the electronic device. Figure 5At 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 recommended method described above. 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.

[0172] 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.

[0173] 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.

[0174] 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.

[0175] 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.

[0176] 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.

[0177] 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.

[0178] 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.

[0179] 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.

[0180] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0181] 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.

[0182] 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.

[0183] 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.

[0184] 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.

[0185] 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.

[0186] 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.

[0187] 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 recommendation method, characterized in that, include: Based on the recommendation request sent by the user, determine a specified number of candidate recommendation objects, as well as a specified sequence of operations that the user has historically performed; The candidate recommendations are combined according to their different positions in the sequence to determine the candidate recommendation sequence; For each candidate recommendation sequence, the candidate recommendation sequence and the specified sequences are used as inputs to the feature extraction layer of a pre-trained click-through rate prediction model. The result is the candidate recommendation features determined by the feature extraction layer based on the content of each candidate recommendation object in the candidate recommendation sequence and the position of each candidate recommendation object in the candidate recommendation sequence, as well as the specified features. The candidate recommendation features and the specified features are used as inputs to the fusion layer of the click-through rate prediction model. The enhanced result of the candidate recommendation features enhanced according to the specified features and the fusion result of the specified features are used as the fusion features. The fused features are used as input to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer. Based on the click-through rate of each candidate recommendation sequence, a recommendation sequence is determined and recommended to the user.

2. The method as described in claim 1, characterized in that, The candidate recommendation sequence is taken as input and fed into the feature extraction layer of a pre-trained click-through rate prediction model. The result is the output of the feature extraction layer, which determines the candidate recommendation features based on the content and position of each candidate recommendation object within the candidate recommendation sequence. Specifically, this includes: For each candidate recommendation sequence, the candidate recommendation sequence is used as input and fed into the feature extraction layer of the pre-trained click-through rate prediction model; For each candidate recommendation object in the candidate recommendation sequence, the object characteristics of the candidate recommendation object are determined based on the content of the candidate recommendation object and the position of the candidate recommendation object in the candidate recommendation sequence; Determine the similarity between the candidate recommendation object and other candidate recommendation objects in the candidate recommendation sequence, and enhance the object features of the candidate recommendation object based on each similarity and the object features of the other candidate recommendation objects corresponding to each similarity. Based on the enhancement results of each object's features, the candidate recommendation features of the candidate recommendation sequence are determined.

3. The method as described in claim 1, characterized in that, The specified operation includes multiple operation types; Based on the recommendation request sent by the user, determine the specified sequences of operations that the user has historically performed, specifically including: Based on the recommendation request sent by the user, determine the user identifier carried in the recommendation request, and based on the user identifier, determine the sequences that the user has browsed in history, as each browsing sequence; Determine the user's historical behavior, and for each browsing sequence, determine the user's operation on the browsing sequence based on the historical behavior, and determine the operation type to which the operation belongs, as a label for the browsing sequence; For each operation type, each browsing sequence labeled with that operation type is determined from each browsing sequence and used as the designated sequence corresponding to that operation type.

4. The method as described in claim 1, characterized in that, For each target sequence, the priorities of the objects contained in that target sequence are not entirely the same; The candidate recommendation sequence is taken as input and fed into the feature extraction layer of a pre-trained click-through rate prediction model. The result is the output of the feature extraction layer, which determines the candidate recommendation features based on the content and position of each candidate recommendation object within the candidate recommendation sequence. Specifically, this includes: Determine the priority of each candidate recommendation object included in the candidate recommendation sequence; The candidate recommendation sequence is used as input and fed into the feature extraction layer of a pre-trained click-through rate prediction model to obtain the candidate recommendation features output by the feature extraction layer, which are determined based on the content of each candidate recommendation object contained in the candidate recommendation sequence, the position of each candidate recommendation object, and the priority of each candidate recommendation object.

5. The method as described in claim 1, characterized in that, The click-through rate prediction model is trained in the following manner: For each user, identify the sequences that the user has browsed in history as target sequences, and determine the labels of the target sequences based on the user's historical behavior. Also, identify the sequences from the target sequences in which the user has performed a specified operation in history as specified sequences. For each target sequence, the target sequence and the specified sequences are used as inputs to the feature extraction layer of the click-through rate prediction model to be trained, and the target features and specified features are obtained from the output of the feature extraction layer based on the content and position of each object contained in the target sequence. The target feature and the specified features are used as inputs to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the target sequence output by the prediction layer. The loss is determined based on the click-through rate and its label for each target sequence, and the model parameters of the click-through rate prediction model are adjusted based on the loss.

6. The method as described in claim 3, characterized in that, The candidate recommendation features and the specified features are used as inputs to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer, specifically including: The candidate recommendation features and the specified features are used as inputs to the prediction layer of the click-through rate prediction model; For each operation type, the probability of a candidate recommendation feature corresponding to that operation type is determined based on the similarity between each specified sequence of that operation type and the candidate recommendation feature. The click-through rate of the candidate recommendation sequence is determined based on the probability of each operation type corresponding to the candidate recommendation features and the preset weight of each operation type.

7. A recommended device, characterized in that, include: The receiving module is used to determine a specified number of candidate recommendation objects and a specified sequence of operations that the user has performed in the past, based on the recommendation request sent by the user. The combination module is used to combine the candidate recommendation objects according to their different positions in the sequence to determine the candidate recommendation sequence; The feature determination module is used to take each candidate recommendation sequence as input, the candidate recommendation sequence and the specified sequences as input, and input the feature extraction layer of the pre-trained click-through rate prediction model to obtain the candidate recommendation features and the specified features output by the feature extraction layer, which are determined based on the content of each candidate recommendation object contained in the candidate recommendation sequence and the position of each candidate recommendation object in the candidate recommendation sequence. The prediction module is used to input the candidate recommendation features and the specified features as inputs to the fusion layer of the click-through rate prediction model, and obtain the enhancement result of the candidate recommendation features enhanced according to the specified features and the fusion result of the specified features fused as the fusion feature. The fusion feature is then input to the prediction layer of the click-through rate prediction model to obtain the click-through rate of the candidate recommendation sequence output by the prediction layer. The recommendation module is used to determine the recommendation sequence based on the click-through rate of each candidate recommendation sequence and to recommend it to the user.

8. 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.

9. 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.

Images